Anthropic Establishes Public Benefit Fund and Releases Employment Research

Anthropic advances architecture and employment impact research, Grok Bot integrates into daily tasks, and open-source models alongside major enterprise foundations achieve new evaluation performance.

The essentials

Anthropic Establishes Public Benefit Fund

Anthropic announced the establishment of a wholly owned subsidiary, Anthropic Public Benefit Fund, with an initial injection of 5% of its shares. Subsequent valuations may add up to 2% for every 50% increase above the baseline, up to a total of approximately 20%.

Anthropic Releases Employment Structure Research

Anthropic released a series of economic impact studies discussing the actual impact of Claude on jobs. Internal data shows that the proportion of entry-level positions in its technical department has dropped from 21% in the early stages to 14%.

Grok Bot Accelerates and Integrates into Daily Tasks

Grok Bot's computer-use task speed increased by about 3 times, and response time was shortened by more than half. Demonstrations show it can place orders via text messages, book restaurants by phone, and connect to bank accounts to audit bills and subscriptions.

In brief

Claude Decrypts Elliptic Curve Seed

Anthropic research shows that Claude took only about 30 minutes to process a 25-year-old puzzle regarding the source of the NSA-generated ECDSA elliptic curve seed.

OpenCode Suffers API Key Distillation Attack

OpenCode's Anthropic and OpenAI API keys were attacked, draining balances, while defensive models triggered safety guards to refuse payload analysis.

DeepSeek V4.1 Flash Evaluation Performance

DeepSeek V4.1 Flash performed excellently in benchmark tests, topping the Code Arena open-source model leaderboard, while significantly reducing the cost of 3D architectural reconstruction.

MAI Develops Self-Researched 3T Parameter Pre-training Foundation

Mustafa Suleyman stated that MAI has fully self-researched a 3T parameter-level pre-training foundation using a native encoder-free VLM, with post-training results to be announced soon.

German Court Rules Google Must Be Held Liable

The Munich Regional Court ruled that Google's AI search summaries constitute its own statements rather than neutral paraphrases, making Google liable for incorrect content, ultimately settling the case.

Source ↗

Browser Use Explains Agent Principles

Browser Use published an article explaining that browser agents consist of three parts: the model, the harness, and the browser, completing tasks through a cycle of observation, action, and checking.