Subscribe
Share
Share
Embed
Anthropic's public benefit fund, minus numbers
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Monday, October nineteenth.
In today's briefing we have Anthropic's five percent share commitment to a public benefit fund as IPO reports circulate, Disney hiring most of Character.AI's workforce, and a reported shift in how Terence Tao works with AI on mathematics.
First up - Today in the big model news;
Anthropic
Anthropic is moving toward public markets, and its public-good commitment now has a number attached. The company will commit five percent of its shares to the Anthropic Public Benefit Fund, subject to stockholder approval. The board may add two percent for every fifty percent rise in company value above a baseline, and that baseline is not specified. The fund is a wholly owned LLC, not a 501(c)(3), and none of its equity can go to lobbying or political giving. Anthropic says roughly twenty percent of shares are dedicated to the public good once founder pledges and the employee match are counted. But it gives no share count and no dollar value. The commitment lands next to unverified reports of an IPO before Thanksgiving. That makes the missing baseline the first number prospective investors and vendor-risk reviewers will look for. Anthropic has promised structure and valuation milestones in the coming months, and those will show whether this is a binding mechanism or a positioning move.
In the harness, tools and orchestration world;
A new paper with Together AI code proposes HybridLink, which switches between latent reasoning and text tool calls inside a single generation. It is trained with two methods called HybridSFT and HybridRL. Across three model scales and eleven benchmarks, the authors claim ten point four percent better performance, thirty-five to fifty-five percent fewer tokens, up to two point six times faster inference, and up to three point seven times lower cost. All of those figures are the authors' own, and they have not been independently reproduced. The authors say the latent thoughts are meant to stay inspectable, which matters to anyone who relies on reading reasoning traces. Agent inference spend is the line item this targets, so it earns a prototype once someone outside the authors reproduces the savings.
In other news…
Disney has agreed to offer jobs to at least sixty Character.AI employees, a majority of the workforce, at roughly sixty million dollars in estimated annual compensation. The filing says the vast majority accepted. It is a stalking-horse bid, so a higher competing offer is still possible. The filing cites high costs, potential litigation-related liabilities, and regulatory engagement as reasons for the sale. That comes four days after the Chapter 11 filing and the twenty million dollar asset purchase. Child-safety suits, including one over a fourteen-year-old's death and a Texas attorney general investigation, are now setting the terms of the exit. Companion and character products should price litigation and regulatory exposure into the cap table from the start.
Separately, the New York Times reportedly describes Terence Tao as barely doing his own research now, instead learning to work with AI. He is quoted saying no expert he has talked to has understood the whole of a reported hundred-and-sixty-five-page Navier-Stokes argument, said to have come from about ten thousand agents over eighty-eight hours. Reported responses include a fifteen million dollar Simons Foundation effort for mathematicians and a Harvard draft that would rest PhDs mainly on in-person defense. In mathematics, where proof is the gold standard, generation has outrun checking. Budget for the cost of independent verification, not the cost of output.
That's the briefing. Have a great day, and don't forget to subscribe.