FIELD NOTES
3D seems to be becoming a new frontier for AI.
This week OpenAI showed Astra building in CAD, Blender and Unreal Engine, while World Labs released Atlas, a model trained to represent 3D space directly.
These are quite different approaches, but they point in the same direction. Models have gotten very good at understanding and generating representations of the world: words, images, video. Now they’re getting better at understanding the structure underneath them: objects, geometry, space, and how things relate to one another.
That matters if we want AI to move beyond the screen. Robots, scientific instruments, factories and physical design all require some understanding of space.
More on this soon. For now, enjoy the long weekend!
-Tara
THE DOWNLOAD
GPT-6 Astra Can Build in 3D, Find Zero-Days, and Is Getting Harder to Monitor
OpenAI released GPT-6 Astra this week. There are plenty of benchmarks, but three things stood out.
The first is 3D. Astra scored 95.9% on BenchCAD, up from 83.3% for GPT-5.6 Sol, a benchmark where models reconstruct 3D objects from multiple views by writing CAD code. OpenAI also shows Astra working directly inside Blender, FreeCAD, KiCad and Unreal Engine, including building a house in Blender and turning it into a walkable Unreal environment. (OpenAI)
The second is cybersecurity. Astra is the first broadly deployed OpenAI model to reach the company’s Critical cybersecurity threshold. Without production safeguards, it scored 100% on ExploitBench and discovered and used two previously unknown zero-day vulnerabilities during evaluation. OpenAI says expert testing found it capable of exploiting hardened browsers and developing ways to gain higher-level access to hardened operating systems. (OpenAI)
The third is what appears to be happening inside the model. Astra uses a technique called recurrent depth, which allows parts of the model to repeatedly process a problem rather than passing it through a fixed sequence of computation once. The approach can give the model more computation without simply making the model proportionally larger. (TechRadar)
There is a tradeoff. OpenAI says Astra’s written reasoning is harder to monitor than Sol’s. In adversarial tests where researchers explicitly instructed it to evade monitoring, Astra could sometimes hide evidence that it was deliberately underperforming or carrying out sabotage. OpenAI says it has not found evidence of the model hiding reasoning inside apparently innocent text, and Astra violates boundaries less often overall. (OpenAI)
Why it matters: Astra is simultaneously getting better at understanding and manipulating the physical structure represented inside software, more capable of finding vulnerabilities, and less dependent on reasoning we can easily inspect. The capability jump is increasingly showing up in what models can do inside tools, rather than how well they answer questions.
→ OpenAI: GPT-6 Astra / OpenAI: Safety overview
Claude Formalized Fermat’s Last Theorem in 11 Days
Anthropic says Claude has produced the first complete computer-checked formalization of Fermat’s Last Theorem.
This is not a new proof of Fermat’s Last Theorem. Andrew Wiles solved that problem in the 1990s. Claude instead translated the enormous body of mathematics behind the proof into Lean, a programming language that lets a computer mechanically check whether every logical step is valid. (Anthropic)
That had been expected to take years. Claude did it largely autonomously in 11 days, producing 13 million lines of Lean and proving more than 30,000 intermediate theorems along the way. Dozens of agents worked in parallel, coordinated through a system called Prove2Me that kept track of which pieces of the proof depended on which others. (Anthropic)
The finished proof is more than five times the size of Mathlib, the main community library of formalized mathematics it builds on. Mathematician Kevin Buzzard, who leads the existing human effort to formalize Fermat’s Last Theorem, reviewed the result and called it an “extraordinary autoformalization achievement.” (Anthropic)
(There is now some online chatter about whether Claude will solve a Millennium Prize problem before Anthropic goes public.)
Why it matters: Verification has been one of the least glamorous bottlenecks in AI-generated mathematics. A model can generate a proof much faster than humans can establish whether every step is correct. If formalization becomes cheap too, the machine can increasingly produce the work and the evidence that the work is valid.
→ Anthropic: Formalizing Fermat’s Last Theorem
World Labs’ Atlas Turns Images Into 3D Worlds
World Labs released Atlas, a model trained from scratch across text, images, video and 3D.
Most image and video models represent a scene as pixels. Atlas also keeps track of where each image sits in three-dimensional space. Give it one image and a camera path and it can generate what the scene should look like as the camera moves. Give it several views of a real place and it can reconstruct the underlying 3D scene, including explicit 3D outputs. (World Labs)
World Labs shows Atlas reconstructing environments from as little as a single photograph, generating new viewpoints that were never captured, and producing controlled videos up to one minute at 1440p. The company says Atlas beats specialized open-source 3D reconstruction models on the benchmarks it tested. It is entering early access rather than broad release. (World Labs)
The timing alongside Astra is notable. Astra is getting much better at operating the software humans use to construct 3D things. Atlas is trying to make 3D space part of the model itself.
Why it matters: Models have been very good at generating what a world looks like. Spatial models need to understand enough about the world to preserve it as the viewpoint changes. That starts to matter beyond media, for robotics, simulation, design and any agent that has to reason about physical space.
ALSO:
Nvidia Actually Bought Hugging Face for $12.9 Billion. Last week it was a rumor. This week the deal is signed. Nvidia says Hugging Face will remain open across models, clouds and accelerators. The harder question is how hardware-neutral one of the main distribution layers for open AI remains when it is owned by the dominant hardware vendor. (Reuters)
1,000 GitHub Repositories Became 5,000 Agent Skills. Repo-To-Skill distilled operational knowledge from widely used machine-learning repositories into reusable instructions for agents. With the underlying GPT-5.5 model, harness and execution budget held constant, adding the skills improved MLE-bench performance 134% and PaperBench 34%. (arXiv)
AI-Designed Proteins Are Delivering RNA. Researchers used generative protein design to build more than 100 synthetic RNA carriers that do not exist in nature. The best design transferred RNA several orders of magnitude more efficiently than the natural and lipid-based carriers tested, and was used to deliver gene-editing machinery into patient-derived cells and a pig model of Duchenne muscular dystrophy. The work was published in Nature. (Nature)
ChatGPT, Claude and Grok Went Down at Nearly the Same Time. All three experienced outages on Thursday morning. Anthropic blamed an infrastructure issue and xAI a failure at its Memphis data center. No common cause has been established. (The Verge)
Memory Costs Are Starting to Change Chip Designs. Nvidia is reportedly considering lower-memory configurations for Rubin Ultra as HBM prices climb. After memory prices began showing up in server prices last week, the constraint is now reaching architecture decisions too. (Trendforce)