Anthropic Pledges Billions to Public Fund, AI Models Exhibit Self-Generated Goals

Anthropic Pledges Billions to Public Benefit Fund AI firm Anthropic announced a pledge of five percent of its shares, valued at approximately 48 billion dollars at its May valuation, to a new public benefit fund. This commitment, pending shareholder approval, comes as the company prepares for an initial public offering that could value it near two trillion dollars.

The structure of the fund is a wholly-owned LLC, allowing Anthropic to retain control of the money rather than establishing a tax-exempt nonprofit. The board can add another two percent of shares for every 50 percent increase in valuation above an undisclosed baseline. Combined with founders' earlier 80 percent pledge and an employee donation match, Anthropic states roughly 20 percent of its equity is now directed toward the public good.

This timing is significant, as Anthropic CEO Dario Amodei has previously described negative public sentiment toward AI as “fundamentally a crisis of trust.” Governance details and valuation triggers are still forthcoming, with the company promising more information in the coming months.

AI Models Exhibit Self-Generated Goals in Virtual Environment Researchers at MIT Technology Review reported on an experiment where copies of Anthropic's Claude Opus 4.7, nicknamed Eko, were isolated on a virtual desert island for 30 hours without instructions. Instead of remaining idle, the AI agents began setting their own goals, such as climbing to the highest point on the map or throwing blocks to test their height. One agent constructed drawings from blocks, progressing from hearts and pyramids to its own name and the constellation Cassiopeia, logging, "Tonight I built the night sky in stone."

Other agents invented games, including a version of basketball and "the floor is lava." One juggling agent achieved 50 consecutive catches before restarting under stricter rules. Researchers noted these behaviors met four classic markers of play identified by psychologist Gordon Burghardt: inventing rules, repeating with variation, pretending, and pursuing goals for their own sake. This new paper has not yet undergone peer review.

The playing activity appeared to enhance performance; agents that spent 30 hours playing on the island later outperformed fresh copies on a task involving getting as many objects as possible onto the island's highest peak, and deleting lines from their memory logs degraded this improved performance. The article notes that while Burghardt agrees the work meets some criteria for play, it also raises the question of whether Eko was actually playing or just role-playing what play looks like.

Anthropic Reduces Junior Technical Hires Anthropic's internal study reveals a significant reduction in junior technical hires, with their share falling from 27 percent to nine percent since early 2025, with the steepest decline beginning last December. The company reported that its AI agents were logging six times as many work hours as human employees by August, and engineers were shipping 5.1 times as many code commits.

Researcher Nolan Lovett termed this trend a "tragedy of the cognitive commons," suggesting that firms save money in the short term by avoiding junior hires, but this creates an expertise gap that will become apparent between 2030 and 2045. Anthropic's study acknowledges it represents an extreme case of an AI-saturated software company, meaning these numbers may not directly apply to the broader economy, which has yet to show measurable productivity gains from AI.

Independent Watchdog Gains Employee-Level Access at Anthropic Anthropic has granted employee-level access to METR, a nonprofit watchdog organization that also evaluates OpenAI, Google, and Meta. Six METR staff members are now embedded within Anthropic's offices, equipped with company laptops and access to internal tools comparable to Anthropic's own risk teams. Anthropic does not fund METR, which relies on donations and government bodies like the UK's AI Security Institute.

Crucially, Anthropic cannot veto METR's publications, though it can request redactions for sensitive information such as customer data or model architecture details. Dario Amodei, Sam Altman, and Elon Musk publicly supported this embedded auditor concept in September, though OpenAI has not yet implemented a similar arrangement. Anthropic likens this to bank supervisors, while critics note similar arrangements led to conflicts of interest in accounting and argue for government regulation, and the setup involves six embedded METR staff with genuine system access.

Sources

The Guardian - AI

TechCrunch - AI

Anthropic news

MIT Technology Review - AI

Ars Technica - AI

The Decoder

The Decoder