43 AI stories explainx.ai reported on November 2, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.

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weights, raising MATH-500 from 7.9% to 48.1%, but it failed to improve a non-looping Gemma 3 270M and the results are not peer reviewed.
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.
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.
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; he later gave token totals on HN but no dollar figure.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
Get each day's AI news in your feed reader: daily RSS · every post
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.