Anthropic Commits 5% of Its Shares to a Public Benefit Fund That Grows 2% for Every 50% Rise in Value -- a Wholly Owned LLC Barred From Lobbying, Pending Stockholder Approval
The October 18 announcement lands as Anthropic prepares an IPO, three days after it gave six METR staff employee-like access to its systems, and two days after a draft federal bill it wrote surfaced on Capitol Hill. The fund is an LLC rather than a 501(c)(3), so it can make direct payments to specific populations as well as grants. The valuation baseline and governance are still to come.
Microsoft Walks Away From Nscale's Flagship Essex Site After the Grid Said Not Before the 2030s -- Weeks After an S-1 That Puts 85% of a $103 Billion Backlog on Microsoft and Anthropic
The Telegraph reported the exit on October 18 and Reuters confirmed the outline. The Loughton site was to hold 23,040 GB300 GPUs for Azure from early 2027 and was billed by the UK government as its largest sovereign AI data centre. For anyone buying compute, the lesson is that a contracted backlog is not capacity, and power, not money, is the gate.
NVIDIA and IIT Madras's AI4Bharat Announce a 2.8-Trillion-Token Corpus for All 22 Scheduled Indian Languages -- Machine-Translated From English, Graded Against a 75-Point Bar, and Already Inside Nemotron 3.5 Super VL
Nemotron-pretraining-Indic-v1, announced October 16 and 'releasing soon', is 2,601 billion native-script and 224.6 billion romanized tokens made by translating 15 Nemotron-CC-v2 subsets with Gemma-4-31B-it and Bodhan AI's Indic Translate. Ten billion of its tokens went into the October 15 Nemotron 3.5 Super VL, a 120B open-weight model licensed for commercial use.
Research Radar: HybridAgents Think in Latent Space and Act in Text for 35-55% Fewer Tokens; CarryOn Replaces Agent History With Reusable Objects; and 114 of 135 Agent-Written Science Papers Fail an Integrity Check
Three October 13 preprints climbed Hacker News and Hugging Face over the weekend: a hybrid latent-and-text agent evaluated on 11 benchmarks, an object-oriented context runtime that cuts OSWorld 2.0 token use by 51.9%, and a Department of Energy lab benchmark on which no frontier agent produced an acceptable follow-up paper. A fourth, BizBench, shows outcome-only scoring hiding invalid execution.
DeepMind's Own Audit Finds 'No Significant End-to-End Acceleration' of AI Research -- but a Quarter of the Tasks in the Benchmark It Used to Rule Out Self-Improvement Could Not Be Passed by Anyone
The Gemini 4 Argon Safety and Security Report puts the frontier capability trend at 0.091 index points a month before April 1 and 0.118 after, not a significant change, while more than 75% of production code at Google DeepMind is AI-written and about 88% of pull requests still need human steering. An October 16 LessWrong analysis shows how thin the benchmark half of the rule-out is.
Trend Watch: The Money Doors Close on AI Agents -- Schwab, JPMorgan, Citi and BofA Block Outside Agents From Moving Funds and Offer Read-Only Access Instead, While Yelp, Delta and Amtrak Shut the Web's Front Door
In one week four US banks said third-party agents may look but not move money, Schwab's CEO said his firm stops them 'tens of times a week', and The Verge catalogued the sites turning agents away. The compromise taking shape is read-only access through MCP or open-banking APIs. The unresolved question is who is liable when an agent acts with your password.
Framework Watch: LiteLLM 1.106 Turns Off stdio MCP Servers by Default, FastMCP Caps Tool-Search Regexes After Security Fixes, Mastra 1.78 Makes Tenant Identity a Run Parameter, and Cline Moves Tasks Into Hosted Sandboxes
The weekend's releases keep narrowing who may run what: LiteLLM's proxy will no longer start local-command MCP servers unless an admin sets an environment variable, FastMCP bounds regex searches at 1,000 characters and about a second, Mastra adds per-call scopes that memory and threads derive from, and Strands fixes subagents that reported blocked runs as success.
Sphinx, a Git Hook That Quizzes You on Your Agent's Code Before the Push Goes Through, Hits Hacker News the Same Day Linear's CEO Asks for 'Someone Behind the Software'
The MIT-licensed hook, five days old, has a separate reviewer model read the diff through read-only Git tools and ask what the change does when things go wrong; a wrong answer earns an explanation and a new scenario. Its 85-comment thread split on whether a quiz can be gamed. Karri Saarinen's essay, 326 points the same day, argues that agent-produced features average a product into nothing.