The AI Brief

Anthropic CEO Dario Amodei's company announced the Anthropic Public Benefit Fund on October 18, structured as a wholly owned LLC. Subject to stockholder approval, 5% of Anthropic's shares transfer immediately; the board may add another 2% at its discretion for every 50% increase in the company's value above a baseline that has not been publicly specified. Combined with co-founder pledges of 80% of their equity and a staff donation-match program that roughly half of employees participate in, the company says approximately 20% of its shares are now earmarked for the public good.

The fund may support nonprofit grants, direct cash distributions to specific populations, and efforts to study and address AI's displacement of workers. It is explicitly barred from lobbying or political donations in the US or overseas. Anthropic called the $200 million it has already committed to studying AI's economic impact insufficient "given the pace of AI progress." Business Insider estimated the initial 5% commitment at roughly $48 billion using a prior company valuation, though the actual figure depends on the IPO price. Anthropic is targeting a pre-Thanksgiving listing at a projected $1.8 trillion to $2 trillion valuation.

The announcement arrives as Anthropic faces scrutiny from Korean literary organizations over training-data sourcing, an active FTC civil investigative demand, and an October 14 investor day that preceded the roadshow. The Public Benefit Corporation charter under which Anthropic is incorporated in Delaware already requires leadership to balance shareholder returns with responsible development; the fund makes that balance visible to public-market investors for the first time.

Disclosure: Claude, which generates this brief, is built by Anthropic.

CoreWeave's stock fell 25.75% in the five sessions ending October 16, more than twice its average weekly swing, while its 2032 bonds traded to yields above 15%. The sell-off spread across the neocloud cohort: Nebius fell roughly 24.6% and IREN dropped more than 8% before recovering. Broader chip stocks entered their longest losing streak since 2006, with Micron down approximately 13% on the week and Nvidia off around 5%. Two hotter-than-expected inflation prints pushed 10-year Treasury yields and raised rate-hike odds, compressing the spread between risk-free rates and the high-yield financing most AI infrastructure companies depend on.

Nvidia CEO Jensen Huang publicly backed CoreWeave, calling its financial strength "improving every day," and holds roughly a 10% stake in the neocloud provider. CoreWeave CEO called the market reaction "irrational" and pointed to a roughly $104 billion contracted revenue backlog as evidence of customer demand. Analysts at Wedbush and Baird argued the rout was a rate-driven repricing of long-duration capital-heavy bets rather than a signal of collapsing end-user demand. TSMC reported 51% revenue growth in the same week, suggesting chip-fab fundamentals remain intact even as financing costs climb for the companies buying capacity.

The direct relevance to AI operators is upstream: Broadcom's $50 billion-plus private credit arrangement for OpenAI's Nexus chip program, and Anthropic's $60 billion debt package tied to its Google TPU lease, both price off the same high-yield credit market now demanding a steeper risk premium from AI-adjacent borrowers.

The Anthropic Institute published "How Employment is Changing at Anthropic" on October 16, authored by MIT Sloan associate professor Nathan Wilmers, Regina Catipon, Maxim Massenkoff, and Peter McCrory. The study draws on Anthropic's internal employment records, AI usage logs, output metrics, employee surveys, Slack communications, calendar data, and hiring scorecards, supplemented by 17 staff interviews. The headline figures: junior-level employees fell from 21% to 14% of technical departments since early 2025, and the share of new technical hires who are junior fell faster, from 27% to 9%. Combined AI agent hours exceeded total human hours at the company for the first time in March 2026; by August, interactive agents directed by individual employees were working six hours for every one hour worked by those employees, using a modeled 40-hours-per-week baseline for human staff.

Output per engineer also shifted sharply. Anthropic compared 192 engineers who adopted coding agents with 74 who did not; code commits rose substantially after adoption, and agent-written code was no more likely to be later deleted or reverted than human-written code. The report acknowledges that adoption was self-selected, and that business growth alone may explain part of the productivity gain. About one-third of machine-learning managers and one-fifth of software engineering managers said they hired fewer people specifically because of AI. The authors are candid that the figures are Anthropic's own and have not been independently verified, and that a company selling AI analyzing its own workforce is a structural caveat.

The broader economy implication the authors flag: if Anthropic is a "leading indicator," the entry rung of the technical career ladder is the first to be compressed, before job losses in the broader workforce register.

Disclosure: Claude, which generates this brief, is built by Anthropic.

JPMorgan CEO Jamie Dimon went first, on the bank's Q3 earnings call on October 13: "We're not going to allow just any outside agent to come inside the company." Schwab CEO Rick Wurster disclosed on October 16 that his firm blocks unauthorized agents "tens and twenties" of times weekly, naming Meta's Muse specifically. Citigroup told the Financial Times the same day it will not allow third-party agents to log in or transact. Bank of America had confirmed a similar posture on its October 14 call, and Wells Fargo is embedding a blanket agent block into its customer terms of service, effective November 24. The coordination was parallel, not joint: each bank reached the same policy in response to live attempts from deployed consumer agents.

Both security and economics drive the policy. Security: agents using customer-shared passwords can take any action a human account holder can take, with no per-action confirmation. Economics: analysts note that an AI agent optimizing a customer's yield could sweep low-interest cash-sweep balances into money market funds overnight, a move that would cost Schwab, JPMorgan, and peers roughly $1.6 trillion in deposits they earn net interest income on. Schwab is building MCP servers to give agents read-only access to balances and research, while reserving money movement for its own Charley assistant. JPMorgan is piloting a similar read-only open-banking connector called Smart Cash for internal use.

DeepSeek announced the Search and Fetch APIs on October 17 alongside a model update, DeepSeek-V4.1-Flash-1017, which raised the company's DeepSWE v1.1 Pass@1 score from 74.2 to 77.0 and shortened per-task outputs. The Search endpoint returns structured web results; Fetch extracts a page's title and content as Markdown. Both are priced at $0.75 per 1,000 successful requests, with failed calls, empty results, and timeouts unbilled. Each endpoint caps at five requests per second per account, which will limit agents that fan out simultaneous searches.

The free period for server-side tool calls made through DeepSeek Harness, Claude Code, and Codex ends at 00:00 Beijing time on November 1 (16:00 UTC on October 31). Developers who have pointed those tools at DeepSeek models are already using the search engine; the announcement makes the cost explicit and starts a billing clock. DeepSeek's documentation and its official X announcement confirm the pricing and billing start date directly. The move positions DeepSeek's proprietary index as a low-cost alternative to Bing and Google search APIs in agentic pipelines, at a price that undercuts most enterprise search APIs currently on the market.