OpenAI’s Astra Just Solved 10 Math Problems No Human Could Crack

🤖 AI Agents

OpenAI’s Astra Just Solved 10 Open Math Problems. The Real Story Is How.

What happened: OpenAI announced August 1 that an internal version of Astra — its next-generation model — autonomously solved 10 open problems in mathematics, spanning sphere packing, group theory, complexity theory, and extremal combinatorics. The total inference cost? Roughly $2,000. The model didn’t just compute answers — it recognized that a decades-old conjecture could be reframed using machinery from an entirely different mathematical domain, then produced formal Lean-certified proofs published on GitHub.

Why it matters: This isn’t a “look how smart our AI is” flex. It’s a signal that AI is crossing from assisting human researchers to autonomously advancing human knowledge. The model disproved Connes’s rigidity conjecture, resolved Ehrhart’s volume conjecture in all dimensions, and closed two Erdős problems (183, 146, and 180). Sam Altman is currently demoing Astra to senators in Washington — the timing is no coincidence. When an AI can do original mathematics at this level for the price of a laptop, the conversation shifts from “should we regulate AI” to “can we afford not to understand what it’s doing?”

What’s next: Astra remains unreleased. It may ship as GPT-6, a GPT-5.x update, or its own tier alongside Sol, Terra, and Luna. Meanwhile, Google DeepMind’s AlphaProof Nexus solved 9 open Erdős problems in May, Anthropic’s Fable disproved the 85-year-old Jacobian conjecture, and Harmonic’s Aristotle claimed a 30-year-old Erdős problem. The AI math race isn’t coming — it’s already here. OpenAI’s attribution approach (crediting mathematical arguments to Astra itself, citing the Leiden Declaration) also previews how AI contributions to science will be credited going forward.

📎 Source: OfficeChai | OpenAI | Ars Technica

⚡ Google DeepMind Drops Gemini Robotics 2 — Full-Body Humanoid Control

Google DeepMind launched Gemini Robotics 2 on July 31 — an AI model that gives humanoid robots full-body reasoning, letting them walk, crouch, balance, and use both five-fingered hands simultaneously. A companion model, Gemini Robotics ER 2, acts as a “project manager” that breaks complex jobs into steps and coordinates multiple robots. The new ASIMOV-Agentic safety benchmark also tests whether robots can refuse unsafe instructions. Why it matters: The race to put AI brains in robot bodies is accelerating — and Google just raised the bar from “can it pick that up” to “can it reason about what to do next.”

📎 Source: WION

⚡ Claude Code Creator: Delete Your CLAUDE.md File for Better Results

Boris Cherny, creator of Claude Code, dropped counterintuitive advice this week: delete your CLAUDE.md system prompts and start fresh. Over-specification hobbles the model — Claude performs better with clear objectives, guardrails, and exit criteria rather than exhaustive instruction files. Cherny also revealed Anthropic uses a three-layer prompt injection defense and an “ablation” process that systematically removes and reintroduces components to optimize performance. Why it matters: As AI coding agents become standard developer tools, prompt engineering is evolving from “write more instructions” to “write better constraints.”

📎 Source: Geeky Gadgets

📰 AI News

AMD Bets $5B on Anthropic as Inference Overtakes Training

AMD CEO Lisa Su announced a $5 billion equity investment in Anthropic and a deal to deploy up to 2 gigawatts of MI450 GPUs — completing a trifecta of partnerships with OpenAI, Meta, and now Anthropic. But the bigger signal: Su declared 2026 the first year global inference compute surpasses training compute. That’s the structural shift that matters — every AI agent deployment, every chatbot query, every automated workflow adds to inference demand. AMD has now committed ~20% of its equity across these AI bets. The first Anthropic gigawatt goes live H1 2027.

📎 Source: CryptoBriefing

The AI Trade War Just Expanded to Humanoid Robots

The FCC moved this week to block new Chinese-made humanoid robots, four-legged robots, and connected power inverters from the US market, adding them to its “Covered List” over surveillance and cyberattack risks. This follows the June introduction of the bipartisan GUARD Act. The context: China operates ~2 million industrial robots (4.5x Japan) and accounted for 54% of global installations in 2024. Meanwhile, Chinese humanoid robot patents surged fivefold. A fragmented global robotics market — US/allied vs. China/developing world — is now the baseline scenario.

📎 Source: Forbes

Mira Murati’s Team Said No to Meta’s Billion-Dollar Offers

Mira Murati revealed that not a single member of her 50-person Thinking Machines Lab accepted Meta’s recruitment offers — despite packages reportedly ranging from $200 million to $1.5 billion per person. This follows Murati’s rejection of Zuckerberg’s ~$1 billion acquisition bid in summer 2025. Meta then launched what’s being called “one of Silicon Valley’s most expensive recruitment failures.” The takeaway: in the AI talent war, mission alignment is proving to be a moat that even a billion dollars can’t cross.

📎 Source: Times of India | NDTV