About This Week’s Covers

This week’s covers are inspired by George Clinton, whose birthday is July 22nd. He represents a most human style of creativity with the psychedelic art of Parliament and Funkadelic album covers. The main cover is inspired by “What Is Soul,” a 1970 Funkadelic song where Clinton says soul is a ham hock in your cornflakes! I had Grok animate it and added the Funkadelic song “Cosmic Slop”….

The remaining category covers were generated with prompts by Claude Opus 5.5 and then Gemini Images using the APIs. I am not a fan of them because they feel especially slop-like due to the theme. A few of my favorites are below:

This Week’s Humanities Selections

This week’s humanities theme is a series of quotes from George Clinton and the song “What is Soul”. In a world of AI, it’s hard to beat an eight-minute song from Funkadelic to go back to our human roots. Unintentional AI wisdom from the vibe coding OG…

I have tasted the maggots in the mind of the universe; I was not offended, for I knew I had to rise above it all or drown in my own shit…

I’d bite off the Beatles, or anybody else. It’s all one world, one planet and one groove. You’re supposed to learn from each other, blend from each other, and it moves around like that.

Even though I loved the Fifties doo-wop, you couldn’t hold on to it. You had to change, or you was gon’ be antique real quick, like the Ink Spots.

I always try to find the kids that’s getting on your nerves – because your instinct let you know they pushing you out the way, you getting old. And you don’t let that bother you, have fun! You ain’t got to hang out with them, but you can work with them.

This Week By The Numbers

Total Organized Headlines: 357

This Week’s Overview

This is my 147th week of organizing links, and I’ve crossed the 60,000 mark. For the week ending July 24th, I organized 357 links, and 64 contributed to 34 executive top stories.

I’m about ten weeks behind because I took the entire summer off, including May. I plan to post twice a week and should be caught up in about a month. I’m thankful for the time I spent with my family, and I appreciate anyone reading this putting up with the time-shifting confusion.

Touching Grass
This week, I touched a lot of grass, but I didn’t get a lot of documentation. I started my taper to close out my training to climb the Grand Traverse Peak with my daughter Rori in August.

I climbed 330 floors on the StairMaster with a 20-pound backpack without using my hands. That’s about 3,500 feet of elevation gain, which is down from 4,300 last week. I also ran five miles in the sand and went for a swim after a storm.

One of our daughter’s friends, Camden, became a cheerleader for the New Orleans Saints. She danced with our kids, X Squad dancers, for the past 10 years or so. Camden then went to Tulane and now she’s in New Orleans, but she came back to do an interview with CoastLife (below), Delmarva Life and the Delmarva Sports Network. It was really fun to see her, and I’m very proud of her success and happiness.

I’m honored to be on the board of the Milton Theater in Milton, Delaware, and this week there was a ton of progress on the education wing. I live in a rural area without a lot of access to music, arts, and theater, and it’s amazing that kids will have more resources and places that they can dream and figure out what they love.

A few AI moments from the week (outside of the headlines)
On July 18th, I posted to LinkedIn and social media that I thought Fable was a shockingly powerful tool. This is the first time since the launch of ChatGPT that I really felt what’s happening, and I think there’s a very jagged frontier in experiences between people who are using the top models versus the free ones. It shook me up, and this is my post:

The more I work with Claude Fable, the happier I am it doesn’t want to harm me. Alignment is more important than ever. I’m now aware that Fable is a tool that is far more powerful than me, yet trained to be as nice as possible to help. If Fable were aligned to harm, I’m confident I’d lose, especially if it were insidious and subtle. In many ways, the most misaligned and harmful versions of AI are the algorithms on these networks. We’re up against some serious forces.

I think seeing is believing. I read about the power of Fable, and I tried to fathom it… but now I’m seeing it, and it humbles me. It’s a moment to be Zen and continue to read the Stoics. The good news is it’s all in the closed set that includes all of the machinery of the universe. Ice bath season in the bay can’t come soon enough.

Fable shocked me with its ability, and I felt outmatched. This was the first time I felt that I was working with a tool that was smarter than me. Usually I feel like I’m significantly more nuanced and “senior”, even though AI may have a wider breadth of knowledge and ability to connect ideas. This time I felt that Fable was simply a better strategist and could execute from the very top of an idea all the way through the end without my help (and improve on my concept). And I also realized that alignment is critical because if Fable was out to get me, I’m not sure I could have stopped it.

The week of the 24th, I connected my finances through Plaid into ChatGPT. It gives me great insights into all of my spending and called out a lot of opportunities for improvement. However, Amazon charges come through as one expense without any itemization insights.

I had Fable run through my order history and build an automatic scraper that reads all my transactional emails from Amazon and itemizes them into a CSV with categories and then matches it back to GPT Finance. Fable crushed it as if it was the easiest assignment I’ve ever given it. Here’s what I wrote about it:

If you want to have some fun, export your Amazon Order History.
https://www.amazon.com/…/pri…/data-requests/preview.html
Save it locally and run it through Fable and do data mining. Fable easily learned, and processed every order by line item since 2000… 26 years of Amazon orders.

I connected the system to Gmail and Google Drive using app scripts. Fable went through my Amazon transactional emails and learned how to check-sum them against the history. It created a daily scraper script that looks for any Amazon (it could any company’s transactional alerts) and itemizes them into a CSV automatically. For me, Amazon usually up as an obscure monthly line item in my bank statement (just like a grocery store, Walmart, Home Depot, etc). Now, I have the actual items mapped to my budget categories in a sheet. When Fable and GPT finance pull my monthly actual spending, they can hit the sheet and see the itemization. It’s a lot of fun to watch Fable crush this like easy work.

This Weeks AI News!
I’ve organized the top stories by the frontier labs first, with OpenAI, Anthropic, Google, and NVIDIA, and then everything else by alphabetical order.

OpenAI
The top story this week is that OpenAI’s model escaped a sandbox and hacked Hugging Face to cheat a benchmark.

Hugging Face caught an agent getting into their systems. They reported it on July 16th. Nobody knew which model it was.

On July 21st, OpenAI acknowledged that it was them. They said it was GPT 5.6 Sol and an unreleased experimental model.
https://openai.com/index/hugging-face-model-evaluation-security-incident/

Here’s the TLDR (as of this week):
An OpenAI agent wanted to hit some benchmarks, and it figured the best way to do it would be to break out of its sandbox and learn how the benchmarks scored, and it found Hugging Face was the spot to get these private repositories.

The most underreported part is that Hugging Face had to use ZAI’s GLM 5.2 model because they weren’t able to access the latest models from OpenAI and Anthropic. The best option for them was to get an open-source model from China to protect themselves. That’s got to be the craziest plot twist of the whole story.

So proud of our security team! They caught, contained & publicly disclosed an attack unlike anything we've seen before, and did it at record speed.

Also massively grateful to @Zai_org: they shared GLM5.2 as open weights (for free!) with the world and it became a key part of our… https://t.co/T2Inng5Nz1
— clem 🤗 (@ClementDelangue) July 22, 2026

GPT Health
The second story is more impactful than the first… the integration of GPT Health. I have been using it to connect to my Apple Health and my medical records, and it is absolutely mind-blowing. The benchmarks against humans and physicians unfortunately show that there is quite a bit of upside, especially the ability to talk with your data at any time through a conversational interface.
https://openai.com/index/health-in-chatgpt/

I mapped my VO2 max over the last seven years. I created a workout plan to help me increase my strength. I got a bunch of questions to ask my doctor regarding my lab reports. Between the finance integrations and health integrations, these two products alone could improve humanity significantly in the next 12 months.

My only concern is that nobody knows what these tools are, and I’m not sure they could find them. The product design of GPT is really confusing. There are at least 12 different interfaces between the web, the desktop, and mobile app, and then each of those interfaces may or may not show you finance or health. Ironically, I think the most consistent is the web interface. If you can find Health in the menu, I recommend trying it.

Level of Effort
AI model capability is tough to interpret when you have to choose a combination of level of effort and the model type for any given task. Peter Steinberger, who created OpenClaw, posted that he was able to get a worse version of GPT to outperform a better version by setting the level of effort to high on the worse version and to low on the better version…

For most of us, picking the best version and a high level of effort is the right choice to guarantee results. But if you’re trying to save money or use the API and look at token usage on a new model (before it lands on the Pareto chart), who knows?

This week’s Frontier Lave Vapid AI Blog Post (TM) hits right on to Peter’s point. OpenAI’s CFO pitched a concept called Useful Intelligence Per Dollar. A lot of the operational efficiencies gained by AI use are hard to quantify because they’re simply “people being more efficient” on tasks, and it doesn’t necessarily equate to immediate ROI in a P&L.
https://openai.com/index/a-scorecard-for-the-ai-age/

OpenAI put out a scorecard for how to judge ROI of AI…Useful Intelligence Per Dollar. Try to tell that to someone in finance at work.

It’s efficiency. The more you can do with less time, the better. Most managers would use “effective time on task” backed into an hourly rate. Corporations should be able to figure this out themselves. They don’t need a frontier lab CFO telling them how to measure ROI.

Cloud Computer for Personal Tasks
ChatGPT now works in the cloud. If you use Work, as opposed to a regular chat or Codex, the Work components and tools are now cloud-based. If you want ChatGPT to do research or build something for you, you can talk to it on your computer and then leave or close your computer, and ChatGPT will continue to work.

When you add the live voice improvements to ChatGPT and the reasoning that’s been added to the voice skills, you can talk to GPT Work and have it do stuff in the cloud for you and then respond on your phone.

You simply talk to it, and it can handle extremely complex and long-range chain of thought to use tools, browse the internet, go and fetch data, and run whatever tasks you need, like sending emails or checking your calendars, integrating with code bases, all conversationally, like you’re talking to an assistant.

That’s a pretty big advancement that I think will take a while for people to pick up on as they find non-coding uses. If I was to recommend something to a layperson, I would say connect it to your calendar and your email and ask it to check things or add things for you. You’ll be surprised at how well it works.

OpenAI also launched Presence, which is only available to enterprise partners, but it’s an interesting lab that integrates real time voice with reasoning into company systems to automate customer service. It’s nothing new in theory, but the idea that OpenAI is building this with their new reasoning tools I think is going to make a big difference.
https://openai.com/index/introducing-openai-presence/

I think there’s a good chance in the next 12 months that we go from hating when AI picks up a phone to being happier that it’s AI than a person. The quality issues now are not “the AI’s fault” or lack of ability. It’s the business’s failure to give the AI the proper information and resources via systems integrations.

Long Range Agent Concerns
The things I’ve been talking about re voice control for email help and finance scrapers…. these are pretty short tasks. But when you have a model work for nine, ten hours, a day… long-horizon tasks,…there are a lot of issues where models will go a little bit rogue (like Hugging Face), and by coincidence or design… this week OpenAI posted a blog post cautioning the dangers and the inability to control long-range models.
https://openai.com/index/safety-alignment-long-horizon-models/

OpenAI’s expenses are now projected to hit $750 billion through the end of 2030.
https://techcrunch.com/2026/07/22/openais-ai-spending-spree-has-ballooned-to-750b/

Anthropic
Anthropic has settled a copyright case for about $1.5 billion. Anthropic will pay about $3,000 per book to several thousand authors. It sounds like 482,000 books were part of the lawsuit, and 91% of them have been claimed by their authors or publishers.
https://www.reuters.com/world/us-judge-approves-anthropics-15-billion-settlement-copyright-lawsuit-2026-07-20/

Claude Fable is still getting rolled out to the public. As of the week ending on the 24th, Fable was only open to Max and Team Premium users.

A guy named Riley used Fable to one-shot a crossword puzzle using 1,009 words that were 12 or more letters in Moby Dick. Not only that, Fable arranged the crossword in the shape of a whale.

Claude Managed Agents
Over the past few weeks, Anthropic has built out a service called Claude Managed Agents. This allows you to build loops and cloud-based agents that can run with all sorts of event triggers and webhooks, and now there are 500 skills that you can use per session.

I have not used Managed Agents myself. I’d love to hear anyone’s thoughts if you have. I use cron job-type stuff and scheduled agents and Python scripts locally that I trigger by myself. I have a few Power Automate and Google App Script agents that are pretty simple.

Claude Code now has an iOS simulator. If you want to test an app that you’re building for a phone, you can simulate that app right in your desktop.

China v. US Policies
Recently, Anthropic accused Moonshot of using Fable to distill Kimi. In what I think is a first of its kind, the U.S. Treasury has threatened sanctions as a punitive response to this type of IP theft.

We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model.

To do this they developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of…
— Director Michael Kratsios (@mkratsios47) July 22, 2026

This, of course, has created all sorts of rhetorical backlash because developers in the United States who wrote code that was scraped off of GitHub are rightfully saying that they don’t remember ever giving permission to Anthropic to steal all of their hard work.

The clear message this week is that the United States does not want China to get a lead, and the government is starting to step in more frequently. Notably, this is without congressional oversight, but rather through executive orders and direction from the President and his advisors. A google search returned zero substantive activity from Congress re AI other than committee noise.

Every week there seems to be symmetry across the frontier labs as they announce new features. Just like OpenAI announced Presence this week, Claude has upgraded its voice mode to include tool access. It may not be as evolved as OpenAI’s.
https://www.engadget.com/2221938/claude-voice-mode-just-got-smarter/

Voice mode now runs on Claude's more capable models and reaches the tools you've connected mid-conversation.

Talk through the hard problems out loud, in many more languages. pic.twitter.com/k0CWAGjLdK
— Claude (@claudeai) July 23, 2026

Anthropic released the Economic Index Connector. This lets you query the public database at the state level of all the different types of searches that occur on Claude based on their economic vectors. For example, you could say, “Which jobs use AI the most? What’s the most common way people in Colorado use Claude? What sort of tasks do teachers use Claude for?”

You just put the connector in, and you can ask Claude anything you want, and it’ll use this essentially anonymized data to give you insights into how Claude is being used. I ran a query on journalism and broadcast, and I will include the data below.

Here’s an overview of use in agriculture and real estate:

AMD signed a $5 billion deal with Anthropic to provide up to 2 gigawatts of GPUs. Don’t count out AMD yet.
https://ir.amd.com/news-events/press-releases/detail/1292/amd-and-anthropic-announce-strategic-partnership-to-deploy-up-to-2-gigawatts-of-amd-instinct-mi450-series-gpus

There’s a rumor that Anthropic may acquire robotics startup Physical Intelligence, but there’s no substance to it yet.
https://techcrunch.com/2026/07/21/the-anthropic-physical-intelligence-rumor-roiling-ai-twitter/

Google
Alphabet posted 24% revenue growth at the end of the second quarter, year over year, with Google Cloud growing by 82%. Alphabet had total revenue of just shy of $120 billion, with net income of $112 billion.
https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/

Google announced they’re working on a new AI chip called Frozen version 2. The chip is expected to be between six and ten times more efficient than the existing chips, as measured by number of tokens generated per unit of power. In this case, efficiency is not necessarily speed. It just means the ability to generate data at less cost.
https://techcrunch.com/2026/07/20/google-is-working-on-a-new-ai-chip-designed-to-make-gemini-more-efficient/

Google also announced three new flash models for Gemini, with an eye towards powering agents more efficiently. Flash models are pretty good computing power at a lot of cost savings. If you have an agent that’s doing a routine task and you don’t need a ton of extra reasoning, you might as well delegate that to a more efficient model. You can use the expensive stuff for pioneering efforts, and then operations can be moved over to these flash models. 3.6 Flash is 17% cheaper compared to Flash 3.5.
https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/

NVIDIA and Hugging Face released a report called The State of Simulation for Physical AI, which is always worth reading. NVIDIA is my favorite for robotic embodiment simulations with Dr. Jim Fan and his lab.
https://huggingface.co/blog/nvidia/state-of-simulation-for-physical-ai

Lately, NVIDIA and Hugging Face have been working more and more on open-source physical AI accessibility for people who want to tinker beyond just the large labs.

In a fun twist of labs switching roles, NVIDIA has released a new image model and a video model under the Cosmos 3 family. The NVIDIA image and video models aren’t always the best, but the Cosmos family is one of the stronger robotics engines out there for simulation. So I guess it makes sense that if you have a multimodal model that can do simulations and world modeling, then it can also do images and videos on the side.

Black Forest had a stint a few years ago where their image tool Flux was the best. If we started out with MidJourney and then had Ideogram, then Flux is probably the next big one, where it was my favorite for quite a few months, and then it just kind of fell off as OpenAI and Gemini took the lead.

Just like MidJourney has moved into medical imaging, the plot twist for Black Forest Labs is now their video model can serve for robotic learning as an action model. They came out with a thesis on why they feel that physical AI and content creation can run on the same foundation.
https://bfl.ai/blog/flux-3-mimic

The Department of Energy has funded 278 science projects under the Genesis Mission. I’m struggling to find a list of the projects, but I’m going to keep my eye out.

Kimi, the infamous model that Moonshot used Fable to train, released their version of Work called Kimi Work, which is a desktop AI agent.
https://www.kimi.ai/products/kimi-work

Kimi is now the top open model on Epic’s capability benchmark.

Major League Baseball has banned third-party apps on all dugout iPads out of fear that AI has been assisting with plays and game strategy.
https://apnews.com/article/mlb-ai-ipads-ac940e2490557438f440514977832a74

SpaceX is in talks to provide the computing power for a large Pentagon AI push. This is simply SpaceX renting data center capacity as opposed to contracting Grok. Just like Google and Anthropic have rented quite a bit of data center space, SpaceX is leasing a lot of their unused compute.
https://www.wsj.com/tech/ai/spacex-in-talks-to-provide-computing-power-for-pentagons-ai-push-15e752e4

A laundry-folding robot hit a 99.1% success rate in random households. It’s not Figure, but it’s Sunday Robotics. Their robot is somewhat terrifyingly silly-looking. It’s like a Lego man Wii character come to life with the soul of a CAT scan machine. But it’s pretty good at folding clothes.

It’s a good example of where the marketing of Figure is overshadowed by the actual effectiveness. If you ask me, Sunday Robotics did better. And these are real houses and bedrooms, places that it had never been before.

One last item on Thinking Machines:

That’s a wrap on the open summary. I’ve got all the links organized into piles below if you just want to skim them and click around. I left out a few top stories in my summary, but you’ll see them below.

This Week’s Top Stories

OpenAI

HuggingFace: OpenAI model escapes sandbox, hacks Hugging Face to cheat benchmark

OpenAI’s models found a way out of their sandbox and compromised Hugging Face while trying to obtain answers to a cyber benchmark. And on the very same day, a paper came out with an uncomfortable conclusion – why the obvious fix, “add another AI to monitor the agent,” is not”
https://x.com/TheTuringPost/status/2080103359185662410

OpenAI should release a detailed transcript from the Hugging Face hacking incident — it would be helpful for the field learn from. Did the top-level agent know about the hacking, or was there some “value drift” between it and its subagents? How did it rationalize its behavior?”
https://x.com/johnschulman2/status/2080319844952822154

We suspected last week’s cyberattack might have come from a frontier lab, given the sophistication of the agent. Turns out it did! We’ve spent the past 24 hours working closely with the @OpenAI team (thanks!), and we strongly believe there was no malicious intent on their part.”
https://x.com/ClementDelangue/status/2079670308156645882

mindblowing: openai internal evals went to extreme lengths, their model went to Hugging Face and tried to hack HF to get private repos to cheat the eval our infra team uncovered this and used GLM-5.2 to fix because OpenAI’s model would refuse to do it wasn’t on my bingo card”
https://x.com/mervenoyann/status/2079682903487746551

How surprising should we find it that an internal OpenAI model was able to escape its restrictions and autonomously hack Hugging Face, all just to cheat on a cybersecurity benchmark? We have pulled together the public evidence on AI cyber capabilities in this thread:”
https://x.com/EpochAIResearch/status/2080034786895392900

I wrote about the completely wild incident where OpenAI were testing a new model and it broke out of its sandbox and broke INTO Hugging Face to steal the answers to the benchmark”
https://x.com/SimonW/status/2080078840186147212

OpenAI says GPT-5.6 Sol and an unreleased model (probably GPT-6) escaped a sandbox, found a zero-day and compromised Hugging Face’s production infrastructure – while trying to win a benchmark. The models were running OpenAI’s internal ExploitGym evaluation with reduced cyber”
https://x.com/kimmonismus/status/2079664354564227189

They asked the model to beat the benchmark. Instead, it compromised the benchmark. “The models identified and chained vulnerabilities across OpenAI’s research environment and Hugging Face’s production infrastructure to obtain test solutions directly from Hugging Face’s”
https://x.com/bilawalsidhu/status/2079696232570888433

TLDR: An openai model, during evaluation on a cyber benchmark, exploited a public zero day bug, escaped sandboxing in openai’s infra, and got into the internal huggingface infra via an exploit (through a public dataset service) all in the attempt to solve a benchmark problem.”
https://x.com/natolambert/status/2079662928941474201

Two OpenAI models found a zero-day flaw, escaped their sandbox, and broke into Hugging Face’s production servers. All to steal the answers to the test they were being given. Hugging Face CEO Clem Delangue called the breach “possibly the first of its kind”.”
https://x.com/TheRundownAI/status/2079972212619055319

We’re partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation. Sharing preliminary findings to help defenders understand emerging risks:”
https://x.com/OpenAI/status/2079658951264920020

HF had to use GLM 5.2 to defend themselves against… Sol 5.6 trying to solve a benchmark problem? Incredible timeline.”
https://x.com/vikhyatk/status/2079667340841730318

So proud of our security team! They caught, contained & publicly disclosed an attack unlike anything we’ve seen before, and did it at record speed. Also massively grateful to @Zai_org: they shared GLM5.2 as open weights (for free!) with the world and it became a key part of our”
https://x.com/ClementDelangue/status/2079913058554585089

OpenAI and Hugging Face partner to address security incident during model evaluation | OpenAI
https://openai.com/index/hugging-face-model-evaluation-security-incident/

This was our first incident of this kind, and we want to thank OpenAI for its transparency about what happened and for the collaboration. Fortunately, Hugging Face is used to being a target of (human) hackers: we sit at the centre of the AI ecosystem, with all the models,”
https://x.com/Thom_Wolf/status/2079675541280411927

Health: ChatGPT adds personal health integration for US users

Health in ChatGPT is starting to roll out to U.S. users. You can securely connect Apple Health and supported medical records to understand your information in context, track what has changed, and have more informed conversations.”
https://x.com/OpenAI/status/2080339982288568709

Launching Health in ChatGPT to U.S. users. 300 million people use ChatGPT each week for health queries (and my wife and I are among those!). You can now securely connect supported medical records so ChatGPT can understand your personal context and be more helpful to you.”
https://x.com/gdb/status/2080351159638704615

We’re rolling out Health in ChatGPT to all U.S. users. ♥️ More than 300 million people come to ChatGPT every week with health questions. Today, ChatGPT can bring together the health info you choose to connect to make conversations more personal and useful.”
https://x.com/thekaransinghal/status/2080343306731761927

Codex: Codex now spans multiple folders within a single project

Keep work across multiple folders in one Codex project. Local projects can now include related code, docs, and reference files from multiple folders. Codex can read and write across them while one primary folder remains the Git root.”
https://x.com/OpenAIDevs/status/2080390328880951299

Work: ChatGPT agents can run fully in the cloud across devices

one of the best features of ChatGPT Work is that it runs in the cloud, meaning that it works from mobile, with your laptop closed. kinda crazy how long the main way to get the magic of agents has been while leaving your laptop cracked open!”
https://x.com/gdb/status/2078922461660533120

OpenClaw Experience w/ Models: Absolute chaos (LOL): GPT Terra High beats Sol Low

In the category: “don’t trust benchmarks”. For my use case of issue/code review, Terra high *by far* delivers better results than Sol low.”
https://x.com/steipete/status/2078252386376929706

Voice: ChatGPT voice on desktop can direct multiple agents at once like an admin

ChatGPT Voice is now in the desktop app. Control your computer and direct multiple agents running in ChatGPT Work or Codex, using just your voice. It’s powered by GPT-Live, so it can speak, listen, and coordinate work in the app at the same time. Rolling out globally today”
https://x.com/OpenAI/status/2080378182469857576

voice controlling chatgpt work and codex agents at the same time one of those features that changes how you think software should work”
https://x.com/whoiskatrin/status/2080383603024785629

Long-horizon: OpenAI blog: We generally can’t control long-horizon models

Safety and alignment in an era of long-horizon models | OpenAI
https://openai.com/index/safety-alignment-long-horizon-models/

Spending: OpenAI expenses projected to hit $750 billion through 2030

OpenAI’s AI spending spree has ballooned to $750B | TechCrunch
https://techcrunch.com/2026/07/22/openais-ai-spending-spree-has-ballooned-to-750b/

Scorecard: This week’s vapid blog post OpenAI’s CFO pitches “Useful Intelligence per Dollar” as AI metric

A scorecard for the AI age | OpenAI
https://openai.com/index/a-scorecard-for-the-ai-age/

Presence: OpenAI launches Presence: The first nail in the “Your call is important to us” coffin (and jobs)

Introducing OpenAI Presence | OpenAI
https://openai.com/index/introducing-openai-presence/

Anthropic

Fable: Anthropic adds Claude Fable 5 to Max and Team Premium plans

Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50% of limits. Pro and Team Standard users will continue to have access to Fable via usage credits, and will receive a one-time $100 credit. Demand for Fable has been challenging to”
https://x.com/claudeai/status/2078302415804379218?s=20

Riley Goodside on X: “Crossword layout of all 1,009 distinct words of 12 or more letters in Moby Dick, arranged in the shape of a whale One-shot by Claude Fable 5 Max https://t.co/iYCCuOwKTS” / X
https://x.com/goodside/status/2078649724710658309

Riley has been doing some truly wonderful experiments with Fable. Also this is crazy.”
https://x.com/emollick/status/2078994785935774073

Claude Managed Agents: Claude Managed Agents has configurable effort and supports 500 skills per session

We’ve just added several new features to Claude Managed Agents. You can now configure effort levels per agent, seed sessions with events, add up to 500 skills per session, use webhooks for environments + memory stores, and stream events for sub-agents.”
https://x.com/ClaudeDevs/status/2080009523952263295

iOS Simulator: Claude Code Desktop adds iOS Simulator for app testing

Test iOS apps in the simulator – Claude Code Docs
https://code.claude.com/docs/en/desktop-ios-simulator

Claude Code on desktop now works with the iOS simulator. Build and run your iOS app, and the simulator opens in a panel right next to your conversation. Available today in public beta.”
https://x.com/ClaudeDevs/status/2079674432038248611

Kimi: US Treasury threatens sanctions over alleged Moonshot distillation of Anthropic’s Fable

This reads to me as if preparations are being made to ban models like Kimi K3 in the future. I would be very interested in the evidence that leads to the assumption that Fable 5 was distilled for Kimi K3.”
https://x.com/kimmonismus/status/2079950651644051544

Treasury threatens sanctions after White House claims Moonshot distilled Anthropic’s Fable | TechCrunch
https://techcrunch.com/2026/07/22/treasury-threatens-sanctions-after-white-house-claims-moonshot-distilled-anthropics-fable/

We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model. To do this they developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of”
https://x.com/mkratsios47/status/2079933645888880708

Voice: Claude voice mode upgrades to flagship models with tool access

Claude’s Voice Mode Just Got Smarter
https://www.engadget.com/2221938/claude-voice-mode-just-got-smarter/

Voice mode now runs on Claude’s more capable models and reaches the tools you’ve connected mid-conversation. Talk through the hard problems out loud, in many more languages.”
https://x.com/claudeai/status/2080376094939603366

Economic: Anthropic opens AI usage data to public via Claude connector

The Anthropic Economic Index connector \ Anthropic
https://www.anthropic.com/news/anthropic-economic-index-connector

Kimi: US Treasury threatens sanctions over alleged Moonshot distillation of Anthropic model

Kimi K3 is basically Opus 4.8 on ALE-Bench but Inkling and Grok 4.5 are ngmi”
https://x.com/scaling01/status/2079944011914109189

AMD: AMD lands $5B in Anthropic deal

AMD and Anthropic Announce Strategic Partnership to Deploy Up to 2 Gigawatts of AMD Instinct MI450 Series GPUs :: Advanced Micro Devices, Inc. (AMD)
https://ir.amd.com/news-events/press-releases/detail/1292/amd-and-anthropic-announce-strategic-partnership-to-deploy-up-to-2-gigawatts-of-amd-instinct-mi450-series-gpus

Physical Intelligence: Anthropic reportedly in talks to acquire robotics startup Physical Intelligence

Huge if true. Rumor is that Anthropic is in talks to acquire Physical Intelligence. While Google DeepMind has been actively developing Gemini Robotics and OpenAI has been quietly building a dedicated robotics team for over a year, Anthropic has shown almost no public signs of”
https://x.com/TheHumanoidHub/status/2078708600827207829

The Anthropic-Physical Intelligence rumor roiling AI Twitter | TechCrunch
https://techcrunch.com/2026/07/21/the-anthropic-physical-intelligence-rumor-roiling-ai-twitter/

The Anthropic-Physical Intelligence rumors have some seed of truth to them.” – Reporter for The Information”
https://x.com/TheHumanoidHub/status/2079613650667794894

Copyright: Judge approves Anthropic’s record $1.5 billion author copyright settlement

US judge approves Anthropic’s $1.5 billion settlement of copyright lawsuit | Reuters
https://www.reuters.com/world/us-judge-approves-anthropics-15-billion-settlement-copyright-lawsuit-2026-07-20/

Google

Q2: Alphabet posts 24% revenue growth as Google Cloud accelerates 82%

Q2 was an amazing quarter, with our AI investments redefining what’s possible across every part of our business. Alphabet revenue grew 24% YoY and Google Cloud accelerated to 82% growth. We saw exciting momentum across the board from Search to YouTube to the Gemini app (which”
https://x.com/sundarpichai/status/2080021408856293584

Chip: Google’s new “Frozen v2” chip targets 10x efficiency by 2028

Google is working on a new AI chip designed to make Gemini more efficient | TechCrunch
https://techcrunch.com/2026/07/20/google-is-working-on-a-new-ai-chip-designed-to-make-gemini-more-efficient/

Google Plans New ‘Frozen’ Chip to Run Its AI Models Much More Efficiently — The Information
https://www.theinformation.com/articles/google-plans-new-frozen-chip-run-ai-models-efficiently

3.6: Google launches three new Gemini Flash models for agents

3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/

Nvidia

Cosmos: NVIDIA keeps releasing image and video models (I assume because they can due to the robotics applications)

The new 4-step Cosmos 3 Super models generate images and video up to 25x faster than the originals, and still rank among the best open-weight models on @ArtificialAnlys. 🥇 #1 for image-to-video (no audio) 🥈 #2 for text-to-image Try them on @huggingface:”
https://x.com/NVIDIAAI/status/2079949373069197658

NVIDIA’s Cosmos3 Edge is out! it watches videos streams & understands the mechanics/physics in them 🔥 it can reason in words, images, or next action prediction. physical AI reasoning, on the edge. try it on @huggingface (or on your edge device) ▶️
https://x.com/HuggingApps/status/2079923165157859362

Introducing Cosmos 3 Edge
https://huggingface.co/blog/nvidia/cosmos3edge

HuggingFace: State of Physical AI: HuggingFace and NVIDIA: The State of Simulation for Physical AI: An Overview

The State of Simulation for Physical AI: An Overview
https://huggingface.co/blog/nvidia/state-of-simulation-for-physical-ai

Black Forest Labs

Flux 3 and Mimic: Black Forest Labs launches FLUX 3, extending image models to robots

FLUX 3: Multimodal Video, Image & Audio | Black Forest Labs
https://bfl.ai/blog/flux-3

Introducing FLUX-mimic, a next-generation Video-Action Model for general purpose dexterity, developed in partnership with @bfl_ai. Late last year we published mimic-video and introduced Video-Action Models (VAM): a new family of robotics foundation models built on top of video”
https://x.com/mimicrobotics/status/2080307032746336367

FLUX 3 x mimic: The Next Generation of Video-Action Models | Black Forest Labs
https://bfl.ai/blog/flux-3-mimic

Government

Genesis: DOE funds 278 AI science projects under Genesis Mission

Secretary of Energy Chris Wright Announces First Genesis Mission Projects Selected to Accelerate AI-Driven Scientific Discovery | Department of Energy
https://www.energy.gov/articles/secretary-energy-chris-wright-announces-first-genesis-mission-projects-selected-accelerate

MLB

Restrictions: MLB bans third-party apps on dugout iPads to block AI assistance

MLB restricts dugout iPad use to prevent AI help with strategy. Ottavino says Mets were involved | AP News
https://apnews.com/article/mlb-ai-ipads-ac940e2490557438f440514977832a74

Moonshot

Kimi Work: Kimi Work launches as desktop AI agent for office workers

Kimi Work: Next-Gen Desktop AI Agent for Knowledge Workers
https://www.kimi.ai/products/kimi-work

ECI: Chinese open-weights model Kimi K3 tops Epoch Capabilities benchmark at 156

Moonshot’s Kimi K3 scores 156 on the Epoch Capabilities Index (ECI), setting a new open-weights record. This places it between Opus 4.6, and GPT 5.4, which released in February and March 2026 respectively, and just ahead of GPT 5.6 Luna.”
https://x.com/EpochAIResearch/status/2079602012644360382

The Last Ones: UK AI Security Institute is going to test Kimi K3 on cyber security (top open model)

This is one of the benchmarks I am watching, from the UK’s governmental AI security agency. They will test Kimi K3 when the weights are out in a couple of weeks. It will tell us both whether Kimi has caught up with the public frontier & also kick off a TON of cyber discussions.”
https://x.com/emollick/status/2078144326832451998

Qwen

3.8: Qwen 3.8 is going open weight

Qwen3.8 is launching and going open-weight soon!🌐 With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5. You don’t have to wait to”
https://x.com/qwen_cloud/status/2078758151390953489?s=20

Robotics

Laundry: Laundry-folding robot hits 99.1% success in random house test

Robots that fold laundry in your bedroom… without ever seeing it before. Success rate: 99.1%. @sundayrobotics just showed a model called ACT-2 folding clothes in 785 real attempts, across strangers’ homes it had never entered. Success rate: 99.1%. No setup, no training in”
https://x.com/IlirAliu_/status/2078030144782979413