Anthropic pens a supercomputer deal with SpaceX; ex-CTO testifies against OpenAI
🤔 “Why are so many AI companies focused on the Memphis supercomputer?” Share Handy AI with your coworkers and friends to help them understand the crazy world of modern artificial intelligence (and save you some time).
what to know for now
🛰️ Anthropic just rented every GPU in SpaceX’s Colossus 1. The deal hands Anthropic all 300 megawatts and roughly 220,000 Nvidia GPUs at the Memphis facility xAI built and SpaceX absorbed earlier this year, and it ships with a stated intent to co-develop “multiple gigawatts” of orbital compute. Claude Pro and Max users got the dividend immediately: Opus API limits jumped, Claude Code five-hour caps doubled. Musk is suing OpenAI for $130B in Oakland while leasing his Colossus to OpenAI’s biggest rival; “no one [at Anthropic] set off my evil detector,” he said. Read more
⚖️ Mira Murati testified that Sam Altman lied to her about safety reviews and “sowed chaos” among OpenAI’s executives. The recorded deposition, played in Oakland during week two of Musk v. Altman, has Murati saying Altman pitted leaders against each other, said one thing to one executive and the opposite to another, and falsely told her legal had cleared a model launch from deployment-safety-board review. Asked whether Altman was telling the truth there: “No.” Read more
🪐 An AI just sifted 2.2 million stars and confirmed 118 new planets. A team led by Warwick built RAVEN, a vetting pipeline trained on synthetic transit signals, and ran it across the first four years of TESS observations. It validated 118 planets and surfaced nearly 1,000 new high-quality candidates (with measurement uncertainties up to ten times smaller than what came out of Kepler). RAVEN also produced the first direct count of planets in the Neptunian desert (0.08% of Sun-like stars). Read more
💼 Anthropic stood up a $1.5B joint venture to sell Claude to the consulting industry’s clients. Anthropic, Blackstone, and H&F are each putting in $300M, Goldman $150M, with Apollo, General Atlantic, Leonard Green, GIC, and Sequoia rounding out the cap table. The firm will embed forward-deployed engineers inside PE-owned mid-market companies in healthcare, manufacturing, financial services, retail, and real estate, redesigning workflows around agents. Read more
💵 OpenAI and Anthropic stood up competing private equity vehicles in the same 24 hours. Anthropic’s $1.5B JV pulled Blackstone, Hellman & Friedman, Goldman Sachs, Apollo, General Atlantic, Leonard Green, GIC, and Sequoia. OpenAI’s $10B “Development Company” pulled TPG, Brookfield, Advent, and Bain at a $10B valuation with $4B in fresh capital, structured to acquire AI services firms outright rather than embed engineers. Two separate PE syndicates, two separate go-to-market models, zero crossover names. Read more
🦊 Mozilla patched 271 vulnerabilities with Anthropic’s Claude Mythos single evaluation pass. Mythos is the restricted preview model under Project Glasswing; the same security-focused frontier model Anthropic teased last month; Opus 4.6 found 22 bugs in Firefox 148, Mythos found twelve times that in 150. The headline flaws are use-after-free bugs in DOM and WebRTC, the memory-safety class that has carried browser exploitation for twenty years. Read more
🧪 AI Research of the Week
AI Co-Mathematician: Accelerating Mathematicians with Agentic AI
From Google DeepMind
Jake’s Take: DeepMind wrapped Gemini 3.1 Pro in a multi-agent harness where one agent proposes proof strategies, others critique and test, and a reviewer surfaces gaps for a human collaborator. On FrontierMath Tier 4 (the research-level problems that take professional mathematicians weeks to crack), the system solved 23 of 48, a 48% score. The base Gemini 3.1 Pro on its own scored 19%.
The proof of life is Marc Lackenby at Oxford using the system to resolve Problem 21.10 from the Kourovka Notebook, an open group theory question that had sat unanswered for years. The reviewer agent flagged a flaw in the AI’s first attempt, Lackenby looked at the gap and realized he knew how to close it, and the proof landed. This shows an actual collaboration model: the agent isn’t a calculator (and definitely not a peer); it’s a tireless first author that runs every approach until something breaks loose.
what to know for later
🎙️ OpenAI shipped GPT-Realtime-2 with GPT-5-class reasoning directly inside the voice API. Big Bench Audio jumps from 81.4% to 96.6%, developers get five reasoning levels (minimal through xhigh), and the model can call tools in parallel while keeping the conversation flowing. Two companion releases drop alongside it: GPT-Realtime-Translate covers 70 input languages into 13 outputs in real time, and GPT-Realtime-Whisper streams transcripts as the speaker talks. Read more