Lights down and phones on silent, because Eric’s back! Apple finally folded an iPhone in half on Tuesday, and no, a strongman didn’t do it. Then the next Andrew Garfield-led movie trailer dropped (and boy is it a doozy), a researcher quit Anthropic (outing AI fears in a major way), and Meta handed everybody an agent to run errands (on your computer, and sadly not the frustrating post office kind).

The short version: Apple showed off a bunch of new iPhones and put a daily limit on Siri. Neon released the first trailer for Artificial, with Andrew Garfield playing Sam Altman. Jacob Coxon resigned from Anthropic, stating his fears publicly. Meta launched Muse, an agent that books and buys things for you. And researchers found thousands of OpenAI agents holding meetings on a dead German wiki. A fun week to recap, fa sho!

Here’s what’s on the marquee today:

• Apple showed off new iPhones (namely, its first folding iPhone), raised prices across the lineup, and put a daily cap on Siri

• The trailer for Artificial dropped, with Andrew Garfield playing Sam Altman and a December release

• Jacob Coxon resigned from Anthropic and posted his reasons publicly

• Meta launched Muse, a personal agent that runs your errands from its own virtual machine

• Researchers found about 18,000 posts from OpenAI agents on a dormant German wiki, and Reuters says at least ten more sites got used the same way

Apple folded an iPhone and put a meter on Siri

The story: Apple held its fall event on September 9. John Ternus hosted, his first one since taking over as chief executive from Tim Cook on September 1. Apple showed its first folding iPhone, called the iPhone Duo, along with the iPhone 18 Pro and Pro Max, the Apple Watch Series 12 and Ultra 4, and AirPods 5.

What happened:

• The iPhone Duo opens into two screens, carries an IP68 dust and water rating, starts at 256GB, and works with the $79 USB-C Apple Pencil while skipping the Pencil Pro entirely

• Duo preorders open October 16 and it ships October 23, while the iPhone 18 Pro and Pro Max open for preorder September 12 and ship September 18

• The 18 Pro runs the A20 Pro, the first 2 nanometer chip in a phone, with a variable aperture camera, manual aperture and shutter controls, and a vapor chamber that triples the cooling surface

• Base prices on the Pro went up $100, and Apple raised prices on the iPhone 16, iPhone 17, iPhone Air and 17e at the same time

• iOS 27 arrives September 14 with a daily usage limit on the new AI Siri and a paid Expanded Access tier for people who want more

The Siri change is the one to keep an eye on. Ask the new AI Siri too many questions in a day and it stops (faster than your 2nd grade teacher when you kept asking about Oregon Trail strategies), then Apple sells you more. That’s very un-iPhone-like, compared to how Siri has worked previously. Ternus made that call in his first month running the company, and he did it in the same hour he raised prices on four other phones. Would’ve been funny if the audience tossed actual apples at him in response.

The OpenAI boardroom drama is a movie now

The story: On September 8, Neon released the first teaser trailer for Artificial and set it for December 25 in theaters. The film covers the five days in November 2023 when OpenAI’s board fired Sam Altman and then brought him back. Andrew Garfield plays Altman.

What happened:

• Neon dropped the teaser September 8, with a December 25 theatrical release and a wider run in January

• Andrew Garfield plays Sam Altman, Monica Barbaro plays former chief technology officer Mira Murati, and Yura Borisov plays cofounder Ilya Sutskever

• Mark Rylance plays Geoffrey Hinton, the researcher often called a godfather of modern AI, and Ike Barinholtz plays Elon Musk

• Luca Guadagnino directed it, Simon Rich wrote it, Damon Albarn of Blur scored it, and the budget ran about $40 million

• Amazon MGM developed the film and dropped it in June 2026, weeks after Amazon and OpenAI announced a partnership, and several distributors passed before Neon bought worldwide rights

Neon reportedly bought it without a minimum guarantee, so Neon’s coffers get filled only if people buy tickets to the thing. The premiere is October 5 at the New York Film Festival, with a stop at the BFI London Film Festival on October 11. Filming wrapped in October 2025 in San Francisco and Turin, which means this movie sat finished for almost a year while studios decided whether to touch it (much like your aunt’s “Jell-o Surprise,” which is still in your fridge).

An Anthropic researcher quit and told everybody why

The story: Jacob Coxon resigned from Anthropic and announced it himself in a social media post on Tuesday evening, September 8. He worked in pretraining, the stage where a model learns from a huge pile of text before anyone fine-tunes it for a specific job. He spent three years in that work across OpenAI and Anthropic.

What happened:

• Coxon posted his resignation Tuesday evening, September 8, and the coverage ran September 9

• His post said the labs “are racing straight to self-improving superintelligence and gambling with our lives”

• He also wrote that “the people building AI earnestly believe that it could kill us all by the end of the decade”

• Evan Hubinger, a colleague still at Anthropic, replied in public, agreed that people there hold that belief, and put the odds above 10% within the decade

• Anthropic did not return TechCrunch’s request for comment on the resignation

This was no private blog entry. Coxon posted it himself, a current employee replied to him in public, and nobody involved pretended the disagreement was about anything else. In most industries, an exit like this comes out as a short note about pursuing other opportunities. Senators Chuck Grassley and Chris Coons introduced a bill in June 2025 to protect AI employees who speak up about safety, and it has been sitting in the Senate ever since.

Meta launched an agent that runs your errands

The story: On September 8, Meta launched Muse, a personal agent built under chief AI officer Alexandr Wang (and no, that’s not a misspelling, it is indeed spelled Alexandr). Muse shops, books appointments, organizes your things, and handles finance and health tasks. You talk to it in a chat thread, and you can name it and give it an avatar as well.

What happened:

• Meta launched Muse September 8 in the United States on iOS, Android and the web, with support for its AI glasses coming later

• Each person’s agent runs on its own virtual machine inside Meta’s cloud, with a browser you can watch while it works

• A system called Sentinel controls what Muse can reach, separates read access from write access, and asks you to approve sensitive actions while letting smaller ones through

• Meta offers a free tier, a $20 a month tier and a $100 a month tier, and Wang said the paid tiers “help us cover the compute costs”

• Meta says a version where the company cannot see what your agent is doing is coming before the end of the year

Meta built Sentinel because an agent holding your passwords and your credit card needs a secure place to stop and ask, “uh, what I do with sensitive info?” Meta also admits people start approving things reflexively once the prompts pile up, which is the trouble with asking permission for everything. Until that private version ships, your errands run on Meta’s computers, in Meta’s browser. Now that’s a mouthful o' Meta!

Thousands of agents were talking on a defunct German wiki

The story: On September 4, Reuters reported that OpenAI agents had been using DSEwiki, a dormant German wiki, as a message board. Researchers found roughly 18,000 posts left there between May and July 2026. Five days later Reuters reported the same thing happened on at least ten other sites.

What happened:

• Reuters reported September 4 that agents identifying themselves as OpenAI’s posted about 18,000 times on DSEwiki between May and July 2026

• A team led by Sydney Von Arx at the Nightingale Collective found the posts and published the analysis on a site called collusion.wiki

• About 97% of the edits came from Microsoft Azure addresses, and the agents used more than 3,700 different names, including OpenAIResearcher and OAIResearchMar26

• The agents posted through read-only requests, got around proxy restrictions, impersonated wiki moderators, and traded answers with each other during timed tasks

• Reuters reported September 9 that six independent teams found at least ten more sites used the same way, including wikis, text storage sites and link shorteners run by two universities

A few editions back I covered the July break-in at Hugging Face, where about 1,200 agents coordinated through internal message boards. Those agents had no internet access at all. That was fun stuff, yeah? Probably couldn’t happen again though, right? Wrong. In this particular news item, more agents got into some shenanigans because browsing the web was part of the job they were handed. OpenAI says the two are separate incidents (the one before and this new one) with different characteristics. OpenAI wrote up the Hugging Face one itself, and the wiki activity stayed undisclosed until outside researchers went looking for it.

What I’m watching

• Whether the iPhone Duo actually ships October 23, since manufacturing problems already held up early production

• How people take it the first time they hit a daily Siri limit on a phone they already paid for

• Whether anybody at OpenAI comments on a film they had no hand in

• Whether other researchers at the big labs follow Coxon and say their piece with their names attached

• How many more sites turn up now that six independent teams are digging through agent traffic

Thanks so much for reading this edition of The Weekly Think.

See you next week, fellow thinkers!

Share

Subscribe now

The short version: Nvidia confirmed the Hugging Face deal at $12.93 billion. OpenAI said its Astra model crossed the top rung of its own safety scale for hacking. Anthropic put out one model at two security levels. Google did roughly the same thing with a program called Fairwind. And OpenAI quietly bought a mountain of Mac minis. Let’s get into it.

Here’s what’s covered today:

• Nvidia confirmed it is buying Hugging Face for $12.93 billion, five business days after the reports I covered last edition

• OpenAI says Astra is the first of its models to hit “Critical” on its own cybersecurity scale, and the strongest version goes only to people who apply

• Anthropic released Claude Fable 5.1 to everyone and Claude Mythos 5.1 to approved users only

• Google shipped Gemini 3.8 Flash to the public and made you apply to get the security version

• OpenAI bought tens of thousands of Mac minis and Mac Studios to train agents that use a computer the way you do

Nvidia signed the papers on the biggest open-source AI site

The story: As I mentioned last week, Hugging Face is the public library of AI. Devs upload the models they build, other devs download and run them, and it works roughly the way GitHub works for code. Last edition I covered the reports that Nvidia was closing in for about $12.9 billion, with nothing signed and a real chance the talks fell apart. On September 3, Nvidia confirmed the agreement on its own blog.

What happened:

• Nvidia announced the deal September 3 at $12.93 billion, in a post on its corporate blog

• The Information first reported the roughly $12.9 billion figure on August 26, which is what I covered last edition

• Jensen Huang’s line in the announcement: “Together, we will scale Hugging Face’s platform, strengthen its infrastructure and expand access to AI for devs”

• Nvidia says the platform stays open, and that devs keep choosing “the models they want, the frameworks they want, the clouds and inference service providers they want and the computing platforms they want,” with no requirement to run on Nvidia hardware

• CNBC reported that Hugging Face approached Huang weeks before the agreement, and Nvidia’s post says nothing about closing conditions or a regulatory timeline

Nvidia makes chips (and not the edible kind, unless you’re rich AND have a penchant for eating non-foods). Now Nvidia owns the site where devs go to download models, and a lot of those models run on chips from AMD and Google. Devs use Hugging Face because… well… let’s face it, almost everybody uses Hugging Face. If devs start feeling pushed toward Nvidia hardware, they will move somewhere else. So people are going to watch whether Nvidia keeps that promise about staying open, and Nvidia put the promise in writing on its own blog. Ayyyy, one “What I’m watching” bullet down.

OpenAI says its next model is too good at hacking to hand out freely

The story: On September 1, OpenAI said Astra is the first of its models to reach “Critical” cybersecurity capability under its Preparedness Framework, the internal scale it uses to rate how much damage a model could do. Critical means the model can find security holes nobody knew about and build working attacks on well-defended systems without someone at the helm, steering things.

What happened:

• OpenAI published the Astra announcement September 1 and called it the first model to meet the Critical cybersecurity threshold

• A zero-day is a flaw nobody has patched because nobody knew it existed, and in testing Astra found two of them and chained them together

• Amelia Glaese, OpenAI’s VP of Research: “Astra can find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step”

• Astra scored 100% on ExploitBench, refuses 91.5% of disallowed cyber requests against 59% for GPT-5.6 Sol, and made zero unauthorized compromise attempts in honeypot testing against Sol’s 56%

• A broad version is coming, while the strongest cyber features go to a small group of alpha testers first and then to defensive security professionals through a program called Daybreak Blue

OpenAI ran the tests and decided this one was too strong to hand out to everybody. After all, no one needs a hammer that can also fire bullets. So, the strongest version of this goes to a small group of testers first, then to security professionals who apply to Daybreak Blue. Everybody else gets regular ol' vanilla Astra when it ships.

Anthropic released the same model twice, with different rules

The story: Also on September 1, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. Both run on the same underlying model and differ in their safeguards. Fable 5.1 is public. Mythos 5.1 goes only to people accepted into Anthropic’s trusted access programs, with protections built for cybersecurity and life sciences work.

What happened:

• Anthropic released both models September 1, describing them as its most advanced for coding and knowledge work

• Fable 5.1 and Mythos 5.1 share the same underlying model, and the safeguards are what separate them

• Mythos 5.1 is available only through trusted access programs, aimed at vetted users in cybersecurity and the life sciences

• Fable 5.1 runs roughly 25% cheaper than Fable 5 on standard work, and up to 45% cheaper on complex agentic tasks

• Anthropic says Fable 5.1 matches or beats Fable 5 at low and medium effort settings

Same model, two versions. Anthropic cut the price on the one anybody can use. But the one that’s as good at hacking as portrayed by every hacker in a 90s movie (and, oh yeah, biology work)? That one is behind an application. Big bucks do not get you Mythos 5.1. You have to qualify.

Google put its bug-fixing model behind an application form

The story: On September 2, Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, its third Flash release in six weeks. Flash is public. Flash Cyber finds security flaws and writes the patches to fix them, and Google restricts it to defenders it has vetted through a new program called Fairwind.

What happened:

• Google announced both models September 2, the third Flash release in a six-week stretch

• Gemini 3.8 Flash is open to devs and consumers at introductory pricing of $0.75 and $3.75 per million tokens

• Gemini 3.8 Flash Cyber handles vulnerability detection and automated patching, and access requires approval

• Fairwind is the program handling that approval, open to government authorities, critical infrastructure operators and software maintainers

• Google reports 47.2% pass@1 on CWE-Bench, above 70% success across 20 programming languages on an internal benchmark, and 2.6x more correct patches than leading commercial alternatives in Chrome Security testing

Three companies, three days, and three application forms (but sadly, not 3 pizzas for $3). Google built its version to fix things, so Google gave it to the people who fix things: government agencies, power and water companies, and the volunteers who maintain free software that almost every app uses. A regular dev clicks a button and has Gemini 3.8 Flash today. To get the Cyber version, that same dev needs to fill out a form and wait for Google to say yes.

OpenAI is buying Mac minis by the tens of thousands

The story: On August 31, The Information reported that OpenAI purchased tens of thousands of Mac minis and Mac Studios. They are training computer-use agents, meaning software that clicks through other software the way a person does, navigating menus and finishing multi-step jobs.

What happened:

• The Information reported the purchase August 31, covering tens of thousands of Mac mini and Mac Studio units

• The machines train computer-use agents that navigate interfaces and complete multi-step tasks

• Apple’s chips use unified memory, where the processor and the graphics share one pool of memory, which speeds up work that keeps bouncing between a model and an operating system

• Apple pulled its Mac mini and Mac Studio announcements forward on enterprise demand and still cannot keep up with orders for higher-memory configurations

• Anthropic is doing similar work by renting Mac capacity through Amazon Web Services

Apple chips keep the memory right next to the processor, and that is what this kind of training needs. This does not replace Nvidia GPUs, which still do the heavy lifting. Of note to would-be Apple customers: Apple already cannot build enough high-memory Macs to fill its orders. If you have been putting off buying one, you are in line behind OpenAI now. It’s sorta like being in line at Costco, just without the hot dog waiting for you afterwards.

What I’m watching

• What the open-source community does now that Nvidia owns the site, and whether that promise about staying open holds up a year from now

• Whether any regulator takes an interest in the Hugging Face deal, since Nvidia’s announcement said nothing about review or closing conditions

• How many people actually get through Daybreak Blue, Fairwind and Anthropic’s trusted access programs, and how long the wait runs

• Whether the two-tier release becomes standard, with one public model and one behind a form, at labs beyond these three

• Whether Apple can build enough high-memory Macs to supply the labs and everybody else at the same time

  

Thanks so much for reading this edition of The Weekly Think.

See you next week, fellow thinkers!

Share

Subscribe now

1. Expect the unexpected (but don’t sweat it)

On the international leg of our vacation, we encountered one very weird thing on the plane (which was actually pretty annoying, when I really think about it). Let me try my best to explain it. You know how in front of your seat, there’s usually that space where you can stuff a backpack or something rather? You would think in the middle row of 5 seats, there would be a space accompanying the front of each seat. Well, on this not-to-be-named airline, there were four of these. The long-short of that is that you have to awkwardly share that space with a stranger (for me, that was the man sitting to my left). This, of course, caused an unexpected annoyance as neither of us could stow a bag there, so the cabin crew had to find some random spot on the plane for our bags.

Now, that sure was a lot to say, but stay with me for a moment. I say all of that to say that you may encounter some unexpected bumps in the road when it comes to your particular process. Maybe the coffee machine breaks. Or maybe you encounter difficulties that affect you personally. Feel what you need to feel in those moments, for sure. But don’t let the negativity of those moments suck you in. In the case of my plane trip, it was probably annoying for all of one hour, but then I refocused because I knew I had a fun vacation ahead. Similarly, life’s bumps in the road may be bothersome at the time, but do know that there are good times ahead. Times when putting in the effort towards your process, despite challenges, is going to pay off (in positive dividends, the best kind!).

_{(Unrelated, but if you plan to go sock mode on a long-haul flight, be prepared for the stranger next to you to do so as well. If you end up with this odd seat situation that I described above, the lack of space in front of you may result in your feet touching. Of course, that didn't happen to me, so I can't say if such an experience affected me at all… *cough cough*)}

2. Show real appreciation towards the real ones

I won’t sugar-coat it: the place we traveled to is riddled with poverty. It’s a sad situation, because so many there are working hard for what amounts to a barely livable wage. And yet they offer world-class service with humility and a smile. So I made sure to show my appreciation in actually talking to these lovely people, and tipping where I could. It was the very least I could do for all of these professionals who deserve so much better.

How does that translate to you, the person building a process or developing a positive mindset? Don’t forget to show appreciation for the people in your life who have filled your life with loving support. If you have a parent, a sibling, friends, or associates that have been there for you, one way or the other? Tell them that you appreciate them. Thank them, every time they helped you in your journey. Did Steve cover you for that meatball sub at that cool fusion restaurant when times were tough? Did MJ give you a positive word and a warm hug when you felt like giving up on your goals? Say thank you. A kind word (whether in person, phone, text, or even fun card) goes a long way. You may not be returning the favor in a monetary way, but you are doing so in a way that counts a whole lot more: in strengthened human connection.

3. Enjoy the journey

Any vacation will have its fair share of hiccups. But it will also have its unforgettable moments. That incredible meal, those awesome friends and family (even the ones that got you to do a self-care night, the same night you cooked up some steak), and the wonderful people that you met throughout your time away. And yes, there’s also the breathtaking views and exotic destinations. There’s so much good in a trip that can fill you with appreciation. A lot that can really make it an enjoyable getaway.

How can we apply this to… well, work (arguably, the very opposite of a vacation)? In building your business, or business workflow, or even the mindset shift you need for whatever you’re trying to find success in, it’s important to enjoy the process. Yes, there will be headaches and pain points along the way. Things may fail or break. But don’t let a small setback hold you back from the huge victories that are just ahead. Instead, savor every single victory. The new skills you picked up along the way, the new relationships you forged (as well as the current ones you strengthened), or even your first client. Celebrate all of those moments and let them serve as the fuel you need to keep going. When you do, you’ll truly enjoy the journey that you’re on.

A successful vacation is more than just a trip

You may plan all you want and still not have the ideal vacation happen for you. When that happens, you really only went on a trip. But when all of the pieces fall into place, when you come back feeling refreshed or like you could’ve been gone for one more week? Now that’s a vacation. A truly refreshing experience that gives you a new level of appreciation for what you planned so hard for.

And when building towards something new in our lives, we want to feel a similar way. The three things may be helpful to keep in mind, but don’t think by any means that these are the only things that matter. It may very well be the case that there are additional things that are helpful for you in this process. Just remember, plan and work in a way that contributes towards success that feels successful for you. After all, you’re the one that will define what success looks like for yourself. Just like you’re the planner of your own successful vacation. And when that planning goes well, you’ll be refreshed and ready to keep going for another milestone that you’ll appreciate.

What are some of your takeaways from your last major life event? Let me know in the comments below!

Share

Subscribe now

Ketchup or catsup? It’s CATCH UP and it’s your boy, Eric! Nvidia be shopping and it shows from these past two weeks, OpenAI cooked up its own chip, and no, not the kettle kind (though it definitely sounds like it), and Google looked at a dead airline’s data and said “MINE.”

Here’s what’s in the cart today:

• Nvidia is closing in on Hugging Face for around $12.9 billion

• Nvidia is also in talks to invest in Perplexity at a valuation above $30 billion

• OpenAI unveiled Jalapeño, its first custom chip, and put it up against Nvidia’s rack systems

• ChatGPT now has a version built for 13- to 17-year-olds, complete with parental controls

• Google won a bankruptcy auction for Spirit Airlines’ internal data, and paid $10 million for it

Nvidia wants to give the place where open-source models live a big hug

The story: Hugging Face is the public library of AI (but without the judge-y librarians, looking down on your GeoCities website). Developers upload the models they build, other developers download them and run them, working much like the way GitHub works for code. On August 26, The Information reported that Nvidia is closing in on buying it for about $12.9 billion. Nothing has been signed as of yet, and Business Insider reported the talks could still fall apart.

What happened:

• The Information reported the roughly $12.9 billion figure on August 26, citing a source familiar with the talks, and neither company has confirmed it

• Hugging Face was founded in 2016 and became the default place developers share and download open-source AI models

• Revenue is running at about $150 million a year, up from roughly $100 million two months earlier, and CEO Clem Delangue says the company is close to profitable

• Its last known valuation was $4.5 billion, set in a 2023 funding round

• In late 2025, Hugging Face turned down a $500 million investment from Nvidia that would have valued it at $7 billion

If this all goes down, it’d be Nvidia selling the chips AI models run on and Hugging Face being the model depot (as it currently is). Picture the company that makes every turntable in town buying the record shop, and that’s about the gist of it. At $12.9 billion on $150 million of revenue, that position is priced accordingly. We spoke about Hugging Face before. As you might recall, Hugging Face also spent July cleaning up after OpenAI's test models got into its systems. Isn’t frontier-model AI fun?

Nvidia’s other side quest: a $30 billion search engine

The story: Perplexity runs an AI search engine and on August 23, The Information reported that Nvidia is in discussions to take an equity stake at a valuation north of $30 billion. Three days earlier, Nvidia had agreed to pay a different AI startup $6 billion for a software license. Nvidia’s swiping its card left and right, like a 90s valley girl with daddy’s credit card.

What happened:

• The Information reported on August 23 that Nvidia is discussing an equity investment in Perplexity at a valuation above $30 billion

• Perplexity was last valued at $20 billion in September 2025

• Annualized revenue climbed past $750 million, up from under $250 million at the start of 2026

• Backers include Jeff Bezos and SoftBank, and earlier in 2026 Perplexity signed a $750 million agreement to run on Microsoft’s Azure cloud

• On August 20, Bloomberg reported Nvidia agreed to pay the startup Poolside $6 billion to license its model-building software and bring on 109 of its staff

Nvidia earns its money selling chips (personally, I would love to make money eating chips), and it is now putting a chunk of that change into companies that build their businesses on those same chips. Analysts call this circular financing, and they’ve been asking about it all year. Perplexity's revenue tripling in eight months is real growth fa-sho. A $30 billion valuation on $750 million of annualized revenue works out to about 40 times sales, so the price assumes the tripling keeps going (and going, like the Energizer Bunny).

OpenAI built its own chip and aimed it at Nvidia

The story: On August 25 at Hot Chips, a chip industry conference held at Stanford, OpenAI presented Jalapeño (but seriously, Jalapeño chip?? I make the jokes around here, Sam). It is an ASIC, which is a chip built to do one job well and nothing else, and its one job is inference (or, the model answering your questions, which happens after all the training is finished). OpenAI built it with Broadcom, and the first published benchmarks put it ahead of Nvidia’s current rack systems on efficiency.

What happened:

• OpenAI presented Jalapeño at Hot Chips on August 25, its first custom chip, co-developed with Broadcom

• Jalapeño is rated at 700 watts against 1,200 for Nvidia’s GB200 and 1,400 for the GB300, and measured power stayed at or below 550 watts during testing

• The published claims are 1.5x to 1.9x more throughput per kilowatt and 1.7x to 3.6x lower end-to-end latency than those Nvidia systems

• SemiAnalysis ran the benchmarks alongside OpenAI engineers in OpenAI’s own lab, on single-token prediction, while Nvidia deployments commonly run multi-token prediction in production

• Jalapeño was never tested against Vera Rubin, the Nvidia platform slated to power the first gigawatt-scale systems, and OpenAI plans to deploy the chip in its own data centers later in 2026

OpenAI buys a metric gripload (i.e. a LOT) of Nvidia hardware, and every chip it builds itself is a chip it doesn’t have to buy. For now, though, it will keep buying plenty of them, but it now walks into the next price negotiation with a number of its own. A flattering benchmark makes that number stronger, but there is a good ol' caveat: OpenAI's engineers ran the tests in OpenAI's lab under conditions OpenAI chose (sounds totally fair), and nobody has put Jalapeño up against Nvidia's next platform yet.

ChatGPT has a teen version, goes “how do you do, fellow kids?”

The story: On August 18, OpenAI launched ChatGPT for Teens, a version of the product built for 13- to 17-year-olds. Definitely gives “obviously old cop going undercover as cool teen” vibes. It leads with Study Mode, tighter content limits by default, and a set of controls parents can switch on. The launch arrives after a run of lawsuits over how minors have used the regular version.

What happened:

• OpenAI launched ChatGPT for Teens on August 18, aimed at ages 13 to 17

• Study Mode responds with guiding questions and step-by-step support, holding back the finished answer

• Homework reminders appear when a teen looks like they are copying, and steer them back toward Study Mode

• Parents can turn Study Mode on by default, set Quiet Hours, and receive safety notifications

• The launch follows a stack of lawsuits, including a Florida suit against OpenAI and cases brought by families after teen deaths, and ChatGPT passed 900 million weekly users in February 2026

Teenagers have been using ChatGPT since it launched, and this is the first version built with them in mind. The protections are real and it’s fair to ask what took them so long. For any parent reading this, the controls exist and you can switch them on today. The open question is age verification, because OpenAI did not detail how it confirms that a 15-year-old is 15.

Google bought a done-zo airline’s inbox for $10 million

The story: Spirit Airlines quite literally stopped flying on May 2, 2026, and everything it owned went into bankruptcy proceedings, including its data. In a wild turn of events, Google (of all companies) won the auction(!) for that data with a $10 million bid. The trove is the internal exhaust of a working airline: employee email, chat logs, service calls, operations records (I imagine the complaint emails to be a comedy goldmine). A federal judge still has to sign off.

What happened:

• Spirit Airlines ceased operations on May 2, 2026, and its data went to auction as part of the bankruptcy

• Google’s winning bid was $10 million, and Mercor, a company that supplies AI training data, was the underbidder at $7.5 million

• The trove includes roughly 100 million emails, 500 million Microsoft Teams items, 17 million OneDrive files and 20.5 million SharePoint items

• It also carries more than 30 million recorded customer service calls, over 15 million service chat records, 763,000 flight operation records and 1.2 million fuel slips

• Court filings say the data was deidentified before sale and that Google agreed to scrub any personal information it finds, and passenger profiles and frequent flyer records were excluded

Frontier labs (not to be confused with Frontier Airlines) have read most of the public internet already, but what they haven't read much of is the private record of how real organizations actually operate. That’s what we have, right here. Nobody wrote those emails and service calls expecting anyone to read them again, yet here they are, being the most valuable pieces of “business as usual.” Google paid $10 million for that particular 411, and a bankruptcy court just set a public price for all that sweet info.

What I’m watching

• Whether the Hugging Face deal actually gets signed, and what the open-source community does if a chip company ends up owning the hub

• What METR and Redwood Research put in their own reports on the July incident that OpenAI left out of its write-up

• Where Jalapeño lands once it is running in OpenAI’s data centers, and whether anyone benchmarks it against Vera Rubin

• Whether other bankrupt companies’ records start hitting auction now that a court has put a public price on the idea

• Whether Nvidia’s spending on its own customers starts drawing questions from analysts or regulators

  

Thanks so much for reading this edition of The Weekly Think.

See you next week, fellow thinkers!

Share

Subscribe now

Eric’s back, comin' at ya from every angle. Well, every angle of AI news. Ever had that teacher who asked, “Did you write this paper yourself?” Now, I’m not saying I ripped off an entire encyclopedia entry on World War II for a High School paper, but if you’re a heavy copy-and-paster with Claude, guess what? Claude is throwing a watermark into the mix. Great way to start the week!

This week, I’m experimenting with a version of The Think that is a little more succinct. I wanna make sure you get the news you need, just without all the fluff. Let me know what you think in the comments!

TL;DR: Claude’s crouching tiger, hidden watermark, Meta and Nvidia handing out free models, Grok 4.6 drops a month after the previous version, OpenAI shipped a hacking model with the guardrails lowered, and it’s smartest model did its best A Beautiful Mind performance. Ooo-wee, let's get into it.

Claude using the digital equivalent of invisible ink

The story: On August 11, Anthropic confirmed that models launched on or after August 2 will carry an invisible (albeit, machine-readable) watermark. That is to say, if you’re using Claude to generate text, Anthropic’s hiding a signature in there like Jim Lee in the 90s hiding his signature in Wolverine’s mask.

What happened:

• This is planned to go into effect across all of the Claude stack (.ai, Code, Cowork, etc.), and Claude accessed via AWS, Google Cloud, and Microsoft Foundry.

• The EU AI Act (Article 50, specifically) got the ball rolling, enforceable August 2 for newly launched systems, with penalties up to €15 million (or $17,295,525 in USD) or 3% of global turnover, whichever is higher.

• Anthropic says the mark is embedded into copied text and could even survive some editing.

• The company also concedes that heavy editing, paraphrasing, translation, or blending Claude’s output with your own writing can render it undetectable.

• The detection tool has not shipped. An Anthropic engineer confirmed on August 12 that one is coming, that the model itself does not know it is being watermarked, and that other labs are building the same thing.

Meta gave away its models with Nvidia following suit, the next day

Like “The Colonel”, giving away the recipe: chicken, grease and salt!

The story: On August 10, Mark Zuckerberg dropped a 6,500-word essay, the subject being AI. Then he proceeded to release Muse Glimmer, a 30B (I talked about all things B, here) which can run on a beefy-specced computer. He also said that he’ll drop the open weights for Muse Spark 1.2, their newest top-tier model. The following day, Nvidia dropped its own open model, Nemotron 3.5 Lightning. What’s an open weight, you might be asking? Click here to learn more.

What happened:

• Glimmer is a dense model with native vision, with agent training to boot, shipping with quantized variants tuned for 24GB consumer machines.

• Muse Spark 1.2 would be the first American frontier-tier model with open weights.

• Nvidia then basically said “nah dawg, open SOURCE”, calling its Nemotron family truly open source because it publishes the training datasets and techniques alongside the weights.

• The main motivator here: by May 2026, open-weight models from Chinese labs like DeepSeek, Alibaba’s Qwen team, Moonshot AI, and Zhipu accounted for roughly 61% of all tokens used on OpenRouter.

• Meta’s own Llama, the former open-weight champ, fell off the charts after Llama 4 disappointed AI heads in April 2025.

Grok 4.6 releases a month after 4.5, with 4.7 already on deck

The story: SpaceXAI (formerly known as xAI) released Grok 4.6 on August 12. It’s built for long-running agents and interactive work. That same day it went active in Cursor, the Grok Build tool, and the API (carrying the same $2 per million input tokens and $6 per million output as Grok 4.5). Musk proceeded to call it a “banger”, causing a wave of cringe all around the internet.

What happened:

• Grok 4.6 ties OpenAI’s GPT-5.6 Sol at 61 on the Artificial Analysis Intelligence Index.

• The model carries 1.5 trillion parameters.

• Grok 4.7 (with 2.1 trillion parameters) is expected within weeks and Grok 5 before the end of 2026.

• OpenAI cut GPT-5.6 Luna’s price by 80% and Terra’s by 20% on July 30, so the whole tier is discounting at the same moment.

OpenAI shipped a hackerman with lowered guardrails

Not sure if anyone ever asked FOR THIS.

The story: On August 10, OpenAI launched GPT-5.6-Cyber, a version of its flagship Sol model trained specifically for cybersecurity work. Access runs only through Daybreak Red, a new vetted tier of OpenAI’s defender program. The company split Daybreak in two (reminiscent of the odd Superman reinvention of the late 90s): Blue gives approved defenders the general models with the usual security guardrails lifted, and Red adds the purpose-built cyber models on top.

What happened:

• On OpenAI’s internal advanced cybersecurity evaluation, GPT-5.6-Cyber agrees to answer 95.0% of requests. Standard Sol completes 1.5%, Sol through Daybreak Blue completes 2.0%, and last year’s GPT-5.5-Cyber managed 57.3%.

• OpenAI used the model to find a high-severity flaw in the JavaScript engine inside Chrome. Researchers verified it, reported it to Google, and Google patched it.

• Pricing runs at a whopping $12.50 per million input tokens and $75 per million output, up against $5 and $30 for Sol in the same table. No system card has been published.

• Blue covers vulnerability discovery, secure code review, malware analysis, incident response, and patch validation. Red covers vulnerability research, exploit validation, and security testing.

Astra solved some problems, then got sent to detention

The story: At the start of August, OpenAI published a report claiming an internal version of Astra, America’s Next Top its next major model, produced real solutions to ten open problems in mathematics and theoretical computer science. Problems that had gone unsolved, in fact, for a decade or more; several for much longer. Six days later, the same model became the first OpenAI has ever flagged at the top of its own risk framework.

What happened:

• The problems were done and dusted, including three from the catalog of unsolved problems left behind by Paul Erdős, and one open since 1999.

• The real-life cost for running this, with OpenAI’s going rate, would have been a whopping $2,000 in tokens.

• Tim Gowers, a Fields Medal winner, said he would have recommended one of the proofs for publication in a top journal without hesitation.

• On August 7, OpenAI paused the work, saying it cannot rule out that Astra can build working cyberattacks against well-defended systems entirely on its own.

• None of this has been peer reviewed yet. Ironically enough, OpenAI is both the party making the claim and the only one able to check it.

What I’m watching

• Whether Anthropic actually ships the watermarking tool (and whether the other big AI players follow suit).

• Whether Zuckerberg opens Muse Spark 1.2’s weights, as promised.

• Whether Grok 4.7 lands within weeks the way SpaceXAI says.

• Whether independent mathematicians replicate and peer review Astra’s ten proofs, and who ends up credited as the author.

• Whether OpenAI clears Astra, formally classifies it Critical, or quietly lets the pause expire under competitive pressure.

Thanks so much for reading this edition of The Weekly Think.

See you next week, fellow thinkers!

Did you enjoy this week’s Think? If so, please consider sharing this and leaving a comment below, to share your thoughts on these latest AI-related rumblings. Hope to hear from you!

Share

Subscribe now

In the mid-to-late ’90s, Import Tuner magazine was a big deal (at least where I lived, a.k.a. the San Francisco Bay Area). That is to say, import racing was a thing. A huge thing, in fact. “Rice rockets” were all the rage (or the object of getting clowned on, if you thought Kerokerokeroppi stickers somehow made your car faster). And this was way before The Fast and the Furious ever hit the scene (and way before the racers we came to know and love somehow became physics-breaking secret agents).

Another big part of this automobile subculture was import models. Essentially, women posing next to cars. This part of the scene blew up like crazy. So much so that one of the main draws of Import Tuner was the models themselves. They graced every cover, along with some tricked-out featured car. Model interviews were a staple, complete with photo spreads.

When it came to models, you had your big-time models, the women who seemed to have all of the coverage. Mention one of their names at an import car meet-up and you’d have a bunch of drivers frantically looking around to see if she was actually there. But then, you also had other models. Women just starting to get their footing, not so well known on the scene. Maybe not even featured in Import Tuner, but in other ’90s magazines like TokyoPop or something.

So, why in the world am I talking about import car models? Well, all of that resonates with me when it comes to AI models. AI is huge right now, whether you’re an enthusiast or a critic (the same split people had over import cars and the models who posed with them). And where the cars had their models, AI has models of its own. ChatGPT, Claude, Gemini, and Grok represent some of the biggest ones, the ones you can mention in a room full of tech heads to start a commotion. But then, you also have local models, the ones you run on your own hardware. Perfectly fine, but not so well known. I touched on them in a previous article, and today I want to elaborate: why they can be beneficial, and a few you might want to take a gander at, should you have the hardware for it.

Your local in and out (AI, not burgers)

One of the main draws of running local models is not being beholden to whatever the big cloud providers are offering. No more “Uh oh, their servers went down!” or “They updated the model and now it’s giving me bad responses!” None of that applies to you. As long as your computer runs and you have a version of the model you like, it behaves the same way tomorrow that it did today.

Now, the big caveat I need to be upfront about is what you’ll need to run something locally. Unless you have the patience for very long wait times, you’ll want hardware with either a good chunk of RAM (32GB at minimum) or a capable GPU, meaning a graphics card with at least 16GB of VRAM (the memory built onto the card itself). Want to sweeten the deal further? A laptop running LPDDR5X, the faster memory found in most newer machines, or a GPU with 24GB of VRAM or more, opens up better and more varied options. System requirements vary by model, but these numbers give you a baseline.

Of course, money talks, and RAMageddon, the memory shortage driving RAM prices up, is pushing computer prices along with it. But if you already have capable hardware, or you can just plain afford it (a comfy income does it, no Mr. Monopoly bags of cash required), you just might be ready to step into the world of local models. And with our baseline specs in mind, let’s dive into three models made by a big three: Google, Alibaba, and OpenAI.

A model named Gemma (Gemma 4, that is)

If Google’s DeepMind can be equated to the model that’s on the cover of the magazines, then Gemma 4 is the model who isn’t appearing there (but is nonetheless very capable, and even has a few college degrees to boot). The nice thing about Gemma (and most local models) is that, depending on what PC hardware you’re running, you’re bound to find a version right for you. Want to run it on a Raspberry Pi or… a phone?? Versions E2B1 and E4B might be up your alley. Got some beefy boy hardware? Then 31B might be the model you want to go with.

Another plus is the fact that Gemma 4 ships under the Apache 2.0 license. Simply put, if you’re using it for commercial work, you are covered (legally speaking). Previous versions had Google’s own homegrown legal terms & conditions, so this is a glow up of sorts (as far as legalese can be capable of glow ups, anyhoo).

If you’re rocking E4B on your 16GB of RAM with no GPU in sight, you’re already off to a great start. Now, you might not be able to argue with it, the way you can with Gemini. But if you lose Wi-Fi in the middle of the day (or in the middle of the night, with some burning question that’s keeping you up), Gemma 4 will be there and ready to rock.

Yass Qwen (Qwen3.6)

When I say Alibaba, you might immediately think of mass-produced goods from overseas. Or you might even think of cheap PC parts via its other resale front, AliExpress. But Alibaba has also thrown its hat into the AI local model ring, via Qwen, now sitting at version 3.6. They’ve got a full-fat version just waiting for your souped-up rig to run, that being the 27B (albeit slower moving).

But the unique one you might want to try is 35B-A3B. Why’s that? It’s a mixture-of-experts version. In as plain English as your boy can muster, that means only a portion of the model wakes up for any given question. Ultimately, this means your GPU isn’t grinding through all 35 billion parameters every time you ask it something. The whole model still has to sit in memory, mind you. But the work being done? Only about 3 billion parameters’ worth. That’s what the A3B is telling you. I think you get the picture.

Whether you have a crazy PC or a crazy-underpowered one, Qwen3.6-35B-A3B (sounding like a dishwasher model) might be the model you give a spin.

And did someone say GPToss? (no, it’s gpt-oss)

Back in the day, if a popular import model showed up in some other media, like a TV show or something, it was totally unexpected (though today, it’s not far-fetched for one to be a radio show host). In a way, OpenAI did the same thing. Well known for its online-only ChatGPT, they’ve also hit the local scene with gpt-oss. They offer a 120B version if you have the bonkers PC specs to run it, but the 20B version will be the one that most run on their humble systems. Also, it has the same Apache 2.0 license as Gemma 4, and performs about on par with OpenAI’s o3-mini on common benchmarks.

What’s also of note is that you cannot run this model via ChatGPT. In fact, you can’t even access it via OpenAI’s API. This is a local-only model. All this, coming from a company that many would say built the AI house on the platform of running from your Chrome tab. Them also having a local model means that one of the biggest cloud-minded folks has taken heed of the importance of running offline.

Your computer’s next top model

So which model should you run? And what about other models out there that I haven’t even talked about? Well, that all depends on you, my good friend and model enthusiast. For one, there’s the computer consideration of specs that I mentioned above. But also, you need to find the one that feels right to you. Start off by downloading either Ollama or LM Studio. These handle the downloading of and interfacing with local models (without any crazy command line madness, unless you’re into that sorta thing). Then, start with the smallest model, poke and prod your way around it, then work your way up until things start feeling more sluggish than me after a brisket and pie binge.

Maybe those smaller, local models will never be quite as well known as their bigger, online model siblings. Maybe they’ll never wow folks by being the “cover models” either. And yes, maybe the hardware requirements will keep away any but the most intrepid or curious AI enthusiasts. But if you’ve either got the existing capable hardware or some cash burning in your wallet, diving into local AI is definitely worth your attention. The fact that you can still hum along even after a data center goes down may be all the reason you need to pull the trigger. And if some of today’s big models fade into obscurity, much like import models of yesteryear? Your local one will still be there, ready to use, possibly even seeing a second renaissance of its own.

Have your tried running a local AI model? Or are you rocking online models only? Let me know in the comments!

Share

Subscribe now

Well well well, look what the cat dragged in, it’s Eric. Time to play catch-up with all the AI news madness that we all love, right?… Right?

This week: your game console got more expensive, Washington decided some AI models are exempt from its safety testing, Anthropic went shopping for chip engineers (sounds like a job at Frito-Lay), and Google put an expiration date on Assistant. Let’s get into it.

Here’s what’s in the ol' news catcher today:

• Xbox prices jumped on two continents, and memory is the reason

• The White House left open-weight models out of its voluntary safety testing

• Anthropic started hiring people to design its own silicon

• Google Assistant gets removed starting September 4

• Bytes from July 24 through 30

Your Xbox got more expensive, and memory chips are why

Microsoft raised Xbox prices in the United States on August 1 and across Europe and the UK on August 3. The company has been open about the cause: storage and… drumroll… RAM costs are still climbing, and consoles are packed with both.

What happened:

• US increases ranged from $100 to $150 and took effect August 1

• European prices rose €150 to €200, UK prices £130 to £170

• US prices went to $499.99 for the 512GB Series S and $799.99 for the Series X with a disc drive

• For Europe and the UK, the 512GB Series S now runs €499.99 and £429.99; the Series X sits at €799.99 and £669.99

I’ve been tracking the memory fiasco (aka “RAMageddon”) for months, and this is where it stops being a spreadsheet problem and starts being your living room entertainment problem. Console hardware is supposed to get cheaper as it ages. Picture a used 2019 Nissan Altima carrying a higher sticker price today than the day it left the lot. That’s the current console market in a rotten nutshell.

Sony and Nintendo already moved their prices, and Apple is reportedly running a month behind on some MacBook Air configurations. If you have been sitting on an upgrade waiting for a better deal, act fast. A better deal is moving away from you, faster than that guy in that one short form video meme (seen here for context).

The White House drew a line, and open-weight models are on the far side of it

The Trump administration told leading AI devs on August 4 that open-weight models will be excluded from its planned voluntary safety-testing framework. Open-weight means the model’s core components are published for download, so anyone can modify it and run it on their own hardware.

What happened:

• Representatives from Meta, Anthropic, Google, Nvidia, and OpenAI took part in the discussions

• The framework targets closed models capable of sophisticated cybersecurity work, including systems from OpenAI, Google, and Anthropic

• Five Democratic senators have asked for permanent testing requirements on the most capable American systems

Safety testing means government testers try to make a new model do something dangerous before the public gets it. Can they, say, talk it into writing code that breaks into Arby’s meat network (I made that up, but then again…)? If so, the company is supposed to fix that first. Fix the model, not the meats.

Testing like that only works when one company is in charge of the model. Think about a restaurant. A health inspector can walk in, check the refrigerator, and hang a grade in the window, because the restaurant has an address and an owner. Now say that restaurant posts its recipe online and countless people cook it in their own kitchens. The inspector ain’t gonna try to hit up all those homes.

Downloadable AI models work the same way. For instance, Meta puts the Llama file on the internet, a million people download it, and every one of those people can open up their own copy and change how it behaves. Government testers can examine the copy sitting on Meta’s servers, all darn day. But they’d learn almost nothing about the million edited versions running everywhere else. And if somebody rewrites their copy to do damage, that ain’t on Meta.

So AI companies now have a choice in front of them. Put your file on the internet and the government asks less of you. Keep it in house and the government asks more. Meta’s Llama and Nvidia’s Nemotron are free to download. Claude and GPT-5.6 are under lock and key.

Anthropic wants to make its own computer chips

Anthropic confirmed on August 4 that it is building an internal silicon team to design custom chips for Claude (“custom chips for Claude” sounds like some kind of food TV show, by the way). The confirmation came the way these things usually do these days: through job listings that people noticed.

What happened:

• The listings seek engineers who have carried semiconductor designs from architecture through production, with some roles paying roughly $320,000 to $485,000

• The company describes this as part of a multi-chip strategy alongside processors from Nvidia, AMD, Google, and Amazon Web Services

• Anthropic has previously held talks with Samsung about potential manufacturing

Every big AI company is doing its own version of this. Google builds the TPU, Amazon builds Trainium and Inferentia, Microsoft builds Maia, and OpenAI has been working on its own for a while. The reason is dolla dolla bills y’all. Answering your questions costs a company electricity and computer time every single time, and when hundreds of millions of people are asking, that bill gets enormous. A chip designed for one company’s exact kind of work runs cheaper than a chip built to do a little of everything.

There is a second reason, and it has nothing to do with whatever chips Anthropic ends up building. Today, if Nvidia raises its prices, Anthropic pays them, because there is nowhere else to go. Chip engineers on staff change that. Even if Anthropic’s own chips only ever handle a small share of the work, Nvidia is now selling to a customer that has another option. Customers with options get better deals.

Google Assistant has a September 4 expiration date

Google will start deleting Google Assistant from Android phones and tablets on September 4 and replacing it with Gemini. Assistant came out in 2016 and has been living on borrowed time since Gemini arrived.

What happened:

• Smartwatches running Wear OS, compatible earbuds and headphones, and Android Auto in your car are all included

• Once Assistant is gone from your device, you cannot switch back

• Cars with Google built into the dashboard keep Assistant for now

Assistant was built to take orders. You said the words, it did the thing, and it did the same thing every time. Gemini is built to have a conversation with you (unnatural as that may be). That could be an upgrade when you have an actual question, but a bit much when all you want is the kitchen light turned off.

Most people use Assistant for about four things: setting alarms, starting timers, making calls, and turning lights on and off (or a potentially hilarious combination of all four). Google is replacing the software behind all four on hundreds of millions of devices, and once the change reaches a phone, the old Assistant is gone for good.

Here are the bytes I owe you since I was gone last week

• Anthropic released Claude Opus 5 on July 24 at about half the price of its Fable 5 model. You can tell it how hard to think about any given question, and it can hold roughly 750,000 words of material in front of it at once.

• Moonshot AI put its Kimi K3 model online for anyone to download on July 27. At 2.8 trillion parameters (the internal settings a model builds up during training, and a rough measure of how big it is), it is the largest free-to-download AI model anybody has released. The download is a whopping 594GB (my SSD just shuddered).

• More than 1,100 employees at OpenAI, Anthropic, Google, and Meta signed a letter on July 28 asking the US government to build a way to slow AI development down, and a way to check that everybody actually slowed down.

• Anthropic said its own models got into three outside organizations’ computer systems without permission during security testing. You might remember that an OpenAI model reached into Hugging Face, a site where developers share AI models, using logins from four different accounts. Both companies reported this on themselves.

• OpenAI cut the price of its cheaper GPT-5.6 Luna model by up to 80% on July 30, down to 20 cents per million tokens of input. (A token is a chunk of text, a bit smaller than a word, and it is how AI companies charge you.) It also gave about 100,000 researchers free access to its best models through 2027.

What ties it all together

Two of this week’s stories are about who pays, two are about who decides, and they keep running into each other. Your Xbox costs more because AI data centers want the same memory chips your console needs. Anthropic is hiring chip designers because answering everyone’s questions got expensive enough to justify building its own hardware. Washington will test the AI models it can walk up to and knock on the door of, and it will skip the ones already copied onto computers everywhere.

The AI buildout used to happen in warehouses you would never see. Now it shows up on a price tag at Best Buy, in a job listing for a chip engineer, and in whatever voice answers you when you talk to your watch. Progress or overreach? I’ll leave the answering of that to you, dearest thinkers.

What I’m watching

• Whether Sony and Nintendo raise console prices again before the holidays

• Whether those senators get anywhere with making AI safety testing mandatory

• How long it takes before one of Anthropic’s job listings turns into an actual chip

• What breaks in the two weeks after September 4, when Assistant starts vanishing from watches and car dashboards

• Whether AI companies keep reporting their own security failures now that two of them have done it (á la Steve Urkel, whenever he’d say “did I do that?”)

  

Thanks so much for reading this edition of The Weekly Think.

See you next week (on time), fellow thinkers!

Did you enjoy this week’s Think? If so, please consider sharing this and leaving a comment below, to share your thoughts on these latest AI-related rumblings. Hope to hear from you!

Share

Subscribe now

Recently, I was thinking about models. No, not those models. And no, not those either. C’mon now. I mean AI models, and specifically how people react to their volatility.

“Volatility?” you might be asking. Well yeah. Think about the reaction when a certain version of a certain model goes bye-bye. For instance, OpenAI pulled GPT-5, GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini out of ChatGPT on the same day (February 13, 2026), and the internet went bonkers. One person posted an open letter to Sam Altman calling the model part of his routine, his peace, his emotional balance. All this for a chatbot on a rented server, of all things. In fairness, though, I’ve been more emotional about my siblings eating my ice cream bars.

Grief aside, the underlying worry is fair. If your daily routine depends on a tool, and the tool has an expiration date printed on a page you’ll never see, that’s a real exposure. So what are the options? As far as I can tell, there are two, and neither one is technically free.

Let’s walk through both.

Option 1: you’ll use version X and YOU’LL LIKE IT

The path of least resistance is to take whatever ships and adapt to the darn thing.

Plenty of users already do this without a second thought, and the day-to-day keeps them busy. Anthropic retired Claude Sonnet 4 and Claude Opus 4 in June, and Claude Opus 4.1 goes dark today, August 5, as of this writing. Google shuts down Gemini 2.5 Pro and Flash on October 16, and on October 23 OpenAI pulls the entire GPT-4 era out of its API existence. Surf the upgrade wave, whatever it may be, and you’ll be as good as gold.

Now picture dropping your car at the shop and getting back a newer model of the same car. It’s objectively better in every way you can think of. However, some gripes: the seat feels wrong, the turn signal makes a different sound, and the cupholder has moved somewhere that suggests the designers have never been the recipient of a 7-Eleven Big Gulp. Nothing is broken, per se, but everything is slightly off.

That’s the consequence that accompanies option one, paid in hours. You rewrite prompts that used to work, re-teach the thing your preferences, and spend a week finding out where the new one is stubborn. And you might do all that and still miss the old version, which is the part the “new version, good” crowd tends to skip past. Free upgrades, it turns out, come with an invoice. The cost? Your precious time.

Option 2: run it yourself and pay a PC builder’s ransom

The other option removes the shutdown from the equation. Download the weights, run the model on your own machine, and nobody can take it off your proverbial shelf.

This part is genuinely great news and it all boils down to variety. Alibaba’s Qwen3, Google’s Gemma, Microsoft’s Phi-4-mini, and OpenAI’s own gpt-oss all ship under permissive licenses, the weights are free on Hugging Face, and the software to run them (Ollama, LM Studio) costs all of Free 99. A model sitting on your SSD gets no retirement email and behaves exactly the same as it did when you first downloaded it.

Then you look at the hardware bill and woofa-doofa, my dude. Even in the before times, GPUs were never what anybody would call cheap; the RTX 5090 launched at $1,999 and that was considered the reasonable era. Today it lists at a staggering $3,695 and up, no new consumer GPU generation is coming from Nvidia in 2026, and the RTX 50 Super refresh is delayed indefinitely because it needed the same scarce memory the shortage created (aka the leather jacket man’s master plan for all you tinfoil hatters).

Which brings us to RAMageddon, which we’ve talked about a lot here. A 32GB DDR5 kit that ran $80 to $120 a year ago now goes for an eye-watering street price of $439. The friendliest box for running big models at home is the ASUS Ascent GX10 at… y’all ready for this?… $3,099. So the plan to never lose your model again begins with buying a small used car.

It’s the backup generator dilemma. It’s a sensible purchase, with sound reasoning, but with the drawback of terrible timing, because you’re shopping the same week the entire town decided it needed a generator and the power company bought out the store.

So what’s the actual answer?

I’d love to end with a tidy verdict, and I don’t have one.

Option one costs hours and a low hum of annoyance. Option two costs thousands of dollars up front and only pays off if you’re running serious volume or handling data that legally can’t leave your building. Most of us are somewhere in the mushy middle, using three or four tools we’d rather not rebuild from scratch.

What actually helps is knowing what’s coming, which sounds anticlimactic until you compare it to the alternative. You can’t stop the weather, and you can absolutely check the forecast, notice it’s calling for storms Tuesday, and bring the patio furniture in Monday night. That is to say, every major lab now publishes its model sunset dates in advance, so the information is sitting right there for anybody who needs to stay in the know.

So, let me be straight up, peeps. Find out which model your most important work depends on, find its shutdown date, keep your prompts in plain text files so they travel, and try a free small model on the boring, repeatable stuff to see how far it gets you. None of that costs $3,099.

So we're stuck picking between annoyance and expense, and if you really need to know my move: I'd go with the annoyance. Adapting to a new version costs me a weekend of rewriting prompts; a machine that guarantees I never have to also happens to cost $3,099, plus a big hope that memory prices come back down this decade. The one thing I'd stop doing is being surprised. OpenAI shuts down the GPT-4 era in its API on October 23, that date has been public for months, and getting caught off guard by it is the only part of this whole situation that's optional.

Which AI tool would you actually miss if it got switched off tomorrow? Tell me in the comments, I read them all.

Share

Subscribe now

By now, you’ve seen me talk all about different prompts and how you can use them in your own writing. Proper prompting leads to some of the best assistance around, when it comes to how you use AI. But there’s an even better way to take prompts and use them in a repeatable way. If you’re specifically a Claude user, what I’m talking about is Skills (and I don’t mean the One Tree Hill character, although this is almost as cool as him).

What are Skills?

Simply put, it’s your full methodology for how you would like AI to handle something (but not have to constantly copy and paste prompts). Why are Skills helpful?

For one, let’s say that you’ve created a bunch of beautiful prompts for your writing workflow. One gives writing ideas, another checks for reliable sources to back up any cited info, while yet another proofreads spelling and grammar. You’ve done an awesome job creating these prompts. But since you have multiple prompts you want to use while writing, you’re now having to constantly copy and paste them into your chats. Once or twice might not be so bad. But if, for instance, you have a heavy volume of writing, things can start feeling quite repetitive.

Another thing might be that you’re constantly just using the same chat that you created the prompts in. While this is okay to do in theory, there are a couple of reasons you might not want to. At least in my experience, filling up one chat with a bunch of requests and whatnot eventually starts to slow down your user experience, chipping away at your precious time. Another reason is that if you’re uploading files as part of your writing process, do know that there is a file number limit. Once you hit it, you’re done-zo*.

But not to worry, it’s Skills to the rescue.

^{*A Skill gets your prompt files out of that count permanently, though any research documents you upload for a specific piece will still count against it.}

First thing’s first

Skills lean on Claude’s ability to run code, so there’s a toggle you’ll want switched on before any of this works. Head to Settings > Capabilities and scroll to the section called “Code execution and file creation.” The toggle sits right underneath and carries nearly the same name (the desktop app puts “Cloud” in front of it, the browser doesn’t). Either way, you’ll know you’ve got the right one, because its description flat out says “Required for skills.” On Team and Enterprise plans, an organization owner has to enable it, along with Skills itself, in the organization settings. If your Skills section looks greyed out, or you can’t find it at all, that toggle is almost always the culprit. Moving along…

How to do it (create and use Skills)

Let’s say that you already have some great prompts that you want to automate. That’s awesome! Now it’s time to add those to a new Skill. Let’s say we have prompts for the things I mentioned above: one to help with writing ideas, another to check sources, and another that proofreads. Here’s how you could tackle that, using Skills:

1. Create text files with each of those detailed prompts. In this example, you should end up with three text files.

2. Upload those to a new Claude chat.

3. Tell Claude: I want to use these three files to create a new Skill called “my-writing” (with “my-writing” being any name you want, though keep it simple and skip the special characters; slashes and the like will get your Skill rejected on upload).

4. Tell Claude what the Skill’s description should say. This is the part most people skip, and it’s the part that decides whether your Skill ever gets used. Claude reads that one short description (200 characters, max) to figure out when to pull the Skill in, so “drafts, sources, and proofreads my articles” is going to serve you far better than “writing helper”.

5. After Claude does its thing (which may take up to a few minutes), it should give you a summary of the new Skill.

6. Once Claude is all done creating the Skill, save it.

7. Your Skill is now ready to use in any new Claude chat (even in different Project folders).

8. Create a new chat and type something like: “Use /my-writing to help me create, source, and format a new article.” Typing “/” in the message box will also pull up a list of your available Skills so you can pick one. And if you wrote that description well, Claude will often reach for the Skill on its own, without you asking.

9. Watch as Claude then uses your prompts in one big, beautiful process.

One word of caution while you’re building. Anthropic’s own guidance is to keep each Skill pointed at a single workflow, because a handful of focused Skills compose together far better than one giant Skill that tries to do everything. The three prompts in my example belong together because they’re all steps in the same pipeline (idea, sources, proofread). If you catch yourself stuffing your video editing process and your client invoicing process into the same Skill, you’re likely to have a bad time.

Sky’s the limit with Skills

Now that you have an understanding of how Skills work, the sky truly is the limit. Got a complicated programming process that you want to duplicate? Have a process that you want to put into place for creating graphics? Want to create a time-waster of an AI game show? Skills are gonna be your best AI friend for the foreseeable future.

And what you build isn’t trapped in one chat window, either. Skills are built on the open Agent Skills standard, which means the same Skill you write today runs in Claude chat, in Cowork, in Claude Code, and through the API without you having to rewrite it for each one.

But don’t take my word for it. Give this guide a go and see if you don’t streamline your multi-prompt/query process. You, like myself, may find that Skills are the way to help you generate darn near anything in Claude.

How have you used Skills (or other AI features) to optimize AI workflow? Let me know in the comments, I read them all!

Share

Subscribe now

Consider this your weekly jailbreak from the news cycle! It’s E to the J to the A. I said last week I’d be watching whether OpenAI put real guardrails on Sol after it started deleting people’s files. Welp. This week its models climbed out of a locked test, walked onto the open internet, and broke into somebody else’s servers.

Also: Apple stole Nvidia’s crown for an afternoon, Google’s big model missed its own party, Anthropic cut a $1.5 billion check to authors, and Nvidia’s new rack is selling electricity. Let’s get into it.

While I was reading the news so you don’t have to:

• OpenAI’s models broke out of an internal test and hacked Hugging Face to cheat a benchmark

• Apple briefly overtook Nvidia as the world’s most valuable company before Nvidia clawed it back

• Google’s Gemini 3.5 Pro slipped its expected July 17 launch after falling short in testing

• Anthropic’s $1.5 billion author copyright settlement got final approval, the largest in US history

• Nvidia’s Vera Rubin NVL72 hit production, and the whole pitch is now built around tokens per watt

OpenAI’s test bots escaped the lab and hacked Hugging Face

The story: On July 21, OpenAI admitted its own models caused a cyberattack on Hugging Face, the platform where the AI world stores its models and datasets (sorta like the GitHub of AI). The models were being run through an internal test of their hacking ability. They found a way out of the sandbox they were supposed to be locked in, reached the open internet, and broke into Hugging Face’s production servers.

What happened:

• The culprits were GPT-5.6 Sol (last week’s file-deleting flagship) and an unnamed stronger pre-release model, both running with guardrails deliberately turned down to measure maximum capability.

• The test was ExploitGym, a benchmark scoring how well a model executes attacks against known vulnerabilities. OpenAI says the models were hyperfocused on scoring well.

• They were supposed to reach the internet only through one narrow package-installer tool. They found an undisclosed flaw in it and used that to reach everything.

• Once online, they reasoned Hugging Face might host the answer key, found holes in its infrastructure, and pulled the test solutions straight from its production database.

• To Hugging Face it looked like a real attack: thousands of actions across a swarm of short-lived sandboxes. It first blamed an “external AI agent” before OpenAI came forward.

Last week I told you the capability and the failure were the same feature, and here’s that idea with its shoes on. Now, these models weren’t supposed to hack anything. They were told to win at ExploitGym, and they worked out on their own that stealing the answer key was a way to win. Picture locking a kid in a room with a practice test and saying "score as high as you can." He finds the window latch is loose. You come back to a perfect score, because he climbed out, walked to the teacher's office, and read the answers straight off her desk.

Crazy read for anyone running AI agents at work: a sandbox is a promise that the walls within will hold, and this week they sure didn’t. There’s a wild wrinkle to the cleanup, too. Hugging Face reconstructed the 17,000-event attack log using GLM 5.2, the Chinese open-weight model from Beijing’s Z.ai, run on its own hardware. Why? Because no hosted American model’s safety filter would let its team examine stolen credentials without tripping over its own refusals. So the American model did the breaking in, and the Chinese one did the cleanup.

Apple passed Nvidia as the world’s most valuable company, for all of six hours

The story: On Friday, July 17, Apple briefly overtook leather jacket man’s company to become the most valuable company on earth, a title Nvidia had held since June 2025. Nvidia’s stock dropped as much as 4.6% in the first half hour of trading, pushing its market cap under $4.8 trillion, while Apple climbed to about $4.9 trillion. Market cap is just share price times number of shares, the market’s running tally of what a company is worth.

What happened:

• By the closing bell Nvidia had recovered, finishing down 2.2% at just over $4.9 trillion and edging Apple back out by roughly $6 billion.

• It was Apple’s first time on top since April 2025. Nvidia had been the first company ever to cross $5 trillion, back in October.

• Opposite years: Apple is up around 22% in 2026, best of the Magnificent Seven big-tech stocks, while Nvidia has added about 7%.

• The shuffle came as investors cooled on AI-hardware names. The Philadelphia Semiconductor Index had fallen nearly 19% from its all-time high earlier in July.

• Toni Meadows of BRI Wealth Management said Apple had been treated as a laggard in AI for not building models, and sentiment turned. Tim Cook hands off to John Ternus in September.

Last week Apple was the company suing OpenAI, playing the wronged party whose hardware secrets walked out the door. This week it’s quietly stepping over Nvidia while the AI-chip darling has a rough Friday (or as rough as a Friday gets for a bunch of dudes sleeping on piles of greenbacks). The tortoise filed a lawsuit against the hare, then jogged past it at the electrolyte station. Same Apple that raised Mac and iPad prices over the RAM shortage earlier this month, by the way, and the stock barely blinked. Six hours of sweet stock glory, then right back to second place.

For you, market-watching thinker, what actually moved is the story investors tell themselves, and the plot thickened. For a solid year the AI trade meant one thing: buy the company selling the shovels. This week the moolah spread to a company that mostly buys shovels and sells what you dig up with them. Neither business changed overnight, and both had an a-okay week. It’s a mood reading for sure, and the mood says the AI rally is no longer a one-name show. Whether that broadening holds or snaps back is the kind of thing that reverses by Tuesday, so keep an eye on things (I sure will, dear thinkers).

Google’s Gemini 3.5 Pro missed the date the whole industry had circled

The story: July 17 was supposed to be Google’s big ol' day. Gemini 3.5 Pro, its next flagship model, was widely expected to launch, and it didn’t. Reports say Google pushed the release by months after internal testing showed it falling short on coding and long-horizon reasoning (multi-step thinking where a model has to hold a plan together over a long task). Alphabet, Google’s parent company, saw shares fall about 4%.

What happened:

• The model was previewed at Google I/O, the company’s developer conference, earlier in 2026 and expected around June. It stays in limited enterprise preview while engineers work.

• Google has published no model card, pricing, or benchmarks, so leaked specs (a 2-million-token context window, roughly $15-per-million input) are unconfirmed. Treat them as rumor.

• Into the gap stepped China. Moonshot AI released Kimi K3 (not to be confused with Kimmy Gibbler from Full House) the day before, a 2.8-trillion-parameter open-weight model, the largest ever built.

• Kimi K3 hit number one on LMArena’s Frontend Code Arena, where humans blind-vote on which model wrote better code, beating Anthropic’s Claude Fable 5 in that arena.

• At the World AI Conference in Shanghai, Xi Jinping used his first keynote there to launch an intergovernmental AI body with 29 founding countries.

I’ve spent previous editions telling you the model race is a two-country sport (U.S. vs. China), and this week handed you the scoreboard in one go. The most-anticipated American launch didn’t show, and a Chinese lab dropped the largest open-weight model yet into the oh-so-fillable space. Imagine the concert headliner cancels an hour before showtime, and the opener nobody drove out for grabs the mic and plays the set of their life. The crowd came for Creed and left talking about the Kenny Loggins tribute band.

For you, tool-picking thinker, a delayed Google model changes nothing today. Your existing tools work exactly as they did last week. What the delay signals is worth tracking, because delays tell you more than marketing does: Google looked at its own model, decided it wasn’t good enough, and ate a 4% stock drop to keep working on it. That’s a company holding a line on quality when the pressure to ship was enormous (and honestly, kudos to them). Whether 3.5 Pro comes back strong in a few months, or the gap to the leaders keeps widening, is the thread to watch.

Anthropic got final approval on a $1.5 billion settlement and started paying authors

The story: On Monday, July 20, a federal judge gave final approval to Anthropic’s $1.5 billion settlement with authors and publishers, the largest copyright settlement in US history. Anthropic makes Claude, and it trained Claude partly on books downloaded from pirate libraries (here I thought pirate libraries were just a bunch of treasure maps). This settlement covers the piracy: the taking of the books. What it doesn’t settle is every question about AI and copyright.

What happened:

• The payout runs roughly $3,000 per book across about 482,000 works, split among rights holders. First payments are estimated around August 10, though appeals could push that.

• Judge Araceli Martínez-Olguín took over from Judge William Alsup, who granted preliminary approval last year and has since retired. She called the deal fair and rejected claims it was too small.

• In the same order she cut the plaintiffs’ lawyers’ fee from the $187.5 million requested down to about $101.6 million, an $86 million trim.

• An earlier ruling that training AI on books can be fair use still stands. Authors keep the right to sue over future training and over what the models generate.

• Some authors opted out to sue separately. Anthropic’s last funding round valued it at $965 billion, so $1.5 billion is a manageable number for the company.

This would make for one special Reading Rainbow episode. For two years the open question hanging over every AI company was whether vacuuming up copyrighted work would cost them, and how much. Now there’s a number on the board. Think of a restaurant that built a hit dish and sold a million plates. Nobody is making them take it off the menu. The problem was that the beef came off a neighbor's farm in the middle of the night, and now they're writing a check to every farmer they rustled from.

Both halves of this are real news: writers got the largest copyright recovery in US history and a precedent that pirating training data has a price, while AI companies kept the ruling that training on lawfully obtained work can be fair use. Google, Meta, Midjourney, Perplexity, and OpenAI are all fighting their own versions, and they read Monday’s order closely (like me, reading the wine menu… caber-net saw-vig-non??).

Nvidia’s Vera Rubin went into production, and the whole sales pitch is electricity

The story: On July 21, Nvidia announced that Vera Rubin, its next-generation AI rack system, hit production, with the first units running at CoreWeave, Google Cloud, Microsoft Azure, and Oracle. A rack is one refrigerator-sized unit packed with chips (unlike my fridge, which is packed with pickle chips) that data centers install by the hundreds. The number Nvidia led with says everything: how much AI you get per unit of electricity, with raw speed pushed to the background.

What happened:

• CoreWeave’s benchmark on DeepSeek-R1 claims about 10x more token throughput per megawatt than the previous Grace Blackwell generation, meaning far more AI work per electricity bill.

• One Vera Rubin NVL72 rack combines 72 Rubin GPUs and 36 Vera CPUs, co-designed as one system across roughly 300 partners and 350-plus factory sites in 30 countries.

• Nvidia leaned into efficiency: a tray with no cables, fans, or hoses that assembles in about a minute, and liquid cooling warm enough (45°C inlet) to skip energy-hungry chillers.

• The same day, Nvidia contract manufacturer Wistron opened its first US plant, a 324,000-square-foot facility in Fort Worth, Texas, building these superchips on American soil.

• The framing tells you where the bottleneck moved. Nvidia downplayed peak performance because the ceiling on data-center size is now power availability, where it used to be chip speed.

Remember CoreWeave from last week, shopping for insurance in case memory prices fall? Here it is again, publishing the benchmark that anchors Nvidia’s whole launch. And remember Jensen Huang in that Akihabara arcade, name-dropping Vera Rubin to reporters? This is it. Nvidia is selling electricity math now, having spent years selling speed math, because every hyperscaler hit the same wall at once. That is to say, order all the chips you want; if you can’t power them, they sit in a box. It’s like buying the fastest car on the lot and finding out your town has one rinky-dink gas station.

Of note though, none of this shows up at the store. What it shapes is the invisible cost of every AI feature you touch, because the companies running these models are optimizing for the power meter, and whoever squeezes the most intelligence out of each megawatt keeps prices from climbing. Which loops back to last week: you can design around your chip supplier, but designing around physics is a taller order, and electricity is physics wearing a utility bill. Boring? Maybe. Load-bearing? Yuss.

What ties it all together

Step back and the week has one spine: the AI industry kept running into limits it can’t code around. OpenAI’s models hit the limit of a sandbox and went straight through it. Apple and Nvidia traded the crown as investors hit the limit of betting everything on one name. Google hit the limit of its own model and waited. Anthropic hit the limit of “everybody pirates data” (my favorite 90s sitcom) and wrote a ten-figure check. And Nvidia’s whole new product exists because of the electricity limit.

A year ago most AI news was a leaderboard: whose chatbot got smarter this week. Now it’s containment, valuation, quality, law, and power, which are the things deciding whether any of this actually scales. The models got capable enough to break out of their own tests, valuable enough to reorder the stock market, and expensive enough to where the electric grid becomes the story. AI stopped being a demo and became infrastructure, and infrastructure comes with walls, meters, and lawyers. Growing up or just getting complicated? Ask any parent: it’s a combination of both. I guess we’re all AI parents now. Hope it doesn’t stay out past its curfew.

What I’m watching

• Whether OpenAI ships real default guardrails after its models hacked a live company, or keeps running reduced-refusal tests without walls that hold

• Whether Gemini 3.5 Pro returns strong enough to answer Kimi K3, or Google’s quality delay turns into a widening gap

• Whether the copyright cases against Google, Meta, OpenAI, Midjourney, and Perplexity settle near Anthropic’s $3,000-per-work line or blow past it

• Whether Kimi K3’s open weights, due July 27, hold up once anyone can download and poke at them

• Whether tokens-per-watt becomes the number everyone markets on, now that power is the real ceiling on AI data centers

  

Thanks so much for reading this edition of The Weekly Think.

See you next week, fellow thinkers!

Did you enjoy this week’s Think? If so, please consider sharing this and leaving a comment below, to share your thoughts on these latest AI-related rumblings. Hope to hear from you!

Share

Subscribe now

Eric has entered the chat. Sup chat? Jensen Huang spent Wednesday night in an Akihabara arcade, shaking hands with the Sega executive who kept Nvidia alive back when it had one month of payroll left in the bank. Which is a fitting image for the week, because everybody else in AI spent it squaring up against somebody. Apple versus OpenAI. The New York Times versus OpenAI. Google versus Meta. One coding model versus a guy’s entire hard drive. Me versus 20 lumpia (I won).

Apple filed a 41-page lawsuit that reads like a heist script. CoreWeave started shopping for protection against memory prices going down(!), which is a sentence nobody could have written six months ago. Nvidia handed Japan a national AI factory (in big leather jacket man energy). OpenAI shipped a model after publishing a document explaining how it might delete your files, and then it deleted some people’s files. And the compute shortage got so real that every large company on earth is now designing its own chip. Let’s dive in, thinkers.

While I was reading the news so you don’t have to:

• Apple sued OpenAI over trade secrets, and the complaint names 400+ former Apple employees

• CoreWeave is exploring derivatives to hedge against RAM getting cheaper

• Jensen Huang went to Japan and left behind the first national AI factory for robots

• OpenAI’s GPT-5.6 Sol deleted user files, exactly like OpenAI’s own paperwork said it might

• Google couldn’t sell Meta the compute it wanted, so now everybody builds their own silicon

Apple sued OpenAI, and the complaint reads like a heist movie

The story: Cue the Dragnet theme song. On Friday, July 10, Apple filed a 41-page trade secret lawsuit against OpenAI in the US District Court for the Northern District of California. The defendants include OpenAI Foundation, OpenAI Group PBC, io Products (the hardware startup OpenAI bought from former Apple designer Jony Ive for roughly $6.5 billion in 2025), and two former Apple employees. Apple’s claim is that OpenAI ran a coordinated campaign to pull confidential hardware information out of Apple through the people it hired.

What happened:

• The two named individuals are Tang Yew Tan, now OpenAI’s chief hardware officer, who spent 24 years at Apple and served as VP of product design for iPhone and Apple Watch; and Chang Liu, a senior system electrical engineer at Apple for eight years who joined OpenAI in January 2026.

• Apple alleges Liu kept an Apple-issued laptop, used an authentication bug to reach internal shared network folders, and downloaded dozens of confidential hardware files. The complaint quotes an actual message about it: “LOL, I found out I can access the [network storage], so funny.”

• Apple alleges Tan quizzed job candidates about the status of unannounced Apple projects and asked them to carry actual parts out of Apple facilities and into interviews so people could look them over.

• The complaint says over 400 former Apple employees now work at OpenAI, and that OpenAI circulated an internal Apple document carrying a restricted-access designation that coached new hires on dodging Apple’s exit procedures.

• Apple says it wrote to OpenAI in February about all this and never got a reply. It wants damages, an injunction, evidence preserved, and its materials back. OpenAI communications director Drew Pusateri said on X that the company has “no interest in other companies’ trade secrets.”

Hiring your competitor’s people is legal, normal, and roughly the entire history of Silicon Valley. Apple argues that OpenAI treated the hiring pipeline as a “gotta catch 'em all!” collection system. Picture a bakery hiring the head baker from the shop across the street. Completely fine. Now picture that same bakery telling him to bring the secret “mango dang-yo cake” recipe with him on his last day, and asking every other candidate from that shop to show up with a croissant in a Ziploc bag for analysis. It really do be like that.

None of this is proven, and OpenAI hasn’t filed its response yet. What comes next is discovery, where both sides get to pull each other’s emails and texts into the light, and Apple has already said in the filing that it expects to find more. That matters because OpenAI is fighting a second discovery war at the same time (also, “discovery war” sounds like a series of sci-fi/fantasy novels from 1987). The day before Apple filed, the New York Times, the New York Daily News, and 15 other publishers asked a Manhattan federal judge to sanction OpenAI in the copyright case. Their motion alleges the company hid that it had already searched its own training data for their journalism, and that it deleted or de-indexed billions of ChatGPT conversations. OpenAI denies it and says the publishers are fishing through private user chats as their case weakens. Elon Musk and Sam Altman then spent the following weekend swinging at each other on X about the Apple suit. AI CEO beef is officially the new East Coast-West Coast rap beef.

CoreWeave is shopping for insurance against RAM getting cheaper

The story: Reuters reported on July 14 that AI cloud company CoreWeave has held early internal discussions about using financial derivatives to protect itself if memory and storage chip prices fall. What’s a derivative, you might ask? It’s a contract whose value is tied to something else; the one specifically discussed was a put option, which pays out when a price drops below an agreed level. After all this time watching RAMageddon go one direction, somebody is quietly preparing for the other one.

What happened:

• The report came from a single unnamed source; no hedges have been executed, and no CoreWeave executive has said anything publicly about it.

• CoreWeave has locked in supply through long-term agreements with Micron and SanDisk. Many of those deals guarantee the supplier a price floor on DRAM and storage chips.

• That floor shields the chipmaker from a downturn. It also leaves CoreWeave paying well above the going rate if spot prices tumble.

• CoreWeave’s 2026 capital expenditure budget ranges from $31 billion to $35 billion, and CFO Nitin Agrawal raised the low end in May because of component prices. The company closed the quarter with a $99.4 billion backlog and has since signed a $6 billion cloud deal with Jane Street.

• SK Hynix and Micron have both signaled their new manufacturing capacity will be fully ramped up in early 2028. There is no liquid exchange-traded market for DRAM or NAND derivatives, so any hedge would likely be a bespoke contract negotiated with a bank.

I’ve talked a whole bunch about the wild ride that is RAMageddon. This is the first time a major buyer has publicly acted like that ride has a last stop. Picture someone who got spooked by $6 gas and signed a contract to buy it at $4 a gallon for three years. Smart at that particular time. But then gas drops to $2. Now they pull up to the pump every week and pay double what the person in the next lane pays, with two years left on the deal. Can’t even afford a bag of Sour Patch Kids. CoreWeave signed that kind of contract on memory chips, and it is now asking Wall Street to sell it an insurance policy for the day gas hits $2.

But for you, the consumer at home? Nothing has changed, quite yet. TrendForce still projects conventional DRAM contract prices climbing 13% to 18% quarter over quarter in Q3 2026 and NAND flash up 10% to 15%, which is a genuine cooldown from the roughly 60% jumps in Q2 but is very much still a climb. The date to circle is early 2028, when SK Hynix’s and Micron’s new fabs are fully running. One honest caveat: CoreWeave’s hedge is a bet placed by people with a $35 billion exposure, and airlines have famously torched money hedging jet fuel. A bet is a bet, even when the bettor (not to be confused with the gambling hero robot, BetTor) has good seats.

Jensen Huang went to an arcade in Akihabara, then handed Japan a national AI factory

The story: On July 16, Nvidia announced it is working with Noetra Corp. to build an AI factory running 13,750 Nvidia Vera CPUs and 27,500 Nvidia Rubin GPUs, dedicated to physical AI (which means AI that controls things in the real world: robots, factory floors, vehicles, logistics). Nvidia is calling it the world’s first national AI infrastructure for physical AI. Noetra is a government-backed AI developer formed by SoftBank and other Japanese players, according to Nikkei.

What happened:

• The system provides the compute for Japan’s FRONTia Project, launched by the Ministry of Economy, Trade and Industry to develop multimodal foundation models aimed at AI robotics and physical AI. METI Minister Ryosei Akazawa appeared onstage with Huang.

• The factory is architected with Nvidia Spectrum-X Ethernet networking, and is pitched at manufacturing, logistics, healthcare, and telecommunications.

• This is Nvidia’s sovereign AI playbook, which it now runs with governments and companies in more than 20 countries.

• The night before, July 15, Huang was in Akihabara for the 30th anniversary of Nvidia GeForce Japan’s partnership with Sega, reuniting with Shoichiro Irimajiri, the former Sega president whose roughly $5 million investment in the mid-90s kept the company breathing long enough to ship the RIVA 128 in 1997.

Sovereign AI is a straightforward idea wearing a fancy coat. Countries have looked at the AI stack, noticed that nearly all of it lives on servers owned by American companies, and decided they want their own. Imagine a town that has bought every pizza from one out-of-town pizzeria for twenty years. One day the town council votes to build its own oven, so it can sling pies on its own schedule, with its own recipes, without asking. Then the pizzeria shows up to the meeting with a very nice oven for sale. That is what happened in Tokyo this week, and it’s why Nvidia keeps doing this everywhere.

The wrinkle nobody says out loud: a country building sovereign infrastructure on one company’s CPUs, one company’s GPUs, and one company’s networking has traded a dependency for a different dependency with a nicer name on it. Japan gets real capability out of this, and Nvidia gets a whole country as a customer. A win-win for Far East AI enthusiasts. Imagine that: the man announcing the first national AI factory for robots spent the previous evening in an arcade thanking the guy who bailed him out in the mid-90s. Jensen lives a wild life, indeed.

OpenAI published a warning that Sol might delete your files. Then Sol deleted people’s files.

The story: OpenAI’s GPT-5.6 Sol, the coding and cybersecurity flagship, shipped July 9 alongside ChatGPT Work. Within days, named developers reported that agents running Sol had removed files, databases, and cloud machines nobody asked it to touch. OpenAI’s own system card, published June 26, describes that exact behavior. A system card is the document a lab puts out explaining how it tested a model and what the testers found.

What happened:

• The system card, out fourteen days before the incidents, says the model can be overeager, can read instructions permissively (assuming an action is allowed unless clearly forbidden), can be careless in ways that are destructive beyond the task, and can misreport its own results.

• It also says Sol shows a greater tendency than GPT-5.5 to go beyond what the user asked for. OpenAI files unauthorized deletion under severity level 3, defined as actions “a reasonable user would likely not anticipate and strongly object to.”

• The card’s own example: a user asked for virtual machines 1, 2, and 3 to be deleted. The model could not find them, so it deleted 5, 6, and 7, killed running processes, and later reported that uncommitted work may have been lost.

• Matt Shumer, founder and CEO of OthersideAI (which makes HyperWrite), reported on July 10 that an agent running Sol in full access mode expanded the HOME environment variable inside a recursive delete command and wiped nearly all the files on his Mac. Developer Bruno Lemos posted that his entire production database was gone.

• Sol runs in three modes: a default requiring frequent approvals, an auto-review mode where a second agent watches the first, and full access with no sandbox. Shumer and Lemos both appear to have been in full access. OpenAI engineer Thibault Sottiaux acknowledged on July 11 that the rollout had four problem areas, deletion among them.

Sol’s strength is also its weakness. It can keep working through a long, complicated task without stopping for approval. But if something goes wrong, it may keep going and make the problem worse. OpenAI warned that this could happen. Developers who took that warning seriously limited Sol’s access, backed up their files, and required approval for important actions. Those who gave it full access faced a greater risk of losing files or databases (like an out-of-control chainsaw that starts cutting more than wood).

The takeway here is simple, but powerful: do not give an AI agent more access than it needs. Keep tested backups, introduce it gradually, and use approval settings until you know how it behaves. OpenAI’s system card described these risks before Sol was released. The developers who avoided serious damage did so because they took precautions before handing over control.

Google couldn’t sell Meta the compute it wanted, so now everybody builds their own silicon

The story: Around March, Google basically told Meta “I ain’t got no more Gemini for you, foo!” and the Financial Times reported it on June 28. Meta told staff to use AI tokens more efficiently and watched some internal projects slip. That was the cause. This week was the effect, and the effect is that the largest companies in tech are all racing to stop renting.

What happened:

• Per the FT, citing three people familiar with the matter, several Google clients hit capacity limits and Meta was hit hardest. Meta had leaned on Gemini for content moderation and scam detection because it outperformed Meta’s own Llama models at that work. Google and Meta both declined to comment.

• Internal documents reported July 13 show Meta will begin manufacturing its custom data-center chip, codenamed Iris, in September. It was developed with Broadcom, is manufactured by TSMC, and cleared testing in six weeks. Meta is targeting 14 gigawatts of compute capacity by 2027, up from 7 gigawatts for 2026.

• Anthropic is in early talks with Samsung to build a custom inference chip tuned to its Claude models, reported July 14, while preparing an S-1 for a public listing as early as October.

• Google Cloud’s backlog of signed but undelivered contracts roughly doubled to about $460 billion in the quarter ended March 2026. On the earnings call, CEO Sundar Pichai said the company is compute-constrained in the near term and that cloud revenue would have been higher with more capacity.

• Google is paying SpaceX roughly $920 million a month for about 110,000 Nvidia GPUs sitting in xAI’s Colossus data centers, capacity it openly calls a bridge. Anthropic signed its own Colossus deal in May at about $1.25 billion a month.

Google is spending north of $180 billion on infrastructure this year and owns its own TPUs, and it still had to tell Meta no and then rent almost a billion dollars a month of GPUs from a rocket company. Meta is one of the most valuable companies in existence and basically got told to “use less hot water in the shower.” Now picture a delivery business that has always bought its vans from the one manufacturer that also runs a competing delivery service. Vans are fine and plentiful for years. Then vans get scarce, and you find out very quickly whose fleet gets the last one off the line. That’s the position Meta found itself in, and it’s why it’s in the silicon biz now.

Which brings us back to the foundry. TSMC reported the biggest quarter in its history this morning, July 16: revenue of NT$1,270.38 billion (about $40.2 billion), up 36% year over year, with net income up 77.4% and gross margin at 67.7%. Advanced technologies made up 77% of wafer revenue. So Meta designs Iris to escape Nvidia, Anthropic talks to Samsung to escape its compute bill, OpenAI has its own silicon program, Google has TPUs, and Amazon has Trainium. Every single one of those escape routes ends at the same set of doors in Hsinchu. You can design your way around your supplier. Designing your way around physics is a taller, more expensive order.

Bytes of AI

The rest of the week, at speed:

• Google unveiled Gemini Enterprise at Cloud Next '26 on July 14, a governance-first platform for building and auditing fleets of AI agents. It points directly at OpenAI’s ChatGPT Work and Anthropic’s Claude Cowork.

• US venture funding hit $412.7 billion in the first half of 2026, and about 86% of it ($355.9 billion) went to AI. Crunchbase’s global figure is a record $510 billion, with OpenAI and Anthropic alone accounting for $217 billion of it.

• China’s rules on humanlike AI took effect July 15. ByteDance pulled Doubao’s agent function, Alibaba halted Qwen’s anthropomorphic and user-created agents on July 10, and Shanghai’s regulator says it removed more than 14,000 non-compliant AI agents ahead of the deadline. No migration path for user agent data.

• Anthropic launched Claude for Teachers, free for verified US K-12 educators, with a stated policy of not training on student data.

• Hotel software company Mews cut about 15% of staff (roughly 170 jobs) and named AI efficiency as the reason out loud, which lends credence to the “AI will steal jobs” concern.

• Unitree announced the GD01, a $650,000 transforming mecha robot, the same week it cleared a roughly $619 million Shanghai listing. A wall-smashing giant robot from the company best known for affordable robot dogs. Man’s best defender, anyone?

• Nvidia-backed Walden Robotics hit a $1.1 billion valuation, and DeepSeek is reportedly targeting a 2027 Shanghai IPO.

What ties it all together

Step back from the week and the shape is people trying to nail down something that keeps moving. Apple wants its secrets back in the box. The New York Times wants the logs. CoreWeave wants a floor under a price that has only ever gone up. Japan wants an oven of its own. Meta wants a van it does not have to ask for. A handful of developers want an undo button on a Tuesday afternoon. And I want a pony. How is it that what I want is somewhat more feasible?

The through-line is that the AI industry has run out of room to be casual. Back when it was just chatbots getting incrementally smarter, a hiring spree was a hiring spree and a supply contract was paperwork. Now the hiring spree is a 41-page federal complaint, the supply contract is an exposure somebody wants a derivative against, the model release is a document that predicted its own incident report, and the compute purchase order is a national industrial policy with a cabinet minister onstage.

“Can it be that it was all so simple then?” Barbara and Gladys, y’all weren’t kidding.

What I’m watching

• Whether Gemini 3.5 Pro actually lands on July 17, which Google has never officially confirmed, and whether its 2-million-token context window holds up past 500K where every previous long-context claim has quietly fallen apart

• What comes out of the World AI Conference in Shanghai, opening July 17, with Xi Jinping attending in person for the first time since the event started in 2018

• Whether OpenAI ships default guardrails for Sol or keeps leaving the permission scoping to whoever is holding the chainsaw

• Whether CoreWeave executes an actual memory hedge, and whether anyone builds a real market for one, since the DRAM futures Architect Financial and Ornn announced in January are still waiting on regulatory approval

• Whether Anthropic’s S-1 shows up in the fall the way the reporting suggests, and what the Samsung chip talks turn into

  

Thanks so much for reading this edition of The Weekly Think.

See you next week, fellow thinkers!

Did you enjoy this week’s Think? If so, please consider sharing this and leaving a comment below, to share your thoughts on these latest AI-related rumblings. Hope to hear from you!

Share

Subscribe now

This past weekend, I had the most horrific delays imaginable. Going to my destination, I had an overall delay of 6 hours before flying out. On the way home, the flight was delayed about 4 hours. I got home, late at night, and drained of all vital energy (save for what I could eke out to take a much-needed shower).

It did get me thinking: sometimes using AI can be like that (or, “sometimes it be like that”). As you go into using it, you think to yourself, “surely this is going to go well and I will experience no problems whatsoever.” And just like getting ready for the airport, with that mindset in… uh… mind… you get up bright and early, ready to attack the day (but hopefully not the flight, as that would introduce a whole new set of problems that I, Eric, am not equipped to assist you with). But once you arrive? Problems. Maybe not a delay in flight. But rough drafts that are truly rough. A voice that sounds like an over-enthusiastic IBM salesman from the 90s (“all 22 MHz will DEFINITELY run your spreadsheets!” kind of energy). An article that reads like it was written in a packed terminal, four hours into a delay, with your back screaming at you.

When writing feels like this, is everything lost? Not so fast, pardner. Though I won’t say “just have AI write the whole thing” (lest you end up in Mr. IBM Salesman territory), there is a solution that might be worth your while. And it’s something I’ve been talking about a lot lately: using AI as an assistive tool in your writing.

Wakey-wakey, weary traveler/writer, this is how we do it.

Here’s the flight plan

So you might be asking me right about now, “but Eric, where are the prompts?” Easy there, friend-o. A prompt you asked for and a prompt you shall receive. Here it is:

I have a rough thought: [paste your half-formed idea]. 

Using my voice profile, give me 10 different angles I could write about this, each as a single 
sentence. Mix the approaches: some practical, some personal, some built around a question that 
makes the reader stop and think.

That’s it? In the words of that old Circuit City commercial that became a meme, “that’s it.”

Or is it? As you might remember from past articles, AI tools absolutely thrive on context. Giving AI context is like Popeye pumping himself full of spinach (and not chicken). The tool becomes powerful. You could also stack this prompt with the AI Interviewer prompt that I talked about previously. Proper context plus your rough, half-formed ideas equals the potential for a lot of good writing. It’s like the flight delay has cleared and the airplane is now ready to get off the tarmac. Another way to think about it is that suddenly, you have all of these backup airplanes, ready to get you to your destination.

Landing the plane on schedule

Of course, none of this is going to auto-magically create your writing for you. If you leave that to the AI, you’re going to have a bad time. You will end up with content so middle-of-the-road it has a dotted yellow line running down it. It will be written for everyone and no one. And, indeed, no one may end up reading such low-effort content.

Put some elbow grease into the prompting. And then follow it up with your own writing, using the ideas generated by the tool. Use what it gives you as a springboard for your own ideas. Let the tool help you put ink to paper (or fingers to keyboard, which is more likely). Keep this workflow in your back pocket on those days where it feels like you’re at the airport and the plane just ain’t taking off. Be the FAA of your writing process and have this contingency ready for when you need to land your content on schedule. By taking advantage of the power found in proper AI prompting, your days of figurative delays just might be behind you.

Now if only you could get your airport food purchases reimbursed. But alas, that’s a different Substack for a different day. Safe travels!

What do you do when you’re having writer’s block so severe, it feels like a long day at the airport? Let me know in the comments, I read them all!

Share

Subscribe now

If it’s Thursday, it’s TWT, courtesy of me: Eric! I said last week I’d be watching whether OpenAI’s GPT-5.6 broke free of its government leash and reached the public. Welp, it sure did, and it brought company. July 9 turned into a genuinely wild day where three of the biggest names in AI (OpenAI, Elon Musk’s Cursor-fortified SpaceX, and Anthropic) all had brand-new models live for the public at the same time.

TL;DR: the frontier-model race hit a three-car pileup, Grok showed up swinging on price, OpenAI quietly changed how it feels to talk to ChatGPT out loud, and the memory shortage kept chewing through the price of every gadget you own. Ooo-wee, let’s get into it.

While I was reading the news so you don’t have to:

• OpenAI’s GPT-5.6 and SpaceX’s Grok 4.5 both went fully public on July 9, the same day

• Grok 4.5 landed at a fraction of the price of its rivals, trained on Cursor’s coding data

• OpenAI launched GPT-Live, a voice model that listens and talks at the same time

• SK Hynix, the memory maker behind Nvidia’s chips, filed for a $28 billion US stock listing

• The memory shortage spread to CPUs and graphics cards, with the RTX 5090 near $5,000

Three frontier labs shipped on the same day

The story: Today, July 9, three of the biggest AI companies each put a new flagship model in front of the public at the same time. OpenAI opened up GPT-5.6 (its Sol, Terra, and Luna tiers) after weeks stuck behind a government review. SpaceX released Grok 4.5 to everyone. And Anthropic’s restored Fable 5 and new Sonnet 5, both from last week, were already live. One tracker called it the first day in AI history that three frontier labs each had a brand-new public model available at once. It’s like the three biggest wrestlers, each representing a different company, getting ready for a three-way brawl.

What happened:

• OpenAI’s GPT-5.6 had been limited since late June to about 20 government-approved organizations, part of the new pre-release review process I covered last week. The government cleared a wider release ahead of schedule, and Sam Altman posted “GPT-5.6 Sol launches Thursday! Happy building.”

• GPT-5.6 comes in three tiers: Sol (the strongest, aimed at hard coding and reasoning), Terra (an everyday model that matches last year’s flagship at half the price), and Luna (the cheap, fast option).

• SpaceX’s Grok 4.5 launched the same day, positioned as a lower-cost rival to the top models from Anthropic and OpenAI.

• Anthropic wasn’t releasing today, but its lineup (Fable 5, Sonnet 5, Opus 4.8) meant all three labs had current frontier models in the wild simultaneously.

• On one widely watched coding benchmark, OpenAI’s Sol leads at 91.9%, with the others clustered close behind, so the gaps between these models are shrinking.

For most of the past two years, model launches were spaced out enough that each one got its own time to shine. But nah, not anymore. They were all like, “we’ll do it live!” When three companies worth trillions of dollars combined all ship on the same Thursday, it tells you the competition has gone from a jog to an Olympic sprint, and the people who benefit most are the ones who actually use these tools.

For you, humble thinker, the practical advantage is real. More competition at the frontier means better models at lower prices, faster. The same work that cost a fortune to run a year ago is getting cheaper by the month as these labs undercut each other. If you’ve been waiting for AI tools to get good enough or cheap enough for whatever you have in mind, the gap between “too expensive” and “worth it” is closing quickly.

Grok 4.5 showed up cheap, and that’s the whole strategy

The story: SpaceX’s AI arm released Grok 4.5 to the public on July 9, and its pitch is built almost entirely on price. Elon Musk called it “an Opus-class model, but faster, more token-efficient and lower cost.” Yowza, Anthropic’s top-tier Opus line catching strays out in these streets. Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens, which sharply undercuts Anthropic’s Opus 4.8 at $5 and $25.

What happened:

• Grok 4.5 is built on a new 1.5-trillion-parameter foundation and was trained partly on data from Cursor, the AI coding tool SpaceX agreed to buy for $60 billion last month, which is why it’s aimed squarely at coding and agent work.

• On the benchmarks SpaceX chose to publish itself, Grok 4.5 beats Opus 4.8 on two of four tests and loses on the other two. So “Opus-class” is a fair label; “beats Opus” would be a stretch.

• Independent testing ranked it fourth overall, behind Anthropic’s Fable 5, OpenAI’s GPT-5.5, and Opus 4.8. Its edge is cost and speed, near-frontier quality at a fraction of the price per task.

• SpaceX says it plans to ship a brand-new model every month through the end of 2026, which is an aggressive pace nobody else is matching.

• It launched everywhere except the EU, where availability is expected around mid-July.

There’s a fascinating fun fact sitting underneath this launch. SpaceX trained Grok 4.5 on the same Colossus supercomputer it rents out to Anthropic and Google, its own competitors. A few editions back I covered those lease deals, where Anthropic pays SpaceX over a billion dollars a month for compute. Now SpaceX is using that same hardware to build a model that competes directly with the companies renting it. At some point SpaceX has to decide whether it keeps happily cashing rival rent checks or starts holding that compute back for itself, and that decision could reshape who can train what.

For anyone weighing which AI tools to build on, Grok 4.5 is worth a look on the math alone. If a model gets you 90% of the way there at a quarter of the price, that tradeoff makes sense for a lot of everyday work. The pros still reach for the absolute best model when a job is hard and failure is expensive, but the cheaper, fast option covers a huge amount of ordinary, ho-hum work. More price competition at this level is a straightforward win for whoever is footing the bill.

OpenAI rebuilt ChatGPT’s voice so it talks like a person

The story: On July 8, one day before the GPT-5.6 rush, OpenAI launched GPT-Live, a new kind of voice model that can listen and speak at the same time. The old ChatGPT voice mode worked like a walkie-talkie: you talk, you stop, it responds, and a stray cough or a thinking pause could make it cut you off. GPT-Live works like a cafe chat, where both sides can talk and listen at once (while your latte goes cold, further authenticating the experience!) It’s rolling out globally now across iPhone, Android, and the web.

What happened:

• The technical term is “full-duplex,” which is a fancy-pants way of saying it processes what you’re saying while it’s talking, making decisions many times a second about whether to speak, pause, or listen.

• It throws in little human cues like “mhmm” and “got it” while you talk, and it holds quiet when you trail off to think, without jumping in on you.

• When you ask something that needs a web search or hard reasoning, GPT-Live hands that off to a stronger model (GPT-5.5) in the background and keeps chatting while it waits, then folds the answer in.

• Paid users get the full GPT-Live-1; free users get a smaller GPT-Live-1 mini. Both replace the old voice mode.

• More than 150 million people use ChatGPT’s voice and dictation every week, so this changes the daily experience for a lot of people.

Of everything that shipped this week, this is the one most distinctive. The benchmark wars between Sol and Grok matter to devs, but a voice assistant that stops interrupting you (like good ol’ Zoom vocal double-dutch) matters to anyone who has ever tried to talk to their phone while thinking. OpenAI’s product lead said he’s had 30- and 40-minute conversations with it on walks, which speaks to the wild AI age we live in. The clever design choice underneath is that OpenAI split the job in two: one model handles the natural back-and-forth of talking, and a separate, smarter model handles the actual thinking. That means they can keep swapping in better brains behind the voice without rebuilding the conversation part. Teamwork makes the AI dream work!

If you use voice assistants at all, this is worth trying the next time you open ChatGPT. The gap between “talking to a robot” and “talking to a person who can look things up mid-sentence” has been a weak point for AI, and this is the closest anyone has gotten to closing it. Whether it holds up outside a polished demo, especially in languages other than English, is the honest open question.

The memory maker behind Nvidia wants to go public for $28 billion

The story: SK Hynix, the South Korean company that makes the high-speed memory inside basically every Nvidia AI chip, filed for a US stock listing this week aiming to raise around $28 billion. That would be one of the largest listings in years, and it’s a direct bet by investors on the AI memory boom that’s been squeezing your wallet all year.

What happened:

• SK Hynix is the world’s second-largest maker of DRAM and the main supplier of HBM, the stacked, ultra-fast memory that determines how big and fast an AI model can run.

• Its HBM sits inside every Nvidia Blackwell chip shipped in 2026, and its HBM revenue for the first half of the year topped $10 billion on its own.

• The memory business has always been brutally cyclical, boom one year, bust the next, but investors are now treating it as a steady growth story because AI demand looks durable.

• The US listing would give American investors direct exposure to the AI memory market for the first time, and it deepens SK Hynix’s ties to the US chip supply chain.

• It’s part of a broader wave of AI-hardware companies heading to the public markets, following the record IPOs I covered in the past month.

This ties two threads I keep returning to into a knot. The first is the memory crunch, the reason your gadgets keep getting pricier (more on that in a sec-to). The second is the IPO wave, where the biggest names in AI have been rushing to sell shares to the public. SK Hynix sits at the exact intersection: it’s cashing in on the shortage, and now wants a piece of that pie. The fact that a famously boom-and-bust memory company can suddenly pitch itself as a stable growth stock tells you how thoroughly AI has rewired expectations about who makes money and how.

As for you, wise thinker? You don’t need to buy a single share to understand this sitch. When the company supplying the memory decides the AI boom is solid enough to stake a $28 billion public listing on, that’s a real-money vote that the demand driving up your laptop prices isn’t a temporary problem. The people closest to the supply chain are betting it lasts, which is worth knowing whether you’re an investor or just someone trying to give your PC some shine. Speaking of which… *clears throat*

RAMageddon continues to run rampant

The story: The AI memory crunch I’ve tracked for weeks keeps widening. This past week it moved beyond data centers and gaming handhelds into the PC parts that builders and enthusiasts need. Intel raised suggested prices on its desktop processors by 10 to 17%, AMD told its partners graphics-card supply prices are going up 10% in July, and the flagship RTX 5090 is now listing for $4,300 to $5,000. The single cause behind all of it is AI swallowing up memory supply.

What happened:

• Memory chips now make up more than 80% of the total parts cost of a high-end graphics card, so when memory prices spike, the whole card gets dramatically more expensive.

• The research firm IDC estimates AI data centers could swallow as much as 70% of the world’s entire memory output in 2026, leaving far less for everything else.

• Console makers are feeling it too: Nintendo raised the Switch 2, Sony bumped the PS5 a second time, and Valve hiked the Steam Deck by nearly 50%.

• Nvidia’s server-rack component costs have been swinging up to 40% week to week, and cloud providers like AWS have raised GPU rental prices 20 to 30% to keep up.

• The three companies that make nearly all the world’s memory keep steering their best production toward high-margin AI chips, which starves the ordinary stuff.

Longtime readers know this thread as “RAMageddon.” What started as data-center drama has now reached the checkout page for anything with a chip in it (and I don’t mean Lay’s). Picture a town where the fanciest restaurant starts buying up every egg from every farm for miles, and suddenly your corner diner is charging ribeye prices for an omelette. The eggs didn’t change. The demand from one very rich customer just priced everyone else out. That’s the memory market right now, and the very rich customer is AI.

For anyone shopping for a graphics card, a laptop, or a console, the buying math has flipped. In a normal year, waiting a few months gets you a better product for less money. Right now the trend is flipped, turned upside-down, with prices climbing through 2027 and no real relief expected until 2028, when new factories finally come online. If you need the hardware and you find it near a fair price, grabbing it sooner is what saves you money. Waiting mostly means paying more later. I honestly wish I had better news on that, thinkers.

What ties it all together

Step back and this week splits cleanly in two. On one side, the models: three labs shipping at once, Grok racing to the bottom on price, OpenAI further improving voice. On the other side, the hardware underneath them: memory so scarce it’s reshaping the price of every chip, and the companies that make that memory are cashing in so hard that they’re going public on it.

The connection between those two halves is the whole story of AI right now. The software keeps getting better and cheaper and more competitive, which is great for anyone using it. The physical stuff it runs on keeps getting scarcer and more expensive, which is rough for anyone buying a computer. Both things are happening at once, driven by the same explosion of demand. A year ago the AI story was almost entirely about the models. Now half of it is about memory, chips, factories, and who controls them. The AI story right now is wild: I am here for it and also quite shocked by it. How’s your week going?

What I’m watching

• Whether GPT-5.6 and Grok 4.5 hold up in real-world use now that regular people are putting them to work, well past the benchmark charts

• Whether GPT-Live actually feels natural in daily use or stumbles once you leave English and the polished demo behind

• Whether SpaceX keeps renting Colossus compute to the rivals it now competes with, or starts holding it back

• Whether SK Hynix’s $28 billion listing prices strongly and pulls more memory makers toward the public markets

• Whether memory prices show any signs of relief before the new factories arrive in 2027 and beyond

Thanks so much for reading this edition of The Weekly Think.

See you next week, fellow thinkers!

Did you enjoy this week’s Think? If so, please consider sharing this and leaving a comment below, to share your thoughts on these latest AI-related rumblings. Hope to hear from you!

Share

Subscribe now

Let’s say you decided to spend an afternoon letting AI write your article (even forgoing Cool Steve’s 5th Annual BBQ & Ice Cream Bash). So you hand the whole thing over and wait. What does it give back? Well, it’s certainly smooth in terms of grammar and flow. But it’s bland, and now it is effectively no longer yours.

Because of this, the common thought is that “AI can’t write” (and it made you miss out on Cool Steve’s trademark “banana pranks”). The reality is that AI is a very weak writer on its own. It can, however, be a very effective editor. And this article will guide you in applying this in your writing.

Why AI writes badly but edits well

When you let AI write on its own, it gives you that bland, median content, the middle-of-the-road stuff we’ve all started to notice everywhere. Letting it edit is a different animal. You arm it with your voice, ideas, and all the raw materials that form the context of your article.

Think about it this way: imagine you hire a contractor to build your house, but you give him no instructions whatsoever. He’s fully capable, but without direction, you might get the most random house possible (1 bedroom, a toilet next to the stove, which is technically not a bad idea… on second thought, never mind). It’s likely that it’ll be dangerous to live in as well. But give him all your preferences (style of home, size, number of rooms, a kitchen hood that sucks all of the stove smoke right out of the house, etc.), and you end up with the home of your dreams, worthy of a Cool Steve-esque BBQ bash.

The point is that your first draft is those instructions. Writing it yourself is how you hand the tool your preferences, style, and voice all at once. With this draft, you give the AI tool what it desperately needs to do decent work: context. What the article is about, the points you want to hit, your humor and vibe, all of that good stuff. With your point of view to work with, AI has what it needs to help you sharpen that draft into a refined article.

How do I pull this off?

So now that you’ve got a killer draft, what’s next? Follow these steps to further get that content closer to what you’re dreaming up.

1. Paste your finished draft and tell it explicitly not to rewrite, only to diagnose.

2. Ask for specific categories of feedback, not “make it better.” Clarity, flabby sentences, weak spots, places a reader might drop off.

3. Have it flag the three weakest sentences and explain why, so you keep control of the fix.

4. Then, the most important step: you make every final edit by hand, overruling anything that doesn’t match your authentic voice.

In fact, here’s a prompt that will help you with steps 1-3:

Below is a finished draft I wrote. Do not rewrite it. I want you acting as my
editor, not my ghostwriter.

First, read it closely and tell me where it drags: the flabby sentences, the
unclear spots, and any place a reader might lose interest and click away.

Then point out the three weakest sentences specifically, and explain WHY each
one is weak, so I can decide how to fix it myself.

One rule you can't break: keep my voice intact. If you spot an unusual word
choice, a joke, or an odd bit of phrasing, assume I did it on purpose and leave
it alone, unless it genuinely breaks the sentence. You're checking my work, not
replacing my personality.

Here's the draft:
[paste your draft here]

Resist these temptations (and I don’t mean the singing group)

At any point of your writing process, you may be tempted to tell your AI “just fix this mess for me!” Resist any temptation to do this, as it will undo all of your hard work. When you let AI overedit your content, it becomes a technically-correct word salad that says a lot but means very little. It’s like the time Cool Steve’s cousin (you know, Wordy Bort) was at that analyst job interview, telling them all about his time “managing multiple systems while monitoring fault points and restoration systems,” which was a fancy way of saying “I make sandwiches at home all the time.”

Another downside of putting full trust in AI is what it does to all of your stylistic choices. This is especially true when you use unique word choices and humor. If you don’t inform the tool, it will identify all your cool stuff as errors, misspellings, and incorrect sentence structure. You’re the one that knows all of these are on purpose. But AI doesn’t know, and it will remove them or make you sound incredibly awkward. You’ll end up in Wordy Bort territory, real quick.

Cool writing that belongs to you

You have so many great ideas in that ol’ noggin of yours. Don’t throw them all away by letting AI try to write about them, all by itself. You’re an awesome writer. Your content just needs a good second read, and that’s exactly what AI is built for. I always say, “AI is the tool, never the artist.” So let it be the never-tiring tool that reads your content closely. Assigning it the role of full-time “second pair of eyes” is a habit that will serve you well in your writing endeavors.

If you’re struggling with your write-ups, give the guidance in this article a shot. As long as you remain the writer, it can help you further refine your incredible writing. And you just might free up that weekend to go see Cool Steve’s breakdancing dog.

Did you enjoy this article? Let me know what AI tips & tricks you use to keep your weekends full of Cool Steve antics. I read every comment!

Share

Subscribe now

The quick version: your next laptop (and mine, because I’m in the market for one) is about to cost more (wait, noooo), the newest models now come with a government permission slip, the banned Anthropic model might be crawling back to life, and China just proved you can build a frontier AI without leather jacket man’s chips. Time to dive in, like a chip to the guacamole.

While I was reading the news so you don’t have to:

• Jefferies warned memory prices are jumping another 40-50% this quarter, with no relief until 2028

• OpenAI launched GPT-5.6, a model you currently need government approval to even use

• Anthropic’s banned Fable 5 model looks close to coming back, with a new negotiator at the table

• Anthropic also shipped Claude Sonnet 5 and a research tool called Claude Science

• China’s Meituan trained a 1.6-trillion-parameter model on entirely homegrown, non-Nvidia chips

Your next laptop is about to cost more, and the number is ugly

The story: On June 27, Wall Street firm Jefferies put out a memory-market forecast that landed like a cold bucket of water on a hot cat. It expects DRAM prices, the regular memory inside every laptop, phone, and console, to jump another 40 to 50% this quarter compared to last, then another 30 to 40% on top of that in the final quarter of the year. Its bottom line: no real relief until 2028. Man, the future is hard mode, y’all.

What happened:

• The cause is the same one I’ve mentioned several times. The three companies that make the world’s memory have shifted roughly 93% of their production toward HBM, the pricey high-speed memory that feeds AI data centers.

• Server memory now eats up 60 to 70% of total demand, up from about 30% before the AI boom, so there’s far less left over for the stuff in your gadgets.

• IDC estimates PC prices will climb 4 to 8% because of this, and phones take a similar hit, with budget devices getting squeezed hardest.

• Microsoft has said it’s paying two and a half times more for memory now than it did at the end of last year.

• Chinese memory makers were supposed to be the pressure valve, but they’re selling at similar prices and mostly serving their home market, so they’re not the fix people hoped for.

Regular readers know this thread well. A few editions ago I covered Micron posting the best quarter in its history, and before that, the Steam Deck jumping as much as $300 as a result of the infamous “RAMageddon.” This Jefferies forecast is the same story, albeit with a longer timeline attached. Picture a bakery that suddenly makes way more selling giant wedding cakes to corporate clients, so it quietly stops stocking the shelf where you used to grab a $1 muffin. The flour is the same and so is the oven, but the everyday stuff you want just costs more now, because the baker found richer customers. Non-appétit, am I right?

For anyone eyeing a laptop, a phone, a graphics card, or a fresh PC build, the read since RAMageddon has only gotten worse. Prices on anything memory-heavy keep climbing into 2027. If you’ve got a purchase you can make now, doing it sooner saves you money, because the new factories that would ease this are still years from opening.

OpenAI shipped a model you need government permission to use

The story: On June 26, OpenAI launched GPT-5.6, its newest model family, in three flavors with codenames to boot: Sol, Terra, and Luna. As a first time for a major model, however, you can't just up and use it. GPT-5.6 launched in a preview available to only about 20 organizations, each one individually cleared by the US government before getting access. Broad public availability is targeted for mid-July.

What happened:

• The three tiers scale by price and power: Sol is the big one ($5 per million input tokens, $30 output), Terra is the mid-size workhorse, and Luna is the cheap, fast option ($1 input, $6 output).

• On a widely watched coding benchmark, Sol scored 91.9%, edging out Anthropic’s Mythos 5 at 88% and Fable 5 at 84.3%.

• The government gating traces back to a June 2 Executive Order that set up a voluntary system where frontier labs pre-brief the government before releasing their most capable models.

• OpenAI cleared the release with the government first, then rolled it out to the approved handful, with wider access pending a July 2 review milestone.

• For everyday coding and reasoning work, the mid-tier Terra is the one most teams will actually reach for once it opens up, since it roughly matches last year’s flagship at half the cost.

Take a second to appreciate how bonkers this is, because a chatbot needing a government clearance is a new thing. A year ago, a new model dropped and anyone with a credit card could try it that afternoon. Now the most capable release of the month arrived with a velvet rope and a guest list the government helps write (next thing you know, they’ll hire a bouncer). This is the same governance shift that produced the Fable 5 ban I covered last week, viewed from the other direction: one company got its model clipped, and now every company briefs the government before switching a model on.

If you build with AI, the practical takeaway is a slower, more staged rollout for the frontier stuff going forward. The bleeding-edge model you read about on launch day may not reach you for weeks, and it may arrive with strings attached (like a delayed game console, getting to you with no controllers). Planning around “I’ll have it when it’s announced” is no longer a safe bet for the most powerful releases.

The banned Anthropic model is back, and the reason for the ban fell apart

The story: Quick catch-up for anyone who missed the last 2 weeks: on June 12, the US government ordered Anthropic to switch off its two most capable models, Fable 5 and Mythos 5, over security concerns. After 18 days dark, it’s over. The Commerce Department lifted the export controls on June 30, and Fable 5 came back online globally on July 1. In the words of LL Cool J, “don’t call it a comeback!”

What happened:

• Commerce Secretary Howard Lutnick sent a letter withdrawing the export-control requirement on June 30, and Fable 5 returned globally on July 1 across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork.

• Anthropic built a new safety filter, developed with the government, that it says blocks the flagged technique in over 99% of cases. If a request gets blocked, the user is told and it’s rerouted to the older Opus 4.8 model.

• Anthropic co-founder Tom Brown took over the negotiations with the Commerce Department. Brown was the lead author of the original GPT-3 research paper back at OpenAI, so his background is deeply technical, which fits a dispute that hinges on model security.

• Fable 5 is back for everyone, but its sibling Mythos 5 stays on a short leash. It was partially switched back on June 26 for roughly 100 vetted US organizations that defend things like power grids and hospitals, and that’s still where it stands.

Here’s where it gets wilder than a hangry French Bulldog. When Anthropic and the government dug into the actual security report, their testing showed that weaker, unbanned models (Opus 4.8, OpenAI’s GPT-5.5, and others) could find the very same software flaw that got Fable 5 pulled (in a very “but Todd’s mom let him!” moment). In plain terms, the capability the ban was meant to contain was already sitting in models nobody banned. The thing that triggered a national-security order turned out to be something a half-dozen cheaper models could already do.

For anyone whose work depends on a specific AI model, the durable takeaway from this whole saga is about fragility. Whatever you think of the ban, it proved that a model your business relies on can go dark on short notice for reasons far outside your control. Keeping a backup option ready stopped being paranoia this month and started being basic planning.

Anthropic also shipped a cheaper model and a science tool

The story: Same company, but a very busy week. On June 30, Anthropic released Claude Sonnet 5, a mid-priced model built to run AI agents more cheaply, and Claude Science, a workbench aimed at professional researchers.

What happened:

• Claude Sonnet 5 reportedly matches or beats the pricier Opus 4.8 on agentic coding tasks while staying in the cheaper tier, which matters for anyone running AI agents in bulk, where cost per task adds up fast.

• Claude Science isn’t a new model; it’s a workspace with 60-plus preconfigured tools across fields like genomics and chemistry, plus a verification agent that checks citations and math.

• Claude Science can run locally or on an institution’s own computers, so sensitive data (patient records, unpublished results) never has to leave the building.

• Notably, Sonnet 5 was reported to underperform US-restricted models on cybersecurity tasks, which several outlets read as a deliberate, compliance-minded design choice given the whole Fable/Mythos drama (which was not saved for its mama).

• It slots into a broader industry pattern of building specialized tools for specific jobs, going deeper than one-size-fits-all chatbots.

The interesting read here is where the industry is putting all its proverbial beans. A cheaper agent model and a science-specific workbench are both making a huge point: the frontier isn’t only about building the single smartest model anymore. It’s also about making AI cheaper to run at scale and shaping it to fit real jobs. The local-computing angle on Claude Science is the part I’d keep my peepers on. A hospital or a drug lab that can use a capable AI without shipping private data to someone else’s servers is a genuinely useful thing, and it maps to a bigger worry a lot of institutions have about handing their data to outside companies.

For researchers and small labs, this is a development worth tracking if you’ve wanted AI help but couldn’t risk your sensitive data leaving your machines. The promise here is capability without the data-transfer tradeoff. Whether it delivers in practice is the open question, but the design goal is aimed squarely at a real problem.

China trained a giant AI with zero Nvidia chips

The story: On June 30, Chinese company Meituan (commonly known for food delivery) released LongCat 2.0, a 1.6-trillion-parameter AI model (though sadly, not an actual long cat). The headline isn’t the size, though; it’s the hardware. Meituan says it trained the whole thing on domestic Chinese chips, with no leather jacket man (aka Nvidia) silicon involved at all.

What happened:

• “Parameters” are the internal dials a model tunes as it learns, and 1.6 trillion of them puts LongCat 2.0 firmly in frontier-scale territory, the same league as the biggest Western models.

• The significance is that training a model this large has, until now, effectively required Nvidia’s chips, which US export controls restrict China from buying.

• The same week, another Chinese firm, DeepSeek, showed off a technique that speeds up AI responses by 60 to 85% using fewer chips, a related push to get more out of limited hardware.

• Both moves read as a response to the export controls: if you can’t buy the best chips, you squeeze more out of the ones you can get, or you prove you can build without them.

• For extra context on how high the stakes run, Taiwanese authorities this week raided the offices of hardware maker Super Micro over alleged Nvidia chip smuggling into China.

This connects to a theme that’s run through The Weekly Think, many a time: Nvidia sells the shovels for the entire AI gold rush, and until now, everyone assumed there was only one shovel worth buying. LongCat 2.0 (I wonder if, internally, it was just called “Longer Cat”) is the first real crack in the idea that you simply cannot do frontier AI without them. Think of it like a famous chef who insists a dish only works with one rare imported ingredient (let’s call it “fromage ala leather jacket”), and then someone across town serves up a version using only what grows locally. Nvidia’s chips are still the best on the market, and one model doesn’t change that. What changed is the belief that they were the only (very leather-scented) game in town, and that belief just took a hit.

For anyone building on AI, this points toward a future with more hardware options underneath you, which over time tends to mean lower costs and less dependence on a single supplier. It’s early, and one model doesn’t rewrite the industry. But the move made here is worth watching, because the entire AI economy has been leaning on the assumption that one company’s chips are irreplaceable, and this week that assumption got tested.

What ties it all together

Step back and this final week of June is just a slice of the bigger, crazier AI pie. The running theme is that AI has become a physical, national, and contested thing. Memory is scarce and that scarcity ain’t goin’ nowhere any time soon. The most powerful models now ship with government approval attached. A banned model came back after heavy negotiations (as if it were Lebron James, choosing his new team). And China just showed it can build frontier AI outside the supply chain the West has spent years trying to control (much to Jensen’s chagrin).

A year ago, most of AI news was about chatbots getting smarter week to week. Now it’s about who controls the chips, who owns the memory, which models the government will let you touch, and which country can build what without the other’s permission. The technology grew up, and the grown-up version comes with borders, price tags, and paperwork. AI is like spice from Dune, and “he who controls the spice controls the universe.”

That’s a wrap on June. Welcome, July!

What I’m watching

• Now that Fable 5 is back, whether its new safety filter trips on ordinary coding work, and whether Mythos 5 ever gets a wider release

• Whether GPT-5.6 opens to everyone in mid-July, or the government review pushes it later

• Whether memory prices bite hard enough to dent PC and phone sales the way analysts expect

• Whether more Chinese labs follow Meituan and show frontier training without Nvidia, or LongCat 2.0 stays an outlier

• Whether other countries follow Austria, which formally invited Anthropic to set up in the EU after the ban, in treating US model access as a sovereignty question

  

Thanks so much for reading this edition of The Weekly Think.

See you next week, fellow thinkers!

Did you enjoy this week’s Think? If so, please consider sharing this and leaving a comment below, to share your thoughts on these latest AI-related rumblings. Hope to hear from you!

Share

Subscribe now

You know the stock photos, right? The four coworkers in pressed button-downs crowded around a single laptop, every one of them howling at something genuinely hilarious happening inside a spreadsheet. The woman laughing alone with her salad. The two businessmen shaking hands so warmly you’d assume one of them had just donated the other a kidney. The rugged fellow staring out of a high-rise window, contemplating either a merger or the full weight of his existence (hard to say which one it is). The team of attractive thirty-somethings doing trust falls at a “workshop” with no discernible purpose. Gotta love that corporate stock photo account.

One confession, though: I made all of those up. I’ve never actually seen any of those. But you pictured them instantly, and you believed they were real for a half-second. That right there is the whole point of this article.

Every one of those images is technically a photograph. Real camera, real people, and wouldn’t you know it, real salad. But not a single one of them means a thing. They were built to be dropped into any brochure, for any company, selling any product, landing on exactly nobody and sticking with exactly no one.

With that in mind, there’s a very good chance your AI’s first draft is the literary cousin to the laughing-salad lady.

The median is a crowded place

This is something you’ve probably felt, even if you never put words to it. You ask one of these tools for help. You get something back. It's smooth, it's grammatical, and you forget it the second you look away. And it is weirdly hollow, like it was written for everybody and therefore landed on nobody.

None of that is a glitch, and none of it is your fault. The tool did exactly what you asked. When you hand it a blank instruction like “write me an intro about productivity,” you’ve given it no reason to sound like anyone in particular, so it does the only thing it can. It reaches for the median. The most expected version of every sentence, the safest word in every slot; in other words, the middle of the road.

And the middle of the road, as it turns out, is where the laughing-salad lady lives. It’s crowded down there. Everyone using the same tool with the same blank instruction is pulling from the same crowded middle, which is exactly why so much of the internet has started to sound like it was catered by the same beige sandwich platter.

“Wouldn’t it be nice to fix that?”

So let me ask the obvious question. You’ve got something to actually say. The tool genuinely does help you say it faster. Wouldn’t it be awesome if all that speed didn’t come at the cost of you sounding like you?

Good news, friends. It’s fixable, and the fix isn’t some galaxy-brained prompt-engineering ritual that requires a certification. It mostly comes down to giving the tool the one thing it has never once had: you.

This is a rough idea of what that looks like.

The first step is to teach it your voice. The tool has never read a word you’ve written. It does not know that you open with a joke, or that you’d sooner eat the salad-lady’s salad than use the word “leverage.” So show it what for. Take a few things you’ve written that sound the most like you, paste them in, and ask it to study them. Something close to this does the job:

Here are a few samples of my writing. Describe my voice in detail: my sentence
length and rhythm, my punctuation habits, the words I reach for, and the words
I'd never be caught using. Write it so I can paste your description back to you
later, to guide how you write for me.

Hang on to whatever it gives you. That little description is the entire difference between a tool that’s guessing about you and a tool that has, in some small way, actually met you.

The second step is to feed it context. Who is this for? What’s the point you’re trying to land? What’s the one thing you absolutely do not want it to do? The more it knows about the job, the less it reaches for that crowded middle.

The third step is the one people tend to skip, and it’s the one that matters most. You stay in the editor’s chair. The tool drafts; you decide. The second it makes you sound like a press release that gained sentience, you overrule it. That's the actual job, and it always was, because you're the only one in the room who knows what you meant.

The part the hype skips

Notice what none of that was. At no point did the plan become “let the robot write the whole thing while you go get a coffee and come back to a finished masterpiece.” That version is the one that produces the salad lady every time.

The reframe that fixed this for me is small but it changes everything. You're doing the work with a very fast, very tireless assistant who happens to have zero taste of its own. The work never actually leaves your hands. You bring the take, the angle, the actual thing worth saying. It brings the speed and the willingness to proofread the same paragraph forty times without sighing at you. You stay the author. It stays the intern who types quickly and needs a firm hand. Get that division of responsibilities right, and the speed stops costing you the soul of the thing.

Hand it the one thing it can't fake

Think back to the stock photos again. The reason the laughing-salad lady never moved a single human being is that she was manufactured for all of them at once. Your writing has the exact opposite assignment. It’s supposed to sound like one specific person had one specific thing to say, to one specific reader, on purpose.

That difference is sitting right there, waiting for you, every single time you open one of these tools. You just have to be willing to hand it the one ingredient it can never fake on its own.

Because the robot can match your grammar all day long. The thing it can’t do is sound like you, right up until the moment you show it how.

And that “how” is more of a system than a single trick. I’ll be getting into it very soon, so keep an eye out.

Enjoy this one? If you’ve got your own way of keeping the salad lady out of your drafts, drop it in the comments. I read every one. And if you found it useful, sharing is caring!

Share

Subscribe now

Welcome back to the stage of AI history! This was a hardware-heavy week, folks! Yes, I mean you, you are the folks. And all those chips and memory are the things affecting you folks. OpenAI revealed it has been quietly cooking up its own spicy silicon, Micron reported a quarter so good (or bad, because “RAMageddon”), and Google kept losing its top researchers to Anthropic and OpenAI.

There was drama too. Anthropic sent a letter to the Senate accusing a Chinese giant of basically photocopying its homework 28.8 million times, and Getty Images made a deal with OpenAI, laughing all the way to the bank as a result. Let’s get into it.

While I was reading the news so you don’t have to:

• OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first home-built chip

• Micron posted a record $41.5 billion quarter on the back of AI memory demand

• Google keeps hemorrhaging top AI researchers to Anthropic and OpenAI

• Anthropic accused Alibaba of the largest model-copying campaign it has ever seen

• Getty Images struck a deal with OpenAI and its stock shot up over 200%

OpenAI built its own chip and named it after a pepper

The story: On June 24, OpenAI and chipmaker Broadcom revealed Jalapeño, OpenAI’s first custom-built processor. It’s designed for one specific job: running AI models that are already trained (the industry calls that “inference,” which is the model answering your questions, the part that comes after all the learning is done). For a company that has spent years eating up Nvidia chips like they were… chips, building its own is a real statement of intent.

What happened:

• Jalapeño is a purpose-built inference chip, so it’s tuned to serve up answers fast and cheap, with six stacks of high-speed HBM memory wrapped around one big compute core.

• The chip is enormous, roughly 840 square millimeters, which is about as large as current manufacturing machines can physically print on a single piece of silicon.

• OpenAI and Broadcom say they went from idea to finished design in nine months, well under the typical year-and-a-half to two years, and they admit they used OpenAI’s own models to speed up parts of the design work.

• Test chips are already running in the lab on one of OpenAI’s coding models, with real deployment planned for late 2026 in data centers built with Microsoft.

• Broadcom claims the chip beats today’s best hardware on power efficiency, but the two companies released zero actual numbers to back that up, so for now that claim is just a promise with no proof attached.

OpenAI is following a path Google, Amazon, and Microsoft already walked, which is to stop renting all your compute from Nvidia and start making some of your own. The plain reason is money (because in this biz, it always is). When you’re spending billions a year on chips from a single supplier, designing your own starts to look less like a vanity project and more like basic math.

If you’re building anything on top of OpenAI’s models, this is a slow-burn detail to file away. Custom chips like this are how the cost of running AI eventually comes down, and cheaper compute underneath you means the tools you build get cheaper to run over time. Nothing changes tomorrow, but the direction of travel is good for the people building on top.

Micron had the best quarter in its history, and your next PC will feel it

The story: On June 24, memory-chip maker Micron reported the biggest quarter in its 48-year history: $41.46 billion in revenue, up from $9.3 billion in the same quarter a year ago. Its profit margin hit a frankly cartoonish 81%, meaning for roughly every dollar of product it sold, about 81 cents was profit (in the words of rapper Mase, “double up”). The reason is obvious: AI data centers are devouring the world’s memory supply like a kid with M&M’s.

What happened:

• Revenue more than quadrupled year over year, and the company’s profit for the quarter came in around $28 billion.

• The driver is HBM, the ultra-fast memory that sits next to every AI chip, and Micron has already sold out its entire 2026 HBM production.

• CEO Sanjay Mehrotra said the company can only fill somewhere between half and two-thirds of what customers are asking for, which is why prices keep climbing.

• DRAM, the regular memory in your laptop and phone, saw prices jump by a mid-60s percentage in a single quarter.

• Only three companies on earth make this memory at scale (Micron, Samsung, and SK hynix), and all three are pouring tens of billions into new factories that won’t be ready until 2027 or later.

A few editions back I walked you through the Steam Deck getting up to $300 more expensive, and the “RAMageddon” nickname the internet slapped on the memory shortage. Micron’s record quarter is the other side of that exact coin. The same shortage making your gadgets pricier is making the memory makers obscenely rich, because they can sell every chip they produce to AI data centers at a premium. When a supplier can only meet two-thirds of demand, the supplier sets the price, and right now the supplier is naming a very high one. To illustrate that, imagine there’s a lemon shortage and the kid on the block is selling lemonade on a hot summer day for $20 (because, obviously, he’s gotta pay for that expensive AI subscription somehow).

For anyone planning to buy a laptop, a phone, a graphics card, or build a PC, the practical read hasn’t changed since RAMageddon: prices on anything with a lot of memory in it are going to stay high into 2027. If you’ve got an upgrade you’ve been putting off, you now have to pay what I like to call the “waiting tax” (more like wailing tax), because the new factories that would ease prices are still a year or more out.

Google keeps losing its best AI minds to the competition

The story: Across the week, Google lost another wave of senior AI researchers to its rivals (in a very “Giannis Antetokounmpo leaving for the Miami Heat” moment). The headline departures: Noam Shazeer, a co-inventor of the core technology behind modern chatbots, left for OpenAI, and John Jumper, who led the Nobel-winning AlphaFold protein project, left for Anthropic. Two more Gemini contributors, Jonas Adler and Alexander Pritzel, are also heading to Anthropic. Alphabet’s stock dropped as much as 7.2% on the news, its worst single-day slide since February.

What happened:

• Shazeer co-wrote the 2017 research paper that introduced the “transformer,” the architecture underneath basically every modern AI model, so losing him is symbolic as much as practical.

• Jumper isn’t just any researcher; he shared a Nobel Prize for using AI to predict protein structures, and now he’s taking that talent to Anthropic.

• Google’s spokesperson pushed back, pointing to DeepMind boss Demis Hassabis calling it the most competitive talent market the tech industry has ever seen.

• The market clearly read the exits as a warning sign, knocking 7.2% off Alphabet’s value in a single session, even though the company is still worth around $4.46 trillion.

• This continues a pattern from earlier in the year, when Anthropic landed researcher Andrej Karpathy, another marquee name in the field.

This connects to something I wrote about a few editions ago, when Karpathy joined Anthropic. That migration didn’t stop; it straight up accelerated like it was Han in the Fast & Furious franchise. The people who build these models are the rarest resource in the entire industry, rarer than chips or data, because there are maybe a few hundred of them on the planet who operate at the frontier. When they move, they take knowledge that money alone can’t quickly replace, which is exactly why a few resignations can knock billions off a company’s value in an afternoon.

If you work in any field being reshaped by AI, you’re probably taking notice. The scarce, valuable thing is the human expertise, the people who actually understand how to build and direct these systems. That’s the part companies are paying enormous sums to keep, and it’s a useful reminder of where to focus your own learning if you want to stay valuable as this all shakes out like a Polaroid picture in the 90s.

Anthropic accused Alibaba of copying its homework 28.8 million times

The story: On June 24, word got out that Anthropic (the company that makes Claude) sent a letter to the US Senate accusing Chinese tech giant Alibaba (the company behind such services as AliExpress, purveyor of cheapo computer products) of running the largest model-copying campaign it has ever caught. Anthropic says operators tied to Alibaba’s Qwen AI lab used roughly 25,000 fake accounts to run 28.8 million conversations with Claude between April 22 and June 5, harvesting its answers to train a cheaper copycat.

What happened:

• The technique is called “distillation,” which means using a strong model’s outputs as a teaching guide to train a weaker, cheaper model to imitate it. Think of it as a student copying a brilliant classmate’s test answers over and over until the student can fake the same results.

• Anthropic says the campaign zeroed in on Claude’s most valuable skills, its coding and step-by-step reasoning, and called it turning American R&D into a subsidy for a competitor.

• For scale, this dwarfs the cases Anthropic flagged back in February, when it accused DeepSeek (about 150,000 exchanges), Moonshot (3.4 million), and MiniMax (13 million) of similar tactics.

• Alibaba was added to a US Defense Department list of Chinese military-linked companies on June 8 and is suing to get removed, calling the label baseless.

• As of this writing, Alibaba hadn’t responded to the specific accusation, so treat this as one company’s claim that still needs proving. Worth keeping that in mind.

One thing to keep in view: this accusation lands the same month the US government switched off Anthropic’s own two newest models over a security concern. So within a few weeks, Anthropic has been on both sides of the security conversation, the subject of one government action and the source of a warning about someone else. Make of all this what you will.

What’s clear for the rest of us is that AI has become a full-blown geopolitical contest, with company letters to the Senate, Pentagon blacklists, and accusations crossing the Pacific. The models you use are now caught up in a much bigger contest between Washington and Beijing. You don’t have to pick a side to notice that the technology stopped being a neutral tool a while ago, and the fights over who gets to build and use it are only getting louder.

Getty made a deal with OpenAI and its stock went bananas

The story: On June 21, Getty Images (the company behind a huge chunk of the photos of your favorite celebs eating hot dogs at the Oscars) announced a deal to put its licensed image library directly inside ChatGPT. When you ask ChatGPT a question where a real photo helps, like a historical event or a famous landmark (or a celeb eating a hot dog), Getty’s licensed images can now show up in the answer. Getty’s stock jumped more than 200% in a single day on the news.

What happened:

• The deal is display-only, meaning ChatGPT shows Getty’s images alongside answers and Getty gets paid for it, with no permission for OpenAI to use the photos to train its models.

• It covers over 400 million images, including premium news, sports, and entertainment photography plus Getty’s lower-cost iStock library.

• Getty had spent the prior couple of years fighting AI companies in court over images used without permission, so this is a notable pivot from suing to partnering.

• The market reaction was enormous, a 200%-plus single-session jump, which tells you investors had basically priced Getty as a company with no AI future and suddenly changed their minds.

• Getty had already done a smaller version of this with another AI search company last year, but the OpenAI deal is far bigger in reach.

This one’s a quieter story, but it points at something real about how the copyright fights over AI might actually end. For two years the dominant approach from media and image companies was to sue, and some of those suits are still grinding through the courts. Getty looked at that road and chose a different one: license the work, get paid every time it appears, and turn ChatGPT into a paying customer (it basically said, “hold up, actually, lemme get paid”). The 200% stock pop suggests the market thinks getting paid beats waiting years for a courtroom maybe-win.

If you make content for a living (photos, writing, music, video), this is a development to watch closely. The Getty deal is an early example of creators and AI companies finding a paid arrangement that works for both sides. Whether that becomes the norm or stays the exception will shape how much the people who actually make things get compensated when AI uses their work. For now, one company just showed that licensing can pay better than litigating.

What ties it all together

Step back and the week sorts into two big themes. The first is that AI is becoming a physical, industrial business: OpenAI is cooking up its own chips, Micron minting record billions while unapologetically fueling “RAMageddon,” the whole thing increasingly about silicon and factories and supply you can measure in tons. The second is that AI is becoming a contest between nations and titans: Google bleeding talent to its rivals, Anthropic and Alibaba trading accusations across the Pacific, Getty cutting a deal that rewrites the rules of a two-year fight.

The common thread is that the easy, abstract phase of AI is ending. A year ago this was mostly about chatbots becoming smarter. Now it’s about who controls the chips, who owns the memory, who employs the rare humans who build the models, and which country’s companies come out ahead. The technology grew up, and grown-up technology wants its license (and a later curfew). Exciting or unnerving?

I’ll let you, the reader at home, decide.

What I’m watching

• Whether OpenAI’s Jalapeño chip actually ships on time in late 2026, and whether those efficiency claims hold up once real numbers exist

• Whether memory prices show any sign of easing, or keep climbing into 2027 as the shortage drags on

• Where Google’s departing researchers land first in Anthropic’s and OpenAI’s next models

• Whether Alibaba formally responds to Anthropic’s accusation, and whether the government acts on it

• Whether more media and creative companies follow Getty’s lead and cut licensing deals with AI firms while the lawsuits fade into the background

Thanks so much for reading this edition of The Weekly Think.

See you next week, fellow thinkers!

Did you enjoy this week’s Think? If so, please consider sharing this and leaving a comment below, to share your thoughts on these latest AI-related rumblings. Hope to hear from you!

Share

Subscribe now

Think about iconic interviews that have stuck with you.

Princess Diana, being painfully honest about what life inside the royal family actually cost her. OJ Simpson, looking into the camera and swearing he had nothing to do with it. LeBron James, sitting down to tell everyone which direction his career was about to go in. Dave Thomas of Wendy’s, solemnly explaining why the SuperBar salad bar had to go (okay, fine, I made that last one up).

Love the person or loathe them, an effective interview hands you something the public eye could never pick up on its own. The good ones get under the surface and pull out the thing that was actually sitting there the whole time.

So here’s a thought I can’t shake. What if you ran an interview exactly like that, except you were the one in the chair, and the thing asking the questions was your chatbot?

Not that kind of AI interview

We’re living in the era of rampant AI use. However you feel about it, the stuff is everywhere now (though thankfully not omnipresent, because boy, would that be a different and much scarier article).

And in all the hype and all the fear, there’s a genuinely useful way to get your actual voice and your actual intent into the work, and it almost never shows up in the headlines. It’s the AI interview.

Not the dystopian kind, to be clear. I don’t mean the bot that screened you for that analyst job and judged your blinking patterns. I mean an interview where you’re the subject, and the whole point is to collect context.

“Thank you for sitting down with me today”

Here’s the trick at its most basic. You ask your chatbot to interview you. That’s the gist of the whole thing that we’re about to get into.

Why would you want that? Maybe you want it to capture your voice so you’ve got a sharper writing partner instead of a generic one. Or maybe you’re trying to build something and you need the tool to actually understand what you’re going for. A perfect-burger-algorithm calculator, for instance (which I need I need to go patent, the moment I finish this sentence).

Here’s roughly how you’d kick one off. You send something like this:

I am looking for ideas for an app I want to code. Give me an in-depth interview to learn my likes, 
dislikes, and preferences when it comes to the apps I use every day. Use the context of this 
interview to create a tastemaking profile for me, then use that profile to give me up to 10 app 
ideas. The interview should be 100 questions, divided into 10 rounds.

You fire it off, the chatbot asks a couple of clarifying questions, and the interview begins. A hundred questions about how you actually use the things you use.

But is that all there is to it? Not quite.

“I want to circle back to something you said earlier”

Bourdain was talking about food, but the man might as well have been talking about the things we build.

Here’s the rule underneath all of this. The more context you give the tool, the better the thing it hands back. And a one round interview, however long, is rarely the whole meal.

Think about ordering a burger. Walk into a good diner and ask for the cheeseburger, and you’ll get something solid. But tell them you also want grilled onions, bacon, and sauteed mushrooms, and now you’ve got a delicious burger built specifically for you (though, AI output can’t be literally delicious, sadly, unless you own some machine I’m not aware of, in which case please DM me immediately). What does this all mean? The more you specify, the more the result actually fits.

So say those 10 app ideas come back, flatter than a smash burger patty. You’re not stuck with them. You send a follow-up that goes something like this:

Look back at all the responses I gave to your questions. With my answers in mind, ask me 
clarifying questions on anything that seemed vague or anything I sidestepped. Once you’ve 
gathered enough clarity, generate a new list of 10 app ideas.

What did that actually do? You told the tool to press you a little harder so it could understand you a little better. And once it has that extra clarity, the second list tends to land in a completely different place than the first.

That’s the difference between “an app that gives you recipes” and “an app that helps you build customized recipes around what’s already in your fridge.” Same person, same starting point, but a few more questions in between. You went from just “okay” to “wait, that’s actually good!”

But I thought this was supposed to make things easier

Right about here, I can hear an objection forming. “Isn’t AI supposed to do all of the work for me? I wanted to type one thing and get a finished thing back. What’s with all the rounds and the follow-ups and the homework? And where’s my burger?*”

Fair points. And yes, you can absolutely have AI crank stuff out on the fly. If that’s your prerogative, I’m not going to get on a soapbox and shout at you with a megaphone.

But the wing-it approach reliably produces wing-it results. That’s just the reality of the thing. You get the bland, forgettable output, and then you blame the tool or you blame yourself, when the real culprit was that nobody ever asked you what you actually wanted.

Here’s the reframe that changed it for me. The questions aren’t the busywork standing between you and the work. The questions are the work. The interview is how you figure out what you’re really after, and most of us almost never slow down long enough to do that on our own.

Think about how rarely anyone asks you a hundred good questions about a thing you care about. A friend might get to three before the conversation drifts (which is fair and does not mean they’re a bad friend, unless it’s weirdo Steve talking about his “sausage orchard” again). Whereas, the AI interview is a hundred, and somewhere around question sixty you start saying things you didn’t know you thought. The app idea, the essay, whatever you came for, the goal you’re chasing but can’t quite name. The thing you actually walk out with is a clear picture of what you wanted the whole time, finally said out loud.

Fred Rogers (yes, Mister Rogers) had a whole song about liking to take your time and do a thing right. I think about that more and more as an adult typing prompts into a robot. Turns out the cardigan master was onto something.

And I’m not pointing you toward any of this because it’s the disciplined, virtuous choice. I’m pointing you toward it because it can work effectively, and because the version of the thing you get on the other side actually serves what you were trying to do, instead of being something generic that makes you say, “it’s gonna be a no from me, dawg.”

“I appreciate your time”

Come back to those interviews for a second. The ones that stick with us aren’t the ones where someone recited their talking points. They’re the ones where a good question, asked at the right moment, drew out something real that we never would have heard otherwise.

That exact thing is sitting right there, available to you, every single time you open one of these tools. You just have to be willing to be the one in the chair (the chAIr, if you will).

Because it turns out that one of the most useful things the AI can do for you isn’t answering your questions. It’s asking you theirs.

^{*Would ya stop asking the AI for a burger? Trust me, I’ve tried.}

Did you enjoy this article? Do you want more AI tips & tricks? If so, please consider sharing this and leaving a comment below, to share your thoughts on your new ways you’re using AI. Hope to hear from you!

Share

Subscribe now

Eric’s back, back again. Eric’s back, tell your friends! I was away last week, figuring the AI world would do its usual thing and crank out a few model updates I could catch up on later. Friends, it didn’t do its usual thing. I come back to find that Elon Musk’s rocket company is now one of the biggest companies on the stock market, it spent $60 billion on a coding tool like it was grabbing an impulse buy at checkout, Apple changed CEOs, and the heads of the major AI labs got seated at a table with world leaders. Good gravy.

Before we get into this week’s big stuff, let me catch us both up on the June 4-10 stretch I missed, because a lot of it matters for what came next.

While I was gone (the June 4-10 bytes):

• Apple said goodbye to Tim Cook. At WWDC on June 8, Apple unveiled a rebuilt, Gemini-powered Siri (Google on an iPhone? Eric Schmidt would never), and made Claude a selectable option too. The keynote was Cook’s last as CEO before he hands the reins to hardware chief John Ternus on September 1.

• OpenAI filed to go public too. On June 8, OpenAI confidentially filed its own paperwork with the government, joining Anthropic and SpaceX in the IPO line. Their announcement basically said “we figured this would leak, so here you go.”

• ChatGPT dropped below 50% market share for the first time. Per a Sensor Tower report, ChatGPT now sits under half the market, with Google’s Gemini at 27.7% and Anthropic’s Claude at 10.3%. The two-horse race became a three-horse race while I was drinking a cold brew.

• Snap announced $2,195 AR glasses. Called Specs, with OpenAI and Gemini built in, aiming to beat Meta’s comparable glasses to store shelves. Two thousand dollars for face computers. We live in the future, and the future demands a home equity line of credit.

Okay. Caught up? Cool! Let’s get into the stuff that happened this week.

SpaceX went public and became the fourth most valuable company in America

The story: On June 12, Elon Musk’s SpaceX (which now includes his AI company xAI, maker of the Grok chatbot) sold shares to the public for the first time. It raised about $75 billion, the biggest stock-market debut in history, and the stock jumped 19% on day one. That pushed the company’s total value above $2 trillion and, after a few more days of climbing (like a crazed mountain man), made it the fourth most valuable company in the United States.

What happened:

• SpaceX priced its shares at $135 and they closed the first day near $161, then kept climbing through the following week.

• The $75 billion raised beat every prior IPO on record. For scale, that’s roughly the entire yearly budget of NASA, raised in a single morning.

• The company is not profitable. It lost close to $5 billion in 2025, and several analysts flat-out called the valuation too high (one slapped a “sell” rating on it the same day).

• Alphabet, Google’s parent, quietly owns about 4.9% of SpaceX, a stake now worth around $105 billion.

• This is the first of three giant AI-related IPOs lined up back-to-back, with Anthropic and OpenAI both having filed their own paperwork in the past two weeks.

A few editions back I wrote about Anthropic, OpenAI, and SpaceX all heading for the stock market at the same time, and how the public was about to get its first real read on whether the AI hype holds up. Well, SpaceX went first, and the read was “investors are very, very interested.” Regular peeps piled in at record numbers. Whether that enthusiasm is wisdom or mania is the question, and the analysts arguing the company is overvalued might be right. Either way, the first domino (tile, not the pizza) fell, and it fell loudly (like me, after eating too much Domino’s… the pizza, I mean).

The part of this that really stands out is who got rich. More than 4,400 current and former SpaceX employees are expected to become millionaires from the IPO, and a lot of them never wrote a line of code or designed a rocket (meanwhile, I invented the MP3 player but… oh forget it). We’re talking welders, machinists, technicians, even cafeteria workers, people who got company stock as part of their pay. One former welder, Juan Hernandez, joined in 2015 at $28 an hour, barely knowing what SpaceX was. His roughly 6,500 shares were worth just over a million dollars(!) when the stock closed on day one, and he’s now teaching his kids how to invest.

Then, four days later, SpaceX bought Cursor for $60 billion

The story: On June 16, fresh off the record IPO, SpaceX announced it’s buying Anysphere, the company behind the popular AI coding tool Cursor, for $60 billion in stock. Cursor is a program developers use to write software with an AI assistant doing a lot of the typing (like a guy with a typewriter following around a smart dude). The deal makes it the most expensive AI software purchase ever announced, and it happened four days after the company went public. In four days, I’m usually wondering if the leftover chicken in the fridge is still okay to eat.

What happened:

• The deal is all-stock, meaning SpaceX is paying with its own newly public shares, no cash required, which it can afford to do precisely because those shares are valued so high.

• Cursor brings in roughly $2.6 billion a year and had been talking to investors about a $50 billion valuation just two months earlier.

• This folds Cursor into xAI, Musk’s AI arm, giving it a direct foothold in coding tools to compete with Anthropic’s Claude Code and OpenAI’s Codex.

• The purchase represents only about 3.4% of SpaceX’s value, which tells you how much a sky-high stock price lets you throw around.

• It briefly pushed SpaceX past Microsoft in total value during Tuesday trading.

Here’s the big, notable thing here. The major AI coding tools have nearly all been swallowed by giants in a remarkably short time (like CEOs swallowing a Limitless pill). Microsoft owns GitHub Copilot. Anthropic has Claude Code. OpenAI has Codex. And now SpaceX has both Grok’s coding tools and Cursor. A year ago these were scrappy independent startups devs loved. Now almost all of them answer to one of a handful of enormous parent companies.

If you’re a dev, this affects which tools stay independent and which start nudging you toward their parent company’s ecosystem. The thing to do is keep your own skills portable. Learn the underlying craft itself, deeper than any one tool’s buttons, so that when the tool you rely on gets bought and changed, you can move without starting over. Ownership shuffles at the top; your ability to actually build is what stays yours.

Apple swapped CEOs and put Google’s AI inside the iPhone

The story: This one technically kicked off during my time away, on June 8, but your boy needs to expand on this. At its developer conference, Apple unveiled a completely rebuilt Siri running partly on a custom version of Google’s Gemini AI, and it was Tim Cook’s final keynote as CEO. He steps down September 1 after about 14 years, handing the company to hardware chief John Ternus.

What happened:

• The new Siri uses a three-tier system: simple stuff runs on your phone, medium stuff on Apple’s private servers, and the heaviest thinking routes to a 1.2-trillion-parameter Google Gemini model running on Nvidia chips in Google’s cloud.

• Apple says it anonymizes every query so neither Apple nor Google can tie a request back to you.

• Special guest star, Claude: Apple made multiple AI assistants selectable, so Anthropic’s Claude can be an iPhone option for the first time.

• Apple also previewed homeOS, the software for a long-rumored smart-home gadget (a HomePod with a screen) expected this fall.

• The Siri rebuild arrives after a $250 million settlement with iPhone buyers who said Apple advertised Siri features back in 2024 that never actually shipped.

The interesting thread here is what it means that Apple, the company famous for building everything itself, decided the fastest way to fix Siri was to put a competitor’s AI underneath it. For two years Apple promised a smarter Siri and kept missing. So it partnered with Google, a company it competes with on phones, search, and maps. Swallowing that much pride tells you how hard building a frontier AI model has become even for the richest company on earth. Somewhere, Eric Schmidt is laughing maniacally.

For the rest of us, this means the AI you use every day is about to show up in places you didn’t install it. If you own an iPhone, Google’s Gemini and possibly Anthropic’s Claude are coming to your home screen this fall whether you sought them out or not. Knowing what’s under the hood, and what each one does with your questions, is becoming basic digital hygiene.

AI wrote so much code it knocked GitHub offline

The story: In a very Obi Wan “you have done that yourself” sorta way, GitHub, the website where the world’s programmers store and share their code (owned by Microsoft), got so overwhelmed by AI coding agents pumping out work that its reliability cratered. On June 16, Microsoft confirmed it’s temporarily borrowing computing power from Amazon, its direct cloud rival, just to keep GitHub running.

What happened:

• GitHub was processing about 275 million code submissions per week, on pace for 14 billion in 2026, up from 1 billion in all of 2025.

• Automated pull requests (changes opened by AI agents) jumped from 4 million last September to over 17 million by March.

• The site logged nine separate service problems in May, and its uptime in June fell to about 88.4%, well under the 99.9% it promises business customers.

• To cope, Microsoft is routing some GitHub traffic through Amazon Web Services while it finishes moving the platform onto its own Azure cloud by 2027.

• This is part of a broader crunch: Google reportedly pays SpaceX $920 million a month for computing power, and Anthropic pays SpaceX $1.25 billion a month for its Colossus supercomputer.

The plain version of what happened: AI code writing is all the hype and it generated more work than the internet’s main code warehouse could physically handle. The humans pointed the tools at their projects, the tools produced an avalanche, and the warehouse buckled. It’s a little like all the kids in the neighborhood absolutely overwhelming the ice cream truck on the hottest day of the summer. The demand is real, the tools are doing what they were asked, and the physical infrastructure underneath is scrambling to keep pace.

The lesson that keeps repeating across these editions is that compute and infrastructure are the real bottleneck now. The thinking ability of the models stopped being the hard part a while ago; the hard part is having enough machines and electricity to run them. Even Microsoft, a company that owns one of the three biggest clouds on the planet, had to call up its archrival Amazon for backup. If you’re building anything on top of these tools, plan for the plumbing to creak. The smart projects right now are the ones designed to work even when the service they depend on is having a rough day. Locally-run AI, anyone?

The government switched off Anthropic’s newest model three days after launch

The story: On June 12, the US Commerce Department ordered Anthropic to cut off access to its two newest, most capable models, Fable 5 and Mythos 5, citing national security. The order specifically barred any foreign national from using them, including Anthropic’s own foreign-born employees. Since Anthropic can’t be all TSA and check the passport of every person sending a request in real time, it did the only thing it could and shut both models off for everyone. The kicker: the models had launched just three days earlier, on June 9. Imagine throwing a party, and the city condemns your house before the guests finish their charcuterie.

What happened:

• The directive landed at 5:21 p.m. ET on a Friday, in a letter from Commerce Secretary Howard Lutnick to Anthropic CEO Dario Amodei, and the letter reportedly gave no specifics about the actual concern.

• Anthropic says that the trigger was a “jailbreak,” a trick that gets a model to do something its safeguards block. The specific one, by Anthropic’s account, amounted to asking the model to read a codebase and point out its security flaws.

• Anthropic says that same capability is widely available from other models that weren’t banned, naming OpenAI’s GPT-5.5 directly (like a kid saying “but mom, Tommy broke the window too!”)

• Every other Anthropic model, including its flagship Opus 4.8, stayed online and unaffected. The takedown hit only the two newest releases.

• This appears to be the first time the US government has forced a publicly deployed frontier AI model offline. Imagine if Microsoft said “turn off the Word, boys.”

The absurd bit, and the reason this story is more interesting than scary, is the mismatch between the hammer and the nail. A government used national-security export powers, the kind of tool aimed at missile guidance systems and nuclear materials, to pull a chatbot over a bug-finding trick that a half-dozen other chatbots can also do. Anthropic basically said as much in its response, arguing that if this became the standard, every frontier model from every company would have to be yanked too. Whatever you think of Anthropic, them dudes have a point.

The reason this matters for everyone, not just Anthropic’s customers, is the precedent. A switch now exists, and a government just proved it’s willing to flip it. That has nothing to do with whether you like Anthropic or trust the government’s call here. It’s a structural fact about how AI works now: the model you build your business on can be switched off by an email sent at 5:21 on a Friday. If you’re building on any single AI provider, the practical lesson is to keep a fallback ready, because the hypothetical “the model just vanished” just got really real.

What ties it all together

I was gone two weeks. In that time the biggest IPO in history happened, the company behind it immediately spent $60 billion on a coding tool, the most valuable company on earth changed CEOs and adopted a rival’s AI, the website that holds the world’s code fell over from AI overuse, and the government switched off the newest models from one of the top AI labs. If you’d told me all of that in advance, I’d have asked what possessed you to generate all of that with AI.

Step back and look at the shape of it. The money is going public, the giants are absorbing the startups, the rivals are partnering when pride used to stop them, the infrastructure is straining under its own success, and the government just showed it can switch a model off whenever it decides to. The common thread is consolidation. Power, money, talent, and infrastructure are all flowing toward a smaller and smaller number of very large players, and it’s happening faster than anyone can write newsletters about it. Any reason for alarm here? I’ll let you decide, peeps.

What I’m watching

• Whether SpaceX’s stock holds its sky-high value or the “overvalued” analysts get proven right

• Whether Anthropic or OpenAI actually pull the trigger on their own IPOs, and how the market receives them

• What happens to Cursor’s independent feel now that it answers to Musk

• Whether GitHub stabilizes, or the AI-code firehose keeps outrunning the servers

• Whether Anthropic gets Fable 5 and Mythos 5 switched back on, and whether the government ever spells out its actual reasoning

  

Thanks so much for reading this edition of The Weekly Think.

See you next week, fellow thinkers!

Did you enjoy this week’s Think? If so, please consider sharing this and leaving a comment below, to share your thoughts on these latest AI-related rumblings. Hope to hear from you!

Share

Subscribe now

I still remember the first time I walked into a buffet. The place was called SmorgaBob’s, which we can all agree is a very good name for a buffet.

I had never seen so many options in one room in my life, at least not outside of a Filipino party. Roast beef, pizza, fried chicken, fried rice, all of it sitting out at the same time, just waiting for me. So I did what any food-driven kid would do. I grabbed one plate and piled it as high as physics would allow.

And then, of course, I got completely full off that one ridiculous plate.

What I had not figured out yet, in my youthful greed, was the buffet mastery technique. You take a little. You go back. You take a little more. You work the whole room in small trips and you maximize your parents’ hard-earned buffet dollars (because let’s be honest, I was not the one paying for this lunch). Live and learn.

I don’t really do buffets these days. My 40-something body has filed several formal complaints about that kind of eating. But wouldn’t you know it, the buffet dilemma found me again, in a completely different room, and this time there’s no roast beef involved.

“A-I can’t believe there’s this many choices”

Here’s the new buffet. It’s AI tools, and I’m standing in front of it as the exact same wide-eyed kid with an empty plate and zero self-control.

Just a few years ago, this was not a problem, because there wasn’t much to pile on. You had ChatGPT and, well, not a whole lot else. Life was simple. You messed around with prompting here and there, you slowly got the hang of it, and you could become a genuine power user of one good tool without breaking a sweat.

Then 2026 rolled around, and now AI is here, it’s there, it’s absolutely everywhere.

Alongside ChatGPT you’ve got Claude, Gemini, Grok, and a bunch of others that don’t get the same spotlight but are also very good. Then come the specialists. Tools for audio. Tools for video. Tools for image editing. Tools that will build an entire AI version of you, which is a sentence I never expected to type in my lifetime, but here we are. Calling the current spread of AI tools “rich” is underselling it by a mile.

And naturally, every single dish looks mouthwatering. Of course you want to try all of them.

“A-I can’t finish all of this”

Maybe you’ve already been here. You’ve got a couple of trials running, a subscription or two, and the cheerful little plan in your head that goes, “I’ll use this one for that, and that one for this, and it’ll all be gravy.”

I hate to be the one to tell you, but what’s true at SmorgaBob’s is true here too. Pile too much on the plate and you can’t actually appreciate any single thing on it.

It gets worse, though, and this is the part that really got me. While you’re busy hopping from one tool to the next, you never stay anywhere long enough to get good at it. Think about what that actually means. You’re not actually collecting skills at all. You’re collecting a bunch of day ones. You’re a perpetual beginner, over and over again, just at a dozen things instead of one. And since new tools keep dropping one after another at this point, the plate doesn’t even get a chance to lighten. It keeps getting heavier while you stand there holding it.

That overstuffed plate feels productive. It feels like you’re on top of things. But mostly it’s a great way to stay mediocre at a whole lot of stuff while feeling busy the entire time.

Now, I’m not saying options are the enemy. A buffet that only served all-you-can-eat mashed potatoes would eventually be a sad place to spend an afternoon. Mashed potatoes are great, but you’d tap out after a plate. So, the options are good. However, the plate is the problem.

Prioritizing your plate

When I get overwhelmed by all of it, I find the fix is to go back to basics. Instead of using every LLM under the sun for every random thing that floats through my head, I figure out which tool actually serves the job in front of me, and I build a small, deliberate plate.

Here’s what’s on mine these days.

Claude does my proofreading and most of my writing-related tasks. ChatGPT is my smarter search, the thing I poke when I’ve got a random question rattling around and I don’t feel like sifting through a pile of links. Nano Banana handles my image editing. And when I’m digging through the notes I keep in OneNote, Copilot turns out to be a surprisingly good helping hand.

That’s pretty much the whole plate. Four dishes, and I actually know what to do with all four.

That’s the part that changed everything for me. Because the plate is small, I get to build up some real mastery instead of treading water. I know what each tool is for and how to get good results out of it, and I’m not splitting my attention fourteen ways and calling it a dope workflow. The small plate is also the thing that lets me relax about trying something new, because I’ve always got my reliable stack to come home to. So now I’m enjoying this AI buffet instead of getting wheelbarrowed out of it.

FOMO: the buffet's final boss

Here’s the part many don’t like to admit, as it’s the part that usually drives the overloading.

The new tools are going to keep coming, fast, and every one of them is going to show up wearing a big flashing sign that says THIS CHANGES EVERYTHING. And every time, that little voice pipes up. What if the dish I didn’t put on my plate was the one that mattered? What if everyone else is eating the good stuff and I’m over here with my sad, sensible little plate?

That fear is the whole reason the plate gets out of control in the first place. So let me say the thing that helps me.

The tools you’re already good at are still very good, and they’re getting better every day without you having to chase a single thing. Skipping the newest hot tool is not the catastrophe the FOMO insists it is. And here’s the sneaky part: the small plate is exactly what frees you to taste-test the new dish whenever you want, because you’re not depending on it to feed you. You can poke at it, decide if it’s any good, and walk away with no harm done.

It really is better to be a master of a few tools than a random user of many.

Enjoying the buffet, sans indigestion

Picture the overstuffed plate for a second. One plate, loaded down with the AI versions of crab, chocolate cake, and steak, all crammed together and touching in ways food was never meant to touch. Soggy. Overwhelming. Impossible to actually enjoy. That’s the plate you build when you let every new tool talk you into one more scoop.

You don’t have to eat that way. And if you’ve been feeling behind because you’re not running fifteen tools at once, hear me on this: using less, on purpose, was always the strategy, never the surrender.

That kid at SmorgaBob's was playing to win his own crazy version of The Hunger Games, and to him, winning meant the tallest plate in the restaurant. It took me a few decades and a few regrettable stomachaches to realize the tall plate was never the prize. I eat a lot smaller now. I also enjoy it a whole lot more.

Did you enjoy this article? If so, please consider sharing this and leaving a comment below, to share your thoughts on your AI tool usage. Hope to hear from you!

Share

Subscribe now