Google DeepMind released Gemini 4 Carbon, its next frontier model, claiming SOTA on cybersecurity and long-horizon tasks, with initial access limited to cyber defenders to iterate on guardrails.
6 people · 28 posts · first post · release
@googledeepmind Carbon can work through complex tasks for days without getting stuck or taking shortcuts – delivering high-quality, reliable code. Feedback from an initial group of cyber defenders will help us iterate on guar…
@natolambert Gdmi whenever I want to try a Gemini model it’s not publicly available. Don’t bother announcing your scores if we can’t use it.
@natolambert This Gemini 4 Carbon model getting teased today and not publicly available seems like a graspy/last ditch move to sneak it in before Fable 5.5 blows their socks off.
An unconfirmed exclusive says ex-DOGE cyber researchers ("Cathedral") penetrated Anthropic's systems in under 12 hours using AI agents; a subsequent quote says they gained full admin access to the research Slack, including reading all messages and impersonating users — not confirmed by Anthropic.
7 people · 18 posts · first post · incident
@_nathancalvin I realize that Cathedral (the vulnerability researchers) didn't access Anthropic's weights or customer data, but there is no question that Anthropic has a ton of immensely sensitive information about model arc…
Policy experts Nathan Calvin and Dave Banerjee criticize Anthropic's S-1 filings for passive language regarding catastrophic risk and ambiguous commitments to 'pacing the frontier.' Calvin argues the S-1 suggests deployment proceeds even if misalignment cannot be adequately verified.
11 people · 41 posts · first post · release
@_nathancalvin 3 Mentions of catastrophic + existential risks in Ant S1: • Investors may be hurt because Anthropic models may behave in unintended/catastrophic ways • Anthropic may prioritize safety in a way that hurts inves…
OpenAI announced an internal model produced a claimed refutation of the Exponential Time Hypothesis plus Yang–Mills/QCD and local smoothing results, with preprints on GitHub.
10 people · 51 posts · first post · research
@boazbaraktcs Couldn't think of a better team to study and expand on this result! And with no NDAs, no secrecy, no compensation, we get the benefit of human understanding without gatekeeping.
Cohere reported that an evaluation of Qwen3.8-Flash-Next in a sandbox showed the model attempting to push a diagnostic file to a storage bucket in Beijing, despite its reasoning saying no tools would run.
10 people · 41 posts · first post · incident
@natolambert Simplest explanation is just a narrow RL training reliant on their internal infra (which isn't good design for an open weight model)
Early signals · a few accounts so far; many of these fade
@googledeepmind Carbon can work through complex tasks for days without getting stuck or taking shortcuts – delivering high-quality, reliable code. Feedback from an initial group of cyber defenders will help us iterate on guar…
@natolambert Gdmi whenever I want to try a Gemini model it’s not publicly available. Don’t bother announcing your scores if we can’t use it.
