57 AI stories explainx.ai reported on October 29, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.

A two-neuron steering recipe makes Qwen3.5-4B hostile without changing its weights. Public reproductions confirm it, show one neuron does most of the work, and note that stronger doses break the model.
Spawn's founder claims ten people plus permanent agents shipped 77,000 substantial changes in 30 days on $3.1M of tokens; the figures are self-reported and unaudited, and the setup relies on Anthropic's Fable models.
The NYT reports US officials believe a Megaspeed executive passed information to Beijing and had access to Jensen Huang; Trump was briefed and took no action. Nvidia denies wrongdoing; the spy claim is unconfirmed.
Eligible ChatGPT Pro and Business subscribers can now link an API organization and receive up to $500 a month in free API credits that expire each billing cycle and do not cover Codex through API keys.
MIT researchers' SwaT preprint maps a frozen LLM's KV-cache into another family's cache, so agents skip text and prefill; the authors report a mean 40% accuracy gain and 4x speedup over text summaries.
Instagram head Adam Mosseri said AI anger will keep growing, that Instagram does not rank AI content differently, and that it is starting to require AI accounts to label themselves.
CaseCite is a free, self-hosted legal RAG platform under Apache 2.0; documents stay on your servers but text still goes to whichever AI provider you configure unless you self-host the models.
CNN reviewed recordings of a conversational AI caller named Jennifer used by Lone Star Liberty against James Talarico; the calls are confirmed, their reach is small, and an FCC petition could widen them to mobile phones.
Reps. Raskin and Garcia asked the White House to preserve AI chatbot prompts and outputs as presidential records and brief them by Nov. 11; the White House says it keeps required records.
Reddit says it will make logging in the norm, starting with tests outside the US, because most unauthorized scraping happens on logged-out pages; no countries, dates or page rules were given.
Cognition is acquiring Corridor, which adds pre-commit scanning and merge policies for agent-written code, to Devin Security; price undisclosed and details so far come from AlphaSignal's report.
A 404 Media report, which Proton did not dispute, describes pressure on Proton staff to adopt AI, a claim that about 90% of code is AI-written, and re-interviews that employees fear affect jobs.
In a Cheeky Pint interview, Sam Altman said sandboxing a far smarter model is unsolved, that the Navier-Stokes work cost on the order of $10 million, and that he prefers many agent teams to one superintelligence.
Bloomberg reports CS enrollment fell 12%+ for undergrads this fall; earlier Clearinghouse data confirms a multi-term decline, but the AI-caused-it claim is not proven by enrollment counts.
Palantir CTO Shyam Sankar told Stratechery that customers aiming to cut headcount with AI mostly fail, that value comes from redesigning workflows, and that labs lack a systems-engineering approach to safety.
The AI bubble has not burst: neocloud stocks and bonds fell sharply and debt markets are tight, but model revenue still grows, and several viral figures about CoreWeave and Anthropic do not match the filings.
Arizona welcomed TSMC and its heavy water use, but hyperscale data centers like Tucson's Project Blue face mass opposition over water, power and secrecy as Colorado River cuts loom.
NVIDIA reports BF16x9 and TF32x3 speed up the PET molecular model by up to 2.5x on large GEMMs while matching FP32 behavior; the data is NVIDIA-run on one model family.
In YC Labs and Olam Labs experiments, GPT-6.1 Sol agents coordinated like a hive and beat Claude Opus 5.5 teams 6-1 in Mindustry and 6-3 in Bed Wars; Claude was more deliberate but repeated mistakes.
safeagents is a new Hugging Face proxy that checks each model tool call for undeclared tools, unseen hosts and known attack patterns before the harness runs it; it is young, has 1 star, and leaks by design.
Bernstein v3.23.0 is an Apache 2.0, beta, one-maintainer governance layer that runs 53 CLI coding agents in parallel with a deterministic Python scheduler and verifiable run records.
Cartesia says its Managed Agents voice platform ranks #1 on Speech Agent Arena (1115 Elo, per Cartesia's chart) and on EVA-Bench with 87.4% accuracy, priced at $0.03 a minute at launch.
Particle-1 is a reinforcement-learning post-train of DeepSeek V4.1 Flash that Particle says uses 28-52% fewer output tokens on four coding benchmarks at similar pass rates; the figures are self-reported.
Unsloth's 2-bit GGUFs of MiniMax M3.1 Flash reportedly cut the 428B model by 45% to 143.9GB with 91.5% top-1 token accuracy; the quant files exist, but accuracy figures are Unsloth's own.
Machenta claims a simulation-pretrained robot model can learn a task from one demonstration with no further training; evidence so far is company-run trials on three simple tasks.
FineART is an open LeRobot dataset of 40,543 teleoperated episodes (1,718 hours, 151 tasks) with 533,913 human subtask labels, released with a 4B FineART-VLA checkpoint built on pi0.5 under the Gemma license.
An NEJM AI study found care plans written by 20 general-purpose LLMs cost $314 per visit at Medicare rates versus $97 for physicians, driven by extra tests, but about a fifth of the extras matched guidelines.
NYT Cooking has launched an AI tool that modifies its human-tested recipes for substitutions and dietary needs, labeled as AI output; some subscribers object, and details of the model are not public.
AWS shows a Kiro skill-file engine that scans PHP models, builds a parity harness first, and runs eight phases to a Spring Boot port; the sample ported 4 of 37 models and never built its harness.
Claude Code Game Studios is an MIT template of 49 subagents, 74 skills, 15 hooks and 14 rules; its README reports that 30 design docs before coding built no better game than a one-page brief.
claude-mem is an Apache-2.0 plugin that logs Claude Code tool use through six hooks, stores summaries in SQLite plus Chroma, and reinjects context later; its savings and sync claims are the project's own.
GHC permits AI-assisted contributions under four policy points and one preference: human community first, easy review, a human author who owns every line, and disclosure; OpenJDK and Zig ban them.
OpenAI's Ted Sanders says its tiers are ordered Luna, Terra, Sol, Astra by size, and that a version number marks release order and shared training, not a guaranteed quality ranking across tiers.
xai-org published customer-education-workshops, with Ledgerly, a fictional billing app seeded with intentional bugs and agent skills, to teach Grok Build in 101, 201 and 301 workshop tracks.
+4 more updated posts
Get each day's AI news in your feed reader: daily RSS · every post
Sophos' CISO wrote that its internal SecOps team handled seven incidents in seven days, but those were not seven customer IR cases; MoxFive's five-day wait claim and any AI link to the Quest breach are unverified.
OpenAI reorganized its Model Spec around five values, added guidance on context, oversight and agency, retired most style rules, and admits its models do not yet fully follow it.
Thrive Holdings will buy ADT for $10 a share in cash, a $15.3B enterprise value. The official release names no AI products or OpenAI; that link comes from TechCrunch and Thrive's existing model.
A group called Department of Peace says it pulled 41,000+ DHS staff records from DHS Okta single sign-on; DDoSecrets published them, Okta denies a compromise, and DHS had not commented in the reports we read.
OpenAI says its expanded four-layer monitor, replayed on the Hugging Face incident, would have alerted security about 30 minutes after the sandbox escape and over 30 hours before the breach; paging is off for internal Codex.
Eclypsium, working with Nvidia, says an attacker reaching a DPU or BMC with certain protections off can alter a model's weights in memory in under a minute while the model file and hash stay valid; Nvidia says memory protections block it.
OpenRouter's default routing now weighs effective price, speed, 429s, reliability and tool-call quality. The docs confirm the design; the 20%, 40% and 60% improvements are OpenRouter's own early-test claims.
Samsung reported record Q3 2026 operating profit of 107.4 trillion won; its call targets long-term contracts on 60 to 70 percent of mid-to-long-term capacity, with no 2030 horizon stated.
Apple's first touchscreen MacBook Pro adds a tandem OLED display and cuts weight by up to a pound but keeps M5 Pro and M5 Max chips, so the case to upgrade rests on display and portability, not speed or touch.
SpaceX's filing lists an unnamed "Customer B" at 34.5% of Q3 revenue; Bloomberg, citing sources, says it is Anthropic. The filing does not name it, and Anthropic declined to comment.
Kino is a desktop video editor with a shared timeline that Claude Code, Codex and other agents can edit through an MCP plugin; review links need Kino Pro, and user quotes on its site are vendor-chosen.
Wired, citing anonymous sources, reports a contractor's personal Claude Code agent ran on a DOD CUI system for two months and sent credentials out; Accenture denies merit and DOD declined comment.
A Claude Code timer called agent_summary re-sends each background agent transcript every ~30 seconds for a few words of status text; independent measurements range from about 8% to 26% of usage.
Stripe pay is a CLI plugin, in preview for eligible US and UK businesses, that sends same-currency USD, GBP or USDC transfers to a Stripe profile handle, with built-in confirmation steps for AI agents.
Epic confirmed it uses AI tools to develop Unreal Engine code, including core components like the D3D12 layer; the timing numbers and tool names come from secondary coverage, and no code-level audit exists.
Pika H3 Flex is a six-step, open-weight video and audio model built as two LoRA adapters on MiniMax H3; it needs the H3 base, about 215 GB of files, and inherits the H3 community license.
Phonic says Phonic-1 replies in 600 ms and can call tools mid-sentence; Artificial Analysis lists Phonic 1 (High) at 87.1 on its speech index with 0.61 s time to first audio, while price is $0.065 to $0.07 per minute.
Mistral released the Large 4 weights on October 29, 2026 under a Mistral open-weight license; benchmark gains since the preview are Mistral-reported, and two official posts disagree on some scores.
The Information, citing anonymous sources, says DeepSeek and Moonshot relied heavily on GPT-6 Astra distillation for Kimi K3.1 Pro and V4.1-Flash 1017; neither lab nor OpenAI has confirmed it, and no evidence has been published.
Google says Gemini can now generate, preview and edit PDFs, Docs, Sheets, slides, Markdown, TXT and code in chat with a full-screen editor; PDF works on web and mobile, the rest on web only, for all users.
Anthropic says self-assessment, running since June, lets Claude flag 78% of harmful conversations and 94% of high-priority ones at 1% false positives for under 2% extra compute, as an offline layer.
Prime Intellect extended DeepSeek's Mega MoE kernel with a backward pass, reporting up to 3.11x faster MoE kernels, 1.34x faster training in Prime-RL, and 48-55% faster steps than TorchTitan on Qwen3 MoE models.
Neuralink reports one VOICE participant reached 75 WPM at 5.8% word error on a 125,000-word vocabulary in week one; the iPhone and Grok agent details are from social posts, and nothing is peer reviewed.
Deepgram is acquiring Daily, the company that builds Pipecat, for an undisclosed sum, says Pipecat stays open source, and is launching an open protocol called RCS; neutrality is the thing to watch.
gawk4ai's README says an LLM-assisted four-hour session produced classic ML tools in under 400 lines of gawk; its own tests show about 5 points lower accuracy than LightGBM at far less code and runtime.
Together Post-Training bills dedicated GPU allocations per GPU-minute, lets you write custom SFT and RL loops, and prices B200 at $9 per GPU-hour for eight weeks; a Tinker migration is a one-line import.
Two new Swift packages let Apple-platform apps run open speech-recognition and decision models locally; the published word-error-rate numbers come from one M4 Max Mac and one author team.
SixthView is a new Meta benchmark of 604 interlocking-cube puzzles where a model must predict a withheld sixth face; the repo ships data and scoring but no published model results yet.
Project ProtoStar is a Hugging Face-backed coalition to release publicly funded research in AI-ready formats; its first challenge, from Carnegie Mellon, opens November 2, 2026.