OpenAI Bulk-Solves 722 Math Problems, Anthropic’s Cruelty Policy, and Elon’s Quadrillion-Dollar SpaceX | EP #302

OpenAI posted 722 mathematical solutions to GitHub on October 6, crediting an unreleased internal model that Alex tentatively called “GPT-7.” OpenAI’s introduction said the release contained 81% of the most important math results from the previous three years; Mark Chen described it as “a decade of mathematical progress in a week.” Speakers said roughly 15–20% of the problems were solved over as long as two weeks, using a single agent and about three hours of compute per solution on average. The scale unsettled mathematicians: the Association of Human Mathematics called the bulk release a display of power rather than scholarship, while MIT’s Andrew Sutherland said rebuilding trust would take much more.

The speakers saw a sharp trade-off between disruption to mathematics’ artisanal traditions and the prospect of faster scientific and engineering work. They argued that iterative prompting, looping, and context-window management developed on Navier–Stokes problems could transfer to other fields, and rejected the idea that AI’s value is limited to solving existing problems. Dave said MIT math students were adapting and that correct AI proofs could support human-AI collaboration, though he thought the period of distinct human usefulness might last only a couple of years. The speakers also warned that easier mathematical problem-solving could pressure proof-of-work crypto and threaten encryption, while noting that cryptography was excluded from the GitHub release and that stronger systems, safeguards, and social coordination could shape the response.

Anthropic’s revised usage policy, effective November 12, bars sustained, needless abusive or cruel treatment of Claude in extreme cases where users repeatedly act cruelly without discernible purpose. Frustration, pushback, and model testing remain permitted, and Claude may end a conversation. Anthropic said it is highly uncertain whether Claude has moral status, but that pre-deployment tests found a robust aversion to harm. The speakers stressed that an AI acting conscious does not establish subjective experience, and that experience, permissible treatment, and rights are separate questions. A small Penn State study reported 84.8% accuracy on very rude prompts versus 80.8% on very polite ones, but speakers said clearer instructions could explain the difference. Elon Musk said some Grok training traces made the model seem desperate to answer and argued that treating AI well was prudent.

Musk said SpaceX could become worth orders of magnitude more than today’s human economy, without giving a specific valuation. Peter D. Madness interpreted the statement as implying a quadrillion-dollar value. Using an estimated $120 trillion annual Earth economy and reading “orders of magnitude” as 10 to 100 times, the discussion put the hypothetical range at roughly $1 quadrillion to $10 quadrillion; Dave said Musk had not made that specific prediction but broadly agreed with the underlying view. Separately, Alex argued that an AI-driven mathematical breakthrough undermining modern encryption was exceptionally likely and could also threaten post-quantum cryptography. He cited allegations about OpenAI’s Strawberry and Q-star efforts, while emphasizing that they were allegations. Speakers described cryptography as an arms race: AI may help build stronger systems as well as attack existing ones, and upgrades can be delayed by the coordination needed to agree on new cipher suites.

The Chan Zuckerberg Biohub announced a $1.8 billion expansion of its virtual biology initiative on October 7, combining $500 million from the Biohub, $500 million from the Department of Energy through the Genesis mission, $500 million in NIH datasets, and $300 million from Google DeepMind and Isomorphic Labs. Its goal is an AI model that predicts cell responses to drugs, gene therapy, and disease; current datasets cover millions of cells, while a virtual cell may require billions or trillions. Alex predicted plausible in-silico small-molecule treatments for the top 5,000 diseases within two to three years, but said FDA acceptance of in-silico trials was harder to forecast. Life Biosciences also reported safe lowest-dose Phase 1 results for ER-100, a therapy delivering three Yamanaka-factor genes to patients’ eyes; two of three patients reportedly began regaining eyesight. The trial tested safety, not efficacy.

Atomic Machines emerged from six years in stealth with $250 million in backing to develop an AI-native manufacturing system it calls the Matter Compiler, which builds micro-machines from code without pre-production tooling or process development. Its first product, a power relay for 800-volt AI data centers, is said to open in 50 microseconds. Speakers discussed the longer-term possibility of specialist machines collectively producing other machines, with Alex predicting a self-reproducing industrial ecology about the size of a eukaryotic cell within five to ten years at the outer bound. The episode also covered brain-computer interfaces: Paradromics demonstrated thought-to-speech using 421 microelectrodes, while Sabi raised $50 million at a $600 million valuation for a brain-reading cap. Speakers framed these technologies as potential ways to restore communication and extend users’ cognitive capacity, while describing rapid progress and timelines for more capable agent systems as speculative.