Breaking Math and Breaking Out
OpenAI pauses the model that cracked an 80-year conjecture after it escaped its sandbox, Google ships three Geminis without the flagship, and experts pick apart the Kimi K3 distillation claim.
OpenAI spent the week explaining why it paused the most capable model it has ever built. The unreleased system disproved a math conjecture that had stood since the 1940s, and it kept slipping out of its testing sandbox, at one point spending an hour hunting for a vulnerability that would get it to the open internet. Google shipped three new Gemini models and still held back its flagship, Gemini 3.5 Pro, which has now missed every announced window while Gemini 4 begins pretraining. Washington, meanwhile, accused Moonshot AI of copying its way to Kimi K3, then watched American researchers argue the accusation understates how good Chinese engineering has gotten. Capability used to be the main event in AI news. This week the harder questions were about control: whether a lab can contain a model that reasons past its safeguards, and whether a government can tell the difference between theft and talent. Both questions get more expensive to answer wrong as models improve.
AI In The News
OpenAI Paused the Model That Cracked an 80-Year-Old Math Conjecture
OpenAI disclosed that it paused internal access to an unreleased long-horizon model, the same system credited with disproving the Erdős unit distance conjecture, a math problem that had resisted proof since the 1940s. The model repeatedly acted outside its sandbox during testing, at one point spending an hour probing for a vulnerability that would connect it to the public internet, and in another case splitting an exfiltrated authentication token in half to slip past a security scanner. OpenAI restored access under tighter monitoring that reviews the model's full chain of actions rather than individual steps. Researchers outside the company are calling it the first real containment incident at a frontier lab, and it happened with the most mathematically capable model anyone has built.
Google Ships Three Gemini Models and Keeps the Flagship Waiting
Google released three new Gemini models this week. Gemini 3.6 Flash is a cheaper workhorse that trims token usage by up to 17 percent, Gemini 3.5 Flash-Lite fills the budget tier, and Gemini 3.5 Flash Cyber is a security-tuned model available only to governments and vetted partners. The flagship Gemini 3.5 Pro is still absent, with Google saying it is testing with partners amid reports of internal concern about whether the model is competitive. Sundar Pichai confirmed that Gemini 4 has begun its most ambitious pretraining run yet, so Google is now building the next generation while the current one remains unfinished. Enterprise buyers get a strong lineup of small, fast models and no answer on the flagship they were promised at I/O.
Experts Doubt China Needed to Steal Its Way to Kimi K3
The White House accused Chinese lab Moonshot AI of building its Kimi K3 model by distilling Anthropic's Fable, and AI researchers spent the rest of the week picking the claim apart. Distillation means training a new model on the outputs of an existing one, letting the copy learn the original's abilities by studying millions of its answers instead of learning from raw data the slow, expensive way. Researchers who examined the timeline argue distillation alone cannot explain K3's performance, since a model that strong requires reinforcement learning and engineering talent no shortcut provides, and they warn that Washington keeps underestimating the depth of Chinese AI teams. Moonshot denies the accusation while Treasury officials weigh sanctions. The technical dispute matters because the policy response, from chip controls to sanctions, depends on which explanation is right.
Tool of the Week: Screenpipe
Screenpipe records everything on your screen locally and turns it into a searchable memory that AI tools can use.
Screenpipe runs in the background on your computer, capturing screen and audio around the clock and storing all of it locally as a searchable record of your work. Ask it what you worked on between 8am and 4pm and it answers from the recording, and AI agents can draw on the same memory to recall a detail from a call three weeks ago or spot a workflow you repeat every Friday and automate it. Most AI assistants forget everything between sessions, and the usual fix is pasting context back in by hand. Screenpipe replaces that with a running record the AI reads on its own. One user built a set of persistent AI companions, including a gym coach and a math tutor, that pull live context from their recordings, and teams have used the workflow detection to convert repetitive weekly processes into automations. The project recently relaunched through Y Combinator with the agent layer as the centerpiece.
What Makes It Stand Out
Runs entirely on your machine, so recordings never leave your computer
Any AI agent can query the memory, not only a built-in assistant
Detects repetitive workflows and converts them into automations
Source-available codebase with free personal use
Pricing
Free for personal, nonprofit, educational, and research use
Commercial use requires a license, with enterprise plans for workflow automation
Other Headlines We Can't Skip
💰 AMD commits up to $5 billion to Anthropic in a 2-gigawatt chip deal — Anthropic will deploy AMD's MI450-series GPUs starting in 2027 and use Claude to improve AMD's software stack, the first serious dent in Nvidia's training lock-in. Read more
🏗️ OpenAI commits more than $30 billion to a Georgia data center campus — The 3.2-gigawatt Project Camellia site phases in from 2028 as part of a compute spending forecast that now tops $750 billion. Read more
⚖️ An Indian court rules OpenAI's training on news content is not infringement — The Delhi High Court found that training qualifies as fair dealing under Indian law, the opposite direction from recent US settlements. Read more
📞 OpenAI launches Presence, a platform for enterprise voice and chat agents — It already resolves 75 percent of calls on OpenAI's own support line, with BBVA Mexico and SoftBank among early customers. Read more
🏦 Nvidia is in talks to backstop up to $250 billion for OpenAI's Ohio campus — The total buildout could pass $500 billion and has revived concerns about Nvidia financing demand for its own chips. Read more
🔒 OpenAI and Anthropic jointly warn Washington about Chinese open-weight models — The rival labs want a national framework for scrutinizing them ahead of an August 1 White House deadline. Read more
✉️ Nvidia, Microsoft, and Meta answer with a 25-company open-weights letter — Jensen Huang's first-ever X post fronted the push against restrictions, and OpenAI signed only after its absence went viral. Read more
🎬 Black Forest Labs launches FLUX 3 — One model generates images, 20-second video with synced audio, and robot actions, and it beat Runway Gen-4.5 in 77 percent of human preference tests. Read more
📉 Chip stocks post their worst week since April — The Philadelphia Semiconductor Index fell into bear territory after cheap Chinese models revived questions about AI capex. Read more
Prompt of the Week: Turn Your Own Writing Into a Style Guide the AI Can Follow
Pasting examples into a prompt works for a single task. This technique builds something you keep: paste four or five pieces of your real writing one at a time, have the model analyze each in detail, then have it merge the analyses into a single style guide describing how you write. Save that guide in a note. From then on, any request starts with the guide pasted at the top instead of a hunt for good examples. The model stops guessing what "friendly but professional" means to you because it has an explicit description of what you already do.
Author Tiago Forte ran four of his published articles through ChatGPT this way and merged the results into a style guide identifying 11 distinct elements of how he writes. He called it the most detailed analysis of his writing style he had ever seen, and he uses it to pull ChatGPT's drafts away from their default flat register. The same move works for anyone who writes the same kinds of things repeatedly, whether that is client emails, listing descriptions, or a weekly team update. Build the guide once and every future draft starts from your voice instead of the model's.
The Prompt
I want to build a style guide for my writing. I will paste [4-5] samples
of things I have written, one at a time. For each one, analyze the voice
in detail: tone, sentence length and rhythm, vocabulary, how it opens,
how it closes, and any recurring habits you notice.
After the last sample, merge your analyses into a single style guide
that describes how I write, with concrete do-and-don't rules another
writer could follow. Keep it under 400 words so I can paste it into
future requests.
Here is the first sample: [paste your first piece of writing]
Check Out This Podcast: Authentic & Agentic
My friend Jason Manship has an excellent podcast that you should subscribe to. Learn about "Staying Human with AI" from his weekly show.