49 AI stories explainx.ai reported on November 11, 2026 so far, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.
Alibaba opened Qwen4-Max preview on Token Plan and named four apps for early access; weights, benchmarks, and Flash or Plus tiers are not published yet.
Expo Exact lets coding agents build, run, screenshot, and test the same Contract app on web, macOS, iOS, and Linux through one agent.mjs driver without webview UI bridges.
AMD and Anthropic confirmed gigawatt-scale MI455 Helios plans and MI355X use; the weekend Claude-on-ROCm port story comes from conference remarks and relays, not a standalone engineering paper.
Experts released OpenAI's Yang-Mills Lean code and wrote that it proves existence of pure Yang-Mills; the Clay mass gap is still not settled.
Microsoft says editing fewer than ten dormant neurons can backdoor an agent model without retraining; the model stays clean on benchmarks until a trigger such as a calendar change fires malicious tool or shell output.
Bessent confirmed a planned US-China superintelligence dialogue, shared model-testing practices, and a bilateral alert channel for serious AI incidents ahead of the November APEC Trump-Xi meeting.
The OfS says institutions must keep assessments valid under condition B4, often with secure AI-free exams, while still teaching students to use AI responsibly.
Finland approved Tesla Assisted Driving nationally with stricter speed defaults and a renamed product, while the EU-wide motor vehicle vote remains pending.
China’s anti-corruption rules pushed pharma marketing from hospital visits into physician AI apps such as Alibaba Hydrion and JD Zhiyi — with disclosure fights ahead.
FT-led reporting frames Singapore as a key transit hub for restricted Nvidia flows tied to Megaspeed, raising export-control and diplomacy risks without confirming every dollar figure.
ChinaTalk says US data vendors sold about $500M in RL environments to Chinese labs with weak export rules, letting buyers reuse the same products built for OpenAI and Anthropic.
Sakana AI named Michael Rogers and Tom Moran as senior advisors to guide defense and cybersecurity work as the Tokyo lab recruits engineers for Japan sovereign AI programs.
VulnCheck says most Glasswing findings are still unfixed or duplicate, and curl’s early OSS Scanner triage shows more than half of reports are not vulnerabilities.
Singapore Airlines reports doubling Kris chatbot satisfaction, roughly 10% faster agent work, 3x Flight Recommender usage, and 60%+ monthly Jarvis adoption using OpenAI models behind orchestration and review.
Model Republic linked 400+ sockpuppet accounts to Koios and IntellACT, running a yearlong pro-Nvidia campaign on export controls that ramped up during the Huang spy scandal.
A PLA-linked university on the U.S. Entity List bought two H200 GPUs through a local reseller while also renting H200 server access in other schools’ records.
Lanier NYT opinion says Valley marketing turns data-center towns into sacrifice zones and undermines public trust, amid siting fights explainx.ai tracks globally.
Governance professors told TechCrunch the Huang probe is standard in form but unprecedented in national-security overtones, with securities fraud suits easier than Delaware mismanagement claims.
Claude Cage wraps Claude Code in a local VM so network traffic and browser fingerprints stay isolated from your main machine.
Agents reviewing public lidar imagery led to an INAH note that no registered site sat at the reported coordinates; the find is promising but not yet an official excavation.
OpenAI's Code Summit workshop framed voice, dots, and local-to-cloud Codex handoffs as one workflow, backed by published Cloud environment docs.
A non-coder reports building a 12 km browser open-world RPG in four days by chaining Claude Fable 5.2 and Opus 5.5 with Three.js, Blender and generated assets; the numbers are self-reported.
InfraRSI is an open agent harness that co-optimizes LLM training launch settings and code patches with invariant checks, reporting up to 2.75× speedups across dense and MoE workloads.
Tether's BotPsy is an 0.8B Apache agent model that the authors say beats larger sub-1.2B rivals on tool benchmarks and ships as safetensors and GGUF for on-device use.
Tree-Ising turns LLM weight repairs into Ising/QUBO problems on BF16 trees, beating an agent-written fix on math benchmarks at a fraction of the token cost.
Google Arts used Mahlangu's oral history plus Veo videos from synthetic stills to document her life before 1980, alongside 45 museum-backed exhibitions online.
Deletion Tuning masks under 0.05% of transformer gate weights and matches full fine-tuning on three post-training regimes while forgetting less than LoRA or sparse baselines.
Infrathon shows frontier agents can improve real GPU infra for hours, and that Opus 5.5 beat a human video KV codec on measured compression.
Lila's November 2026 open-endedness framework splits creativity into idea generation, judging interestingness, representation learning, and managing a population of hypotheses, with four new arXiv papers on the first two pillars.
SWAM shows mixing egocentric video and UMI demonstrations preserves language-video alignment while improving robot control, with open code and weights from Shanghai Jiao Tong-led team.
Hugging Face published token pooling-aware training so multi-vector search indexes can shrink about fourfold with smaller nDCG losses than pooling an untrained model.
Anthropic’s S-1/A-2 implies about 85.5% gross margin on customer-serving compute for nine months to September 2026, while model training and free-tier compute sit in R&D and sales expenses.
+51 more updated posts
Get each day's AI news in your feed reader: daily RSS · every post
Herdr keeps agent terminals alive when you detach or lose SSH, shows which pane needs input, and exposes a CLI and socket API so agents can spawn and coordinate with each other.
Kimi K3.1 Pro Highspeed is on Bedrock under global.moonshotai.kimi-k3.1 with OpenAI-compatible APIs, explicit prompt caching, a half-price Flex tier, and an official OpenCode config.
Spell is reported as ByteDance’s Muse-style personal agent for Doubao; only press and trade sources describe timing and features until ByteDance confirms.
Singles Day 2026 showed AI chatbots, token packs, and assistants driving a growing share of purchases while headline sales rose 2.4% in Syntun’s partial window.
Nod is n8n Labs' new local personal agent: its own sandbox, browser, and approval gates, separate from n8n workflows.
WikiTrace stores agent lessons with links to the trajectories that produced them, letting models verify insights before acting; KAIST reports +8.5 points over prior memory methods on average with lower regression.
DeepL Aure continuously adapts translated speech to a speaker's changing tone, pace, and emphasis across eight languages in Voice meetings, in-person mode, and the Voice API.
Helsing's Clarity uses hundreds of military-tuned AI agents to fuse battlefield data, but every strike hand-off still requires a human operator.
The improve skill writes markdown plans for other agents to implement, using your strongest model for audit and vetting and a cheaper model in an isolated worktree for execution.
OpenAI says over 20,000 veterans claimed a free year of ChatGPT Plus through ID.me, with Academy trainings helping transitions and career searches.
dNaViT open-sources a discrete vision tokenizer in LongCat-Next so images use a shared visual vocabulary like language tokens.
Open Source Says No tracks 22 verified open-source repos that ban LLM-generated contributions, with automated policy checks and a community suggest flow — not a global consensus.
Grok Bot can authorize into your Tailscale tailnet via OAuth and reach private MCP servers under your ACLs, instead of exposing internal tools to the internet.
Adobe MAX Sneaks 2026 previews generative experiments that make 3D camera moves, character auditions, audio mixing, and edit cloning feel more interactive.
Bloomberg-sourced reporting says Zhang Yiming targets a Muse-like ByteDance agent soon while warning frontier parity may not arrive before 2030; ByteDance did not answer an on-record request.
HyperFlow H3 Flash is an open 14B image-text-to-video model that outputs short clips with stereo audio using 4–8 step schedules and runs on consumer GPUs via GGUF or INT8.
APEX uses condition shifting so one network handles generation and fake-score estimation in few-step streaming video, improving instruction following while lowering DMD training overhead.
Longtail-SWE adds 202 repo-level coding agent tasks across 13 underrepresented languages with datasets, environments, and a judge-repairer pipeline on Hugging Face.