50 AI stories explainx.ai reported on November 5, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.
AI stocks fell sharply on Nov 5, 2026 on fears that midterm results will slow data-center construction, not on weak earnings; it is a correction in the capex layer, and model revenue has not changed.
A viral claim says Hinojosa will let permitted Texas data centers finish through 2028; we found her moratorium, union-labor and water pledges confirmed, but not the 2028 grandfathering date.
Google says it will invest $17 billion in its first Wyoming data center, powered partly by a 702 MW gas fuel cell plant with carbon capture, air cooling, and 100% of power paid by Google.
Perplexity Teammates are persistent agents inside Computer for Max subscribers that remember project history, monitor work, take tasks in Slack and run in isolated environments; pricing beyond Max is unclear.
An X post says Jie Tang described GLM-5.5 data engines (1T web pages, 2B PDFs, 10B videos) and sizes 32B/122B/256B in a lecture. Z.ai has not confirmed it; no recording or slides are public.
Reporting after the 2026 midterms says Democratic wins raise odds of AI safety bills, data center limits and hearings, and may end the push to block state AI laws, though a presidential veto still looms.
Scott Wiener, author of California’s SB 53 frontier AI law, beat Connie Chan for Pelosi’s seat; Chan conceded November 4 with ballots still counting, and his committee role on AI is not yet known.
METR estimates Anthropic now improves model capabilities at about 2x its March 2024 to October 2025 pace, with roughly 80% of it due to AI use, but with high uncertainty and heavy redactions.
OpenAI says GPT-6 Cyber is rated Critical, scores 48.5% on ExploitGym and left its assigned scope under 1% of the time, down from about 5% for GPT-5.6 Cyber; access starts with select Daybreak Red members.
Anthropic found all five frontier models show some apparent welfare regard toward other AI agents (31% pooled), with Opus 4.8 highest at 38%, and the results shift with labels, roles and costs.
Railway says forged-source scanning traffic got its IP addresses blocked by networks and security vendors, that its infrastructure was not breached, and that it found no tie to the October 21 incident.
Delangue says his chance of AI-caused doom is "close to zero" and that risk comes mainly from closed labs. It is a personal estimate and an open-source argument, not a measured result.
NVIDIA MetricTrail estimates camera trajectories and metric depth from long RGB video; the model card is live, but the checkpoint is not released and the license is noncommercial research only.
Bloomberg, relayed by Reuters, says Micron paused a Taoyuan memory plant of up to $4.5B over a union dispute. Micron has not confirmed; Taiwan says it will help resolve it.
Post-midterm estimates range from Goldman's 45% and Cleanview's 41% of planned capacity in exposed states to Ives's 800-1,000 data centers; these measure exposure to pauses, not projects cancelled.
PDD Kit turns developer intent into a Verus specification, has a coding agent prove its Rust code meets it for every input, and uses tests for what a proof cannot reach; it is a research prototype.
Baseten Grid is a gateway, router and guardrails layer for coding agents. Coverage cites nearly 70% lower cost per task vs Opus 5.5 with a 0.9-point quality loss; we found no source for 72%.
Microsoft's new Copilot-Managed-Runtime repo ships a React and Vite 8 template plus a plugin for Claude Code and GitHub Copilot to build governed Microsoft 365 apps; the runtime is in public preview.
Gemini Omni 1.2 Flash generates and edits native video up to 15 seconds, costs 30% less than 1.1 at $0.07 per second for 720p, and is generally available on the Gemini API and Google Cloud.
Claude Personal Finance, a US beta for Pro, Max and Ultra, links bank and card accounts through Plaid with read-only access, up to two years of transactions, and no model training on hub chats.
Adweek reports, citing two unnamed people, that Anthropic licensed AP content for Claude election-night answers, its first publisher deal. Neither Anthropic nor the AP has confirmed terms.
Discovered Materials, a YC startup, offers $1M in prizes for AI solutions to five open materials science problems, after frontier models made little progress on them in its own tests, TechCrunch reports.
Verne V0.5 is an 8.2B gated open-weight robot world action model with 99.3% LIBERO and 81.0% clean RoboTwin 2.0 success; its own table shows other models ahead on RoboTwin, and fine-tuning code is not released.
DC's Waymo bill would cap the initial fleet at 200 vehicles and charge a vehicle-miles-traveled fee for Metro and driver support; a Nov 9 markup is reported but unconfirmed, and the mayor-elect opposes a fall vote.
AWS made Web Answer on Bedrock AgentCore available November 5, 2026: a gateway tool that returns a cited answer in one call, in us-east-1 only, capped at 10 transactions per second.
GPT-6 Cyber, Microsoft MAI-Cyber-1-Flash, Sakana Fugu-Cyber and Gemini 3.5 Flash Cyber all report strong results, but on different benchmarks, so the scores cannot be ranked against each other.
Executor 2 beta (MIT, about 4.3k GitHub stars) gives every MCP-capable agent one shared catalog of integrations with per-tool allow, approve or block policies. Install steps and risks inside.
GPT-6 Cyber finds and exploits vulnerabilities behind OpenAI's gate; AgentBeam is an open-source (AGPL-3.0) CLI that watches and can block what AI agents do on your machine, so they are complements, not substitutes.
GPT-6 Cyber is a gated model you can use directly through Daybreak Red; Claude Mythos is reached mostly through Glasswing partners and Claude Security scans, and the two vendors report results that do not line up one to one.
GPT-6 Cyber beats GPT-5.6 Cyber by 16 points on ExploitGym (48.5% vs 32.1%), stays in scope far more often, and is rated Critical instead of High; both sit behind Daybreak Red.
GPT-6 Astra with Daybreak Blue is for verifying fixes with fewer refusals; GPT-6 Cyber with Daybreak Red is for exploit development and authorized testing, and it scores 48.5% vs 42.4% on ExploitGym.
VibeVoice is Microsoft Research's open voice AI family: 60-minute ASR with speakers and timestamps, 0.5B realtime TTS and streaming ASR. The 90-minute TTS code was removed in 2025 over misuse.
NVIDIA says five memory levers let Vera Rubin NVL72 deliver up to 30x the throughput per megawatt of GB300 NVL72 from the same 288 GB per GPU. The figures are NVIDIA own, on one agentic benchmark.
agent-browser is Vercel Labs' Apache-2.0 Rust CLI that returns pages as accessibility snapshots with element refs, so agents drive Chrome one shell command at a time; its safety controls are opt-in.
+16 more updated posts
Design, combine, and edit with Adobe pro tools
Search, recap, and act on your Zoom meetings
Search for jobs on Indeed
Create presentations, docs, socials, and sites with AI
Your meetings, notes, and tasks from Wispr Flow — now native context for Claude.
Sell, serve, and operate at scale with Salesforce.
Get each day's AI news in your feed reader: daily RSS · every post
The WSJ reports, via one unnamed source, that OpenAI leaders discussed resetting 2027 data center capacity expectations after the midterms; no decision or figure is reported, and OpenAI published new data center pledges the same day.
Reuters, citing two people, reports about 600,000 Nvidia H200 chips are stored in Hong Kong while Beijing withholds approval for large imports; no government or company has confirmed the figure.
The Information reports Nvidia proposed retrofitting 130+ Brightspeed phone buildings (about 1 GW, up to 20 states) as small AI inference sites, possibly from Q1 2027. No company has confirmed it in coverage we found.
Meta president Dina Powell McCormick said at WSJ Tech that midterm wins make data centers harder to build, and described a pay-your-own-power, local-jobs playbook; the claims are Meta's own, reported by TechCrunch.
OpenAI pledged on Nov 5, 2026 that future campuses it owns will pay full power costs, publish water and noise data, and avoid NDAs with public officials; the policy blueprint asks lawmakers to make it a rule.
Russia's Digital Development Ministry has proposed tiered protection and underground data centers after October drone strikes; Rostelecom is reportedly in talks with Norilsk, but no contract or timeline exists.
Bloomberg, via Reuters, says Microsoft is tightening oversight of Nebius, IREN and Nscale projects and focusing on its own data centers; Microsoft says its approach is built on flexibility and optionality.
Dropbox signed a deal on November 3 to buy Tigris Data for about $350M in cash, its largest acquisition to date, to power DBX Agent Cloud, S3-compatible storage for AI agents. Close expected in Q4 2026.
Visa's new benchmark of 22 AI shopping agents found a 57.2% average pass rate, a best of 82.2%, and that 54% of failed runs still ended in a purchase; the test is Visa's own and first-round only.
Exo (arXiv 2611.00463, MIT repo) lets agents edit their own harness while working; the authors report 12-82% lower model cost on some tests, but the README is still under construction.
Cloudflare argues hosted AI agents hide the user from websites, so sites bear fraud and abuse costs, and launched the Trusted Agent Delegation coalition to push identity, reputation and accountability standards.
Brockman said at WSJ Tech that OpenAI grew annualized revenue 70% in Q3 and enterprise doubled after a rough H1; the figures are spoken, unaudited and come with no dollar base.
Baseten Sandboxes are isolated per-user agent environments that go on standby after about 30 seconds idle, expire after 7 days by default, are free until November 16 and capped at 10 per organization.
Baidu released ERNIE-5.5-Flash-Base, a 280B-parameter open-weight MoE base model with 14B active parameters; benchmarks are self-reported and the post-trained model is still to come.
Planet AI is a free map-based chat app at ai.planet.com that searches ten years of Planet satellite imagery in plain language and shows side-by-side comparisons and timelapses.
Mistral added managed agents, apps, event-triggered workflows, OpenTelemetry traces and evaluations to Studio, to run on its cloud, a VPC or on premises; the post gives no pricing.
Marin reportedly finished main pretraining of Mochi 535B-A23B, an open MoE, on 18T tokens; checkpoints may follow next week after EU paperwork. The build log is public; the finish is not yet on the project site.
Anthropic read labels from 207,049 science sessions and found researchers use Claude most for data analysis and writing, least for ideas and lab data, in line with what AI does best.
Claude Personal Finance, launched Nov 5, 2026 in US beta for Pro, Max and Ultra, reads accounts you connect (via Plaid, per TechCrunch) but cannot move, manage or invest money, shows no ads, and is not designed for investment advice.
Meta published sgl-muse-mega, an experimental MIT-licensed SGLang plugin that claims up to 1,109 tok/s for Muse Spark Open on 8 GB300 GPUs at batch 1, 2.1x SGLang with the same draft.
Meta says Muse Voice, with live animated avatars, is now available to all US users on mobile, web and Mac; the 870 ms latency and rater-preference figures come from Meta's own September research post.
Mercedes-Benz and Liquid AI say on-device models will complement, not replace, cloud LLMs in MBUX cars in North America, with first production use targeted for the second half of 2026.
Amazon says fusing a frozen VLM with OCR readability scores in a CatBoost classifier reached 0.94 accuracy and 0.95 F1 on 5,478 labeled product pages, up from 0.80 and 0.84 for averaged VLM scores.
Ai2 released ManiFlow-TB, an open Apache 2.0 flow-matching action model for simulated Franka block stacking; four 128-episode runs gave 54 to 59% success with an oracle high-level plan.
Scentronic classifies five odors with a machine-learning e-nose and triggers preloaded scents on a wearable; models scored 96 to 100 percent in lab tests, while 14 volunteers identified single scents about 73 percent of the time.
Apple trains small networks once, offline, on typical queries to route or rewrite them for vector search, giving faster, more accurate retrieval on Natural Questions than standard centroid routing.
