Scale, autonomy, and AI where people already chat

🧩 The Gist

OpenAI and SoftBank Group partnered with SB Energy to build multi‑gigawatt AI data center campuses, including a 1.2 GW site in Texas for the Stargate initiative, a clear signal that compute and power are being scaled in tandem. Terence Tao highlighted that AI tools “more or less autonomously” solved Erdős problem #728 after some feedback, suggesting measurable progress in machine‑assisted reasoning and proof search. New applied tools landed across governance and UX, from an offline‑first EU AI Act compliance project to a single‑image 3D scene demo and iMessage‑native agents. Developers also saw pragmatic upgrades, including Datadog using Codex for system‑level code review and a client‑side GitHub recommender.

🚀 Key Highlights

• OpenAI and SoftBank Group will develop multi‑gigawatt AI data center campuses with SB Energy, including a 1.2 GW Texas facility supporting Stargate.
• Terence Tao reported an Erdős problem (#728) was solved “more or less autonomously” by AI after some feedback, with no known prior identical result in the literature, although similar results by similar methods exist.
• EuConform launched as an open‑source, offline‑first EU AI Act compliance tool, covering risk classification (Articles 5–15), bias checks using CrowS‑Pairs, Annex IV‑oriented PDF reports, and local operation via browser and Ollama.
• A demo turns a single image into a navigable 3D Gaussian splat with depth, built on Apple’s SHARP research model for non‑commercial use.
• Flux lets users spin up iMessage‑native AI agents in about two minutes, no app download required for people who text the agent, reflecting a messaging‑first interaction thesis.
• Datadog uses OpenAI’s Codex for system‑level code review, pointing to continued adoption of AI in software delivery pipelines.
• A client‑side GitHub recommender computes cosine similarity from your starred repos (for example against Karpathy’s), builds an embedding, and suggests repositories, plus a Skill Radar.

🎯 Strategic Takeaways

• Infrastructure and energy: Large, power‑dense campuses tied to specific AI initiatives show compute and energy planning moving together, which is essential for scaling training and inference.
• Governance built in: Local, auditable compliance checks (risk classification, bias evaluation, Annex IV reports) can bring regulatory requirements closer to day‑to‑day engineering.
• Agentic UX: Placing agents inside default communication channels like iMessage reduces friction, which can increase real‑world usage compared to standalone apps.
• Research automation: AI solving an Erdős problem with limited human feedback hints at growing capability in mathematical reasoning, while human oversight and claims of autonomy remain active discussion points.
• Dev productivity: From code review with Codex to local similarity recommenders, lightweight AI add‑ons keep improving developer workflows without heavy infrastructure.

🧠 Worth Reading

• Terence Tao’s note on Erdős problem #728: he describes an instance where AI tools produced a solution “more or less autonomously” after feedback, with no known identical prior result, though related results via similar methods were found. The practical takeaway is that orchestration of AI tools, plus targeted human feedback, can now yield novel results in rigorous domains like mathematics.