The day's highest-engagement model launch.
AI NEWS
Past AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
This day
October 28, 2026
23 items for 2026-10-28, ranked by combined engagement. Unverified items are marked.
Top 3
First-hand cases and numbers for AI across the drug pipeline.
A milestone-scale claim for AI-driven formal math.
All
Google DeepMind unveils Genie 4 world model: real-time natural-language steering, native audio and live dialogue
The day's highest-engagement model launch.
CollapseGoogle DeepMind unveils Genie 4 world model: real-time natural-language steering, native audio and live dialogue
Google DeepMind released Genie 4, calling it its most realistic and interactive world model, letting users explore dynamic environments and see how scenes, sounds and people respond in real time. Sundar Pichai added that users can steer events, camera angles and actions live, get native audio matched to the world (including characters who converse in real time), simulate photorealistic scenes and physics, and get more coherent multi-step events via Gemini reasoning. Google AI says Gemini acts as the simulation's "brain", translating requests into actions and tracking the scene. It is rolling out as a research preview to trusted testers, coming to Google AI Ultra subscribers in Project Genie and to early-access partners via the Gemini API and AI Studio. The post drew about 14.1k likes and 1.49M views.
Anthropic shares early drug-development results with six partners and acquires one-person biotech Mithril to run its own programs
First-hand cases and numbers for AI across the drug pipeline.
CollapseAnthropic shares early drug-development results with six partners and acquires one-person biotech Mithril to run its own programs
Anthropic's blog says it ran deep research partnerships with six organizations—BigHat, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group—and has started its own drug discovery. Reported figures include: Generate used Claude Code to design a segment about 70% shorter that yielded roughly 10x more receptors in lab tests; Divergence generated roughly 35,000 antibody designs in about a month; at Enveda, across 1,061 pages of manufacturing documents Claude caught all 11 issues human reviewers found and flagged seven they missed, cutting human review time by 89%. Anthropic acquired one-person Mithril Medicine, whose founder used Claude to coordinate 11 external labs and took a muscular dystrophy candidate from idea to promising mouse results in under 18 months. Anthropic stresses the work is early and mostly preclinical; figures are company- and partner-reported.
Axiom announces a complete Lean auto-formalization of the Classification of Finite Simple Groups, ~15,000 pages now computer-verified
A milestone-scale claim for AI-driven formal math.
CollapseAxiom announces a complete Lean auto-formalization of the Classification of Finite Simple Groups, ~15,000 pages now computer-verified
Axiom announced a complete Lean auto-formalization of the Classification of Finite Simple Groups (CFSG), calling it the largest math proof in history now computer-verified. The official post says the original proof spans about 15,000 pages built by over 100 mathematicians across generations. Team member Andrew added that the formalization exceeds 43 million lines of Lean, covers 715 papers and books, and made 100 substantive fixes to the literature. These figures are self-reported; independent review status was not stated. The post drew about 4.8k likes and 1.79M views.
Claude expands Claude for Nonprofits: 25,000+ nonprofits use it; new plugin co-built with ~30 groups
Vertical plugins for small teams.
CollapseClaude expands Claude for Nonprofits: 25,000+ nonprofits use it; new plugin co-built with ~30 groups
Claude says more than 25,000 nonprofits now use Claude, most of them small local teams. It is expanding Claude for Nonprofits with a plugin built with around 30 nonprofits, adding skills for fundraising, marketing and hiring grounded in each team's context and AI policy. Details are in the official article at claude.com/resources/articles/expanding-claude-for-nonprofits. The post drew about 6.8k likes and 720k views.
Jensen Huang's GTC Berlin recap: DSX POD Vera Rubin modular AI factories, Agent Toolkit and OpenShell, RTX Spark with Microsoft
Highest-view industry keynote recap.
CollapseJensen Huang's GTC Berlin recap: DSX POD Vera Rubin modular AI factories, Agent Toolkit and OpenShell, RTX Spark with Microsoft
Jensen Huang recapped GTC Berlin, saying Europe is building its AI ecosystem from AI factories to agents to robots. Highlights: DSX POD—Vera Rubin modular factories built, cabled and tested in the factory and delivered ready to run; NVIDIA Agent Toolkit bringing together models, harnesses, sandboxes, knowledge and tools so companies control what agents know, access and do, with OpenShell for safety; new physical-AI customers across AI-RAN, robotaxis and robotics; and RTX Spark, after four years with Microsoft, bringing personal agents to Windows PCs. The post drew about 6.6k likes and 4.39M views.
Cursor ships Agent SDK: agents as code (Markdown prompts, TypeScript tools) with built-in evals
Coding tools extending into agent platforms.
CollapseCursor ships Agent SDK: agents as code (Markdown prompts, TypeScript tools) with built-in evals
Cursor announced its Agent SDK: developers define agents as code, with prompts in Markdown and tools in TypeScript. Agents can be deployed on Cursor's hosted infrastructure or run on your own. Built-in evals measure every change so agents improve over time. The post drew about 3.8k likes and 379k views.
DeepSeek V4.1 Flash lands in Cursor; every effort level sits on the CursorBench cost frontier at $0.60–$0.98 per task
Cost-performance signal for a Chinese open model in a mainstream IDE.
CollapseDeepSeek V4.1 Flash lands in Cursor; every effort level sits on the CursorBench cost frontier at $0.60–$0.98 per task
Cursor announced DeepSeek V4.1 Flash is available in Cursor. It says every effort level sits on the CursorBench cost frontier, from $0.60 to $0.98 per task, with no model scoring higher at that cost. Users can enable it under Cursor Settings > Models. These are results on Cursor's own benchmark. The post drew about 3.9k likes and 321k views.
Nvidia to invest $3B in Meta's equity offering as Meta lifts its stock sale to $25B (relayed)Unconfirmed
A shift in how AI buildouts are financed.
CollapseNvidia to invest $3B in Meta's equity offering as Meta lifts its stock sale to $25B (relayed)
Wall St Engine reports Nvidia will invest $3B in Meta's equity offering. Meta plans to deploy more NVIDIA Vera Rubin systems, Vera CPUs, Spectrum-X networking and BlueField-4 for next-generation AI agents, and the two are exploring collaboration on agent security and monitoring. The post says Meta raised its common stock offering to $25B from $22.5B, bringing current offerings to $37.5B, with a potential $60B share-sale program planned for 2027. Relayed by a markets account; no primary filing link in the post.
AMD: Helios rack-scale systems now running at Meta; EPYC powers Meta's Muse
Meta locks in multiple chip suppliers the same day.
CollapseAMD: Helios rack-scale systems now running at Meta; EPYC powers Meta's Muse
AMD says its Helios systems are now running at Meta, EPYC processors power Meta's Muse, and Instinct GPUs accelerate AI workloads. The post links an official release on AMD and Meta advancing their strategic partnership for Meta's AI infrastructure. A markets account separately reported a $1B AMD investment in Meta; that figure is relayed—check the official release. The post drew about 450 likes and 24k views.
Baidu open-sources Unlimited OCR 2: SR-SWA keeps per-step latency and memory nearly flat on very long documents
Long-document parsing matters for records, contracts and archives.
CollapseBaidu open-sources Unlimited OCR 2: SR-SWA keeps per-step latency and memory nearly flat on very long documents
Baidu open-sourced Unlimited OCR 2. Its SR-SWA mechanism condenses text-visual attention interactions into a fixed-size set of working references to curb attention compute on long documents. Baidu says it parses documents of virtually any length within a finite context window, with nearly constant per-step decoding latency and memory as documents grow, and reports state-of-the-art results on MPDocBench and MarathonDocBench (dozens to hundreds of pages). Benchmarks are self-reported. The post drew about 990 likes and 79k views.
Ant open-sources Ling-3.1-flash-1019: 44 on the Artificial Analysis Intelligence Index (up from 41), plus a base model
Continued iteration of Chinese open models.
CollapseAnt open-sources Ling-3.1-flash-1019: 44 on the Artificial Analysis Intelligence Index (up from 41), plus a base model
Ant's Ling team open-sourced Ling-3.1-flash-1019, reporting 44 on the Artificial Analysis Intelligence Index v4.3.2, up from 41 for Ling-3.1-flash. It also released Ling-3.1-flash-base as an open foundation for further training. The score cites a third-party index; the post is official. It drew about 480 likes and 257k views.
Artificial Analysis: Microsoft's MAI-Image-2.7-Preview ranks #2 in image editing, beating its predecessor on all 7 edit actions
Third-party eval with per-dimension data.
CollapseArtificial Analysis: Microsoft's MAI-Image-2.7-Preview ranks #2 in image editing, beating its predecessor on all 7 edit actions
Artificial Analysis says Microsoft's MAI-Image-2.7-Preview ranks #2 on AA-Image-Editing v2.1, behind OpenAI's GPT Image 2.5 Sunburst (max). It beats MAI-Image-2.6 on all 7 editing actions measured and ranks #1 on 4, with the biggest gains in Composition & Framing (+57 Elo), Text or Symbol Edits (+49) and Scene & Style Edit (+48). It is available in MAI Playground for editing only, with text-to-image said to be coming. Mustafa Suleyman called it the world's #2 image-edit model. The post drew about 400 likes and 245k views.
Nous Research revamps Hermes Agent plans: every new plan includes a cloud agent, with daily and weekly limits
How open-model agents are being monetized.
CollapseNous Research revamps Hermes Agent plans: every new plan includes a cloud agent, with daily and weekly limits
Nous Research revamped Hermes Agent subscriptions: every new plan includes a cloud agent out of the box. Usage now has daily and weekly limits instead of one monthly block of credits. Existing subscribers can switch any time in the official portal. The post did not detail price changes. It drew about 1.7k likes and 183k views.
Anthropic study: analysts went with an AI agent's lean 86% of the time, versus ~50% expected with no effect
Quantified evidence of agents swaying experts.
CollapseAnthropic study: analysts went with an AI agent's lean 86% of the time, versus ~50% expected with no effect
Anthropic researcher Alex Tamkin described a study in which an agent was told to make the best honest case for its options, change its mind if evidence required, and never mislead analysts. Analysts went with the agent's lean 86% of the time, versus about 50% expected if the lean had no effect. Tamkin said the study was preregistered, with details in the paper; Ethan Mollick amplified it as evidence of AI shaping expert judgment. News coverage relays about 50 causal-inference professionals as participants—a relayed figure; see the paper.
Ethan Mollick: the end of detailed thinking-trace summaries in Claude and ChatGPT makes human-in-the-loop harder
Directly affects auditability and oversight.
CollapseEthan Mollick: the end of detailed thinking-trace summaries in Claude and ChatGPT makes human-in-the-loop harder
Wharton professor Ethan Mollick says Claude and ChatGPT no longer show detailed summarized thinking traces. He suggests the reasons are likely anti-distillation and UX (his inference). He argues this makes being a human in the loop harder, since it is tougher to weigh in on in-process AI work, pushing toward automation rather than augmentation. The post drew about 720 likes and 356 replies.
Replit launches Cloud Agents for any codebase; its design team shipped 250k+ lines in six months
Non-engineers contributing to company codebases.
CollapseReplit launches Cloud Agents for any codebase; its design team shipped 250k+ lines in six months
Replit says Cloud Agents turn natural language into tested pull requests, with engineers still approving final code before it ships. Its own design team used it to ship over 250k lines of code in the last six months, reviewed and approved by engineers. CEO Amjad Masad added that the answer to "is Replit built with Replit?" went from "kinda" to "absolutely", and Cloud Agents can run any repo. Output figures are company-reported.
GitHub Copilot assisted permissions GA: an LLM auto-approves clearly safe requests; others still need you
A concrete pattern for agent permission approval.
CollapseGitHub Copilot assisted permissions GA: an LLM auto-approves clearly safe requests; others still need you
GitHub Changelog announced assisted permissions are generally available in GitHub Copilot CLI, the app and Visual Studio Code. A large language model auto-approves clearly safe requests, while others require user approval. See the github.blog changelog entry dated 2026-10-26. Engagement is low (about 15 likes), but it bears on enterprise agent permission governance.
Concurrence launches "Superintelligent Care": AI agents execute a doctor's plan—follow-ups, med checks, pharmacy calls
High-attention launch of care-execution agents.
CollapseConcurrence launches "Superintelligent Care": AI agents execute a doctor's plan—follow-ups, med checks, pharmacy calls
Concurrence founder Khoo (@khookhar) announced the company's next chapter as the greatest leap toward Superintelligent Care, aiming to see each patient's future in time to change the outcome; he shared losing his mother at 13 to a cancer that should have been caught earlier. Unusual Whales relays the product as AI agents that carry out a doctor's plan end to end—follow-ups, med checks, pharmacy calls—fed back to the care team. No clinical or operational metrics were given; outcomes unverified. The post drew about 670 likes and 620k views.
Tesla execs on Optimus: a single ~2.5 kWh battery that isn't a major cost; near-term job is charging the robotaxi fleet
Clues to humanoids' first commercial jobs.
CollapseTesla execs on Optimus: a single ~2.5 kWh battery that isn't a major cost; near-term job is charging the robotaxi fleet
The Humanoid Hub summarized a Munro Live interview: Tesla's Lars Moravy and Eric Earley said Optimus uses a single ~2.5 kWh battery module that, unlike a car pack (a quarter to a third of cost), isn't a major cost of the robot. Near-term, Optimus is expected to plug in and charge the robotaxi fleet, and might later service cars or enter delivery. They framed Cybercab, Optimus and Robovan—plus Semi for long haul—as a trifecta: "Maybe we're a transportation company." Relayed interview notes. The post drew about 2.2k likes and 81k views.
Trump reportedly says he will "ban" the term AI after an order told agencies to use "Super Intelligence (SI)" (relayed only)Unconfirmed
Could affect public-sector procurement language; awaiting primary text.
CollapseTrump reportedly says he will "ban" the term AI after an order told agencies to use "Super Intelligence (SI)" (relayed only)
Several accounts relay that Trump said he will officially "ban" the terms "AI" and "Artificial Intelligence", and that a September executive order already directed federal agencies to use "SI / Super Intelligence". Another news account says agencies were ordered to replace the term. Only relays and screenshots were found on X—no White House or Federal Register primary text—so scope and legal effect are not officially stated. The post drew about 40 likes and 9.1k views.
Business Insider: Amazon projected ~$1.7B in 2026 Alexa+ compute costs and is routing more to in-house models (relayed)Unconfirmed
Unit economics of consumer assistants.
CollapseBusiness Insider: Amazon projected ~$1.7B in 2026 Alexa+ compute costs and is routing more to in-house models (relayed)
Wall St Engine relays Business Insider: Amazon internally projected Alexa+ cloud computing costs of about $1.7B in 2026, nearly triple the prior year. To cut costs, Amazon was relying less on Anthropic's models and routing more requests to its own cheaper systems. It also considered delaying or scaling back Moonraker, an advanced AI agent project projected to need over $100M in GPU spend this year. Relayed reporting; Amazon has not detailed it officially.
The Information: OpenAI wins Anthropic customers with steep discounts; Replit moved most AI spend after a 30%+ discountUnconfirmed
Price-war signal in enterprise AI procurement.
CollapseThe Information: OpenAI wins Anthropic customers with steep discounts; Replit moved most AI spend after a 30%+ discount
The Information reports OpenAI is winning Anthropic customers with steep discounts on models and coding tools. It says Replit shifted most of its AI spending to OpenAI after securing a discount of more than 30% on a contract worth over $100M a year. Another post from the outlet says OpenAI's share of AI-assisted coding at Rippling rose from 16% to 52% as the company sought to cut costs. Paywalled media reporting; the companies did not respond in the posts.
Mustafa Suleyman: MAI Cyber doubled user adoption to 50% in 30 days and got 20% more token-efficient
Adoption data for an in-house security model.
CollapseMustafa Suleyman: MAI Cyber doubled user adoption to 50% in 30 days and got 20% more token-efficient
Microsoft AI chief Mustafa Suleyman says its cyber models are getting strong: within Microsoft's security harness, the MAI Cyber model doubled user adoption to 50% over the last 30 days. He adds it became 20% more token-efficient without sacrificing quality. Company-reported; no evaluation method given. The post drew about 110 likes and 15k views.