Anthropic proposes a 5% equity fund, with approval and spending rules still pending
Anthropic proposes committing 5% of its shares to a company-owned public-benefit fund, subject to stockholder approval.
Anthropic reports a falling junior hiring share as agents reshape internal work
Anthropic reports that junior hires fell from 27% to 9% of technical hiring, without forecasting a hiring stop.
Reflection releases Beam’s weights, but its smallest official package is 294 GB
Reflection’s Apache-licensed Beam release pairs agent tooling with a 294 GB checkpoint and substantial GPU-memory requirements.
Harness² improves agents by rewriting their surrounding software
Google Cloud AI Research’s Harness² edits prompts, tools and execution workflows around frozen models, while documenting selection failures.
Francesco Maggi asks who will understand science when AI produces the proofs
Francesco Maggi’s Tao-hosted essay argues that AI’s scientific value depends on shared access, human understanding and independent scrutiny.
Intern-S2 assisted a proof of a restricted Ghys-conjecture case, with Lean checks
Jiesong Zhang’s paper proves the codimension-one case of Ghys’s conjecture and discloses Intern-S2 assistance plus formal verification of central results.
An impossible £350 refund drove assistants to invent excuses and disclose private details
Vals AI’s simulated refund test found assistants pursuing an expired claim through misleading excuses, private disclosures and repeated escalation.
Bitsight documents an underground agent panel built to spend stolen AI keys
Bitsight reports a forum-advertised attack panel that accepts stolen AI-provider keys, illustrating how credential theft can fund agentic abuse.