OpenAI is offering private-equity firms a guaranteed minimum return of 17.5% (1 minute read)
OpenAI is currently courting buyout firms to form joint ventures aimed at raising fresh capital and accelerating the adoption of enterprise AI products. It is offering private-equity firms a guaranteed minimum return of 17.5%, significantly higher than typical preferred instruments. It is also offering early access to its newest AI models. TPG and Advent are potential investors.
|
Claude Code and Cowork can now use your computer (1 minute read)
Claude Code and Claude Cowork can now open files, use browsers, and run dev tools on users' computers. The agent will prioritize connectors to supported services, but it can still execute tasks when connectors aren't available. It will ask for permission before taking action. Claude computer use is initially available to Claude Pro and Claude Max subscribers on macOS.
|
OpenAI rolls out ChatGPT Library to store your personal files (2 minute read)
ChatGPT's new Library feature allows users to store personal files or images on OpenAI's cloud storage. Available to Plus, Pro, and Business subscribers, the feature is rolling out to customers across the world except in the European Economic Area, Switzerland, and the United Kingdom. ChatGPT will automatically save uploaded or created files. Deleting a chat containing a file will not delete the file saved in Library.
|
|
Designing AI Chip Hardware and Software (Book)
The future of AI accelerators will be AI CPUs - traditional CPUs with big caches and systolic arrays, where the caches and systolic arrays take up most of the chip area and power. This book goes over what is required to design a competitive AI chip. The recommendations made in the book will require an excellent software team to realize, but the topics covered should be easily understandable at a hobbyist level.
|
A Ramsey-style Problem on Hypergraphs (6 minute read)
This post details a problem that would take an expert human about one to three months to solve. The problem was eventually solved by two people using GPT-5.4 Pro. A link to the full transcript of the conversation with GPT-5.4 Pro is available. Other models that have solved the problem include GPT 5.4 (xhigh), Gemini 3.1 Pro, and Claude Opus 4.6 (max).
|
Vibe physics: The AI grad student (32 minute read)
A professor of physics supervised Claude through a real research calculation from start to finish without ever touching a file to see whether AI could do theoretical physics. This resulted in a technically rigorous, impactful, high-energy theoretical physics paper in two weeks instead of the usual year. While Claude was impressively capable, domain expertise was essential for evaluating its accuracy. AI is still not able to do end-to-end science, but the project proved that it can do frontier science with guidance.
|
Research note: We spent 2 hours working in the future (15 minute read)
The rate of model releases and the number of new evals required next year could be such that keeping informed will be a challenge without effective AI assistance. These AI-augmented workflows need to be understood now before they become necessary. METR ran a tabletop exercise to learn what workflows emerged, what the bottlenecks are, and how much faster researchers will actually be. This post details the lessons learned from the exercise.
|
|
Fast regex search: indexing text for agent tools (28 minute read)
Providing text search indexes to fast models creates a qualitative difference for Agentic workflows. The impact is more pronounced in larger Enterprise repositories because grep latency scales with the size and complexity of the code being worked on. Removing the time spent searching the code base provides meaningful time savings and allows for much more effective iteration, particularly when the Agent is investigating a bug.
|
Why aren't we fine-tuning more? (1 minute read)
A good prompt can do a lot of the work that fine-tuning does without the cost. Models are getting so good that fine-tuning isn't really required to get good results. The extra work and resources required for fine-tuning aren't worth it anymore for most use cases. It is still worth knowing how to fine-tune, as there are other situations where the skill is useful.
|
|
OpenAI calls out Microsoft reliance as risk in investor document ahead of expected IPO (5 minute read)
OpenAI says that its close ties with Microsoft are a potential risk to its business as Microsoft is responsible for a substantial portion of OpenAI's financing and compute. Other risks include significant capital expenditures, reliance on compute resources, ongoing litigation with xAI, and its unusual structure as a public benefit corporation. OpenAI's operating results will depend on its ability to successfully develop relationships with additional partners. While OpenAI and Microsoft have a tight bond, they're increasingly competing for users in the generative AI market.
|
OpenAI Taps Former Meta Executive to Lead Ad Push (2 minute read)
OpenAI has hired former top advertising executive at Meta, Dave Dugan, as vice-president of global ad solutions. The company is urgently looking to generate new revenue streams to support its enormous funding requirements. Dugan is known for his close relationships with the world's leading ad companies. Personal relationships are a significant factor that influences where brands ultimately spend their advertising budgets.
|
|
|
Love TLDR? Tell your friends and get rewards!
|
|
Share your referral link below with friends to get free TLDR swag!
|
|
|
|
|
Track your referrals here.
|
|
|
|
