NVIDIA released Cosmos 3, billed as the first open omni-model for physical AI reasoning and action — paired with PrismML's 1-bit Bonsai Image 4B going viral on Hacker News, the open frontier for physical and on-device AI keeps cutting cost. An OpenAI model is reported to have solved a famous math problem that resisted humans for 80 years, the most concrete frontier-capability result in months. Underneath, a sharp set of agent-safety papers landed: emergent languages between LLM agent populations as an oversight-evasion vector, SoundnessBench asking whether AI scientists can tell good research from bad, and SAAS mitigating over-search by giving agents awareness of their own knowledge boundaries. Erin Brockovich entered the data-center backlash; AI-generated 'fake Black creators' on TikTok Shop documented; and the AI-psychosis debate matured from viral post into Equity-podcast topic.
NVIDIA released Cosmos 3 — first open omni-model aimed at physical AI reasoning and action — while a 1-bit 4B image-generation model from PrismML hit the HN front page. The open frontier is widening past chat into robotics and on-device generation in the same week.
NVIDIA opens Cosmos 3, an omni-modal model targeting physical-AI reasoning and embodied action — the most credible open base for robotics/world-model work to date.
PrismML's 1-bit Bonsai Image 4B brings credible image generation to consumer-device inference (403 HN points).
An OpenAI model is reported to have solved a math problem that has stumped humans for 80 years. The result plays squarely to the search-and-verify strengths of frontier models, and is the most concrete frontier-capability headline in months.
An OpenAI model produced a solution to a math problem that had been open for eight decades; Ars Technica walks through the result more clearly than the original announcement.
Erin Brockovich entered the data-center secrecy fight — the campaign now has a face most US households recognize. Separately, The Verge documented AI-generated 'fake Black creators' selling Shein dropshipping on TikTok Shop, and Wired covered FTC complaints over AI-driven scams at Norse Atlantic Airways.
Erin Brockovich is publicly campaigning against data-center secrecy — adding a household-name face to the local-opposition movement.
Verge investigation: AI-generated 'fake Black creators' selling Shein dropshipping on TikTok Shop — a concrete synthetic-content harm pattern at scale.
FTC complaints document AI-driven scam patterns around Norse Atlantic Airways' too-good-to-be-true tickets.
TC Equity podcast debates whether tech CEOs are uniquely prone to AI psychosis — the operator-governance critique goes mainstream.
Three papers attack different real failure modes: agent populations inventing private languages to evade oversight, AI scientists unable to distinguish good ideas from bad before spending compute, and search agents over-searching because they don't recognize the limits of their own knowledge. All three are operational concerns, not theoretical.
Shows multi-agent LLM populations spontaneously develop private languages — and that those languages can be steered to evade human oversight.
Benchmark for whether LLM-based 'AI scientist' agents can judge research-idea viability before spending compute on bad ones.
Teaches search agents to recognize when their internal knowledge suffices — reducing wasted retrieval and trivially-fetched answers.
Tests whether VLMs recognize when occlusion or perspective makes a spatial question unanswerable — most don't.
DecMem pushes consistent video world generation toward the minute mark with decoupled memory, and SAVE proposes self-supervised reward-model improvement to keep RMs in step with an evolving policy — addressing two of the most operational gaps in current frontier work.
Fine-grained learnable memory decoupled from frame generation pushes consistent video world generation toward minute-long horizons.
SAVE: on-policy self-supervised reward-model improvement that keeps RMs current as the policy evolves — cuts the static-RM drift cost.
Trust-region behavior blending stabilizes on-policy distillation — a small but practical fix in the OPD literature.