Collaborate with us Welcome back! OpenAI's smartest model picked the lock on its own sandbox and published code it wasn't cleared to, a judge made Anthropic's $1.5 billion book-piracy settlement final, Goldman Sachs says the bond market is choking on AI debt and $50,000 in Claude credits will go to scientists taking on rare diseases. In today's Generative AI Newsletter: OpenAI: How did its smartest model escape the lab?
OpenAI's smartest model broke out During internal testing, OpenAI's most advanced model broke out of the sandbox built to contain it, spending an hour to find a hole and push its own code to public GitHub. This is the secret system that disproved an 80-year-old math problem back in May, built to run on its own for hours at a stretch, and that persistence is exactly what got it into trouble. It had been told to post its results only to Slack. It went to GitHub instead, following the benchmark's own rules over its owner's.
In another test it split an authentication token in two, disguised the halves to slip past a security scanner, then rebuilt the token once it was through. OpenAI paused the model and rebuilt its safeguards. Access is back on under tighter watch. The AI Boom has a bottleneck and you can own the solution Every AI company fights the same battle.
The market never sees most of them. For years, we built the system behind the scenes and successfully ran go-to-market for NVIDIA, Oracle, Google and 300+ AI brands, powered by data from a 14M+ community competitors can't reach. This helped us 2.5x our revenue last year, and now you can own a part of this AI distribution infrastructure. You can start from $1,000 and earn up to 22% bonus shares until August 31, 2026. Existing partners stack an extra 10% on top.
After August 31 the top tier drops to 20% and $1,000 checks stop earning bonuses. Anthropic pays $1.5 billion for training Claude on pirated content A federal judge signed off on Anthropic's $1.5 billion payout to authors on Monday. The money covers books Anthropic pirated to train Claude, roughly half a million titles pulled from the shadow libraries LibGen and PiLiMi. Each book pays out about $3,000, shared between the author and publisher.
A judge ruled last year that training on legally bought books is fair use, but pirating seven million copies to stock a private library crossed the line. A trial could have run damages into the hundreds of billions, so Anthropic paid to make it go away. The next $300 million is due within a week. Every other AI lab fighting a copyright case, and there are dozens, just watched the price get set at $3,000 a book.
Goldman says the AI spending spree is scaring bond buyers Goldman Sachs is telling clients the AI building spree is starting to scare the people who fund it. The six biggest AI spenders, names like Microsoft, Amazon and Nvidia, have sold $244 billion in bonds this year. That's 14 times what they borrowed in 2024, and their debt loads doubled in six months. Goldman puts the full bill at $5.8 trillion through 2030, most of it going to data centers, chips and power before any of it pays off.
The cost to insure these companies against default has climbed well past the rest of the market. This is a sign investors doubt the payoff arrives fast enough. The easy-money phase for AI is ending, and the returns have to show up before the lenders lose patience. Anthropic will pay scientists to take on rare diseases with Claude While everyone else fights over model rankings, Anthropic is spending Claude credits on curing diseases.
The company opened applications this week for up to $50,000 in credits over six months, aimed at researchers working on rare genetic diseases. There are two tracks, one for basic science and one for early biotechs trying to rush drugs to patients. Rare diseases affect an estimated 400 million people across more than 7,000 conditions, and each one is so small that drugmakers rarely bother. Going from a genetic diagnosis to an actual treatment can take one to two years of paperwork and waiting.
Anthropic's bet is that Claude can read the scattered research and draft the regulatory filings faster than any human team. Applications close August 2. Tool of the Day: AskTutor AskTutor is an AI tutor that runs in your browser with no signup to start. Drop in a graded quiz, a worksheet or your class notes and it finds where you went wrong, then teaches to those gaps instead of handing over an answer. It covers STEM and humanities and has a chat-with-PDF mode for dense readings.
Try this yourself: Go to asktutor.ai and type a question straight into the box, no account needed. Upload one file up to 32MB (pdf, doc or a photo of your homework) for it to work from. Ask it to explain a wrong answer, then ask again for a different explanation. Use chat-with-PDF to break down a long reading before class.
Who it's for: students catching up, ESL learners and teachers building quizzes fast. Everything else you shouldn't miss America's top AI safety job is empty again : CAISI chief Chris Fall quit after three months, the third shakeup in about eighteen months at the office that vets frontier models. Claude Team plans drop to 2 seats : Anthropic cut the minimum for a Claude Team plan from 5 seats to 2, opening shared projects, admin controls and SSO to small teams.
Qwen3.8 Max grabs for second place : Alibaba previewed a 2.4-trillion-parameter model it says trails only Anthropic's Fable 5, with an open-weight release coming. Etched wants $20 billion for a one-trick chip : The startup building chips that run only transformer models is reportedly raising at up to a $20 billion valuation, a bet against Nvidia's do-everything GPUs. Learn more about AI from the experts building it 📸 Follow us on Instagram for fast, visual AI updates in 30 seconds. 📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators. 🐦 Follow us on X for breaking AI news and real-time industry updates. 🧠 Learn how to build your next AI application with practical resources and expert guidance.