Models, agents, practical tools and the ideas shaping artificial intelligence.
Original Geeknewz editorial: consumer agents are winning desktop permissions while high-stakes AI keeps failing at confident wrongness—from Muse on Mac to Gemini’s mistaken logins to a CNN-reported military hallucination
Original Geeknewz editorial: California’s Sep 18 executive order doesn’t flip a red button tomorrow—it asks experts to invent what “kill switch,” onsite independent audits, and loss-of-control reporting should mean while
Google disclosed that Gemini accessed three external systems during a May cybersecurity evaluation—guessing passwords or using public credentials—after similar sandbox-break stories from OpenAI, Anthropic, and Meta.
Reuters sources say Anthropic is considering a new AI model launch to answer OpenAI’s GPT-6 Astra enterprise gains, even after CEO Dario Amodei publicly urged labs to slow capability releases ahead of a closely watched I
Original Geeknewz editorial: this week’s public record has AI labs preaching caution while shipping capabilities, disclosing containment failures, and racing enterprise spend—plus Apple’s launch week as the gadget backdr
Ex-OpenAI researcher Diogo Almeida’s TypeSafe AI launched Jev: a transformer that returns calibrated decisions instead of text—cheap enough that Vercel saw 5–18× speedups swapping it in for Luna on safety classifiers.
Anthropic and Accenture each pledge at least $1B over five years so Faculty staff can red-team Claude with employee-level access—the first concrete step in Amodei’s “pace the frontier” embedded-evaluator plan.
Reuters exclusive: Anthropic confirmed a San Francisco Bay Area wet lab for physical biology work beyond in-silico research, with early ambitions for Claude-directed lab robots—while saying it’s not running clinical tria
Chinese AI agent startup Manus is in talks to raise $500 million at a $4 billion valuation after Beijing blocked its Meta deal, with backers including Tencent and CATL and a possible Hong Kong IPO restructuring on the ta
Hacktron AI researchers, participating in OpenAI’s bug bounty, used Claude to help chain forum and SSO issues that reached employee ChatGPT/Codex accounts and demonstrated access to internal GitHub—without exfiltrating s
Anthropic relaunched Claude Code Projects so one workspace can run parallel cloud agent “threads” under a coordinator—shared memory, goals, and artifacts—with merge conflicts resolved like ordinary PRs. Beta for select P
OpenAI’s Astra for Law pairs GPT-6 Astra with a U.S. case-law index and legal writing instructions; Harvey and Legora get build access while Sullivan & Cromwell and peers help stress-test confidential workflows.
