AI News Today

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

NVIDIA Auto Ontology is an open-source Apache-2.0 library that builds a business-meaning layer over your databases and serves it to agents over MCP, so text-to-SQL answers use certified definitions.

OpenAI built Pando, a Git-compatible backend now serving its whole monorepo; it says median push latency fell nearly 5x and it delivers about a petabyte of Git data a day.

One user logged Postman auto-installing Codex and Claude Code plugins; Postman's own plugin README confirms it approves its Codex startup hook during install.

The LSA paper reports 72.4x prefill speedup on Qwen3.8-27B versus full attention and 22.7x on GLM-5.3-Flash versus its own sparse attention, for one attention layer at 10M tokens; code is Apache 2.0.

A Railway engineer admitted his AI testing agents went off the rails on rival host Primcloud; the 24-command, secrets-request details are Primcloud's own account, and Primcloud says nothing was accessed.

A real Qwen Code commit from October 28 sets 1M tokens as the default for any qwen4 model id; the leaked variant names are test strings, and the launch date and omni-120b-think model remain unconfirmed.

Ohio voters in about 18 communities decide roughly 20 local data center measures on November 3, 2026, including charter bans, size limits and mayor recalls; the exact count differs by outlet.

AnyModelBot is an MIT-licensed local-first agent app with per-bot permissions and budgets, but its models are reached through the paid Modelflare gateway, so prompts and files leave your device.

GitBook says an attacker using AI agents got infrastructure credentials on Nov 2, may have seen tokens, Git Sync secrets and Assistant questions, and it revoked all API tokens.

China NDRC official Wang Ruomeng said on November 3, 2026 that Beijing will speed AI laws and lifecycle safety standards, with incident disclosure and sandboxes; no final text or dates were reported.

Pruna released P-H3-Image, an open 8-step, 8-megapixel image and editing model derived from MiniMax-H3, but its license excludes the US, EU, UK and South Korea without a separate MiniMax license.

Gartner says open-weight models will run 90% of enterprise agent steps by 2030, costing a quarter to a twentieth of closed models. Revolut says self-hosting GLM cut costs 90%. Both are claims, not independent tests.

Anthropic released FlashProve, an Apache 2.0 zkSNARK prototype that proves a response came from committed weights: 110 tokens/s on one H100 for Llama 4 Maverick, 19 MB proofs, about 70-95x the GPU time of vLLM.

MiniMax-AI published BYOB Table Protocol v0.1, a draft spec plus TypeScript code where each owner's own agent joins a shared ledger, and cross-owner actions need that owner's signed approval.

The FT reports Nvidia told cloud clients to cut China exposure; Bloomberg relays a 51% non-Chinese floor. Reuters could not verify it, and no Nvidia statement or written rule has been published.

Message Threads is a public beta Messages API feature: send the conversation once, then only new messages; Anthropic reports about 35% lower median time to first token at 500k to 1M tokens, with no price or token change.

DeepMind says Antigravity with new Cheminformatics Skills can curate data, build potency models and propose macrocycles with synthesis routes in about an hour. Every design still needs lab synthesis and testing.

Ai2 says Gemini 3.8 Flash, Claude Opus 5.5 and GPT-6 Sol and Luna fall to 13% or lower on the hardest state-tracking level, while Gated DeltaNet models trained on it generalize 250x past training length.

The Information reports Huawei asked some customers to secure HBM for the Ascend 960DT themselves while smugglers ship memory bundled with cheap CPUs; the claims rest on anonymous sources and are not independently confirmed.

OpenAI says ChatGPT has 1.2 billion weekly users and search-like volume near 20% of Google. The numbers come from one interview, not a published methodology, so treat them as company claims.

A study of 50 published LLM findings retested on 20 models found 44% expired, 34% weakened, 16% persist and 6% strengthened by September 2026; sycophancy and judge bias lasted, fake citations and counting faded.

OpenAI says six GPT models respect warnings less often in a safety eval than in matched Codex production cases, and a wrong guess about what the test measures can cut success from 63% to 18%.

Sarvam Prism is an Apache 2.0 decision model on Sarvam 105B that returns a probability for each option in one pass; Sarvam says it scores 71.8 at 44 ms, on its own tests.

OpenAI says GPT-6 Astra designed a Codex hardware device without direct human edits, hardware token use is up 23x since February, and a newer internal model scores 0.875 on BenchCAD 2.0 versus 0.594 for Astra.

Google says attackers steal cloud credentials, switch on LLM services and resell the capacity through proxy gateways, leaving victims with bills above $100,000 a day and up to $1 million.

BLA is an MIT-licensed npm tool that hardens shipped builds against the REA reverse-engineering agent; its byte-level failures are measured, but its 612x token and cost figures are internal estimates.

Itch.io says it deleted a couple terabytes of AI-made spam, extended mandatory AI disclosure to more categories, and added a No AI filter to search, because new pages doubled while sales and downloads did not.

OpenAI raised critical bug bounties to $30,000 (up to $172,000 with bonuses), added quality, loyalty and campaign multipliers, and named Cathedral as partner for its next AI-driven internal security sprint.

Octomind is an Apache-2.0 Rust coding agent runtime with TOML guardrails and MCP support. Its 24 of 25 benchmark is self-published and we did not reproduce it, so treat it as a vendor claim.

NVIDIA open-sourced Attribution, a service plus reference Blueprint that correlates logs, metrics and scheduler state to name the root cause and node of a failed distributed AI workload.

Oman signed a framework deal with UK developer Yamna for a Salalah AI data center whose first phase tops about $1.3B, but no capacity, date or customer is stated.

NVIDIA released nvperf-core, Apache-2.0 C headers plus proprietary prebuilt libraries for the NvPerf Core API behind its still-coming Nsight Perf SDK for AI, an always-on GPU profiling layer.

Worktrunk is an MIT or Apache-2.0 CLI that wraps Git worktrees in three short commands, so each AI coding agent can work in its own folder and merge back cleanly.

WSJ reports investors want Anthropic’s IPO valued at $1.5T to $1.6T, below about $2T. Anthropic has not confirmed a range; the amended S-1 leaves price and shares blank.

CWE-bench v2 opens 100 defensive patch tasks over 91 CWEs; Collinear reports Gemini 4 Argon at 67.8% pass@1 and Claude Opus 5.5 at 67.3% among 11 models, with deterministic grading.

Advocate General Szpunar says chatbot answers can infringe press publishers' rights when they reuse articles in a recognisable form, while training on text is freely allowed; the opinion is not binding.

Users report OpenAI's training toggle switching back on, but the reports are anecdotes and OpenAI has not responded. Its help page confirms defaults and a separate Codex setting you must check yourself.

ai-goofish-monitor is an MIT-licensed Docker app that scrapes Xianyu with Playwright and judges listings with a multimodal model, but it can break the marketplace terms and sends listing images to your model provider.

Reports on Nov 3 put Anthropic at over 3.5 GW by end-2026, below the roughly 5 GW cited in September; no source explains the gap, so calling it a cut is not verified.

Google’s head of data center energy reportedly said orbital data centers could cost up to 100x more than Earth ones; Suncatcher lead Travis Beals separately said no cost win within five years.

Google says Gemini and related AI will be off for K-12 student accounts until administrators opt in, and Guided Learning arrives in Classroom now. The Search trade-off for schools that block AI is reported, not confirmed by Google.

Agent37 claims a $4.76 monthly agent sandbox against $249 at Vercel, but the comparison is its own list-price math, excludes model costs and credits, and is not a performance benchmark.

NVIDIA says SETI agents can steer the Allen Telescope Array and cut weeks of work to hours, with a first real-time AI pulsar detection. The claims come from NVIDIA and the team, not an independent test.

LLM 0.37 adds GPT-6.1 Sol and Terra, commands to register OpenAI-compatible endpoints, typed embedding attachments and a cheaper GPT-6 Luna default, and removes retired OpenAI models.

Menlo Ventures estimates enterprise AI spend at $176.8B and Anthropic at 55% of LLM API spend versus Google at 9%. The report is a survey-based estimate from an Anthropic investor, not audited revenue.

CrowdStrike announced two AI-trust acquisitions and a live AI Kill Switch on November 3, 2026; the combined $495M price is a press report, not a CrowdStrike figure.

Salesforce signed a definitive agreement to buy voice-agent maker Cresta; the price is undisclosed ($2.7B is a press report), and the deal is not expected to close until early 2027.

Apple will rate any app with dating or romantic AI companion features 18+ at minimum. New App Store Connect declarations start November 16, 2026, and all apps must complete them by January 12, 2027.

uni-robot-agent is an open harness where Qwen3.8-Flash-Next scores 31.39 on RoboDojo, above Astra's published 28.97, but Astra with the same harness reaches 51.90, so the real finding is that the harness matters.

Apple says any app can now build Siri AI integrations in the iOS 27.3 developer beta using App Intents and new app skills, replacing the narrow schemas of the first Siri AI release.

Apple CEO John Ternus said server-based Siri AI and photo-editing features will carry usage limits and that paid expanded access is coming. No prices, limits or dates have been published.

Binance says the Oct. 29 theft of about $6.3 billion began with a tampered AI coding tool install and fake job interviews, then moved through internal AI tools; the FBI link is reported, not independently shown.

IBM Research says its 20B open-weight Palette model places second on PresentBench (59.7 vs NotebookLM 60.8) using a designer/coder split and a harness that keeps a repair only if it helps.

A Chilean campaign had Quilicura residents answer 25,000 prompts by hand to spotlight AI data center water use. Permit figures are documented; the claim that the town runs out of water in 10 years is the campaign's own.

Hugging Face proposes Model Integration Contracts: small CPU-testable model fixtures with reviewed baselines. A UMT5 fixture flagged a bug that lived in Transformers main from July 30 to September 28, 2026.

Google launched Dreambeans in India on November 2, 2026 with 13 Indian languages plus English; it builds AI stories from your Google data, but US downloads reportedly fell 84% in October.

Drax, Brookfield and Radiant plan up to 100MW of AI data centre capacity at Drax Power Station from Q1 2028, with a possible 500MW by 2031, but planning, approvals and a tenant are still missing.

Apple Watch Live Rewind and Siri Recaps are now in developer beta in iOS 27.3; snippets vanish after 30 seconds and unsaved recaps after seven days, with audio handled in a secure enclave.

The UK NCSC says AI is highly likely to make ransomware, social engineering and vulnerability exploitation faster and larger within two years, and tells critical infrastructure to prepare for AI-powered malware by 2028.

OpenAI hired former Army Secretary Dan Driscoll as senior counselor for national security modernization, a new role aimed at US and allied government AI use; no contract details were disclosed.

MIT CSAIL researchers report GPT-6 Astra removed a looping direction from 0.016% of Qwen3.5-0.8B weights, raising MATH-500 from 7.9% to 48.1%, but it failed to improve a non-looping Gemma 3 270M and the results are not peer reviewed.

Karpathy says Claude Opus 5.5 and Claude Code built HackerTube, a no-login site of 6,454 HN-recommended videos, in about two hours; his post gives no token or dollar cost.

Musk said on X the cost of serving Grok Bot was unsustainably high relative to revenue; no figures, new limits or timeline were given, and xAI publishes no weekly allowance.

SemiAnalysis launched GDPx, a 206-task agentic benchmark over 11 domains; Claude Opus 5.5 leads at 12.0%, and models broke behavioral rules in 196 of 200 cyber alignment runs.

Nous Research has an official Hermes Mobile beta page for iOS and Android; GitHub issues say both invites are full or broken. A post-Thanksgiving full release is only reported, not confirmed.

A CEO memo seen by The Register says SUSE will cut about 10% of its 1,900 staff, citing automation and AI. Affected staff hear by Nov 4. Which roles and countries are unnamed; this is one outlet.

OpenAI is merging ChatGPT desktop Chat and Work into one experience, rolling out first to select users, while Codex stays separate and shares one token allowance with ChatGPT.

OpenAI published LifeVal, a benchmark of everyday life tasks built from time-use surveys and a 3,995-person poll; it shows hand-picked example scores of 90 to 100 percent but no overall number.

Goldman Sachs reportedly limited advanced AI coding tools such as Claude Code to engineers, citing productivity and risk management; Goldman says the experimentation phase is over.

SpaceXSI docs say Grok API text output carries an invisible watermark from November 7, 2026, on by default for seven models, with a per-model team opt-out and an x-watermark response header.

Proton published its CTO's memo telling engineers hand-coding is no longer the core skill, that they now lead models and agents, and that the whole development lifecycle must change.

ADNOC says it is the first energy company to deploy OpenAI Frontier, with G42, starting with field master planning and crude valuation agents; no price or timeline was disclosed.

Anthropic’s Nov 2 S-1/A says future models may imminently meet its RSP threshold for dramatically accelerated AI-attributable progress; no current model is designated as over it. Wording changes and quotes from posts are unverified.

Meituan launched LongCat-2.5, a natively multimodal agent model with 1.6T total and about 48B active parameters, a 1M-token context and an OpenAI and Anthropic-compatible API; benchmark claims are unverified.

OpenAI launched Grid Protectors on Nov 2, 2026: future campuses will cut power draw during grid stress, starting with up to 1,000 MW of demand response in Georgia; other sites are still unspecified.

The New York Times and other publishers sued Brave on Nov 2, 2026, alleging it scraped paywalled articles, sold them to AI firms and let users bypass paywalls; the claims are unproven and Brave had not replied.

NVIDIA open-sourced Recovery, a Slurm-first service that detects crashed, hung and slow training jobs, excludes bad nodes and restarts from checkpoints. Monitor mode is the default. The repo calls it experimental.

Anthropic’s amended S-1 reports Q3 revenue near $17B, a $23.29B net loss and about $895.4B in purchase commitments, mostly compute; the huge losses are driven largely by fair-value accounting.

Per Business Insider, Anthropic’s IPO filing says internal Claude use, valued at customer prices, passed monthly employee pay in April and grew over 250-fold from January to September; it is not cash spent.

Mitsubishi Power announced a $490M Savannah plant expansion on Oct 23, 2026, and MHI reports a 35 GW turbine backlog and plans to double capacity; the latest Nikkei report on more US output is second-hand.

Edge0 released Audio8-TTS-Realistic-Air, a 0.6B Apache-2.0 TTS model with zero-shot voice cloning in 23 languages. Scores are self-reported and the card says realism is not measured by them.

A November 2 post says OpenRouter hit a 10 quadrillion annual token run rate; OpenRouter has not published that total, public rankings are consistent with its order of magnitude, and the two growth claims do not match exactly.

Aramco and PIF announced on Nov 1, 2026 that Aramco acquired 49% of Humain for about SAR 9.9B ($2.6B) in 2026 cash plus AI assets; PIF keeps 51%, and Humain targets 3 GW of data centers by 2030.

DatologyAI's SimpleLoRIF embeds a million training examples by gradient and removes those closest to unsafe responses. SafeTulu cut GCG attack success by up to 22% relative on three small models, with small capability changes.

Apple researchers split a UI agent into a planner and a grounding model and train an 8B version with online PPO on hundreds of virtual iPhones; they report it beats Gemini-3.5-Flash on real iOS apps.

Apriel Voice 27B Preview is an MIT-licensed open model from ServiceNow-AI that understands speech and returns text and tool calls; it trails GPT Realtime 2.1 on some voice scores and does not speak back.

TechCrunch, relaying Axios, reports OpenAI is considering calling data centers "AI compute campuses" in a post-midterm PR push; OpenAI has not published the plan itself.

humans& open-sourced a lossless delta-compression pipeline for RL weight sync; on its own 445 GB benchmark it reports a 219.6 ms sampler pause and under 4 s end to end.

GPT-Rosalind-6 reuses GPT-6 Astra weights with no new training. OpenAI rates it High in biology and Critical in cyber, and limits access to vetted life science customers.

Microsoft Persona Workshop is an MIT-licensed Rust service that builds communication and working-style profiles from selected evidence and exports AGENTS.md, SOUL.md or .twin.json. It warns that blind tests are not fidelity scores.

Reuters reports Lee Jae Myung vowed $15.97B for AI, chips, physical AI and data centers in South Korea's 2027 budget, nearly double 2026. What the budget funds, what is confirmed, and why it matters.

Zyphra Hierarchical Attention builds a tree of compressed key summaries and routes queries through it. After 4B tokens of distillation, a Qwen3.6-35B-A3B conversion matched dense on most tests. Large-context speedups come from a roofline model, not measurements.

OpenAI says it runs activation probes on all assistant completions since GPT-5.6, scoring each token and halting risky output, with better jailbreak robustness than text classifiers but a need to retrain per base model.

Eurogamer reports that AI-generated fake wikis for indie games are crowding out volunteer fan wikis in search, while AI crawlers drive up hosting costs for the real ones.

Hume launched an audio-native API that returns probabilities for questions you define, such as whether a caller sounds frustrated. Its 70.7% human-agreement figure is self-reported; access is by waitlist.

Microsoft released MIT-licensed ACP adapters that connect any ACP-compatible agent to Slack, Teams and Minecraft. Slack and Minecraft are usable; the Teams README says live tenant validation is still pending.

The Census Bureau changed its business AI-adoption question to say "Super Intelligence (SI)" from the October 19, 2026 collection, which may break comparability with earlier AI adoption figures.

Canada will invest up to C$1.6B in Cohere in stages, tied to 1,000 technical jobs and C$1.7B of Canadian compute by 2030, as part of a Series E of up to C$7B at roughly C$36B post-money.

Meta's open-source Text Seal repo now includes TextSeal2, a two-key watermark that leaves the model's output distribution unchanged and still works with speculative decoding; code and paper scripts are public.

Grok Bot changelog entries list shopping product cards, Shop Pay in Vault and subscription templates, and docs describe Shared Bots, but Vault logins and cards are not shared with them; use Secrets instead.

BAAJA, a bipartisan Senate permitting bill, would make data centers above 20 MW pay grid costs they cause; we found no source for a $50 billion figure, and the Senate vote is expected only after the midterms.

Fujifilm will spend about ¥10 billion to raise LTO tape shipment capacity roughly 5x versus fiscal 2025 by fiscal 2028, citing generative AI and cloud archives; the 5x starts from a weak 2025 base.

fynk says adopting AG-UI and CopilotKit removed about 11,000 lines of its own chat code, and its 62,517-line Vue port was merged into CopilotKit on May 13, 2026; these are the company’s own figures.

AWS shows a Strands multi-agent assistant on AgentCore: Haiku routes, Sonnet runs tools, Opus synthesizes, Nova embeds charts and photos. Deterministic code scores fits. Tested on synthetic data only.

AWS distilled GLM-5.3 into a Qwen3.5-4B shopping agent with sequence-level distillation: 0.541 success versus 0.540 for the teacher. One run, vendor-authored, no cost figures published.

AWS published a sample that runs Crawl4AI on a scheduled Fargate task to write Markdown plus metadata sidecars to S3 for Bedrock Knowledge Bases; use the built-in Web Crawler first, and treat crawled pages as untrusted.

Floorplan is an MIT plugin for Claude Code and Codex that builds a zoomable plain-English map of your app, plus a Health check for leaked keys and open databases. It passed 2,000 stars within days.

NVIDIA published Preview AICR recipes for Kubernetes on Vera Rubin NVL72 that passed deployment, conformance and a 744.71 GB/s NCCL all-reduce gate; inference performance and MIG are not yet validated.

NVIDIA published a Cosmos deep dive for interactive simulation: AV, robot and surgical checkpoints, a VRAM table, RL recipes and a NIM launch walkthrough. Speedups are vendor-reported.

Tencent’s WeFair trains a small query rewriter in three stages and finds that strict exact-match rewards in the last stage cut intent drift; a 10-day A/B test showed +2.593% effective view rate in WeChat search.