69 AI stories explainx.ai reported on October 20, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.
Free My TV is an MIT-licensed, Claude Code-built replacement home screen for Google TV that uses on-device ADB to disable stock apps; it is popular but relies on developer mode, an accessibility service and self-reported performance numbers.
OpenFilm is a free MIT-licensed tool in which a coding agent writes a film as HTML pages that you can then edit by hand in a local Studio and export to MP4, GIF or editor timelines.
Ornith-1.5-122B is an MIT-licensed 122B MoE with a 262K context that runs on one 80GB GPU; its 84.5 SWE-bench Verified score is vendor-reported and uses custom harnesses.
earshot is an MIT-licensed CPU toolkit and MCP server that lets coding agents score and describe sound using CLAP and AST models; its README claims 98% strong-match precision and lists limits.
Graphify builds a local tree-sitter knowledge graph of your code and docs that coding agents query instead of grepping. Its README reports benchmark results; the 80 percent token saving is a claim for its paid cloud product.
Spotify says it removed 50 million-plus slop and spam tracks from recommendations in 2026 and labeled thousands of AI Personas, but it did not say how accurate its detection is.
Postmedia, Torstar, The Globe and Mail and The Canadian Press sued Anthropic in Ontario on Oct 6, 2026, alleging unlicensed scraping for Claude; the claims are unproven and Anthropic has not responded.
Bloomberg reports four Z.ai founders worked on a CGTRI co-funded project; MI5 named CGTRI as tied to China's spy agency but not Z.ai, and nothing reported shows the GLM models are compromised.
NYC Mayor Mamdani said the federal government has abdicated on AI law and urged New Yorkers to hold AI companies accountable, pointing to the city's school AI ban and a proposed AI-listing disclosure rule.
The Intercept reports the Army is soliciting AI agents that move between devices like a worm and neutralize targets with minimal human oversight; it is still at the white-paper stage and the Army has not answered questions.
Microsoft reportedly credited 23 October 2026 flaws to "Yuki Chen using GLM from z.ai" before Chen joined Zhipu; how much the model did versus the researcher is unknown, and the figures come from secondary reports.
AISafetyHot-Hub offers 30 curated AI safety papers a day, six agent skills and a keyless MCP endpoint, but the README does not explain how news and summaries are produced or reviewed.
Meta paused Muse sharing after Claroc reported 6,800+ share links discoverable through public DNS records because each hostname contained the chat title; Meta has not confirmed the count.
The WSJ reports a draft Anthropic prospectus showing $16.3B H1 2026 revenue and an $80.7B net loss, about $72B of it non-cash; the draft is not a public filing, so every figure is secondhand.
NVIDIA released an NVFP4 quantization-aware-distilled build of Qwen3.8-Flash-Next (125B total, 6B active) at about 96 GiB, losing under a point on its two reported benchmarks, but it needs community vLLM containers.
NVIDIA Flare is an Apache-2.0 local inference server for SM120/SM121 GPUs like DGX Spark, using a Python build step (Ignition) and a Rust runtime (Helios) tuned for long conversations and coding agents.
The FT reports big-tech AI data centers in Shaoguan are angering some locals; public sources confirm the buildout and its scale, but not the local grievances, which rest on the FT reporting.
Steam players want a filter for AI-generated games during Next Fest; Valve already requires AI disclosure, but a filter would need a defined tag and enforcement, which Valve has not announced.
Exa 3.0 adds a larger embedding model, free Dynamic Highlights and a $4 per 1,000 Lite tier; its first-place claims on eight benchmarks come from Exa itself and publish no competitor scores.
Sensor Tower estimates Meta spends 39% of its ad budget advertising Muse, up from 23%; that is Meta buying ads for Muse, not ads inside Muse, and Meta says Muse carries none.
Battle Drones is a browser RTS built in 12 iterations and 321 commits mostly by Claude Code, with audio-based two-device multiplayer; its author says he reviewed almost none of the code.
Skyfall AI's SKY-001 is a hierarchical world model for CAD and BIM that, on the company's own benchmarks, roughly ties GPT-6 Astra on IoU while running about 50x to 120x faster and 150x to 185x cheaper; waitlist only, no independent tests yet.
mg-styles-15 releases 15 ten-second films that Claude Opus 5.5 wrote as code, with prompts, source and an MIT license; the same prompt gives a different film each time and few soundtracks match exactly.
The Information reports an Anthropic model that better infers unstated user goals, plus Gemini 4 Carbon and a possible OpenAI Terra model this month; the report is paywalled and unconfirmed, with no names, dates, prices or benchmarks.
Electric Plant says Bombadil interprets electrical signals from about 125 plant species and can cut water use 20% or more; the claims come from the founder via TechCrunch with no published data yet.
Radical Numerics says Omnii-1 designed phages given to a patient under compassionate use; the evidence is the company's own posts, with no peer-reviewed patient data or outcome reported yet.
Self-HPO alternates between the agent rewriting its own harness and RL training the model inside it, lifting Qwen3.5-35B-A3B from 57.8% to 64.8% on average across three SWE-bench benchmarks without an external model.
AWS Agent Registry is a managed catalog where agents search approved MCP servers and A2A agents at runtime; search returns connection metadata but never credentials, so auth stays separate.
Reports describe a seawater-cooled, mostly wind-powered underwater data center off Shanghai, but the viral post overstates it: "almost no one is talking about it" is false, the image is unverified, and the water and fines comparison is unsupported.
Clairic returns extracted, computed, estimated and forecast values, while Jev, OpenAI Decisions, Clef, pplx-decider and Microsoft-Decision-1 choose among listed options; they work best together, and every number here is vendor-reported.
A Reddit user says Claude Code found an xmrig cryptominer using about 12 cores while building a temperature app; the account is unverified, but the checks it used, and the safe way to run them, are worth knowing.
Claude Code can debloat a Fire TV over ADB, but old Fire OS 5 blocks disabling Amazon apps, so the biggest win on our 2016 stick was a YouTube-only home screen, not removing bloat.
The Granny Benchmark is a satirical, unverified viral test in which Claude reportedly declined to recite Baby lyrics to save a fictional grandmother; it shows policy adherence under pressure, not a measured benchmark result.
OpenAI added a pair-matched metric and a post-control risk metric to g-mean², plus a lower bound for interventions, to avoid misleading monitorability comparisons between models.
SKY-001 is a design-and-build planner for CAD and BIM, not a robot controller or video simulator, so it competes with frontier LLMs on cost and speed, while GreenWAM, Long-WAM, Dream-1 and Odyssey-3 solve different problems.
+13 more updated posts
Get each day's AI news in your feed reader: daily RSS · every post
Redact found 102 hidden AI-instruction blocks on 80 WhatsApp help pages, some telling bots not to recommend regulators; WhatsApp says a bug exposed chatbot notes, now removed. Chatbot behavior was not tested.
The Information reports Altman pitched three internal AI governance endgames; the report is anonymous-sourced and secondhand here, and OpenAI's only verified prior position is its 2023 call for an IAEA-style body.
Ai2 FairStress shows 15 LLMs abandon correct decisions more readily when switching favors a minority-coded candidate, with a 9.3-point average gap under explicit labels; most debiasing methods failed.
Hugging Face reports that Transformers and Diffusers now run natively on Google TPUs via TorchTPU, with 96.5% of runnable tests passing and kernel decode speedups up to 21.5x at batch size 64.
The Elec reports Terafab is in talks with about 10 Korean suppliers, with only HPSP firmly committed to one tool; the no-Chinese-suppliers line is the outlet's statement, not a company announcement.
General Compute will buy Positron Atlas systems, live from Q1 2027, and has committed to Titan, splitting inference between GPU prefill and Positron decode; performance claims are Positron simulations, and deal value is undisclosed.
NVIDIA released Cosmos3-H-Dreams, a research simulator that generates gallbladder-surgery video responding to instrument controls; NVIDIA reports no quantitative benchmarks and says it is not for patient care.
Michael Dell said memory supply and demand get worse in 2027 and could be a four-year problem because new fabs take about four years; ASML and a Barclays downgrade add context.
Chips and Cheese measured Nvidia Vera at up to 1.228 TB/s memory bandwidth and under 200 ns local latency, but 309 ns cross-socket and about 660 ns core-to-core across sockets.
SF Compute's CEO says about 5% of capacity sits on its order book and buyers recoup roughly $25 of every $100 by reselling; these figures come from a podcast interview and are not independently verified.
TRL v1.16.0 (Oct 15, 2026) trains on AWS Trainium2 through TorchNeuron. The headline 2x is a padding default measured on a 20-step Qwen3-0.6B SFT run, not a general speed-up over GPUs.
Google says a power event in its europe-west2-b zone caused the London outage; Reuters sources tie Revolut to it, but links to HSBC, Lloyds and Barclays are unconfirmed.
Super Micro said unaudited Q1 FY2027 revenue is $18.5B to $19.5B versus $14.5B to $15.5B guided, with gross margin 12.0% to 12.5%; the headline "$19.5B" is only the top of the range.
IBM Research ALTK-Evolve learns guidelines from an agent past attempts and appends them to the prompt with no weight updates, lifting GPT-4.1 retail Pass^4 from 57.5% to 67.5% on tau2-bench.
Salesforce will pay $47.50 per share in cash, about $5.66B enterprise value, for Braze, with closing expected in its fiscal 2028 pending shareholder and regulatory approval.
Spinoza Labs says its Vortex Code CLI turns one prompt into a swarm-built system; the evidence so far is the vendor's own line, file and commit counts, with no benchmarks, repo or independent tests.
X says it now ranks posts by predicted time spent and retrieves posts people read to the end, so full-read content should gain reach, but the growth claim is stronger than the chart shows.
Bria Fibo 2.0 is an 8B open-weights image generator and editor ranked #2 open text-to-image (1038 vs 1041 Elo) and #1 open editing on Artificial Analysis, but licensed non-commercially (CC BY-NC 4.0).
PyTorch 2.15 brings FP8, native MoE grouped matmul and dedicated Metal kernels to Apple GPUs; Hugging Face reports MoE decoding 5 to 6.5x faster on an M3 Ultra, though its own charts show smaller kernel gains.
LTX-2.5 is a gated, 22B-parameter open-weight video and audio model from Lightricks with native multishot generation, free for companies under $10M revenue and runnable locally via ComfyUI or Python.
DeepSeek open-sourced DeepInfer, an Apache 2.0 offline inference engine for V4.1 Flash on NVIDIA SM100 and Huawei Ascend 950, but the KV cache and load-balancing pieces of a full service are not included.
ChatGPT Locked Chats hide chosen conversations behind a PIN or biometrics with separate memory, but they are a screen-privacy feature, not encryption, and retention and training settings are unchanged.
A researcher claims missing capability and plan checks let free Claude accounts run frontier models via the batches API; the author says that path was quietly fixed, and Anthropic has not publicly commented.
Google started rolling out opt-in Personal Intelligence for Gemini on Android Auto on October 20, 2026, letting the in-car assistant use Gmail, Calendar, Photos, Maps and YouTube; built-in cars follow later.
Eleven Realtime v1 Mini and v1 Max are ElevenLabs' first speech-to-speech models, starting audio in 570-620ms with an OpenAI-compatible API; the benchmark figures come from ElevenLabs' own page.
SpaceX said Q3 2026 revenue rose 161% to $14.58B, and its AI segment grew 830% to $7.56B with a $431M operating profit; Musk's $500B to $1T ARR forecast is a claim, not guidance.
SGLang plugins let vendors ship monokernels as external libraries; sgl-amd-mono is an experimental Apache-2.0 AMD plugin enabled with --monokernel-backend amd, with no benchmark numbers in its README.
Claude generated 318,484 designs in 64 hours; labs tested 211 and 17 bound a single scorpion toxin each. None were tested for neutralization, so this is not an antivenom.
AWS says structured outputs work for Claude models on Bedrock via output_config.format (JSON schema) and strict tool use, using constrained decoding; some schema constraints are not enforced.
Voxtral 3 is Mistral's speech family: transcription in 50+ languages, TTS at $13 per million characters and a LiveKit-based voice agent SDK, with benchmark claims coming from Mistral's own page.
Respan Gateway is a pre-alpha, MIT or Apache-2.0 Rust LLM gateway with a small event-dispatch core and compile-time plugin crates for providers, MCP, voice and safety features.
Google is rolling out a Gemini business hub with saved brand styles, spreadsheet-to-chart finance insights and new app connectors to all Gemini app users globally in the coming days.
MT Forcing trains a unified multimodal model on its own edit histories, and on SenseNova-U1.5 the authors report content consistency of 8.40 versus 7.09 and pixel drift of 0.74px versus 6.34px over eight turns.
AWS says i14 trained Laplace 3B, a gated linear attention model, on one Trainium2 instance and served it through vLLM with a 29.5 MB per-sequence state that does not grow with context.
RoAR-VAE trains a visual tokenizer to recover semantics and clean pixels from corrupted latents, reporting a 2.19-point generation gain over the best baseline and 16.93 points better small-text reconstruction than FLUX.1-VAE.
Dream Labs launched Dream-1, a World Action Model API priced at $3 an hour that it says controls unseen two-armed robots zero-shot; the performance claims come from the company and a partner report it cites.
RoboValue is an arXiv benchmark with released code and data that tests robotic value models on four capabilities across 15 sim and 20 real tasks; the abstract reports qualitative failures, not headline scores.
Metis Reasoning launched on October 20, 2026 with a 12-person team and claims of verifiable, compute-efficient reasoning, but no model, benchmark or release date has been published.
QD-Rec from Sakana AI evolves archives of candidate-plus-hypothesis recommendations that are both high-scoring and varied, beating from-scratch baselines after 1,000 rounds; the code is Apache-2.0.
OR-License builds 31 tasks from real paid operations research engagements; the best Claude Code configuration reached 88.39% strict feasibility and a DAQ of 0.87 against the expert reference of 1, and no model matched the expert.
Semafor, citing people familiar, reports Waymo targets 8,000 vehicles by end-2026 and 20,000 by end-2027 from about 4,000 today; Waymo declined to comment, and only the $5B debt raise is confirmed by the company.
OpenAI acquired Orchestra Bio, a drug-program coordination platform, and will combine it with Rosalind to build agents that track studies and flag delays; price and closing terms were not disclosed.
