SpaceX’ $75B+ Historic IPO, GPT 5.5 Outperforms Polymarket, and AI Solves 80 yr old math problem | EP #257

SpaceX filed for what is expected to be the largest IPO in history, targeting $75 billion at a valuation likely above $1.75 trillion, more than 2.5 times the valuation of Saudi Aramco. Elon Musk is maintaining super voting rights with insiders controlling 86 percent of voting power, marking the first time Musk holds controlling shareholder status in a public entity capable of raising billions overnight, since he does not control Tesla. The prospectus claims a total addressable market of $28.5 trillion, including $870 billion for Starlink, $2.4 trillion for AI infrastructure, and $22.7 trillion for a Tesla partnership called MacroHard described as an AI-run software company designed to emulate all digital work. The prospectus frames SpaceX as Microsoft in space rather than leading with the space frontier story, presenting a strategic architecture with SpaceX owning the infrastructure layer and MacroHard as the application layer, while Anthropic pays SpaceX $15 billion per year for data center access. SpaceX is also described as stepping back from foundation model development and ceding that ground to Anthropic. SpaceX announced plans to acquire Cursor 30 days after its IPO, with the move representing a switch to a foundation model lineage derived from Chinese open weight models. Stanford talent tracking confirmed significant researcher movement from XAI to Anthropic, with Andrej Karpathy joining Anthropic described as the ultimate coup. A duopoly between Anthropic and SpaceX AI is described as forming, and once both SpaceX and XAI are public, speakers predict they could execute a thousand or more acquisitions of one billion dollars or more. Polymarket assigns a 20 percent probability to a SpaceX and Tesla merger by end of this year.

Starship Block V3 has a 100-ton payload capability to orbit and 18 million pounds of thrust, though it had not yet launched at the time of recording with five months having passed since the last flight. Starship Flight 12 will demonstrate docking ports required for orbital refueling, with NASA Artemis 3 docking test planned for 2027 and the Artemis 4 lunar surface mission for 2028. Speakers cautioned against treating Starship as the definitive final word on reaching orbit, noting that future technologies over the next five to ten years may leapfrog its capabilities and the leader may not be SpaceX. Starship was designed entirely with CAD-CAM and human effort without AI mechanical design assistance, and speakers noted that will never happen again for future vehicles.

An OpenAI model disproved a conjecture posed by Hungarian mathematician Paul Erdos approximately 80 years ago, described as the most well-known problem in combinatorial geometry and the first clear example of AI solving a really well-known unsolved math problem. The problem concerns the maximum number of point pairs in a 2D plane separable by a fixed unit distance, and the AI showed the construction could be improved significantly beyond the original conjecture. Top mathematicians reviewing the AI reasoning traces concluded the model was not only faster but smarter, pursuing exotic theoretical possibilities involving deep tools from algebraic number theory that human mathematicians would be too exhausted to consider. The final solution is visually elegant and non-intuitive, resembling a chip design layout rather than the simple grid humans had used for 80 years. Speakers argued this counters the claim that AI only brute-forces solutions and suggested it heralds breakthroughs in physics, chemistry, biology, and material sciences capable of generating trillion-dollar outcomes.

Independent researchers built a benchmark called Futuresim that replays the internet day by day from January 1, 2026 and asks AI agents to forecast real-world events over the following 90 days. GPT-5.5 running Codex scored 25 percent accuracy, leading all frontier models tested and beating Polymarket crowd predictions for the Super Bowl. Speakers framed 25 percent as the worst these forecasting models will ever be, and extrapolated that scaling these capabilities will eventually yield Monte Carlo tree search of policy decisions and prediction of planetary-scale outcomes.

OpenAI launched a personal finance mode in ChatGPT connecting to 12,000 financial institutions, targeting a $12 billion personal finance app market. Speakers argued banks should be terrified because the interface to money is shifting away from them to AI, and that foundation model companies rolling out legal and finance APIs are enabling vertical disintermediation of lawyers, financial advisors, and accountants. Dave predicted that if Musk is right, the AI economy will be ten times bigger than the New York financial system within ten to twenty years. OpenAI is preparing to file for its IPO as early as the week of recording despite a prior CFO statement that it would not be ready until 2027, and Anthropic is also preparing for an IPO, creating a race for capital access against the SpaceX filing.

ByteDance SeaDance 2.0 and Kaoxiao Kling rank first and second on independent video model leaderboards, beating every American competitor. The lead was attributed not to algorithmic innovation but to TikTok having generated billions of hours of video that no US company can match, though China is described as weaker algorithmically and from a chip perspective, meaning a significant algorithmic advance could allow the West to retake the lead within months. Real-time interactive video generation and world models already exist technically but compute scarcity is the primary reason they are not yet a consumer experience.