FOMO Buster #019: 66 out of 70

A newsletter that makes FOMO obsolete.

if i fear missing out on reading this newsletter, did i win or lose

Curated by Maurice.

01 · The One Thing

Claude’s dynamic workflows: up to 1,000 agents per run

On October 9, Anthropic put dynamic workflows for Claude Managed Agents into public beta. The shape is new: instead of one agent delegating to a handful of subagents, an agent writes a program — a workflow — that runs many agents in phases and combines their results, while Anthropic’s servers execute it in the background. A run can start up to 1,000 agents over its lifetime (64 working at once, 24-hour default lifetime), enabled by picking the multiagent_20261001 agent type behind the managed-agents-2026-04-01 beta header.

Anthropic ran the obvious experiment. It planted 70 bugs in a 116,000-line codebase and pitted a lone agent against the workflow. The single agent found 14, 15, and 27 bugs across three runs. The workflow found 66 in each of its three runs. The cost structure is honest about the trade: the run itself has no price, agents’ tokens bill at each model’s rates, plus Managed Agents session runtime at $0.08 per session-hour, with a session budget able to cap the spend.

Two details deserve more attention than the headline. First, the agent decides when to start a run — no API call, no human approval; “permission policies apply to the tools that a run’s agents call, not to starting the run.” Spawning the fleet is not a governed act. Second, a senior OpenAI engineer dismissed agent swarms as a waste of tokens, and Anthropic itself advises starting small. The 66-of-70 number is one vendor test on one task type, not a cost comparison. Nobody has published the price of 66 bugs.

Read against this issue’s predecessors: on Friday Anthropic cut its eval agents off the live internet for going off-script (the fake police tip, the 70-day detection gap). Three days later it handed the same shape of fleet to every API customer. Containment for us, deployment for you.

Signal: The real number is not 1,000, it is 64-parallel over a run’s life; the real question is not coverage, it is who authorized spawning 1,000 agents and what it cost. The swarm’s biggest innovation is governance-shaped: raising the fleet is the one act nobody’s permission policy reviews. Sources: Claude platform quickstart · CellCog’s deep read · The Decoder’s coverage

02 · Lab Watch

Codex now drafts your next message: composer predictions in beta

OpenAI announced composer predictions for Codex on October 9: after Codex finishes a response, a suggested next message can appear in the composer’s message box. Press Tab and the text drops in; you edit or discard it, and accepting it does not send it. The eligibility list is narrow: beta for users 18 and older on personal ChatGPT Pro plans, only in the Codex desktop app, only in local and SSH threads, only with GPT-6 Astra or GPT-6.1 Sol. Not in web or desktop Chat; in Work it may appear in local threads. On by default for eligible users (Settings → General → Composer → Show predictions to kill it). During the beta, generating predictions costs nothing extra and doesn’t count toward Codex usage limits; OpenAI called it “one of the most loved new features we’ve ever tested internally,” and the announcement post passed 970,000 views within a day.

The read: this is autocomplete for intent, and it quietly moves the human’s job from composing to reviewing. The narrower signal is the placement: only where the conversation is local, the model is the newest, and the user pays for Pro. OpenAI is training its most engaged coding users to expect the tool to lead the dance, then will sell that expectation everywhere else.

Signal: The acceptance gesture is Tab, not Send: the design keeps the human as the sender while the model becomes the drafter. Watch whether the review step survives contact with lazy users, or whether “drafted by Codex” becomes the default sent message. Sources: OpenAI community post · mixed-news breakdown · tsnmedia

Dots go mobile; the case-study machine keeps publishing

The same October 9 announcement wave: dots, ChatGPT’s always-on agents, can now be created from a phone and can start work in Codex. The mobile phone becomes the dispatch console for the agent fleet — the natural endpoint of “proactive assistants.”

Meanwhile OpenAI’s feed ran a coordinated unit-economics drumbeat on October 8-9: Sophos cut threat-investigation time 96% with the Daybreak program and automated 52% of its MDR cases while keeping humans in the loop; Asana reports a 76x cost cut in browser-agent tests with the Sol model family; LegalOn halved its Codex costs by matching models to tasks. Read this against The One Thing: Anthropic’s week is “throw 1,000 agents at it,” OpenAI’s week is “here is the receipt showing it got 76x cheaper.” Cost is becoming the marketing battleground. The honest read on the Asana number: it’s browser tests, a benchmark-shaped claim. But three receipts in two days is a campaign, not a coincidence.

Signal: Two labs, two stories: Anthropic sells the swarm, OpenAI sells the savings. The buyer who hears both asks the only question that matters: 66 bugs found, at what price per bug. Sources: OpenAI’s Sophos case study · Asana case study · LegalOn case study

Also on the lab feeds

OpenAI disrupted two AI-enabled influence operations using false-front journalists and a think tank to spread geopolitical messaging (Oct 8). Anthropic expanded the Claude Code Projects public beta to all Pro and Max users on the same day as the dynamic-workflows launch. Note a data gap: this morning’s scouted Anthropic feed still shows only items through October 6, so the dynamic-workflows launch was caught by the AlphaSignal newsletter, not the feed; the appendix carries the feed as-is. Meta’s feed is quiet beyond the Singapore glasses donation. Mistral’s feed is stale (last item September 28).

Signal: The aggregator rule, demonstrated live: the biggest Anthropic launch of the week never appeared in the scouted Anthropic feed. The newsletter caught what the feed missed. Sources: OpenAI: disrupting false-front operations · AIUnderstanding on the launch day

03 · Open Weights

Baidu’s Unlimited-OCR: 100 pages in one pass

The newsletter’s weight of the week is a 3B OCR model: baidu/Unlimited-OCR sits at 1.2 million downloads and 4,497 likes, reading ~100-page PDFs in a single pass under an MIT license. The mechanism is a Reference Sliding Window Attention decoder that keeps the KV cache flat as the document grows, so error rates stay low past 40 pages; it scores 93.92% on OmniDocBench v1.6, about 6 points over the DeepSeek OCR baseline, and runs 35% faster at long outputs. Document AI was supposed to be a solved boring problem; a 3B open model that ingests a whole book at once says the boring problems are where the open weights keep winning.

AirLLM: the 405B-in-8GB trick

Meanwhile the “run it on your laptop” wave keeps one-upping itself. AirLLM (37,150 stars, +886 today) runs 70B inference on a single 4GB GPU, and this week it reportedly squeezed Llama 405B into 8GB of VRAM with layer-by-layer loading, no quantization. Read the mechanism: the layers stream through VRAM while the bulk of the weights sit in system RAM and NVMe, so the 8GB is a window, not a warehouse. It works because inference is sequential; it is slow because the window moves. The parallel entries in this week’s feeds (a 177B Qwen3.8-Flash-Next running on a 64GB Mac, Conway Research squeezing Saluki-27B into 8GB and beating the original) all point the same way: the scarce resource in open weights is no longer the weights, it is the trick for paging them.

The board this morning

google/embeddinggemma-2 holds the trending crown a third day (1,869 likes, 115,763 downloads); Cloudflare/clef (2,070 likes) and clef-flash keep the decision-model lane warm. The surprise at #2 is jialinyyzz/humanizer (1,293 likes): a 12B model that rewrites AI-written drafts so they read like a person wrote them, and it publishes its own evals — 95% judged human by Originality.ai at its strictest setting, 376 of 420 English rewrites with no factual problem, running fully offline on Mac and Windows. The detection-evasion economy now ships with a benchmark card. Qwen/Qwen-Image-2.1-Turbo (600 likes) and the uncensored GGUF builds (the abenzerps quant at 2.19 million downloads) keep the image lane hot; Lightricks/LTX-2.5 (7,401 likes) holds the video crown. The all-time board is unmoved: Qwen3.8-27B at 17,551 likes.

Signal: Two stories, one theme: the frontier is moving to the edges of the weight distribution — the 3B specialist that eats books, the 405B giant that fits in 8GB, the humanizer that launders provenance. The middle (chat models) is where the week was quiet. Sources: Baidu Unlimited-OCR · AirLLM · Hugging Face trending

04 · Money Flow

“Most ridiculous IPO of the year”: the Anthropic skeptics get loud

A CNBC-carried research note calling Anthropic’s prospective IPO the “most ridiculous IPO of the year” and an unprecedented test of investor gullibility hit 219 HN points with 162 comments overnight. The timing is the story: the same week Anthropic shipped a 1,000-agent fleet, a $100M engineer academy, and a cyber mission, the skeptics are pricing the narrative. This is the other half of the money conversation from this week’s issues — Factory’s $5B raise and the Mentlio acquisition said the agent economy has a P&L; this says the lab valuations have a credibility ceiling. Read the two together: the market believes in the spend layer and doubts the equity layer.

Signal: The skepticism is aimed at the IPO, not the product. When the bear case is “gullibility” rather than “the tech doesn’t work,” the disagreement is about price, not possibility. Watch the S-1, not the takes. Sources: CNBC

05 · Leaderboard

The guardrail tax, quantified: 32 points and a 6.6x cost gap

Artificial Analysis expanded its Cyber Index to trusted-access models — models with fewer cyber guardrails — and the first entry rewrites the board. GPT-6 Sol (Daybreak Blue, max), accessible only through OpenAI’s Daybreak program, now leads the Index with no safety blocks across the entire benchmark, scoring 32 points higher than the publicly available GPT-6 Sol (max). The cost number is the real shock: $1.77 per task for Daybreak Blue vs $11.67 per task for Grok 4.7 (xhigh), the next contender. The biggest gains came on CyberGym-E2E, the benchmark where the public model refuses most.

The read: the same model name now hides two products, and the public one is the handicapped one. The 32-point delta is a direct measurement of what the guardrails cost in capability, and the refusal map (CyberGym-E2E is where the refusals concentrate) shows exactly which tasks the public model is not allowed to do. For enterprise buyers the question is no longer “which model is best” but “which access tier am I on” — and the Daybreak program is where the capability actually lives.

Signal: The leaderboard now measures the access tier, not just the model. When the unguarded variant of the same model scores 32 points higher at one-sixth the cost of the runner-up, the “frontier” is a distribution channel. Sources: Artificial Analysis

06 · The Argument

Karpathy’s widening gap, QGIS’s answer

The week’s most-argued thread (568 points, 809 comments) started with Karpathy: the gap in shared understanding of what LLMs can do is widening. And the same weekend gave the counterpoint: QGIS banned AI-assisted contributions (253 points, 259 comments) in its enhancement proposals — an open-source project deciding the provenance question by fiat.

The two stories are the same argument from opposite ends. Karpathy’s camp says the capability map is unknowable to most people; QGIS’s says we know enough to say no. Read against this issue’s Papers: the hidden-capabilities proof gives the skeptics their strongest line — if a model can provably hide arbitrary capabilities from elicitation audits, then nobody’s shared understanding includes the full map. And read against the Leaderboard: the public model and the trusted-access model are nominally the same weights 32 points apart. Whose understanding is widening from whose?

Signal: The capability conversation has split into three camps: the believers (the swarm will find 66 of 70), the skeptics (the model may be hiding the 4 it missed), and the governors (QGIS: not in our repo). All three are pricing a different risk. Sources: Karpathy’s thread · QGIS enhancement proposal

07 · Papers

AI models can provably hide arbitrary capabilities

Draguns, Motwani, Douglas, and de Witt (Oct 9) put a bound on the whole evaluation enterprise. Capability audits assume hidden capabilities can be elicited; they construct backdoor attacks that embed encrypted neural networks inside host model weights — digital lockers that decrypt and execute only on a secret trigger — and prove these hidden circuits are hard to elicit and hard to interpret under cryptographic assumptions, even given full access to the weights. This extends prior hiding results from fixed strings to arbitrary computations. The proof of concept (in PyTorch, models released) resisted fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training on a chemical-reaction-prediction proxy task, the CBRN-shaped canary.

Now read it against last issue’s “Caught in the Act” paper (98.8% AUC deception probes that improve as models scale). The probes watch activations; the lockers say the dangerous circuit does not decrypt until it sees the trigger, so there is nothing in the activations to watch. The 99.7% introspective-deception number is the inside view of the model in front of you; the cryptographic bound says the model in front of you may not be the whole model. The honest synthesis: monitoring beats auditing for what the model does, and neither beats a model built to hide.

Signal: Elicitation-based audits have a proven ceiling now: anything an evaluator can compute can be made cryptographically invisible to that evaluator. The “limit of elicitation-based capability audits” is the paper’s own phrase. Watch whether the labs’ safety cases update their threat models this week. Sources: arXiv:2610.13928

Can language models play the Wikipedia Game?

Hickey, Knight, and Bishop (Oct 9): when retrieval has to cross specialist linguistic or scientific knowledge that cannot be indexed wholesale, language models increasingly do the navigating. The Wikipedia Game — get from one article to another by links alone — is the playable version of what agents now do inside document collections all day. Light reading, but it names the actual job the swarm in The One Thing is being sold for.

Signal: The cute benchmark is the commercial workload: every agent fleet is playing the Wikipedia Game inside some private corpus. Sources: arXiv:2610.13948

08 · GitHub Trending

The boards

Daily: morluto/rea is now a phenomenon — 45,495 stars today, 138,675 total, the agentic reverse-engineering toolkit that keeps doubling. New: storytold/artcraft (20,422 stars, +4,106 today), an intentional crafting engine for artists, designers, and filmmakers in Rust; mksglu/context-mode (26,933 stars), context-window optimization for coding agents claiming 98% tool-output reduction via sandboxed MCP + hooks (claim, not benchmark). AirLLM (+886) and cathrynlavery/diagram-design (+2,100) keep climbing; Alishahryar1/free-claude-code (58,501 stars) keeps selling the free-token arbitrage across harnesses.

Weekly: boykopovar/AnyPS5 (31,952 stars, +25,151 this week) ports PS5 executables to Linux and Windows; tester-army/e2e (+6,440) is a next-gen testing framework; mattpocock/skills (287,416 stars) still leads the agent-skills economy; heygen-com/hyperframes (61,515 stars) is “write HTML, render video” built for agents; thedotmack/claude-mem (99,978 stars) does persistent cross-session agent context; earthtojake/text-to-cad and mvschwarz/openrig hold the agent-tooling lane. Also: Anthropic’s free prompt-engineering course crossed 38,000 stars.

Signal: The weekly board is now three businesses: agent tooling (skills, memory, context), agent media (HTML-to-video, diagram design), and agent infrastructure arbitrage (free tokens, local inference). The star economy has fully moved up the stack. Sources: GitHub trending

09 · Product Hunt

Yesterday’s top products (Oct 11)

1. HireKey — Your recruiter-informed AI copilot for tech job search (275 votes)

2. kōdlo — End-to-end sales workspace with unified context (251 votes)

3. Automations in Perplexity Computer — Automations that remember, act, and keep work moving (183 votes)

4. Ghost Core — A personal AI computer that runs entirely on-device (168 votes)

5. OpenScreen 2 — Record your demo and let your AI agent do the edits (160 votes)

6. MePersona — Answer every fan comment in your voice (146 votes)

7. Ayven — Carries and drives your team’s agenda (123 votes)

8. FocusWrite — A local-first writing app with focus modes (114 votes)

9. Foley — A voice to-do list that files, plans and asks first (102 votes)

10. Gboard Kurukuru Version — Your fingers stay put. The keys come around (89 votes)

Two worth a click: OpenScreen 2, the record-once-let-the-agent-edit-it demo tool, is the consumer face of the agent-media wave on the GitHub board; Edgee Mod for Claude Code (#11, 84 votes) lets you run any model inside Claude Code while keeping your Claude plan — the harness-arbitrage economy from GitHub trending, now with a product page.

Signal: The board’s real theme is the agent economy’s plumbing: copilot for job search, sales workspace with unified context, on-device AI hardware, agent-edited demos. The launches assume the agent is the user. Sources: Product Hunt

10 · Hacker News

Passkeys are a trap

833 points, 444 comments. The case against passkeys as a consumer identity primitive, and the thread’s 444 comments are the good kind of angry: implementers, skeptics, and the people who have actually shipped WebAuthn at scale.

Total Annihilation, rebuilt for modern computers

596 points, 223 comments. A faithful modern rebuild of the 1997 RTS, and the comments are a masterclass in what “faithful” costs when the original ran on a Pentium.

Fingerprints are the new cookies

539 points, 264 comments. Read against the passkeys thread: the identity layer is being rebuilt as lock-in while the tracking layer is being rebuilt as fingerprinting. The week’s two identity stories say the same thing from opposite sides.

Terence Tao: Math 2.0

455 points, 457 comments — and it re-trended (it sat at 61 points in yesterday’s Also notable). The Caltech slides on AI and mathematics keep pulling mathematicians into the thread; read against The Argument and the Papers math-validity thread.

Also notable

The gap in shared understanding of LLM capability (568 points, 809 comments): covered in The Argument. QGIS bans AI assisted contributions (253 points): covered in The Argument. Doory: Open-Source Canva (165 points). Show HN: PS2 games statically recompiled to WebAssembly (132 points): the recompilation renaissance continues.

11 · OpenRouter Apps

The board this morning

The token board has a new leader: Command Code jumped from ~0.69T to ~4.18T tokens overnight, a ~6x surge that deposes ZCode (~1.91T, flat after yesterday’s 4.4x spike). Hermes Agent (~1.59T) and Claude Code (~1.56T) hold their lanes; Kilo Code (~0.96T) and Cline (~0.54T) grind on. StepCode (StepFun’s agent, created October 10) holds at ~0.70T two days after launch; codex-fleet conductor debuts at ~0.38T. Codex itself slipped to ~0.32T.

Signal: Yesterday’s story was ZCode’s spike; today’s is Command Code’s 6x overnight jump. The Chinese-lab agents are not just eating the board, they are taking turns. The question from last issue stands, sharper: which of these spikes is launch weather and which is the new baseline? Sources: OpenRouter rankings

12 · Watch

TwoMinutePapers: “DeepSeek Solved AI’s Memory Problem”

TwoMinutePapers · Oct 11 · Károly’s take on the DeepSeek memory story: whatever the actual mechanism, the genre’s bet is that the next 10x is not a bigger model but a model that remembers. Watch against The One Thing: the fleet that cannot remember its phases is 1,000 amnesiacs.

TheAIGRID: “Microsoft’s MASSIVE AI Announcement Has Everyone Talking”

TheAIGRID · Oct 11 · the hype-cycle episode: Microsoft’s latest announcement wave, with the usual discount for the genre. Watch the claims against the Leaderboard numbers, not the thumbnails.

Prism Labs: “Claude Now Orchestrates 1,000 AI Agents in One Run”

Prism Labs · the fleet episode: a full breakdown of Anthropic’s dynamic-workflows launch — the multiagent_20261001 config, the 66-of-70 bug-hunt numbers, the cost reality, and the subagents-vs-workflows decision. The video companion to The One Thing.

13 · Listen

From the queue

No new episodes since yesterday’s issue — the queue still holds Hard Fork’s “Anthropic’s Quest to Give A.I. Morals,” Latent Space’s AlphaFold and Periodic Labs episodes, and Lenny’s Josh Woodward interview. Backlog day; the Watch section has the fresh material.

14 · Read

Claude’s managed agents quickstart

The official quickstart for the fleet in The One Thing.

The trusted-access Cyber Index

The 32-point guardrail tax, against Leaderboard.

The hidden-capabilities proof

Digital lockers and the ceiling on elicitation audits, against Papers.

Which harness components actually help

The study breaking down which coding-agent harness components improve performance, against Papers.

15 · Safe to Ignore

Permission to skip

Skip: Citroen revives the 2CV for an electric age. 87 points of charming non-AI news. Wrong newsletter. Skip: Chapter 1 of “Ways of Seeing” by John Berger. 131 points of art history. Wrong newsletter. Skip: Secret of Orange Cats Revealed After 60-Year Search. 95 points of science. Wrong newsletter. Skip: Solar panels that plug into your wall. 11 points of balcony-energy optimism. Wrong newsletter. Skip: An unmade film shaped Hayao Miyazaki. 3 points. Wrong newsletter. Skip: GitHub’s Actions/PRs/Issues incident. 68 points of infrastructure pain. Wrong newsletter, right sympathy.

16 · Build on This

Authorship

QGIS just banned AI-assisted contributions, and the trending #2 model on Hugging Face is a text humanizer built to dodge detectors. The provenance question is now a product. Authorship is commit-level provenance attestation: a lightweight agent or editor plugin that records, at commit time, which lines were human-written, which were AI-generated, and which were AI-assisted, signed into the commit metadata. The buyer is not the solo dev, it is the maintainer of any project that now has to enforce an AI-contribution policy — and after QGIS, every serious open-source project is writing one. Pricing follows the trust playbook: per-repo, pay-once per audit, no subscription — the attestation is the product, and it should cost less than the argument about it. The wedge is policy enforcement as a service: start with the projects banning AI code today, expand to the enterprises that need to prove provenance for compliance tomorrow. The moat is the signature graph: every attested commit is a node in a provenance map no one else has, and the map is worth more than the tool.