Six AI Updates: Model Costs, Genie 4, and an IPO

Matt Wolfe surveys six AI developments and tries Google’s Genie 4 world model.

IPO estimates are not proceeds

Anthropic filed for a proposed IPO; the company says the offering price and share count are not yet determined.
Wolfe discusses a $1.5–2 trillion valuation and employee wealth estimates as projections, not completed sales; he has not checked the employee calculations.
His interpretation of reported losses includes financing-related accounting, which he explicitly says he struggles to explain.

Wolfe questions where GPT-6.1 Terra fits against Sol, comparing benchmark cost and performance at different effort settings.
He says Gemini 4 Carbon is available to selected testers, not broadly available, and cites Artificial Analysis results rather than his own test.
Moonshot says Kimi K3.1 Pro’s weights are planned for November 4; Wolfe notes its large total parameter count makes local deployment demanding.
His SVG test uses an LLM judge and one specific image task, not a general model ranking.

Wolfe prompts a neon city, a ninja, and Godzilla; the session freezes after a further interaction.
In a candy world, objects change position or appearance after the camera turns; a session lasts about two minutes in his test.
He suggests scene planning as a possible use, while noting the demo is not a coherent long-form game.

OpenAI says a model produced a refutation of the exponential time hypothesis, a conjecture about the difficulty of Boolean satisfiability; the company reports a Lean formalization.
OpenAI has released the artifacts for examination and assembled an independent expert group; practical algorithms and broader implications require further work.
Earlier results conditional on the hypothesis do not all become false merely because the condition is challenged.

“I cut out some of the rapid fire.” — Matt Wolfe, 28:08