FOMO Buster #009: The decision goes open

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

Clef: Cloudflare’s open-weight decision models make the 38-millisecond call

Cloudflare released Clef and Clef-flash: open-weight, multimodal decision models that return calibrated probabilities over typed schemas instead of free-form text — 38ms typed decisions for AI agents (204 comments on the HN thread). They’re already climbing Hugging Face: Cloudflare/clef sits at #4 trending, with Cloudflare/clef-flash at #16.

This is the open answer to the week’s running theme. OpenAI shipped decisions as an API (Decisions API), YC shipped them as an engine (Magnitude), and now the open-weights world ships them as weights you can run yourself. The same day, LiquidAI’s d1 landed on Vercel’s AI Gateway with zero output-token costs — “a cheap drop-in for classification, routing, and scoring work.” Two decision models, two distributions, one message: deciding is becoming a model category distinct from generating.

The bet both are making: most of what agents call “thinking” is actually sorting — route this, score that, classify the other thing — and sorting doesn’t need a frontier model burning tokens per word. If they’re right, the cost floor of every agent stack just dropped.

Signal: The week’s theme closes its loop: decisions are the new primitive, and now they’re open weights with a price of zero tokens. Watch which frameworks make Clef the default router — whoever wires it into the skills stack first owns the cheapest path to “good enough” judgment. Sources: Cloudflare: Clef decision models · AlphaSignal: 38ms typed decisions · AlphaSignal: LiquidAI’s d1 on Vercel

02 · Lab Watch

Pi 1.0

The open agent toolkit shipped 1.0 — 506 comments on HN, the day’s biggest fresh thread. The same day came Pi Durable (earendil.com), its answer to the oldest agent complaint: runs that die lose everything. Pi is now a serious gravity well: ~111K GitHub stars, ~451B tokens on OpenRouter. The open harness is no longer the scrappy alternative; it’s infrastructure with a version number.

Claude Code Mods let developers rewrite the agent from the inside

Anthropic now lets developers write TypeScript mods that intercept Claude Code’s events, redraw its UI, and replace built-in features through the plugin system. The IDE’s most successful agent is becoming a platform you can reprogram — the same week its OpenRouter token count keeps climbing. The agent you can rewrite beats the agent you can only prompt.

DeepSeek Harness Desktop for macOS and Windows

167 comments on HN. DeepSeek is shipping its harness as a desktop app — the lab’s agent stack escapes the terminal and the browser. Read it next to the OpenRouter board, where DeepSeek Harness holds ~268B tokens: the labs aren’t just training models anymore, they’re shipping the operating layer around them.

03 · Open Weights

Qwen-Image-2.1 owns the image week; laya holds the like-crown at zero downloads

Qwen/Qwen-Image-2.1 is everywhere at once: #5 on trending (81K downloads), an uncensored GGUF at #3, Viggle’s turbo variant at #9, a Comfy-Org build at #26 — and it just topped two open-weight image leaderboards, sweeping 16 of 36 categories and dethroning Ideogram and HunyuanImage among open weights. The image-model replacement cycle is now measured in days, not quarters.

convaiinnovations/laya still leads trending with 4,921 likes and zero downloads — six weeks of attention without adoption, the purest “like, don’t run” model on record. Meanwhile Qwen/Qwen3.8-27B remains the rare double: #1 on the all-time liked board (16,756 likes) and #8 trending — gravity well and climber at once.

The delta list — trending climbers not yet established: Cloudflare/clef and clef-flash (the One Thing), Contrastive-LM/CLM-v0.1-8B, nvidia/Nemotron-3-Diarization, prism-ml/Ternary-Bonsai-2-27B-gguf, SupersonicLabs/Julia-1, PSRben/VisionHOPE, TaichuAI/ZDTaichu5.0-9B, Edge0/Audio8-ASR-Infinite, and Altworld/Hemmingway-1. And the churn exhibit: DeepSeek-V4.1-Flash fell out of trending entirely last week and is already back at #22 — while Apple’s LensVLM-9B, last week’s curiosity, has vanished from the board. Spike, vanish, return: the trending list is weather, not climate.

Signal: Two patterns are hardening: image models now rotate weekly (plan your stack for swapability, not loyalty), and the decision-model tag is appearing in trending metadata for the first time — Clef may be the start of a whole category on the board. Sources: HF trending · Qwen/Qwen3.8-27B · AlphaSignal: Qwen-Image-2.1 sweeps leaderboards

04 · Money Flow

Amazon shops $8B of Nvidia chips to investors; the RAM squeeze runs through 2028

Amazon is seeking to offload $8B of Nvidia chips to investors, per FT reporting via Reuters — the hyperscaler recycling silicon to keep the buildout going without carrying all the capex. The AI boom’s financing layer is getting creative: sell the chips, keep the capacity, let someone else hold the depreciation.

On the supply side the picture darkens: memory executives now expect the RAM shortage to continue through 2028 (Ars Technica), extending the Micron warning from earlier in the week. Intelligence gets cheaper while the substrate gets scarcer — and now the biggest buyer is renting its own hardware back.

Signal: Watch for the financing innovation to become the story: offloaded chips, structured capacity deals, memory as the binding constraint. The next 18 months of AI economics will be written in procurement contracts, not benchmark tables.

05 · Leaderboard

Codex is the only thing on the board still accelerating

Codex went from ~699B to ~1.11T tokens in 24 hours — up 59%, now just 2% behind Kilo Code (~1.14T) for the coding-agent crown. It’s the only board leader still accelerating: Hermes Agent (~2.03T) and Framer (~1.33T) both slipped, which reads like a rolling window — and in a rolling window, Codex is eating everything else’s lunch. The trio also reshuffled: Cline (~1.07T) passed Claude Code (~1.06T) for the first time.

Elsewhere on the leaderboards: Claude Sonnet 5.5 reached #2 on the Artificial Analysis Intelligence Index, and Microsoft’s MAI-Transcribe-2-Streaming topped 38 speech models with a 2.5% word error rate delivered 0.13 seconds after speech ends — streaming transcription is now a solved-enough problem that the leaderboard is about latency, not accuracy.

Signal: Two days, two doublings-ish: Codex’s curve is the distribution story of the quarter. If it passes Kilo Code, the “open coding agents” era ends not with a benchmark but with a token count. Sources: OpenRouter rankings · AlphaSignal: MAI-Transcribe-2 tops 38 models

06 · The Argument

Git 3.0’s upcoming SHA-256 default will be a costly mistake

444 comments, the day’s fiercest infrastructure debate. Git is preparing to flip its default hash from SHA-1 to SHA-256, and GitButler’s argument is that the migration cost — broken tooling, hardcoded hash assumptions, content-addressed pipelines, years of scripts that assume 40 hex characters — dwarfs the theoretical security gain. The counter-argument: SHA-1 is broken, the transition has been a decade in the making, and “costly” is what progress looks like.

The honest version of the fight: defaults are the most expensive API in software. Changing one moves cost from the maintainer to every downstream user simultaneously — and the thread is really about who should pay for a security upgrade nobody asked for this quarter.

Signal: The SHA-256 flip is a forcing function with a deadline, which makes it a market: every org with SHA-1 baked into CI is about to need a migration audit. See today’s Build on This.

07 · Papers

KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux

Agents using real security tooling on Kali, measured fine-grained. Lands the same week Google is marketing Gemini 4 Argon as the defensive-cyber model — the benchmark layer is arriving right behind the marketing. Cyber agents are about to be as measurable as coding agents were in the SWE-bench era.

TACO: Ternary Absolute-max Column-wise One-sparse Optimizer for LLM Fine-Tuning

A ternary optimizer for fine-tuning — the quantization wave keeps moving up the stack from inference weights into the optimizer itself. Fine-tuning is getting cheaper at every layer simultaneously.

Embedding Prediction Helps Image Generation

Predicting embeddings as an auxiliary signal improves image generation — the kind of quiet architectural trick that shows up in next quarter’s open image models. File it next to the Qwen-Image-2.1 sweep: the image-model churn has a research pipeline feeding it.

Signal: The paper layer keeps rhyming with the product layer within days — KaliBench behind Argon’s cyber launch, embedding tricks behind the image-model rotation. arXiv is the leading indicator; the trending page is the lagging one.

08 · GitHub Trending

The skills stack fills in: runtimes, memory, and design join the board

The daily board is still the agent-skills takeover — DietrichGebert/ponytail (+1,429 today, 151K stars) leads again, mattpocock/skills (+955) right behind — but the new entries show the stack filling in around the skills: NVIDIA/OpenShell (+584, “the safe, private runtime for autonomous AI agents”) — Nvidia shipping an agent runtime is the enterprise tell. pbakaus/impeccable (+717, “the design language that makes your AI harness better at design”). heygen-com/hyperframes (+584, “Write HTML. Render video. Built for agents”). colbymchenry/codegraph (+241, pre-indexed code knowledge graph that auto-syncs for Claude Code and Codex). mvschwarz/openrig (+691, build agent teams from Claude Code, Codex, and Pi). mksglu/context-mode (+276, sandboxes tool output for a claimed 98% context reduction). Plus google/skills — Google’s official agent-skills repo, the incumbent blessing the pattern.

The weekly board: vectorize-io/hindsight (+17,403, “Agent Memory That Learns”), debpalash/VoiceStudio (+16,114, fully-local ElevenLabs alternative with voice cloning), paperclipai/paperclip (+14,335, manage agents at work), and google/ax (+2,919, Google’s open agentic orchestration runtime).

Signal: Skills were the land grab; now come the layers around them — memory, runtimes, orchestration, design systems. When Nvidia and Google both ship into your category in the same week, it’s a platform, not a trend.

09 · Hacker News

Pi 1.0

506 comments — the day’s biggest fresh thread, and the comments are doing what HN does best: stress-testing whether 1.0 means “stable” or “marketing.” Covered in Lab Watch.

Git 3.0’s upcoming SHA-256 default will be a costly mistake

444 comments of migration-cost arithmetic. Covered in The Argument.

Clef: Open-weight decision models, and new RL fine-tuning platform

204 comments — the decision-model debate, with the RL fine-tuning platform as the sleeper detail. Covered in The One Thing.

DeepSeek Harness Desktop for macOS and Windows

167 comments on the lab’s desktop play. Covered in Lab Watch.

Also notable

Gemini 4 Argon (1,157 comments and still climbing): yesterday’s One Thing keeps growing — the Fairwind gating debate has legs. Several vulnerabilities discovered in the Linux kernel (309 comments): real security news, wrong newsletter, but the KaliBench paper above is its AI-adjacent cousin. Frog and Toad and the Increasingly Capable Machines (80 comments): the children’s-book parody the agent era deserved. Using Opus 5.5 to discover a new eyewitness record of the dodo (59 comments): covered in Read. Vote on which of Hacker News’ challenges for AI have been met (215 comments): the goalpost-moving, quantified.

10 · OpenRouter Apps

Codex +59% overnight; Cline passes Claude Code

Codex (~1.11T tokens) jumped ~59% in 24 hours and now sits 2% behind Kilo Code (~1.14T) — the coding-agent crown changes hands within days at this pace. Cline (~1.07T) passed Claude Code (~1.06T) for the first time; the trio is now a three-way commodity race with no moat but momentum. Hermes Agent (~2.03T) still leads overall, Framer (~1.33T) holds #2. Behind: pi (~451B), omp (~422B), Command Code (~417B), Freebuff (~324B), DeepSeek Harness (~268B), OpenClaw (~203B).

Signal: Verify-before-citing number of the week: Codex’s two-day run from ~314B to ~1.11T is the fastest share grab in the board’s recent history. If OpenAI’s harness takes the #1 coding spot, the “open agents” narrative gets a lot harder to tell. Sources: OpenRouter rankings

11 · Watch

The one OpenAI announcement that can actually make you money… — Fireship

Fireship · Oct 1 — the money angle on the week’s OpenAI announcements, in the usual hundred-miles-an-hour format. Watch after this week’s Decisions API and Codex coverage: which of the DevDay stack actually has a business model attached.

Signal: The fastest filter for “which OpenAI launch matters” — if Fireship can’t find the money angle, there isn’t one.

Googles New Gemini 4 Argon is Now The Worlds Smartest AI — TheAIGRID

TheAIGRID · Oct 1 — the video version of yesterday’s One Thing: the benchmarks, the Fairwind gating, the pricing, decompressed into ten minutes.

ASTRA 6.1 too dangerous to be released… — WesRoth

WesRoth · Sep 29 — the capability-maximalist take on GPT-6.1 Sol, which replaced GPT-6 Sol after just 7 days with near-Astra intelligence. Pairs with the AA note in Leaderboard: the labs are now shipping replacements faster than reviewers can review them.

12 · Listen

Academia is for Ambition — Alex Zhang, MIT — Latent Space

Latent Space · Oct 2 — fresh this morning. The episode’s framing (RLMs — the reasoning-model research conversation) lands right as the decision-model layer goes mainstream: the research frontier and the product frontier are converging on “how much thinking is enough.”

A.I. Agents: Cute, Cuddly and Maybe Catastrophically Dangerous? — Hard Fork

Hard Fork · Oct 2 — the agent-safety discourse goes mainstream-audience. Listen for the gap between how the industry talks about agent risk (benchmarks, monitors, the Goodfire piece in Read) and how the public hears it.

Signal: When the NYT’s tech podcast titles an episode “catastrophically dangerous,” the Overton window on agent risk has moved. The industry’s answer had better be better than a launch bullet.

13 · Read

Epoch tracked 8.3 million ChatGPT messages to reveal how usage deepened

Epoch AI released a public explorer built on 8.3 million ChatGPT messages from 5,000 YouGov panelists, tracking how real usage intensified from 2022 to 2025. This is the rare dataset about what people actually do with AI instead of what labs say they do. The deepening-habits finding is the one to sit with: usage isn’t just growing, it’s compounding per user.

Signal: The usage data says the market isn’t saturating — it’s deepening. Every “AI fatigue” take has to survive this chart.

Using Opus 5.5 to discover a new eyewitness record of the dodo

A researcher used Opus 5.5 to surface a previously unknown eyewitness account of the dodo — 59 comments on HN. The capability story that isn’t about benchmarks: frontier models as research instruments for fields that will never buy an API subscription. The humanities are quietly becoming power users.

Goodfire builds sequence-aware biosecurity monitors to guard AI biology agents

Goodfire’s protein-embedding monitors catch more harmful biology requests with fewer false refusals, staying 3–5x more robust against adversarial attacks. The biosecurity layer is being built as interpretability products, not policy documents — which is exactly how it should work.

AIHOT opens its full AI news curation stack to developers

An open-source TypeScript framework that scrapes six source types, double-scores every item with an LLM, clusters duplicate stories, and publishes a daily brief. Yes, this is uncomfortably close to describing this newsletter’s own pipeline — which is precisely why it’s worth reading. The curation stack is becoming a commodity; the taste isn’t.

14 · Safe to Ignore

Permission to skip

Skip: Several vulnerabilities discovered in the Linux kernel. 309 comments of real security work. Real, important, and not this newsletter. Skip: StreetComplete on iOS is now in public beta. 261 comments of open-street-map joy. Not this newsletter. Skip: Ask HN: Who is hiring? (October 2026). 229 comments, monthly ritual. The jobs are real; the FOMO is not. Skip: Cloudflare K2: serverless event streams. Interesting infra, wrong week — the decision models ate Cloudflare’s news cycle. Skip: SvelteKit 3 is here. 143 comments. A fine framework release. Not this newsletter. Skip: How to speed up the Rust compiler in September 2026. 154 comments of compiler craft. Admire from afar. Skip: ArXiv’s updated rate limit policy. The library posted new rules. Read the papers, skip the policy.

15 · Build on This

HashReady: the Git 3.0 migration audit every org is about to need

Git is flipping its default hash from SHA-1 to SHA-256, and the HN thread is 444 comments of migration dread: hardcoded 40-character hash assumptions, CI scripts that parse object databases, build caches keyed on SHA-1, vendored tools that never got the memo. Somewhere in every mid-size org is a pipeline that will break on upgrade day — and nobody knows which one.

Product: point HashReady at your org’s repos and get a migration bill — which repos assume SHA-1, which tools break, what to fix first, in what order. Static scan of code and CI configs plus a knowledge base of known-broken tools. Sell it as a one-time per-org audit, no subscription — migrations are events, not services, and the forcing function (the default flip) does the selling for you. The wedge: start with the public repos of the loudest HN complainers and publish anonymized “state of SHA-256 readiness” reports. Nothing sells an audit like a leaderboard of who’s unprepared.

16 · Appendix

Everything below didn’t make the cut. Got a better pick? The next issue gets sharper.