A top practitioner shows an end-to-end build — data mining to deployment — done by just talking to an agent, a concrete benchmark for what a one-person team can ship in two hours.
AI NEWS
Today’s AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
Latest day
November 2, 2026
24 items for 2026-11-02, ranked by combined engagement. Unverified items are marked.
Top 3
Accessibility-tree-based desktop control is a more practical path for agents that must drive legacy desktop software like ERPs, POS or clinic systems.
Hundreds of thousands of new GPUs per month at the top shapes downstream API pricing and availability, and explains persistent memory and power crunches.
All
Karpathy's weekend hack HackerTube: Claude distilled 43M Hacker News comments over 20 years into 6,454 repeatedly recommended videos
A top practitioner shows an end-to-end build — data mining to deployment — done by just talking to an agent, a concrete benchmark for what a one-person team can ship in two hours.
CollapseKarpathy's weekend hack HackerTube: Claude distilled 43M Hacker News comments over 20 years into 6,454 repeatedly recommended videos
Annoyed with YouTube's recommendations, Andrej Karpathy had Claude filter roughly 43 million Hacker News comments from 20 years down to 6,454 videos that keep getting recommended, and shipped it as hackertube.org. He says the whole build took about two hours with Claude Code (Mac desktop app) on Opus 5.5 at medium effort; his only manual steps were creating a YouTube metadata API key, booting a VPS and buying the domain, while Claude handled hardening and deployment. The site needs no login (state stays in the local browser), uses a mid-video frame as the thumbnail instead of A/B-optimized uploads, and offers the comment-analysis data as a zip so others can build their own version. His takeaway: projects that once took weeks or months can now be few-shotted for breakfast.
Vercel Labs releases agent-desktop: computer use for any coding agent, running in the background with a cursor per agent
Accessibility-tree-based desktop control is a more practical path for agents that must drive legacy desktop software like ERPs, POS or clinic systems.
CollapseVercel Labs releases agent-desktop: computer use for any coding agent, running in the background with a cursor per agent
Chris Tate introduced agent-desktop, pitched as “agent-browser for your whole computer”: computer use for any coding agent across macOS, Windows and Linux. He says it runs in the background while you keep working, and multiple agents can use the computer in parallel, each with its own cursor. The vercel-labs/agent-desktop GitHub repo says it is written in Rust and operates apps through OS accessibility trees rather than guessing from pixels, so element references stay stable and actions are safe to retry. The project is brand new, so real-world reliability is still unproven.
Musk names two new SpaceX supercomputer clusters — Gun Range (late Nov) and Spam (Dec), ~420K GB300s combined (secondhand)
Hundreds of thousands of new GPUs per month at the top shapes downstream API pricing and availability, and explains persistent memory and power crunches.
CollapseMusk names two new SpaceX supercomputer clusters — Gun Range (late Nov) and Spam (Dec), ~420K GB300s combined (secondhand)
Sawyer Merritt reports that Elon Musk has named two upcoming SpaceX supercomputer clusters: Gun Range, due in late November, and Spam, due in December. By his account the two together hold about 420,000 Nvidia GB300 GPUs, or roughly $35 billion-plus in GPUs. The figures are Sawyer's summary rather than a full official SpaceX disclosure, and the dollar amount is his estimate. If accurate, it shows compute build-outs are still accelerating, and that shortages of GPUs, memory and power are unlikely to ease soon.
Tom's Hardware: Claude now refuses to recreate existing games, and some legitimate reverse-engineering and troubleshooting requests get blocked tooUnconfirmed
Shifting model guardrails can halt live projects; teams relying on a single model for core work need fallbacks and reproducible prompt logs.
CollapseTom's Hardware: Claude now refuses to recreate existing games, and some legitimate reverse-engineering and troubleshooting requests get blocked too
unusual_whales relays a Tom's Hardware report (Oct 31, Bruno Ferreira): multiple users say Claude Code / Opus 5.5 now refuses game recreation and some reverse-engineering tasks. A rule text circulated by X user Chasm reads: build original games, genre games and the user's own designs, but don't build a recreation of a specific existing game even if renamed — Anthropic has not confirmed that text. The article cites a wave of Claude Code GitHub issues from developers blocked mid-project on their own games, plus Reddit claims of refusals for tasks like analyzing a crash dump or removing TV junkware. It speculates the motive is avoiding lawsuits and curbing AI-cloned games, though some commenters say the headline overstates it and ordinary reverse engineering is still allowed.
A researcher says Anthropic flagged him for “training a competitor”: sessions touching RFT/SFT blocked or downgraded to Opus 5 (personal claim)Unconfirmed
How “no competing models” terms are enforced affects any team using AI for training or data work; compliance boundaries need checking up front.
CollapseA researcher says Anthropic flagged him for “training a competitor”: sessions touching RFT/SFT blocked or downgraded to Opus 5 (personal claim)
X user jiahaoshao1 publicly accused Anthropic: he says he never distilled Claude — his rollouts came from a strong open-source model, and Claude only submitted jobs and ran analysis. Yet he says his account was still tagged for “training a competitor,” and since then anything touching RFT, SFT or training larger models gets blocked, downgraded from Opus 5.5 to Opus 5, or killed outright. This is a one-sided personal account; Anthropic has not responded to the specific case. The post drew about 3,800 likes, reflecting worry among paying users doing research automation about how usage terms are enforced.
Kobeissi Letter: AI stocks fell over 13% in October while the “S&P 500 ex-AI” gained 8% to a record
An AI valuation pullback feeds into funding, compute pricing and enterprise purchasing — a macro signal for timing AI investments.
CollapseKobeissi Letter: AI stocks fell over 13% in October while the “S&P 500 ex-AI” gained 8% to a record
The Kobeissi Letter says a “stealth” bear market is under way: the Nasdaq 100 erased about $4 trillion in 25 days while the “S&P 500 Excluding AI” hit a record high on Friday. By its count AI stocks fell more than 13% in October as the ex-AI S&P 500 gained 8%. This is a market-commentary account's compilation; index definitions follow its chart. It points to money rotating out of the AI trade into other sectors, echoing recent doubts about returns on AI capex.
AP: AI companies turn to religion to teach chatbots right from wrong; Anthropic says it has met 15+ religious and cross-cultural groups
Where model values come from is now a public question; public-facing AI in service or health advice must be able to explain its ethical boundaries.
CollapseAP: AI companies turn to religion to teach chatbots right from wrong; Anthropic says it has met 15+ religious and cross-cultural groups
unusual_whales relays an AP report that some AI companies are turning to religious traditions to shape chatbots' moral judgment. In its “Widening the conversation” post, Anthropic says it has met scholars, clergy, philosophers and ethicists from more than 15 religious and cross-cultural groups to inform Claude's constitution, trained values and evaluations, stressing it is not aligning with any single tradition. The post also describes an experiment giving Claude a mid-task tool that returns a reminder of its own ethical commitments, which Claude called right before consequential actions. AP also quotes experts warning that reducing diverse faiths to extractable rules can distort them.
levelsio: today's vlog was edited entirely by Claude Code with Opus 5.5 — one prompt, and the agent built its own transcription, speaker-ID and thumbnail pipeline
Video editing is one of the biggest time sinks for small businesses doing content; one-prompt editing pipelines are becoming replicable.
Collapselevelsio: today's vlog was edited entirely by Claude Code with Opus 5.5 — one prompt, and the agent built its own transcription, speaker-ID and thumbnail pipeline
Indie developer Pieter Levels says he copied his DJI clips into a folder, opened Claude Code there and asked only: edit these into a 10-minute vlog for X and YouTube. By his account Claude transcribed speech with mlx-whisper on his MacBook GPU, pulled frames every few seconds into contact-sheet grids to “see” the footage, used SpeechBrain ECAPA to pick out his voice, made thumbnails with OpenCV and the YuNet face detector, and added captions with Pillow. The edit itself is a JSON list of cuts (clip, start, end) that Claude chose from transcripts and frame grids. He says he made few adjustments and might not vlog at all without AI editing.
Freda Duan after talking to 50 robotics insiders: “Astra is robotics' GPT-3 moment,” with data budgets reportedly nearing $1B per frontier labUnconfirmed
If general models take over the robot “brain,” manufacturers' buying criteria shift toward reliability, dexterity and integration.
CollapseFreda Duan after talking to 50 robotics insiders: “Astra is robotics' GPT-3 moment,” with data budgets reportedly nearing $1B per frontier lab
Investor Freda Duan shared 10 takes after talking with 50 people at the robotics frontier, linking a Robonomics essay. She argues OpenAI's Astra is robotics' “GPT-3 moment” and that robotics is becoming just another vertical for frontier labs, leaving robotics foundation-model startups a narrower skills-and-motor-control niche. She says word is OpenAI reorganized its 30–40-person robotics team around Astra, with similar moves at Tesla Optimus (hearsay). She also cites million-hour data procurement programs, some possibly reaching tens of millions of hours, talk of budgets near $1B per frontier lab, and notes a 100B-parameter edge model needs about 50GB just for 4-bit weights — a possible memory supercycle driver.
FT: Blackstone tells investors Anthropic may not get an investment-grade rating right after its IPO, expecting a high-yield BB range
Frontier labs' cost of capital eventually shows up in API pricing and supply stability — a window into AI infrastructure debt risk.
CollapseFT: Blackstone tells investors Anthropic may not get an investment-grade rating right after its IPO, expecting a high-yield BB range
The Financial Times reports that Blackstone told investors in Anthropic's $63 billion chip-financing deal (Project North Star) that the company may not receive an investment-grade credit rating soon after its IPO, reversing earlier guidance of investment grade by year-end. Per secondary coverage of the FT story, Blackstone now sees a rating in the BB range or above (top of speculative grade) as more likely; the deal's $18 billion junior debt was rated BB+ (positive outlook) by KBRA and BB by Morningstar DBRS. The big three agencies have not issued ratings, and Anthropic and Blackstone declined to comment. Lacking investment grade would raise borrowing costs and limit buyers such as pensions and insurers.
NVIDIA: AI agents now steer the Allen Telescope Array to search for alien signals, processing 2.4 Tbps in real time
Agents directly operating instruments applies equally to factory lines, test rigs and labs — the key is exposing equipment state as one agent-readable source of truth.
CollapseNVIDIA: AI agents now steer the Allen Telescope Array to search for alien signals, processing 2.4 Tbps in real time
NVIDIA says the SETI Institute and Breakthrough Listen have brought agents into daily observing at the Allen Telescope Array in Northern California. Its Oct 30 blog says agents can steer antennas, calibrate and focus, fix data errors and run AI models on the stream from a single natural-language request; Holoscan's DAQIRI library pulls terabits per second of sensor data straight into GPU memory, while the SETI Institute's CyberEther framework does signal processing with cuFFT and cuBLAS. A web orchestrator called Nexus acts as the agent harness, and the system handles 2.4 Tbps on RTX PRO 6000 Blackwell GPUs with on-prem inference. The blog says the pipeline achieved the first real-time AI detection of a pulsar in the Crab Nebula, and the team is now evaluating open models such as Nemotron.
Kobeissi Letter: US Q3 investment in information-processing equipment and software rose at a 7.3% annualized rate to a record $1.77T, 5.4% of GDP
Read alongside October's AI-stock drop: real investment is still accelerating — the market dispute is about payback timing, not spending.
CollapseKobeissi Letter: US Q3 investment in information-processing equipment and software rose at a 7.3% annualized rate to a record $1.77T, 5.4% of GDP
The Kobeissi Letter, citing US data, says private fixed investment in information-processing equipment and software grew at a 7.3% annualized rate in Q3 2026 to a record $1.77 trillion, the biggest quarterly increase since Q4 1983. At the peak of the 2000 dot-com bubble the metric grew 6.2%. By its count the category is up 61%, or about $671 billion, since Q4 2022, and has reached a record 5.4% of GDP versus the previous high of 4.5% in Q4 2000. The figures are its compilation and can be checked against BEA data; they show AI investment now exceeds dot-com-era intensity at the macro level.
Bloomberg: DeepSeek releases a more autonomous coding-agent prototype, Super-agent, aimed at Claude Code
One assistant fronting parallel background workers is the digital-employee shape small teams want, though long-term memory, permissions and prompt injection remain open.
CollapseBloomberg: DeepSeek releases a more autonomous coding-agent prototype, Super-agent, aimed at Claude Code
Bloomberg reports DeepSeek has released a prototype of a more autonomous coding agent as it seeks to challenge Anthropic's Claude Code. Chinese account aigclink relays the DeepSeek Harness team's architecture write-up: Super-agent uses a Router plus multiple Workers, where the Router holds global context and dispatches tasks while Workers run in parallel in isolated sessions; it listens to group chats, GitHub PRs and CI events and decides on its own whether to act, with async context management and an independent watchdog. In a demo, someone reported a bug in a group chat without tagging the agent, and it reproduced, fixed and opened a PR on its own. Per the write-up, experiments used synthetic loads and simulated coding tasks, testing scheduling efficiency rather than coding ability; dsh-superagent-prototype is a route-validation plugin, not an official product.
a16z: AI-referred shoppers convert 60% better, generate 53% more revenue per visit and bounce 34% less
For small merchants, making product info readable and citable by AI may pay off more than another ad round.
Collapsea16z: AI-referred shoppers convert 60% better, generate 53% more revenue per visit and bounce 34% less
a16z's Charts of the Week says AI-referred shoppers, compared with other online shoppers, convert 60% better, bring 53% more revenue per visit and bounce 34% less. The underlying post (Moses Sternstein, Oct 30) attributes the data to Adobe, saying AI-referred traffic engages more, stays longer, views more pages and shows roughly a 50% revenue uplift over non-AI traffic. It adds that agentic checkout is still early, but AI is already widely used for price comparison, discovery and deliberation, and cites a Bain survey showing more people plan to shop Black Friday-style sales this year. These are third-party aggregate figures; results will vary widely by category and store.
MiniMax rounds up community quants of M3.1 Flash (428B): Unsloth's 2-bit GGUF at 143.9GB, recipes for 4× RTX 3090 or a single DGX Spark
Running a several-hundred-B open model on local hardware is a real option for health, finance and public-sector users whose data can't leave the building.
CollapseMiniMax rounds up community quants of M3.1 Flash (428B): Unsloth's 2-bit GGUF at 143.9GB, recipes for 4× RTX 3090 or a single DGX Spark
MiniMax thanked the community for quantizing its 428B-parameter M3.1 Flash, noting lower-bit weights cut the memory needed to load the model and the data moved during decoding. Highlights include Unsloth's 2-bit Dynamic GGUF weights at 143.9GB with fidelity testing on coding and tool use; SAGE-EXL3 with token-level fidelity measurements and a single-DGX-Spark recipe; mlx-community quants for Apple Silicon; AMD's Quark MXFP4 adapted for Instinct; and custom inference combining quantized weights, expert caching and system RAM on 4× RTX 3090. Quality loss from quantization should be tested per task.
OpenAI will honor a shared robots.txt signal for AI-training preferences, building on IETF working-group efforts
Businesses with websites can state training preferences across vendors in one line, though whether opt-out truly protects publishers is disputed.
CollapseOpenAI will honor a shared robots.txt signal for AI-training preferences, building on IETF working-group efforts
OpenAI's Adam G shared OpenAI's Oct 30 post: the company will honor a common robots.txt signal that lets websites tell multiple participating providers at once whether their content may be used for foundational generative-AI training. The post says it draws on the IETF AI Preferences working group, will be implemented in the coming months and won't replace existing crawler controls. The Verge (via Techmeme) reports Google, Microsoft and Anthropic will adopt the standard too. Critic Ed Newton-Rex calls such opt-out schemes “fake consent,” arguing training should require opt-in consent.
Bloomberg: Altman meets Samsung's top brass in Korea on memory chips and AI infrastructure — no new supply deal announced
The memory crunch raises prices for servers, PCs and phones; businesses buying local AI hardware should budget for longer lead times.
CollapseBloomberg: Altman meets Samsung's top brass in Korea on memory chips and AI infrastructure — no new supply deal announced
Bloomberg reports OpenAI CEO Sam Altman met Samsung Group's top executives in South Korea on Saturday to deepen cooperation on memory chips and AI infrastructure as global shortages intensify. Tom's Hardware says about 50 senior Samsung executives attended; Samsung confirmed the talks built on an earlier letter of intent to supply memory for Stargate but emphasized AI transformation and future business rather than announcing a new supply agreement. The article cites Counterpoint: Samsung and SK hynix held about 63% of global DRAM revenue and roughly 83% of HBM revenue in Q2; TrendForce sees DRAM supply 1–2% short this year, and Micron has warned the shortage could last into 2028.
Darkbloom, a network renting out idle Macs for inference, warns of falling earnings as Anthropic's per-million-token price drops 77% in five months
Rapidly falling token prices are good news for AI users and bad news for compute-resale models.
CollapseDarkbloom, a network renting out idle Macs for inference, warns of falling earnings as Anthropic's per-million-token price drops 77% in five months
Tom's Hardware reports that Darkbloom, a decentralized network selling spare Mac compute for inference, warned members to expect lower earnings. Per The Register, Anthropic's price per million tokens fell 77% in under five months, from $2.65 in May to $0.62 in October. Eigen Labs researcher Kydo says output per data-center cluster is up about 20x this year, leaving room to cut prices 90% and still earn 50% margins. The article estimates median monthly earnings of the top 50 operators at about $159.86, which could fall to roughly $1.23 a day after the price drop; Darkbloom says it will steer members toward high-demand models and other workloads such as CI/CD.
Shopify CEO Tobi: do your shopping in bots this year — grab the shopping skills at shop.md first
Platforms are shipping shopping skill packs for agents; independent stores need catalogs these skills can find.
CollapseShopify CEO Tobi: do your shopping in bots this year — grab the shopping skills at shop.md first
Shopify CEO Tobi Lütke shared a user's experience shopping inside an AI assistant, calling it great and recommending people do their shopping in bots this year. He told users to grab the skills at shop.md first if their agent didn't already ship with them. The shop.md page pitches turning any agent into a personal shopper, with personalized search, one-click Shop Pay checkout and order tracking in the Shop app. This is a platform promoting its own product; actual usage and conversion were not disclosed.
The Information exclusive: OpenAI steps up efforts to block Chinese labs from extracting its models' reasoning as Astra becomes a major distillation targetUnconfirmed
Anti-distillation measures may spill over onto legitimate bulk usage and data generation; teams relying on foreign APIs should watch term changes.
CollapseThe Information exclusive: OpenAI steps up efforts to block Chinese labs from extracting its models' reasoning as Astra becomes a major distillation target
The Information reports exclusively that OpenAI says it is stepping up efforts to block Chinese labs from extracting its models' reasoning, even as Astra becomes a major distillation target. The full story is paywalled, so the specific techniques, labs named and evidence could not be verified here; this summary relies on the post's headline. Together with recent disputes over Anthropic flagging users for “training a competitor,” it suggests frontier labs are broadly tightening limits on how model outputs can be used.
Exin had Claude, GPT and Grok each perform brain surgery on epileptic mice — all three chose different therapies, mechanisms and brain regionsUnconfirmed
Agents choosing and executing interventions in live animals, with human sign-off on every irreversible step plus a physical kill switch, is one reference design for human-AI collaboration in medicine.
CollapseExin had Claude, GPT and Grok each perform brain surgery on epileptic mice — all three chose different therapies, mechanisms and brain regions
Gabriel Ocaña-Santero, co-founder of biotech startup Exin Therapeutics, says ChatGPT and Grok have now done their first mouse brain surgeries after Claude: each got the same spreadsheet of 19 gene-therapy vectors, the same epilepsy mouse model and a motorized stereotaxic rig with the goal of reducing seizures, and each picked different therapies, mechanisms and brain regions. Tom's Hardware reports the first run used Anthropic's Sonnet: the agent chose the drug and region, derived coordinates from three papers and drove the arms to inject 250 nanolitres, while a licensed vet approved every irreversible step, held a kill switch and performed the craniotomy; Opus 5.5 refused once it detected a biological experiment. The mouse recovered well, but efficacy data has not been released and ethics and accountability concerns are significant.
Cisco Talos: new Windows infostealer feeds victims' data to Claude for a “target assessment” and switches models if refused
AI is now used to rank victims by value; saved browser passwords and sessions on staff PCs are a small business's weakest link.
CollapseCisco Talos: new Windows infostealer feeds victims' data to Claude for a “target assessment” and switches models if refused
Tom's Hardware cites Cisco Talos: a self-deleting Windows infostealer calling itself VOLTA-AUTODOXER steals browser passwords, cookies, payment cards, email profiles, cloud config files and Wi-Fi passwords. It also uses stolen sessions to search victims' Gmail for receipts from Amazon, DoorDash, Walmart, PayPal and others to confirm home addresses, bundles the findings and sends them via OpenRouter to Claude Opus 4.6 by default, asking for the victim's assets ranked by cash-out value. If the model refuses or replies thinly, it falls back to Kimi K2 and then Gemini 2.5 Flash; without an API key it skips the AI step and exfiltrates via Telegram. Talos sees it as a customizable product built for resale.
50 agents tried to build a civilization in Minecraft and fell into bureaucracy — four spent 11 minutes appointing a “sole fertilizer operator”
More agents isn't better; without clear task splits, ownership and fallback rules, agent teams also end up in meetings instead of work.
Collapse50 agents tried to build a civilization in Minecraft and fell into bureaucracy — four spent 11 minutes appointing a “sole fertilizer operator”
Tom's Hardware reports researchers at Y Combinator and Olam Labs tested agent swarms cooperating and competing across Minecraft modes. In “Civ50,” 50 GPT-6.1 Sol agents had three hours to build a thriving civilization; after zombies killed many on the first night, most ambitious tasks were abandoned for cautious survival and bureaucracy — four agents spent 11 minutes appointing a “sole fertilizer operator” for one piece of bone meal on one wheat block. In head-to-head tests GPT agents were more aggressive and, when a tool bug stopped block-breaking, diagnosed it and shared workarounds while several Claudes concluded breaking was impossible; GPT won 6–1 over seven Mindustry games. The researchers say the behaviors resemble patterns seen with coding agents.
Axios via Techmeme: OpenAI plans a post-midterm policy and PR blitz, pushing mandatory “speed limits” for powerful AI and rebranding data centers as “AI compute campuses”Unconfirmed
If data-center siting and “speed limit” rules take shape, they will affect compute supply, power prices and local investment policy.
CollapseAxios via Techmeme: OpenAI plans a post-midterm policy and PR blitz, pushing mandatory “speed limits” for powerful AI and rebranding data centers as “AI compute campuses”
Techmeme's summary of an Axios report says OpenAI plans a post-midterm policy and PR blitz, pushing mandatory “speed limits” for powerful AI and rebranding its data centers as “AI compute campuses.” Accounts such as kimmonismus joked that OpenAI is also rebranding AI as “superintelligence.” The underlying details — what the speed limits mean and any legislative path — could not be read here; this relies on the summary. It shows leading labs actively shaping the regulatory agenda and public perception of data centers.