FOMO Buster #017: Judgment for sale

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

Deno joins Cloudflare: the standalone runtime has a year to live

Ryan Dahl’s October 9 announcement is phrased as “the entire Deno team is joining Cloudflare,” and the coverage calls it what it is. The numbers that matter: the Deno runtime gets one more year of monthly bug-fix and security releases, then company development ends (it stays open source, and community forks are welcome); Deno Deploy shuts down in six months, with migration support moving paying customers onto Cloudflare Workers; JSR keeps running, infrastructure moving to Cloudflare; rusty_v8 continues, with work toward folding it into Cloudflare’s workerd. The team merges with the Workers and Durable Objects teams to make the celld/Workers programming model the default, whether on Cloudflare’s network or self-hosted (celld launched in August). No price or terms disclosed.

The read: the edge-runtime war ended in consolidation, not standards. WinterCG is never mentioned in the announcement; the winner is the programming model, not the committee. And the AI angle is explicit: Dahl invites people building agents at scale to email him directly, because Durable Objects (cheap serverless execution, persistent state, WebSockets, a JS interface) is the runtime agent harnesses want. Cloudflare did not just acquire a runtime, it acquired the developers who would otherwise have built the agent layer elsewhere. Yesterday’s 187-point card becomes today’s 1,287-point confirmation.

Signal: When the company that runs the edge absorbs the runtime built to fix Node, the “run anywhere” JavaScript pitch consolidates into one company’s programming model. Watch the forks: Deno stays open source, and a year of security patches is a long runway for a community takeover. Sources: Deno announcement · RuntimeWire coverage

02 · Lab Watch

Anthropic opens the cyber arsenal in three tiers

The Cyber Verification Program now gives vetted security professionals reduced blocking classifiers on Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1, in three tiers. Defense Access covers defensive work (SOC/IR, malware reverse engineering, vuln analysis); individuals with a track record of reported vulns can qualify. Red Team Access adds authorized penetration testing against systems the org may test, with real-time blocks remaining for physical harm or mass disruption; organizations only. Specialized Access has the fewest blocks, for testing systems where failure costs lives or markets (flight operating systems, power grids, interbank transfer); every org is reviewed in depth with the US government. The CyScenarioBench numbers: without CVP every task was blocked on the first prompt; in Defense, 46 of 50 trials were blocked at some point and 4 succeeded; in Red Team, 34 of 50 completed, matching the no-safeguards baseline. The Glasswing ledger: partners uncovered 129,000+ verified vulnerabilities Apr-Jul 2026, Anthropic’s own open-source scanning added 5,500 more, 33,000+ rated critical or high, with true impact estimated at 5x. Separately (Oct 9): dynamic workflows in Claude Managed Agents run up to 1,000 agents in parallel and found 66 of 70 hidden bugs in a 116,000-line codebase (a single agent caught 14-27; the model was not named).

Signal: The labs are deregulating cyber capabilities for the vetted while the usage policy re-regulates everyone else: offense is becoming a credential, not a taboo. Watch who qualifies for Specialized Access; that list is the real story. Sources: Anthropic announcement · The Decoder on dynamic workflows

Asana cuts model costs 76x in browser tests with GPT-6.1 Sol

New on OpenAI’s case-study feed: Asana cut model costs 76x in browser tests by moving to GPT-6.1 Sol. Read against the Ultrafast tier from yesterday’s issue (8x speed, 6x price): the labs are selling a menu of price-speed tradeoffs, and the enterprises are finally publishing which line items they picked. The case study is the continuation of the cost-per-hour theme from the DeepSeek 4.1 Flash essay: the frontier premium is fine when the work justifies it, but a 76x line-item cut moves budgets. The enterprise AI story has fully moved from pilots to published unit economics.

Signal: The judgment of an API is now the receipt, not the benchmark: enterprises publishing 76x cost cuts and 96% investigation-time cuts (Sophos Daybreak) are the new benchmark suite. Sources: OpenAI case study

Meta: Muse for Small Business

Meta’s announcement feed surfaced an older but unmissed launch: Muse for Small Business, extending Meta’s AI gadgets program to small businesses. The Muse line (the dongle giveaway from the September issues) is finding its enterprise-distribution channel. The strategic read stays the same: hardware as ecosystem conscription, and the small-business tier is where that conscription pays for itself in usage data.

Signal: Meta’s giveaway hardware is moving from influencers to install bases: small businesses get the gadgets, Meta gets the usage. Sources: Meta announcement

03 · Open Weights

The board this morning

google/embeddinggemma-2 takes the trending crown (score 1,377; 1,444 likes), displacing the decision models: autotrust/JEV-27B-VL and GEV-26B-Decide drop off the trending board entirely. Cloudflare/clef holds at #2 trending (882; 1,970 likes), Cloudflare/clef-flash #9. New climbers: jialinyyzz/humanizer (#3, 833), Aleph-Alpha/Kolibri-1 (#4, 680), the 200T-to-20T training-data filter model, still climbing. Image generation stirs: Lightricks/LTX-2.5 at #7 trending with 7,126 likes and Qwen/Qwen-Image-2.1-Turbo at #8. The all-time board: Qwen/Qwen3.8-27B at 17,391 likes, unmoved; FLUX.1-dev, DeepSeek-R1, Kimi-K3 all still absent from trending. The decision-model moment cooled; embeddings and video are this week’s gravity.

Underdog Saluki 27B: 7.89 GB that beats the 54 GB original at tool calling

Conway Research released a 2-bit mixed-precision GGUF of Qwen3.8-27B: 7.89 GB against 54 GB for the BF16 original, about 7x smaller, running on stock llama.cpp with full GPU offload, Apache 2.0 licensed. The build stacks three quantization passes: the Qwen3.8-27B base, ISTA-DASLab’s GSQ/RCO build (their smallest was 8.4 GB), then Underdog’s own pass to 7.89. The numbers: 96% average retention across 9 benchmarks, and the headline is real: parallel tool calls 42 vs 35 for the full model, 120% retention. The compression was deliberately tuned to protect tool calling (the agent skill) at the expense of competition math (AIME 2025: 79.2 vs 96.7). Read against this week’s judgment theme: at 2 bits, the model keeps what the agents need and drops what the leaderboards want. Quantization is becoming a statement about which capability matters.

Signal: A quantized model can now beat its full-precision parent where it counts for agents. The compression knew what to keep. Sources: Model card on HF · MarkTechPost writeup

04 · Money Flow

Typesafe AI raises $870M at $7.5B for a model that doesn’t generate text

The round: $870 million at a $7.5B valuation, a16z leading, Sequoia and DCVC participating, Martin Casado joining the board. And it is a Series A, closed roughly three to four weeks after the product launch: Jev shipped September 15 alongside a $40M seed from DCVC, so total funding is about $910M. Jev is a “System One model”: transformer-based but not an LLM, no text output, returns structured, typed decisions (choice, score, probability) with an attached confidence score, sampled in parallel rather than token-by-token. Trained with “Reinforcement Learning for Calibrated Decisions.” The company claims: a third of Fortune 500 use it (no customers named), 1M users within days of launch, millions of dollars saved in production. Founders: Diogo Almeida (CEO, ex-OpenAI, InstructGPT-era), Sasha Sheng (ex-Meta research), Erik Gafni.

The read: this is the judgment-layer thesis with a $7.5B receipt. Yesterday’s issue noted three sellers at the $0.10-per-million judgment spot price; today Microsoft undercut it at $0.042 (see The Argument), and Decision-1 claims its best score on JevBench, Typesafe’s own benchmark. The decision-model stack now has a benchmark, a price war, and a nine-figure-funded pure play, all inside one week. Note the round came in below the rumored $1B+ at $10B+ valuations from earlier coverage: even in a hot week, someone negotiated.

Signal: When the biggest AI raise of the week goes to a model that outputs decisions instead of text, the agent economy’s unit of sale is officially the judgment, not the token. Sources: Typesafe announcement · TechCrunch

05 · Leaderboard

Artificial Analysis: Mistral Large 4, Gemini 4 Argon, and the eval gaming turn

The changelog’s notable entries: Mistral Large 4 released Oct 7, with AA’s headline that it makes “France home to the most intelligent model”; Gemini 4 Argon (Sept 30) puts “Google back as one of the top three labs in intelligence achieved”; plus Ling 3.1 Flash, Grok 4.7 (Low), Solar Mini 4, and the full Haiku 5.5 effort ladder. Two days after this issue argued that benchmarks are adding trust dimensions faster than capability dimensions (Harvey LAB-AA v1.1’s hallucination checks), the papers section below lands the other shoe: the psychometric audit of HarmBench says the safety numbers never measured a single thing.

Signal: The leaderboard and the audit are converging: capability scores keep climbing while the trust scores get interrogated. Read the rankings as “who is fastest” and the papers as “who can be checked.” Sources: Artificial Analysis changelog

06 · The Argument

Microsoft’s Decision-1: 35x faster judgment at $0.042 per million

Microsoft’s purpose-built decision model (announced October 9, via Satya Nadella on X) is a post-trained Qwen3.5-9B doing single-pass scoring: a fixed set of options in, calibrated probabilities out, with yes/no, multiple-choice, rating, and rubric-grading modes. The vendor numbers: 35x faster than GPT-6 Sol at P50 latency, 4.5x faster than the runner-up; $0.042 per million input tokens, output tokens free. Xbox Research labeled 10,000+ feedback items at GPT-6-Sol-competitive quality, 14x faster and 200x cheaper. Highest accuracy in a 36-benchmark, ~150,000-question comparison, and best on JevBench across 36 benchmarks. Available now in Microsoft Foundry, coming to OpenRouter; Microsoft plans to rebase future versions on Microsoft AI and OpenAI tech.

The read: the judgment price war is real, and it opened below the spot price. At $0.042 per million input with free output, Decision-1 undercuts the $0.10-per-million judgment tier (OpenAI’s Decisions API, Anthropic’s Haiku 5.5) by roughly 58% on input alone. And the benchmark politics are delicious: the best-scoring model on JevBench is not Jev, it is Microsoft’s. Watch for Typesafe’s response and for OpenAI to move Decisions API pricing off $0.10.

Signal: Judgment is a commodity with a spot price, and the spot price just fell 58%. Every agent decision now has a cost floor you can calculate. Sources: TestingCatalog overview

Thomas Hales on Tao’s blog: never believe a Lean proof until the kernel checks it

Thomas Hales, guest-posting on Terry Tao’s blog (October 9), delivers the measured credibility check the autoformalization wave needed. The setup: Lean’s mathlib is ~300,000 theorems, 2.5M lines of code; 2026 was the year autoformalization became practical (Anthropic’s FLT: 13 million lines of Lean in 11 days; OpenAI’s Navier-Stokes, accompanied by an autoformalization). The warning: the “Summer of Soundness Bugs”. In July and August, frontier AI models in security researchers’ hands found Lean soundness bugs, including one that produced an illicit disproof of the Collatz conjecture and another with a short illicit proof of the Kepler conjecture. Hales’ demands: Lean proofs “should never be believed until they have been checked by the kernel,” and no proof accepted until a human audits statement fidelity (does the Lean Navier-Stokes match Fefferman’s Millennium statement?). The defenses: ~25 Lean kernels for cross-checking (Navier-Stokes confirmed by 12+), Joachim Breitner’s Con-Leche verified kernel (September 10; code and consistency proof generated by Claude), and fixing the metatheory (Carneiro’s 2019 thesis has an error and targets Lean 3).

Read against the SCI-infinity paper from this week: two fronts in the math trust war. Bastounis et al. attacked translation fidelity (the bridge between English and Lean); Hales attacks kernel soundness (the checker itself). The honest position: machine-speed math needs both, and Hales explicitly credits the AI-assisted bug hunting, including OpenAI’s Dan Selsam, as a positive development. (Authorship note from the post: fully human, with AI used as a tool for search, fact-checking, and proofreading.)

Signal: The green-checkmark era is over before it began: a Lean proof needs a kernel check, a human fidelity audit, and a soundness story, in that order. Sources: The full post

07 · Papers

Searching for “Harmful Refusal”: a psychometric audit of an AI safety benchmark

Stewart et al. (Oct 8) do the thing nobody does to safety benchmarks: they ask whether the number measures anything at all. Using item response theory and differential item functioning analysis on HELM Safety, they find HarmBench does not measure a singular attribute: models with similar overall scores have very different attribute profiles, three of four candidate datasets are saturated, and the single-score aggregation hides it. The striking bit: models from different developers with the same estimated refusal ability still score differently on 13 items, and most of that developer signal collapses when items are scored in their proper scopes (standard, contextual, copyright), which is exactly what aggregation artifacts look like. Read against yesterday’s issue: the Harvey LAB-AA v1.1 hallucination checks add a trust dimension, but this paper asks whether trust dimensions can be scored at all. The honest line: a score should earn its single-attribute reading before anyone compares models with it.

Signal: The benchmark war’s trust turn has a measurement problem: if the safety score conflates distinct behaviors, “who can be checked” is unanswered. Psychometrics, not more benchmarks, is the missing discipline. Sources: arXiv:2610.12409

GeoReform: formalization as an optimizable policy

Wang et al. (Oct 8) treat converting geometry diagrams into formal representations as a policy to be optimized, not a parser to be built: run the full reasoning pipeline, collect failed rollouts, diagnose defects in the representation, mutate the policy. The non-obvious detail: naive “extract more facts” formalization is net-negative on Geometry3K, structure injection fixes 28 errors but introduces 13 new ones across 200 examples, because redundant relations distract and ambiguous references mislead. The policy loop takes Qwen3VL-2B from 42.0% to 56.0% on Geometry3K. Read against The Argument: faithful formalization is the bridge everyone is fighting over, and the winning move is optimizing the representation itself, not the parser.

Signal: In autoformalization, more facts are not better: the representation is the product, and optimizing it beats extracting more of it. Sources: arXiv:2610.12391v1

08 · GitHub Trending

rea: four days at the top, 63,832 stars

morluto/rea added 25,784 stars today to reach 63,832, its fourth straight day on top. Reverse-engineer anything with agents, from app behavior down to native binaries. The same story hit HN this morning at 536 points (see below). In every enterprise with a legacy codebase, “explain this to me” beats “write this to me,” and the market is voting with stars.

The boards

Daily: boykopovar/AnyPS5 (25,096; +5,831 today), automatic PS5 executable porting to Linux and Windows. storytold/artcraft (13,291; +3,217 today), an intentional crafting engine for artists, designers, and filmmakers, the anti-slop entry. cathrynlavery/diagram-design (48,581), 44 editorial diagram types, still no Mermaid slop. mksglu/context-mode (26,103), context-window optimization for coding agents claiming 98% tool-output reduction, session memory, and routing across 17 platforms via MCP + hooks. Weekly: thedotmack/claude-mem (99,133; +3,850 this week), persistent context across sessions for every agent. EpicGames/raddebugger (8,309; +654 this week), a native user-mode graphical debugger from Epic. heygen-com/hyperframes (60,170; +4,003 this week), HTML-to-video for agents.

09 · Product Hunt

Yesterday’s top 10 launches

1. Busabase — The general system of record for AI agents (343 votes)

2. Zernio — Marketing infrastructure for products and agents (322 votes)

3. Playground by Google Labs — Create, play, and share your own games with Google AI (217 votes)

4. Together Link — Use open models in your harness (184 votes)

5. Odyssey 3 — A world model that simulates physics in real time (158 votes)

6. Gemini Agent for Google Cloud — The work agent that runs on both Gemini and Claude (149 votes)

7. Staffcoder — Build like a real engineer (108 votes)

8. Refs — Give Claude better references for your videos (105 votes)

9. AgentSDR — Open-source AI SDR (104 votes)

10. OpenPilot — Open-source desktop AI agent for any model you choose (96 votes)

Three worth a click: Busabase, the system-of-record-for-agents pitch lands the same week agents got receipts in this issue’s Build on This. Odyssey 3, a real-time physics world model is the infrastructure play behind the next wave of simulators. Gemini Agent for Google Cloud, the work agent that runs on both Gemini and Claude: the multi-model harness is now a Google product, not a hack.

10 · Hacker News

Cloudflare acquires Deno

1,287 points, 663 comments. Covered in The One Thing. Up from yesterday’s 187-point card: this is now the confirmed acquisition story, not a rumor.

Typesafe AI raises $870M at $7.5B

407 points, 321 comments. Covered in Money Flow.

Triple-A Minesweeper

1,147 points, 228 comments. A gloriously overproduced Minesweeper. The week’s best argument that craft still gets clicks, no AI required.

REA Reverse - Engineer Anything

536 points, 233 comments. The GitHub juggernaut’s front-page moment. 63,832 stars and counting; the comments are where the legacy-codebase crowd tells you what they’d point it at first.

Telegram Desktop one-click account takeover: CVE-2026-107181

BeakSec’s writeup (October 3) chains two defects: an IPC injection (Telegram Desktop forwards clicked links to its running instance over a local socket using ; as the command separator, without escaping it) and a legacy internal interpret: URI scheme (a release-publishing helper that reads an instruction file and sends the file it points to, with no authorization or confirmation). The attack: add the victim to a supergroup, post instruction files (auto-downloaded to a predictable path), get them to click one crafted link; the injected commands exfiltrate files with no dialog. Three files equal the account, and with no local passcode set (the default), the encryption key derives from an empty string. Severity 8.1 High; fixed in 7.2.9, which shipped quietly in September with only a rendering fix in the changelog. Mitigations: upgrade, turn on “ask where to save each file,” restrict group adds to contacts, set a real local passcode.

Signal: The scariest bugs are the boring seams: a URL separator nobody escaped and a debug URI scheme nobody removed. Sources: BeakSec writeup

‘123456’: the Danish CPR breach ran through a delegate, and the password claim is unconfirmed

The scale is confirmed: roughly 8.8 million CPR records (names, addresses, numbers) were pulled through a private company’s legitimate access to Denmark’s Central Person Register, exceeding the ~6M living population because the register holds the deceased and emigrated. The register itself was never breached: it is a compromise of delegated trust, through authorized channels. The ‘123456’ detail is a hacker’s unconfirmed claim (three accounts at a Funen-based IT firm, per Politiken), not an official finding; authorities have not confirmed it. The ministry’s warning stands regardless: CPR numbers are permanent, so treat any caller who knows yours as untrusted. People with registered name-and-address protection were not exposed.

Signal: The breach pattern of the year is not breaking the vault, it is abusing someone’s key to it. Password hygiene at the delegate is the whole game. Sources: The Brief on the 123456 claim

Also notable

Talorys, a self-hosted personal AI agent on Cloudflare’s free tier (102 points): read against The One Thing, the edge is already the agent’s cheapest datacenter. Apple/macOS silently removed from official Unix registry (101 points, 104 comments): the certification lapsed, the memes didn’t. Computers Cannot Make Decisions (150 points, 126 comments): a philosophical counterweight to the judgment-model price war, read against The Argument. Show HN: Carrier-Explode (367 points): iPhone, Pixel, and Galaxy carrier settings decoded. Pointing AI at archives found a forgotten meteorite, lost rhinos, and more (168 points): the best use of retrieval this week. Jane Street on autoregressive diffusion for market data (134 points): the quant side of generative modeling. What mathematicians should know about the Lean Theorem Prover (154 points): covered in The Argument.

11 · OpenRouter Apps

The full board this morning

By tokens this morning: Hermes Agent (~1.99T, back to #1), Claude Code (~1.55T), Nous Research API (~1.29T, the 3.7x jump still deflating), Command Code (~1.11T), Codex (~0.71T, still climbing), Kilo Code (~0.63T), Cline (~0.55T), Freebuff (~0.45T), ZCode (~0.43T), pi (~0.39T), codex-fleet conductor (~0.36T, back on the board), LangChain (~0.35T, new to the board), OpenClaw (~0.21T), DeepSeek Harness (~0.17T), OpenHands (~0.06T). Three straight weeks of reshuffling: Hermes and Claude Code hold on persistence, Nous fades, Codex climbs, and the new entrant is the oldest name on the list, LangChain, which is either nostalgia or a genuinely new harness play.

Signal: Still a sentiment market: when the board’s newest entrant is the original orchestration framework, you are watching launch weather. The durable positions remain the agents with session persistence. Sources: OpenRouter rankings

12 · Watch

TwoMinutePapers: “DeepMind’s New AI Just Cracked The Code Of Life”

TwoMinutePapers · Oct 7 · the science episode. Watch against Money Flow and the Lab Watch theme of scientists as users.

Fireship: “A $6.3 billion open-weight model just got embarrassed by the French…”

Fireship · Oct 7 · the Mistral Large 4 episode. Watch against the Leaderboard: France is home to the most intelligent model, and the price of open-weight pride.

TheAIGRID: “Anthropic Has Officially Lost Its Mind…”

TheAIGRID · Oct 9 · the dashboards-and-motion reaction. Watch against The One Thing and Lab Watch: the workspace-era critique, volume up.

13 · Listen

From the queue

The podcast feeds were quiet this morning, so two still worth your ears: Latent Space on Periodic Labs’ synthesis superintelligence (semiconductors to superconductors, against the Money Flow physical-substrate theme) and Lenny’s Podcast with OpenAI’s head of ChatGPT on the new era (the consumer strategy behind GPT-6 for everyone).

14 · Read

The Deno announcement, in full

One year for the runtime, six months for Deploy, and the programming-model consolidation, read against The One Thing.

Typesafe’s “Series AI” post

The $870M at $7.5B round and the System One pitch, read against Money Flow.

Hales on Lean reliability, in full

The summer of soundness bugs, the kernel-check rule, and the Con-Leche verified kernel, read against The Argument.

15 · Safe to Ignore

Permission to skip

Skip: A coffee machine used 1TB of data in 10 days. Still the wrong newsletter. Skip: Theranos.world. 178 points of domain-name schadenfreude. Wrong newsletter. Skip: The role of cat eye narrowing movements in cat-human communication. 109 points of adorable science. Wrong newsletter. Skip: Bitwarden’s dual license model. 23 points of licensing news. Filler. Skip: Clinical trial of a prion disease drug candidate. Great news. Wrong newsletter. Skip: “I hired an illustrator” home dashboards. Already in the archive. Move on.

16 · Build on This

The agent tax audit

This week the agent economy got its first itemized bills: Vals AI found multi-agent coding teams cost 5x more while rarely scoring better; OpenAI’s case studies published a 76x model-cost cut and a 96% investigation-time cut; the judgment models landed at a $0.10-per-million spot price. Nobody sells the audit. Agent Tax is a drop-in spend-control plane for AI-agent harnesses: sit between the harness and the model APIs, meter dollars per task, enforce per-project budgets, auto-route to the cheapest model that clears the eval bar, and flag the 5x waste with receipts. Pricing follows the week’s logic: a cut of the savings, pay-as-you-go, no subscription, because the product proves its price every invoice. Start with the agent-heavy shops burning through OpenRouter credits (the board above is your lead list); expand to any team where “the agents ate the cloud bill” is the CFO’s new complaint. The moat is the routing dataset: every task logged is a data point on which model is actually cheapest at a given quality, and that map is worth more than the software. When the agents get expensive, sell the audit.