OpenAI Releases GPT-6.1 Terra as Runway and Nvidia Push Agent Tooling

Key Takeaways

1. OpenAI released GPT-6.1 Terra, pitched for screen understanding, cross-app actions and natural writing at a lower price; Sam Altman and Greg Brockman both framed computer use as crossing a threshold.
2. OpenAI and Nvidia each announced three discoveries in harmonic analysis, complexity theory and quantum field theory; separately, a claimed refutation of the Exponential Time Hypothesis is circulating as disputed/unverified.
3. Runway launched App Studio, which turns natural-language descriptions into runnable creative tools using its image, video and audio models plus Brand Kits.
4. Nvidia detailed European AI-factory and agent infrastructure at Berlin GTC and expanded a Muse model collaboration with Meta.
5. Hugging Face open-sourced 1,300+ 3D models (39 GB, CC BY 4.0) and released Papers with Code under Apache 2.0.
6. A benchmark audit of 15 AI benchmarks found 4 validated, 9 flawed and 2 with insufficient information, drawing a defense of the audit from Hamel Husain.
7. John Carmack argued robot demos overstate near-term capability, citing the difficulty of a pumpkin-lantern battery swap.

1. Models and Agent Products

1. OpenAI released GPT-6.1 Terra, described as understanding screen content, executing actions across applications and writing naturally for all-day work, at a lower price 1. Sam Altman called it strong across many tasks and highlighted a new era of computer use 2, while Greg Brockman said computer use with Terra has crossed a qualitative threshold and feels faster than his own unaided computer operation 3. The release posts and the quoted descriptions come from the same announcement, so the performance claims are vendor statements rather than independent results.
2. Runway launched App Studio, which lets users describe a creative app in natural language and generate a runnable tool without writing code 1 2. The tool calls Runway's image, video and audio models and combines them with users' Brand Kits, with outputs usable by individuals or entire teams 3.

2. Research Claims and Benchmark Scrutiny

1. OpenAI announced three discoveries in harmonic analysis, complexity theory and quantum field theory, in mathematics, computer science and physics respectively 1; the same three claims were relayed by Sam Altman and Greg Brockman, with Brockman saying AI is beginning to help deepen understanding of the universe 2 3. Ethan Mollick quoted the announcement and commented that it is "two and a half years away" 4. Separately, Brockman said P may still not equal NP 5.
2. A claim that the 27-year-old Exponential Time Hypothesis has been refuted is disputed/unverified: the account states that on October 16 OpenAI and Anthropic models independently solved 3-SAT with the same 2^(n/√log n) bound, which it says does not prove P=NP but exceeds prior expectations 1. No primary result is provided in the material, and Mollick separately noted that even AI models struggle to predict when the hypothesis will be resolved 2.
3. A Benchmark Reviews audit of 15 AI benchmarks found 4 validated, 9 flawed and 2 with insufficient information to review 1. Hamel Husain rejected criticisms that benchmark authors already knew the issues, that critics should contribute instead, and that verdicts are too binary, noting the benchmarks drive millions of dollars in token spending and that many people trust headline numbers for generalizability 2. He listed checks including whether failed tasks are real, whether harnesses and scoring are correct, whether user intent is disambiguated, whether model and API configurations are consistent, and whether scores align with real outcomes, and called the exercise a "red pill" for building private evaluations 3.

3. Infrastructure, Open Assets and Robotics Reality Checks

1. Nvidia used Berlin GTC to describe its European AI ecosystem: DSX POD is a Vera Rubin modular factory built, cabled and tested in a factory before delivery for production intelligence 1. The Agent Toolkit combines models, runtime frameworks, sandboxes, knowledge and tools so enterprises can control what agents know, access and execute, with OpenShell for added security; physical AI added AI-RAN, robotaxi and robotics customers, and RTX Spark brings personal agents to Windows PCs 2. Nvidia and Meta will expand their Muse model collaboration to improve inference and training performance, explore OpenShell for secure agent deployment and Sentry for monitoring agent behavior, and Meta plans broader deployment of Vera Rubin and Vera CPU 3.
2. Hugging Face open-sourced more than 1,300 3D models totaling 39 GB for free download, including input images, prompts and settings, GLB models with textures up to 8K, native Blender files for Opus and Astra, under CC BY 4.0 1. Papers with Code was also open-sourced under Apache 2.0 with the goal of jointly improving the papers interface 2.
3. John Carmack argued that flashy robot demos can misleadingly suggest robot doom is near, while no robot can currently complete a task like changing the batteries in a pumpkin lantern: opening the battery package, taking out a precision screwdriver, pressing and unscrewing four small screws and removing the old battery are each difficult 1. He acknowledged a purpose-built, trained robot could do this flow today but remains far from replacing the full range of human ability, so near-term doom scenarios should stop at "secretly controlling humans" rather than "eliminating humans" 2.
4. Perplexity researcher Adel Karimi observed a PHP2Shell exploit successfully used on a honeypot server, with the attacker executing shell commands to fetch additional payloads; no second-stage payload had been recovered 1. Aravind Srinivas said the exploit is confirmed in a real environment and that Perplexity Computer can run simulations before real-world testing 2 3.