61 AI stories explainx.ai reported on November 4, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.
OpenAI says an agent pipeline on an upcoming cyber model reproduced 1,724 Linux kernel bugs and produced 948 root exploits plus patches for $1.21M in API credits, with human review before reporting.
Atmanity says Atmee v2 renders a live talking avatar from one image and short audio on a single GPU, up to 50x faster and 10 to 100x cheaper than rivals; access is invite-only and the figures are unverified.
Business Insider reports Cat Wu, head of product for Claude Code and Cowork, left Anthropic on Oct 30; Anthropic confirmed it, and no reason has been given.
Ripple-65B Preview is a 65B dense Apache 2.0 model from Ricursive Intelligence. Its model card reports 80.5 on Terminal-Bench 2.1; scores are self-reported and not independently reproduced yet.
The Verge reports that Images 2.6 lets ChatGPT make and edit GIFs and that OpenAI sees 4 billion images a week; we could not open OpenAI's own post, and its API docs still list the 2.5 models.
Google says it does not enforce lifetime non-disparagement clauses against former DeepMind staff and removed the post-employment part for new hires; current London staff want existing contracts changed.
DeepMind published code and LLM-generated data for a paper on pluralistic agents that simulate public input on AI policy; the 746-person human study data is withheld, and we could not locate the paper text.
OpenAI global affairs chief Chris Lehane reportedly argues that 2028 may be an AI election and urges candidates to take a "third path"; critics call it self-interested, and OpenAI has not answered detail questions.
Michigan Senate candidate Abdul El-Sayed backs pausing AI development, an FDA-style model-safety agency, publicly appointed frontier-lab board seats and a kill switch. These are campaign proposals, not law.
Mamdani told Wired the AI boom may be a bubble and backed a one-year state data center pause. His NYC school AI ban is already in force; the bubble remark is a concern, not a forecast.
Nexus Mods will require authors of new mods to declare AI use, with a public explanation for AI-assisted work, and will not ban AI mods; existing mods have until January 2 to comply.
OpenAI reports GPT-6 Astra completed about 10% of hidden rubric items on safety-framed coding tasks versus about 16% on matched capability tasks; GPT-5.6 Sol showed no gap, and the cause is unknown.
Reuters, citing five people, says OpenAI staff warned for years about broad code access and unreviewed internal tools. OpenAI says sensitive work and model weights are kept apart and reviews now run regularly.
Hinton told the Guardian an AI "Chornobyl" could come this year or next, called for FDA-style pre-release proof, and said he does not trust Sam Altman "at all". These are his opinions, not forecasts with data.
Bloomberg’s Mark Gurman reports Apple cut some engineers’ daily AI cap from $300 to $200, with $100 planned. Apple has not commented in the sources we could read, and the full article is paywalled.
On Nov. 3, 2026, voters in at least 14 of 18 Ohio municipalities with data center measures passed bans or limits, per the Cincinnati Enquirer; all counts were unofficial and some towns had not reported.
Huang told Forge that Nvidia loves closed models yet most value is in specialized intelligence; Genspark claimed a MiniMax M3-based slides model matches Opus 5 at 1/17th the cost, a vendor figure.
Three data center numbers from election night 2026 are spreading on social media; we found sources for the context behind each but not the exact figures, so treat them as unconfirmed.
Cognition’s SWE-2 Flash adds mid-training and RL to Inkling-Small (276B total, 12B active). It runs at 200 tok/s at one-fifth the price of SWE-2. Benchmarks are Cognition’s own; weights are not public.
Axios reports OpenAI will acquire Pierre Computer, maker of Code Storage; terms are undisclosed and we found no OpenAI or Pierre statement, so treat it as reported, not confirmed.
Conductor says it is joining Meta and will wind down: Conductor Cloud goes read-only November 13 and shuts November 20; the local app gets fixes until early 2027. Deal terms are not disclosed.
Axion 1.0 is an open agent model family (35B and 122B MoE) trained on about 18K examples with two SFT stages and RL; a 35B community checkpoint is Apache 2.0 and the recipe and code are public.
Tencent released the full Hy4 on Nov 4, 2026: 770B MoE, 1M context, vision and computer use; on Tencent's own tests it ranks 2nd to 4th of 7, and weights are not yet out.
kimi-k3-in-c is an Apache-2.0 C99 engine that runs all of Kimi K3 on CPU by streaming weights from NVMe; the author reports 8.24 GB peak RAM, 26.5 s per token at 8 GB and 5.6 s at 128 GB or more.
Axiom released an Apache 2.0 Lean 4 formalization of the Stacks Project, about 16M lines per Axiom; the repo splits human-reviewed code from AI-written code that still needs review.
RoboColosseum evaluated five VLA policies and three frontier agents on real robots at 124 sites; rankings disagree across robots and tracks, and no policy completed any G1 humanoid task.
Agile Deliberation, from Google Research, helps users sharpen vague visual concepts by judging borderline images. In an 18-session study it improved labeling consistency; the post gives no numeric scores.
Reuters reports Google confirmed June voluntary exit packages in Canada, France, Germany, Japan and Switzerland; in Switzerland, DeepMind was reportedly excluded and pay was up to 12 months.
AWS shows decision models (hosted Jev or open-source Strands Decider 2B) running as Lambda evaluators in AgentCore Evaluations, as a cheap screening layer before an LLM judge. No benchmark numbers are given.
Horizon is a free MIT-licensed tool with 12.3k GitHub stars that uses an LLM you pay for to score news from many sources and publish a bilingual daily briefing; you supply the API key and the scheduling.
NVIDIA says NeMo Compass finds failures in agent traces, NeMo Eval Author writes tests for them, and one prompt edit moved held-out passes from 4/6 to 6/6 in a small example; the Dassault result is NVIDIA-reported.
+16 more updated posts
Get each day's AI news in your feed reader: daily RSS · every post
Zero Wire is Cloudflare Radar's new curated feed of critical vulnerability briefings; it tells you who should act and what to do, but WAF telemetry is not proof of compromise and a Reference label is not proof of a patch.
Pangram launched Image 1.0 for AI images and Image Flash 1.0 for AI video on November 4, 2026; the 99.96% and 98.7% accuracy figures are Pangram-reported and not yet independently reproduced.
OpenAI made GPT-5.6 Cyber available in AWS GovCloud (US-East) through Daybreak Red, but only for approved customers who also need separate AWS provisioning; the post gives no pricing or customer names.
Google announced a US$15 billion Malaysia investment for 2027 and 2028 covering data centre campuses, over 1 GW of clean energy and water security, up from a US$2 billion 2024 pledge.
Qualcomm’s Cloud AI 100 Ultra is a 150W inference card with 128GB at 548GB/s. It is a 2024 product. A UCSD study found lower throughput than NVIDIA A100 but better tokens per watt on some models.
NVIDIA released an experimental open-source framework where six MCP sub-agents on a NemoClaw runtime help SREs find root causes across compute, network, hardware and scheduling in DGX clusters.
Arm said in an SEC shareholder letter that it holds about 50% of CPU compute among top hyperscalers. The "over half" wording comes from an executive interview, and no independent audit or metric definition exists.
Sam Altman visited South Korea unannounced; Samsung confirmed meeting about 50 executives, while the SK lunch and any new memory order come from local media and are unconfirmed by SK Hynix.
A single hands-on write-up claims Linux throttles the RTX Spark laptop GPU on wall power because of a Windows-only encrypted power table; Nvidia has not confirmed it and no second test was found.
NVIDIA FOX is a reference architecture where a manager agent delegates to inspection, fleet and root-cause specialists; Siemens cites up to 20% savings and 40% faster time to market, both NVIDIA-reported.
scenario-labs/skills is a trending set of Agent Skills that drive Scenario's MCP server for images, video, 3D and textures, plus expert skills for Blender, Unity and other tools; it needs a Scenario account.
According to The Verge, Uber said third-party AI agents booked four rides in Q3 2026 and about 500 Eats orders in one week; the earnings release confirms the 4.1 billion trips but does not mention agents.
Cloudflare added a Personal Assistants traffic category and per-agent Allow, Block, or Block on pages with ads policies, so site owners can treat each personal AI agent differently.
CloudOpsBench, an Apache 2.0 suite of 45 AWS tasks, reports Claude Fable 5.1 at 60% and GPT-6 Astra at 53%; the team says about 80% of failures are reasoning errors such as stopping early.
Factory announced on Nov 4, 2026 that HumanLayer and its founders Dex Horthy and Kyle Mistele are joining. HumanLayer core is Apache 2.0 open source; hosted users get a transition period. Terms undisclosed.
Agent Autoloop is an MIT-licensed Node hub that triggers Claude Code or Codex from events, runs one round per event, and asks you for approval by card; most integrations are only partly tested.
Anthropic confirms that Opus 4.6 and Sonnet 4.6 may not carry the text watermark when temperature, top_p or top_k is non-default. Four more older models get marks from Nov 16.
LoCo trains a reasoning model with DPO on state-conditioned pairs to stop overthinking, cutting completion length 41.7-45.4% on math and 23.4-36.3% on code; authors released LoCo-7B and MIT code.
MoviEdit (Adobe, KAIST) edits multi-shot clips up to 25 seconds from text or an edited frame, using 87K auto-built training pairs; the MoviEdit-Bench benchmark is public, but no weights were linked.
llama.cpp v0.9.0 ships with ggml 0.29.0: native FP8 E4M3, a flash-attention rows op, batched copies and NCCL multi-node tensor parallelism, plus kernel work on CUDA, Metal, Vulkan, SYCL, OpenCL and Hexagon.
Google's help pages say Gemini chats are used for training and human review when Keep Activity is on, including for personal subscribers; turning it off keeps chats for 72 hours.
Anthropic says Claude helped its biology team solve the first 3D structure of a prokaryotic vault: 96 copies of one phage protein. Purification fell from days to 90 minutes. Function is still unknown.
webAI’s TwIL-8B is an 8.8B open-weight model from Granite 4.2. Its card reports MATH-500 98.2 and LiveCodeBench 81.6 but trails on GPQA and HLE. Scores are vendor-run; license is non-commercial.
Together AI says its new Together Compiler uses agents and a persistent engine to re-optimize inference continuously, reporting 1.3-3x higher TPS on open models; the figures are the vendor's own.
Dharma-AI says its 2B Field Extractor scores 0.834 on its own 1,638-case benchmark, ahead of Claude Opus 5 at 0.816, and costs 6 to 147 times less; the benchmark and cost model are the company's own.
MoneyPrinterTurbo is an MIT-licensed tool that makes short videos from a topic via LLMs, stock footage and TTS; it has about 138,500 stars, and users bear API costs, key handling and platform-rule risks.
PR-Agent, an MIT-licensed AI pull request reviewer, is now community-owned after Qodo donated it; it runs as a GitHub Action or CLI with any LiteLLM model, and /help_docs is disabled over a credential issue.
Hydrogen-1.0 is a science model post-trained from GLM-5.3 with open weights. Its wins over GPT-6-Astra come mostly from Deep Principle's own tests, so treat them as claims until others reproduce them.
Snowflake Cortex Training, in preview since Oct 30, 2026, runs SFT and RL for open-weight LLMs on Snowflake GPUs, billed per second in AI Credits, AWS regions only; model list and rates are not in the release note.
SenseNova-Loop-9B-Preview is an Apache 2.0 multimodal model that loops a shared 24-layer block twice, giving 56 layer executions from 32 stored layers; its technical report is still to come.
LiberAI reports that Liber-0 succeeded on 36% of 600 trials across 60 unseen tasks, without task-specific robot demos; the benchmark, the tasks and the comparison to GPT-6 Astra are the company's own.
Perceptron Egocentric 1.0 is a self-service API for annotating egocentric video; Perceptron reports 0.787 on its own EgoHome benchmark against 0.797 for GPT-6 Astra at $9 versus about $76 per video-hour.
Uber’s Q3 2026 call set targets of about 700 driverless vehicles by year-end and 2,000-plus by end of 2027. The CFO said “10 thousand+ in 2028”; some summaries say tens of thousands. The financing is meant to be third-party.
