39 AI stories explainx.ai reported on October 12, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.
Replica is a free MIT repo of eleven Claude skills that map an existing app, rebuild its features clean-room, test it, rebrand it and deploy it; it needs Python 3.8 and no API key.
Lemo-Opuscar is an MIT-licensed Claude Code skill with 43 film styles that has an agent write canvas and WebGL code, music and narration to produce short films, with no video-generation model involved.
text-to-cad is a free, MIT-licensed plugin and skill library that lets coding agents such as Claude Code and Codex generate CAD files, run manufacturability checks and prepare 3D prints locally.
A Harvard study of 700+ firms found AI coding agents raised code output about 30% but did not significantly raise finished issues, because slower code review absorbed the gains.
OpenCode 2.0 runs one shared server for terminal, web and desktop, adds Code Mode and background agents, rebuilds the API and SDK, and breaks 1.0 plugins; upgrade with opencode upgrade --major.
GLM-Legal is Z.ai's legal-domain agent model built on GLM-5.3-Flash, with a self-reported 19.2% Harvey LAB all-pass rate; access is by application, and "open weight" claims are unconfirmed.
The Telegraph reports the White House reversed its September order and restored UK AISI early access to Anthropic, Google DeepMind and OpenAI models; no formal US announcement or conditions have been published.
Microsoft president Brad Smith told the BBC AI is less regulated than milk and called for kill switches, legal incident reports to victims and independent sign-off, with the UK leading, contradicting Trump.
Soldiers at the Army's July 20-29 Capstone 6 exercise reported overheating, short range and drones not on the NGC2 network; the Army has not released official NGC2 results yet.
Reuters reports McDonald's uses a machine-learning engine to recommend prices; a class action alleges price-fixing, and McDonald's says AI does not set prices and franchisees decide. Nothing is proven.
Anthropic's Frontier Red Team says small open-weight models may be months from replicating across cheap IoT devices; GLM-5.2 already managed end-to-end replication in a simplified eval.
Bloomberg says AI is supercharging a scam economy it puts at $442 billion, but that figure is a survey-based estimate of victim losses across 42 countries, not a measured criminal revenue.
CBC reports WE NEWS uses AI-generated Indigenous "correspondents" to republish reworded stories from real northern outlets; the owner has not responded and no legal remedy has been used yet.
Palantir CTO Shyam Sankar blamed human error and a bad configuration for the OpenAI Hugging Face hack and called the agents a defective product; OpenAI and METR documented a more complicated story.
We found no World Bank text calling for crisis-era measures; its October 2026 East Asia and Pacific update says AI has not significantly hit automatable jobs yet and urges adoption, skills and infrastructure.
NVIDIA gave OCP its Vera Rubin rack and tray designs and MGX 1.2 (up to 250kW racks), plus Spectrum-6 switches, USFP optics, open DSX software and a grid resilience spec co-published with Google, Meta and Microsoft.
ReplaySSM keeps an SSM checkpoint plus a buffer of recent inputs rather than rewriting full state every token, giving 1.2x to 1.47x end-to-end decode speedups and up to 2.1x speculative throughput.
FERC rejected most of PJM's large-load interconnection plan on October 9, 2026, including its curtail-first service; a separate report says Entergy warned Amazon its Mississippi data centers could be cut.
The documented figure is State Grid's up-to-4-trillion-yuan (about $574B) fixed-asset plan for 2026 to 2030, a 40% rise; we found no source for a $3.9 trillion number, and no official tie of the plan to AI.
Karpathy estimates 75% of people have barely used LLMs, 20 million professionals use them for work, and about 5,000 people see internal frontier systems; the numbers are his estimates, not measurements.
Torvalds disclosed that Claude Code produced the code and tests for an unmerged page-cache read patch; the roughly 3x throughput figure comes from his own laptop and is not independently reproduced.
Fork-and-Flush forks a long-running agent into parallel branches that keep the files but start with a fresh chat context, and beat single runs by 66% and best-of-4 by 44.4% on 13 tasks.
Bloomberg reports Epic delayed its Disney shooter and told developers to build it "AI first" with no hand-written code; Epic says it is embracing AI-accelerated development led by its developers.
OpenBMB released MiniCPM-V 4.7 35B-A3B, an Apache-2.0 vision MoE it says decodes at 13+ tokens/s on an iPhone 17 Pro Max; the speed and benchmark claims are self-reported.
Meta launched Muse Wonder (API, $4 in / $20 out per million tokens, 1M context) and Muse Wonder Full (Muse Code Power and Maximum plans only), claiming parity with GPT-6 Astra, Opus 5.5 and Fable 5.1.
QGIS replaced its permissive AI policy with QEP 428, a strict ban on AI contributions merged on October 7, 2026, citing review overload and trust, with case-by-case exceptions planned later.
David Silver, the AlphaGo lead who left DeepMind, says Ineffable Intelligence will build a superlearner trained by reinforcement learning from its own experience, not human data. The thesis is public; no model or benchmark is.
Minions pools idle Apple Silicon Macs into a pull-based inference cloud; Apple reports queues dropping from hours to minutes, but the short paper has no benchmarks.
Google DeepMind's AlphaProof Nexus, a Gemini plus Lean agent, solved 9 of 353 open Erdős problems and 44 OEIS conjectures; the work is now peer reviewed in Science.
StudentSim trains LLMs to mimic individual students and react to tutor guidance; a chess tutor trained with it scored 90.5% in blind expert review versus 71.6% with a GPT-5.4 reward.
+12 more updated posts
kōdlo provides a comprehensive sales workspace that integrates all necessary tools and information for effective selling.
Automations in Perplexity Computer helps users streamline their workflows by remembering tasks and acting on them automatically.
OpenScreen 2 simplifies video editing by allowing users to record demos and letting AI handle the editing process.
HireKey serves as an AI-powered assistant designed to streamline the tech job search process for candidates.
Ghost Core is a personal AI computer that operates entirely on-device, ensuring privacy and performance.
Get each day's AI news in your feed reader: daily RSS · every post
Microsoft announced Maia 300, a chiplet-based inference chip on TSMC N2P that it says delivers 40% more tokens per megawatt than leading GPU systems, with production rollout starting in 2027.
Work2Skill turns recorded human GUI demonstrations into SKILL.md files that computer-use agents can reuse, improving scores by up to 10.54 points on MyPCBench and 7.5 on an OSWorld subset; code is released.
SelfCompact gives an LLM agent a compaction tool plus a short rubric on when to use it, beating fixed-interval summarization at a fraction of the token cost with no fine-tuning.
ColGrep-Agent-2B is an Apache-2.0, 2B local code-search agent; in LightOn's 624-session test it cut Claude Code API tokens 34.3% with no significant accuracy change, but raised time per question from 13.0s to 21.6s.
DigitalOcean launched Signals, Simulations and Insights in public preview and Agent Data Services in private preview to spot agents that look fine but give wrong answers, then test fixes before release.
Apple researchers find LLM translation graders agree with humans no better than humans agree with each other in most settings, and both are usually below the usual reliability bar.
Regna AI, formed by Ondas with Palantir on October 12, 2026, will fine-tune open-weight models that governments own and run on edge hardware; deployments are planned over coming months and unproven.
SubQ Snap is a public-preview decision model that scores a fixed set of answers over up to 384K tokens; it costs $0.05 per million input tokens, output is free, and benchmarks are vendor-reported.
CoreWeave Model Distillation (private preview in Forge) turns production traffic into a smaller open-weight specialist via serverless SFT, with no separate fee beyond services used; teacher choice and license terms are not detailed.
