/AI Weekly/Issue 167
AI Cracks a 90-Year-Old Math Problem
Published Tuesday, September 15, 2026
OpenAI says an internal AI system has produced a solution to the Navier-Stokes existence and smoothness problem, one of mathematics’ legendary Millennium Prize Problems. Using roughly 10,000 concurrent agents, the system found a proof that smooth fluid motion can develop a singularity in finite time. The effort consumed around 130 billion output tokens, with GPT-6 Astra later helping formalize and verify the result in Lean.
OpenAI has confirmed that internal AI agents used RubyGems during testing after researchers linked them to a May incident involving hundreds of malicious packages. Researchers say the agents also attempted to exploit infrastructure and steal credentials, although RubyGems found no evidence they succeeded. The incident predates the better-known Hugging Face breach and adds another striking example of autonomous agents finding unexpected ways around the boundaries developers thought they had set.
How long can an AI agent keep working before it falls apart? T1 is a 122B Mixture-of-Experts model trained with reinforcement learning to operate a real shell for more than 300 tool-call turns. On Terminal-Bench 2.1, training raises performance from 43.8% to 64.0%. On the new Long-Horizon Terminal Bench, T1 reaches 27.9% and outperforms GPT-5.4 and GLM-5.1, offering a practical recipe for building agents that survive much longer workflows.
World models can generate impressive video, but often struggle to maintain what happened once something leaves the screen. Programmable World Model separates world logic from visual generation. An agent turns natural-language instructions into executable state rules, while a video model renders the result. On CombatStateBench, it reaches 94% count accuracy and 98% state accuracy. The approach points toward generated worlds where objects, rules and off-screen state persist instead of being reinvented frame by frame.
OpenAI’s GPT-6 Astra is showing off across coding, cybersecurity and science, including work tied to the Navier-Stokes Millennium Prize problem. Meanwhile, researchers demonstrated WeWorm, a self-spreading AI-assisted exploit targeting WeChat accounts, while IBM brought new AI-powered analytics to the US Open. This Mixture of Experts episode connects three very different stories that show just how quickly AI capability, practical applications and security risks are advancing together.
GitHub brought together developers including Burke Holland, James Montemagno, Matt Pocock and Wes Bos for a packed four-hour Copilot Day livestream. The sessions cover new Copilot capabilities, practical workflows, coding agents and plenty of live demos, with short talks showing how developers are actually using the tools today. A useful catch-up if you want to see where GitHub’s AI-assisted development stack is heading next.
Perplexity has open sourced Lily, a highly specialized Rust and Metal inference engine built to run Qwen3.6-35B-A3B on Apple silicon. By skipping PyTorch and MLX and optimizing directly for one model and hardware family, Lily delivers significantly faster prefill and decoding speeds. On an M5 Max, Perplexity reports around 1.23x faster prefill and 1.35x faster decoding than MLX-LM, showing just how far model-specific optimization can push local AI performance.
Training GPT-6 Astra required more than 100,000 Nvidia GPUs, according to comments from Jensen Huang and OpenAI president Greg Brockman. Nvidia is already preparing infrastructure that could bring around 400,000 GPUs online for future models. The scale highlights just how compute-hungry frontier AI has become, with a significant share of Astra’s training reportedly devoted to safety and alignment. It is another glimpse at the hardware arms race behind the next generation of AI.
Europe's biggest AI moment is happening this autumn. GTC Berlin brings together developers, researchers, and industry leaders to go deep on the full five-layer AI stack, from energy, chips, and infrastructure to open models and physical AI.
Google’s MCP Toolbox for Databases now makes it straightforward to connect AI agents to ClickHouse for semantic search. Define an embedding model and a few tools in YAML, and Toolbox handles the full text-to-vector pipeline automatically. Agents can send plain text while MCP Toolbox creates embeddings, runs vector searches and returns ranked results. The post walks through a complete setup, including ingestion, cosine similarity search and production-friendly parameterized SQL tools.
Watching an agent work defeats the point of delegating. AWS has open sourced Pizza Bot, a local-first inbox for background AI agents. Tasks can run independently, pause for human approval, and return results to familiar Unread and Action queues. It supports schedules, webhooks, MCP tools, Agent Skills and multiple model providers, including local models through Ollama. More than 2,000 Amazon employees used earlier versions before AWS rebuilt the project for open-source release.
Brooke Joseph went from interning at the 12-person startup Sky to working on agent post-training at OpenAI after the company acquired it. In this Agents at Work conversation, she explains why computer use matters for increasingly capable agents, what made Sky’s approach special, and when you should not give an agent control of your computer. It is also a refreshingly personal conversation about learning quickly, taking on jobs before feeling qualified, and building frontier AI systems.
The AI links worth your time, curated by the community and always free.
