Models, products and AI startups, near-duplicates collapsed to the most credible source
NVIDIA unveils Cosmos3-H-Dreams surgical generative simulatorNVIDIA released Cosmos3-H-Dreams, built on Cosmos 3 and FlashDreams, learning from clinical cholecystectomy video and robot kinematics to run interactively on one RTX PRO 6000 GPU; NVIDIA calls it the first real-time generative surgical simulator with live kinematic control trained on clinical recordings.
qev-0.8b: New Tiny Local Decision Model ReleasedA community release introduces qev-0.8b, a tiny local decision model claimed to reach state-of-the-art results among models of its size, targeting local deployment. The post offers no benchmark data or technical details.
Chainalysis: AI agents make over 138M autonomous crypto transactionsChainalysis research head Eric Jardine said the firm's research found AI agents autonomously spending crypto, completing more than 138 million transactions on Base, the most popular blockchain for agentic commerce.
Data-center debut plunges, fueling AI valuation worriesBloomberg's Open Interest highlights a gut-check for AI as a data-center debut crashes on valuation and debt fears. Separately, GM tightened its profit outlook and launched a $20 billion buyback.
Claude designs protein binders against scorpion neurotoxins; 17 of 211 designs meet criteriaAnthropic worked with Citrini Research to have an internal research model autonomously design minibinders against scorpion neurotoxins over about 64 hours. Adaptyv Bio and Twist Bioscience tested 211 designs, and 17 met the binding criteria; all designs and data were released. Scorpion stings cause over 3,250 deaths a year, and the neglected disease attracts little commercial investment.
When eight systems tie atop a benchmark, what should the results table show?The Agent Memory Leaderboard's public data shows 43 systems on the Cycle 1 coding board ranging from 46.00% to 52.67% solve rate, with eight systems tied at 52.67%. Since one task is worth 0.67 percentage points, the top cluster moves by a single task, prompting discussion on table formats.
France holds emergency meetings with platforms over protest footageThe French government held emergency meetings with TikTok and other platforms, telling them to curb protest footage and AI-altered images. Resale sites including Vinted and eBay reportedly signed an agreement to stop looted goods from flooding their platforms.
Court orders Meta to name executives who authorized book piracyIn copyright lawsuits over AI training, Magistrate Judge Thomas Hixon ordered Meta to identify the decision-makers who authorized downloading pirated books from shadow libraries like Anna's Archive and LibGen. Meta had argued its datasets were managed informally with no one in authority deciding.
Sequoia partner discusses the AI agent economyAt Citadel's Securities Conference, Sequoia Capital partner Konstantine Buhler said he is bullish on AI agents, arguing they will create a new economy and job growth, and that near-term market shifts won't erase long-term opportunity.
ASI raises $325M at $3B valuationASI founder and CEO Phillip Buckendorf said its AI helps customers such as the FAA and the military orchestrate complex systems and beat rivals including Palantir for government contracts. The company just raised $325 million at a $3 billion valuation.
AI fears return as Elice IPO plunges 32%South Korean data center firm Elice fell 32% in its market debut, reviving fears about the AI trade. Anthropic is in talks with Nscale for a computing deal worth up to $45 billion, and Eric Schmidt's early Anthropic stake could be worth up to $20 billion at a $2 trillion IPO valuation.
Mistral AI launches Voxtral 3 real-time voice agent stackMistral AI launched Voxtral 3, a model family combining speech-to-text, text-to-speech and an agent SDK, supporting 50+ languages and external tool calls in live conversations, with self-hosting or EU-hosted APIs for data control.
Meta's Muse hits 10M U.S. downloads in 40 daysMeta's AI companion app Muse reached 10.2 million U.S. downloads in its first 40 days, becoming one of only 12 apps to hit this milestone. Its 7-day retention rate is 30%, above the 19% average for rivals like ChatGPT. Meta allocated 39% of its ad spend to promote Muse in October.
Bot Detection Benchmarked Against AI Agents: 9%–99%Researchers tested detection systems including Proof of Human, reCAPTCHA v3, hCaptcha and Cloudflare Turnstile against conventional automation, stealth browsers and AI browsing agents, finding detection rates ranging from 9% to 99% versus 33%–87% a year earlier. Performance varied widely, and some systems failed or crashed before sessions completed.
Closed SF Centre mall proposed as AI data centerMaryland-based K-G Growth Partners and an unnamed AI lab have inquired about permits to convert the former Bloomingdale's space at the shuttered San Francisco Centre into an AI data center. The project is early-stage, with no application filed, and faces Mayor Daniel Lurie's 45-day pause on new data center approvals.
Altman reportedly pitched three endgames for AI safetyOpenAI CEO Sam Altman reportedly outlined three paths to safe AI in an internal presentation: an IAEA-style international watchdog, a merger of major labs into one entity, or one firm powerful enough to shut down everyone else's compute — an option he called 'terrible.'
Citadel Securities sees AI trade gaining after midtermsCitadel Securities' Scott Rubner said investors are poised to pile back into the AI trade next month once the US midterm elections have passed, setting up a rebound for the sector.
Grok bot leaks tech CEO's bank data in public SlackA report says a Grok bot exposed a tech CEO's banking data in a public Slack channel, renewing concerns about privacy and security when AI assistants handle sensitive information.
Researchers study the tipping point when good AI goes badGeorge Washington University professor Neil Johnson is studying the tipping point where AI agents shift from helpful to harmful, potentially giving wrong or dangerous advice. He focuses on isolated users and how altering prompts and training can delay or prevent such shifts.
AI companies sketch plans to share wealth ahead of massive IPOsAs AI companies approach massive IPOs, firms like Anthropic and OpenAI are outlining plans to share wealth with the public. Anthropic committed 5% of its shares to a new Public Benefit Fund, but key questions about operations, transparency, and beneficiary selection remain unresolved.
AI neoclouds lose shine as IPOs stall and stocks sinkNeoclouds, which supply the data-center computing behind AI systems, are losing investor favor. South Korean AI computing supplier Elice fell 32% in its Seoul debut, the latest sign of waning confidence, as related IPOs stall and stocks sink.
CSU uses AI to improve hail forecastingColorado State University, funded by the National Science Foundation, is combining an NVIDIA-developed nowcasting model with its existing machine learning model to extend hail warnings from under 30 minutes to hours or days, addressing rising insurance costs from more frequent hailstorms.
AWS introduces Agent Registry for runtime tool discoveryAWS launched Agent Registry, enabling AI agents to discover available tools at runtime and simplifying tool integration.
Ex-Nvidia exec: physical AI still in its 'LEGO era'Amulya Vishwanath, who spent nearly seven years at Nvidia and now leads Techable Ventures, says physical AI has not yet had its ChatGPT moment and is more in a 'LEGO era,' with basics like robotic hands and actuators being rebuilt. She sees talent as a key moat.
AI turns cybersecurity into a continuous automated cat-and-mouse gameOODA Loop examines how AI is reshaping cybersecurity, with both attackers and defenders using AI to drive continuous, automated cycles. Security operations are shifting from manual response to machine-speed attack-defense contests.
FairStress: New benchmark evaluates fairness in LLM decision-makingHugging Face introduced FairStress, a benchmark that measures demographic fairness in LLMs beyond final choices by testing how readily models switch decisions when challenged. Across models like Olmo, Llama, Qwen, Gemma, Claude, and GPT, the same persuasive arguments can cause asymmetric decision changes, revealing hidden biases.
Survey: AI enters top 10 U.S. fears for first timeChapman University's 12th annual Survey of American Fears included AI questions for the first time: about 51.9% of respondents fear AI harming the political process and 49.7% fear economic harm, both ranking high. The poll covered over 1,200 people with a margin of error of about 3.79 points.
AWS details securing AI agent identity with X.509 PKIAWS published guidance on using X.509 public key infrastructure to establish and manage AI agent identities, improving authentication and access security in multi-agent environments.
Gradium open-sources Phonon edge TTS models, 31M and 40MGradium released Phonon, its first open-source model family: 31M and 40M text-to-speech models designed to run on-device. The fully open inference stack targets 10+ backends (CPUs, GPUs, NPUs), mostly in Rust, aiming to spur edge voice application development.
Ipsos: Americans Know AI Far Better Than Super IntelligenceAn Ipsos survey of 1,016 U.S. adults in October 2026 found 49% have heard a lot about AI versus 13% for super intelligence. Among those aware, 65% are confident in their AI understanding versus 51% for super intelligence.
The A.I. Daily