Anthropic establishes public benefit fund, initially committing 5% of its shares
要闻
Anthropic announced a public benefit fund that would receive 5% of the company’s shares, subject to stockholder approval, with a mechanism for additional equity as its valuation rises. The company says roughly 20% of its shares will be dedicated to public benefit when existing founder and employee commitments are included.
Anthropic announced the establishment of the Anthropic Public Benefit Fund in its announcement, with an initial allocation of 5% of the company’s shares pending stockholder approval and no specific date provided. Under the proposed structure, the board may allocate an additional 2% of shares for every 50% increase in the company’s valuation above a set baseline. The fund will be established as a wholly owned LLC rather than a 501(c)(3) or another organizational form. The company says this structure allows adjustments between forms of giving, and that Anthropic equity held by the fund will not be used for lobbying or political donations in the United States or elsewhere. Details of the valuation milestones and the fund’s governance are scheduled for release in the coming months.
According to the company, the fund could make large donations to traditional nonprofits or distribute funds directly to specific populations, with labor market impacts cited as one possible use. Anthropic has already committed $200 million to studying and addressing that issue. At the company’s founding, its co-founders pledged to donate 80% of their respective equity holdings. The company also operates an employee equity donation matching program, with approximately half of employees participating. Anthropic says that, together with the new arrangement, roughly 20% of its shares will be dedicated to public benefit, with potential increases upon reaching specified valuation milestones. The company also plans to publish a public benefit report detailing progress toward its mission.
The compute behind open-source model training: Wolf estimates cumulative costs at nearly $1 billion
技术与洞察
Hugging Face co-founder Thomas Wolf estimates that the compute used to train models made freely downloadable by about 35 organizations is worth nearly $1 billion at current GPU rental rates. The corresponding figure was below $40 million at the end of 2023.
Hugging Face co-founder Thomas Wolf published a rough estimate of the training compute behind open-source models on the Hub, covering 72 model families and yielding a central estimate of approximately $950 million at current GPU rental rates. The roughly 35 organizations involved include Meta, NVIDIA, Alibaba, DeepSeek, Moonshot and Mistral, and all of these models are free to download. About half of the estimate is based on compute data disclosed by the organizations themselves, with the remainder coming from informed estimates; the corresponding valuation at the end of 2023 was below $40 million. Wolf treats the result as a lower bound because it excludes about 900,000 community fine-tunes, as well as failed training runs, research and development team expenses, and datasets. He drew an analogy with Wikipedia, comparing volunteers contributing time to an encyclopedia with the use of GPU hours and data to build public resources for AI.
Multiple X users reported that 3 new model IDs briefly appeared in the OpenAI API, and TestingCatalog said the references were subsequently removed. Another user reported encountering one of the models in Arena, but OpenAI has not officially confirmed their identities or availability plans.
Three new model IDs briefly appeared in the OpenAI API, according to findings by multiple X users, with TestingCatalog reporting that the references were subsequently removed from the API. The IDs were aurora-red-alpha, aurora-green-alpha, and aurora-blue-alpha. OpenAI has not officially confirmed the models’ identities or availability plans.
User reports about the models have also surfaced in Arena. X user Lumina said they encountered aurora-red-alpha and claimed that all 3 models were being tested in Arena and appeared to be intended for image generation. Screenshots they shared showed aurora-red-alpha and aurora-green-alpha generating images from cat-related prompts. In a subsequent reply, however, Lumina clarified that they had still not encountered the blue version.
Anthropic reportedly in talks with law firms over rules for AI negotiations
前瞻与传闻
Anthropic is consulting law firms on fair rules for AI Agents negotiating on users’ behalf, according to the Financial Times. The company is also recruiting lawyers for four months of Claude testing, with compensation of $100,000 to $167,000.
Anthropic is seeking advice from law firms to develop fair negotiation norms for AI Agents acting on users’ behalf, according to the Financial Times, covering price negotiations, complaint resolution and improved contract offers. The company’s previous experiments showed that more capable models secured better deals, which could widen societal inequality. Related research also addresses how to evaluate models’ judgment: Anthropic is recruiting lawyers as Conceptual Reasoning Fellows to spend four months testing Claude on questions requiring judgment in law, philosophy, cognitive science, economics and political science, with compensation of $100,000 to $167,000. Unlike mathematics and coding, where verifiable data is plentiful or easy to generate, the humanities often lack clear standards for right and wrong answers, making comparable data difficult to provide. Anthropic therefore needs experts to build test cases and evaluations to improve Claude’s capabilities in these fields.
