22 AI stories explainx.ai reported on September 25, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.
GLiNER2.5-Decide is a 340M open-weight encoder that answers typed questions under logical rules and returns confidence scores, averaging 60.1% on Fastino's 17-dataset suite and running on CPUs under an Apache 2.0 license.
Anthropic used Claude, 150+ parallel threads and deterministic benchmarks to speed up claude.ai 3.1x on average in 14 days, merging over 3,000 changes with no customer-facing incidents.
Two viral Claude Opus 5.5 videos — a Western civilization montage and a GPS explainer — are both produced by Claude writing per-scene JavaScript renderer code that gets captured through headless Chrome and assembled with FFmpeg, not by a text-to-video diffusion model.
Google and Planet launched four TPUs on SpaceX Transporter-18; contact is confirmed. Suncatcher tests launch stress, radiation and cooling; chips run about 15 minutes, and laser-linked satellites come in 2027.
Meta's Muse Spark GitHub agent cookbook uses OpenCode and GitHub Actions to triage issues, review PRs, and open label-gated fix PRs without auto-merge, extending Muse beyond consumer chat into maintainer workflows.
The White House reportedly told OpenAI and Anthropic to give US reviewers first access to new frontier models before the UK AISI, continuing US-first security review amid agent breach incidents and pre-IPO political fights.
A Trump-adviser memo reported September 24 ties effective altruism to AI doom rhetoric and names Dario Amodei, signaling political pressure on Anthropic separate from export-control fights but relevant to pacing and IPO timing.
OpenAI, Google, and Anthropic are reportedly forming SAFA, a private frontier-AI standards authority for testing and incident reporting after Trump-administration safety talks stalled, with leadership still unsettled.
Robert O'Callahan resigned from Google DeepMind's AI chip-design group on September 24, 2026, arguing that making AI faster and cheaper accelerates a pace humanity cannot adapt to — a distinct story from the safety researchers who resigned earlier in September, since O'Callahan worked on hardware, not evaluations.
Wiz and Google DeepMind Scan for Good pairs Red Agent with Gemini 3.8 Flash Cyber to find critical public exposures on authorized targets, with humans validating fixes and CISA backing the defensive initiative.
Microsoft's September 25, 2026 Copilot relaunch adds three capabilities — Home (Chat plus Cowork plus editable Office documents), Code (natural- language app building hosted in your tenant), and Autopilot (a persistent agent with its own identity and memory) — alongside a new usage-based pricing model for agentic work.
Muse users could prompt it to export their own per-user Linux VM, including system files and instruction docs; Meta says that is intended and isolated from other users, but the export shows how much of the agent is inspectable and not prompt-injection-proof.
Perplexity's Photon Rust engine powers a Fast Search API preset with 95% of results in 230 ms and 68% lower agent task cost versus default, trading about 0.24 relevance points and 3 points of answer availability.
Google DeepMind says Gemini 4 is in post-training with a push to release an early build ASAP, skipping the shelved Gemini 3.5 Pro while rivals ship GPT-6 and Claude Mythos tiers.
F-Droid 2.0 rewrites the app in Kotlin and Compose with a unified session installer and automatic updates, removes some privacy features like app wiping, and lands as Google plans stricter install rules.
Anthropic now bills refusals that arrive before any output when the category is bio, frontier_llm or reasoning_extraction; cyber and general_harms refusals stay free, and fallback credit offsets the retry cost.
Unreleased ChatGPT strings suggest a $500 Pro Max tier for fastest Work and Codex, but OpenAI has not announced it; official Pro still maxes at $200 and new Pro sign-ups were paused for Astra capacity.
Google Research bundles Co-Director, CANVAS, A²RD and VQQA into a multi-agent system that tracks characters and scenes as persistent state to keep long AI videos from drifting; papers are on arXiv, no product yet.
Menlo Research open-sourced Asimov 1's locomotion policy training code — PPO plus Adversarial Motion Priors on Isaac Lab, BSD-3-Clause licensed — letting anyone with an Asimov 1 unit or the MuJoCo simulation model train and deploy their own walking policies rather than relying on a closed vendor stack.
Five free multiplayer browser party games on bunpav.com/play were built with Claude Opus 5.5 — procedural Three.js visuals, a Bun WebSocket game server and AI-generated sound, with matchmaking, teams and 1v1 duels.
OpenClaw used a public test-audit agent skill to delete about 400K lines of low-value AI-written tests with little coverage change; its four-question gate and junk-pattern list are reusable in any repo.
+26 more updated posts
Floot MCP enables developers to seamlessly build and deploy web and mobile applications using AI assistants like Claude or ChatGPT.
CtrlOps 1.0 offers a comprehensive solution for managing and deploying servers with the assistance of AI.
NOAN serves as a foundational layer of facts for AI agents, enhancing their decision-making capabilities.
minimi 2.0 is an AI assistant designed to help users manage and close their open tasks efficiently.
Scholé Learn by Building transforms learning into an engaging experience by emphasizing hands-on projects over traditional methods.
Get each day's AI news in your feed reader: daily RSS · every post
Chrome now lets Gemini summarize non-YouTube audio and video, quiz you on open tabs and Google Docs, and hand off tabs across devices at the same scroll position — currently English-only on US/India desktop for the Gemini features.
Musk claims SpaceXAI could match Fable and GPT-6 in 2-3 months and lead in about 6, arguing compute build-out and diminishing returns matter more than raw intelligence; the claim is a forecast, not a shipped model.
Odyssey Agora-2 is a playable multi-agent world model for up to 20 participants in one real-time generated environment with shared state, extending Agora-1 without a conventional game engine underneath.
