FOMO Buster #010: The distillation war goes public

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

OpenAI names Moonshot: the distillation war goes public

OpenAI published a September 30 post disclosing that it disrupted a coordinated campaign to extract hidden reasoning from its models — and pinned a core cluster of the activity on individuals associated with Moonshot AI, the developer of Kimi. The campaign ran through July: a slow start on July 1, then a spike to 16,000 extraction-pattern requests from 4,000+ users on July 24–25, with related prompt-pattern activity across more than 15,000 users total before OpenAI says it shut the operation down on July 28.

The technique was blunt rather than sophisticated: operators copied encrypted reasoning from one conversation and asked another model instance to decrypt and transcribe it. OpenAI says the encryption was not broken, no databases were touched, and no stored user conversations were exposed. It calls the behavior “adversarial distillation” — the systematic, unauthorized use of one model’s outputs or reasoning to train, reproduce, or improve another. The response: banned accounts, tightened sign-up verification, new protections on hidden reasoning, and the findings shared through the Frontier Model Forum and government channels. Caroline Zier, who leads OpenAI’s strategic national security policy work, drew the line Bloomberg quoted everywhere: “Our concern is about violation of our terms of service, not open models or legitimate distillation.” Anthropic previously leveled similar accusations at Moonshot, and Google has made its own.

This is the first time a frontier lab has publicly attributed an adversarial-distillation operation at this scale, with names. The hidden chain-of-thought is now officially crown-jewel IP, policed like theft rather than weather.

Signal: Three things harden at once: labs will sell “anti-distillation” as a feature (reasoning encryption, output monitoring, account-graph abuse detection); every big open-weight release will be read as a distillation event until proven otherwise; and the Frontier Model Forum just became a threat-intel sharing club. Watch which lab is next to publish its own takedown — silence is now a signal. Sources: OpenAI: disrupting the campaign · AI Weekly: Moonshot users linked to the 16,000-request push · AI Affairs: the full timeline

02 · Lab Watch

Meta open-sources Muse Gadgets: solder your own AI hardware

Meta’s Superintelligence Labs launched Muse Gadgets on October 2 — open-source ESP32 firmware and a Linux SDK that let anyone build hardware for Meta’s Muse assistant. Grab an API token from gadgets.muse.ai, point a coding agent at the GitHub repo, and wire Muse to your displays, buttons, sensors, and actuators. The reference build is Muse Home Link: a USB-C dongle that puts Muse on your home network to talk to TVs, speakers, and anything exposing HTTPS. Meta manufactured 5,000 of them and is giving them away free to Muse subscribers while supplies last, shipping in weeks. Nat Friedman’s post about it pulled ~30,000 views in hours — the freebies may already be spoken for. Suggested community builds include color e-ink displays, HDMI sticks for TVs, and touchscreen gadgets in the spirit of the Muse Charm. There’s a Discord for builders, and it follows this week’s Muse for Small Business launch (free with usage limits, wired into Shopify, Dropbox, and Slack).

The frontier-lab playbook has a new page: ship the assistant as firmware, let the community build the devices. Meta isn’t selling hardware — it’s recruiting it.

Signal: The assistant is escaping the phone. When a lab gives away 5,000 dongles, it’s buying a developer ecosystem, not selling units. The interesting question is what the community builds that Meta wouldn’t: the e-ink pocket display (already done, via MagSafe) is the shape of things. Sources: TechCrunch: build your own Muse gadget · iPhone in Canada: the Home Link details

ChatGPT Sites hits HN: the prototype-vs-production debate

311 points, 297 comments — HN finally stress-tested ChatGPT Sites, which has quietly been in public beta since June. The thread forced the practitioner questions: is this a toy, a prototype host, or something you’d put a real audience on? The facts that survived: public publishing for Plus, Pro, Business, Enterprise, and Edu (not Free or Go, not EEA/UK at launch); custom domains where available; Sign in with ChatGPT for visitor auth; D1 up to 10 GB and R2 for storage; every deploy URL is a production deploy. And the week’s sharpest detail came October 1, when OpenAI staff said Sites can host MCP servers — the hosted answer to “running a backend is hard,” and a free hosting tier for the plugin directory in disguise.

Signal: The thread is really about whether a chat app is becoming a PaaS. The MCP-hosting claim is the tell: OpenAI wants the next million plugins to have somewhere free to live, and Sites is that somewhere. Sources: ExplainX: the HN-thread read · OpenAI help: Sites in public beta

The coding agents add security scans and charts

Two small releases, one pattern — the agents are growing up into tools with jobs, not demos with prompts. OpenAI upgraded Codex Security Cloud to scan entire GitHub repos automatically (announcement) — the security story moves from “review this diff” to “watch this codebase.” And Cursor added inline chart and diagram generation directly in chat (announcement) — the IDE renders what the agent means, not just what it types. Security as a background service, visualization as a first-class output: the agent IDE is filling in the affordances the old IDE took for granted.

03 · Open Weights

Kolibri: Germany’s sovereign model, timed for Unity Day

Aleph Alpha released Kolibri on October 3 — German Unity Day — and the symbolism is the product: an English-German mixture-of-experts (78B total, ~3B active per token), 1M-token context, weights on Hugging Face under Apache 2.0, trained from scratch on infrastructure in Germany and Finland on 20 trillion tokens filtered down from over 200 trillion. 21.3% of the training text was German, and Aleph Alpha says it used machine translation sparingly because translated text “tends to carry the cultural fingerprint of its source language” — the pitch is a model bilingual by design, not an English model that read some German. Its own evals put it at 70.8 on German against 69.8 for Qwen3.5-35B-A3B (the dense Qwen3.8-27B scores 79.9, with ~8x the active parameters). Hardware bill: ~78 GB of GPU memory, so two 80GB A100/H100s or one H200/B200/B300.

Signal: Sovereignty is becoming a spec sheet, not a slogan: built-under-EU-law, open weights, Apache 2.0, commercial-use allowed. Regulated European industries just got a model they can run inside their own trust boundary without asking anyone. Watch procurement, not benchmarks, for the adoption signal. Sources: Aleph Alpha: Kolibri has landed · Tejas Kumar: how the sovereign German LLM works

FLUX 3 Image ships; the open-weights clock starts when the files land

Black Forest Labs released FLUX 3 Image on October 1 (383 points on HN): bounding-box layouts on a normalized 0–1000 grid, up to 10 reference images, multi-turn edits that lock unchanged pixels, and output up to 4K — through the Playground and API at 50% off through October 8 ($0.0205/image at 768px, roughly doubling after). Commercial weights are licensable now; open weights are “coming in weeks.” The same week, Ideogram 4.5 shipped its own precision-edit model, also headed for open weights. Layout-as-structured-input is the new front in image gen: coordinates replace prompt wrangling, and the layout becomes project state an agent can revise.

Meanwhile on the board: Arena’s FLUX.2-dev beats every open-source image model with a 66% win rate (AlphaSignal), via a 5M-vote preference model plus rubric-based rewards — the recipe behind the leaderboard highs.

The trending delta: LTX-2.5 climbs, laya still likes without running

Lightricks/LTX-2.5 is the week’s climber: #5 trending (6,070 likes, 1.63M downloads) and already #13 on the all-time liked board — video models are having a moment. Qwen/Qwen3.8-27B just took the all-time like crown (16,854, passing FLUX.1-dev) while holding #8 trending: gravity well and climber at once. Cloudflare/clef holds #2 trending (924 likes) with clef-flash at #9 — the decision-model tag is sticking. Qwen-Image-2.1 is still everywhere: the uncensored GGUF at #3 trending with 1.46M downloads.

New climbers not yet established: Contrastive-LM/CLM-v0.1-8B (#4), orcarouter/OrcaSAQ-2-Cyber-27B-Uncensored-GGUF (the uncensored-cyber GGUF economy is alive and well on HF), ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF, and prism-ml/Ternary-Bonsai-2-27B-gguf. And convaiinnovations/laya still leads trending with 5,044 likes and zero downloads — the purest “like, don’t run” model on record, seven weeks running.

One more from the newsletter: Mercor’s research team post-trained Qwen3.5-397B-A17B on expert data and raised Pass@1 by 11.2 points on APEX-Agents — the big open MoE keeps getting better after release. RL post-training as a publishable result is becoming its own beat. (Mercor’s write-up)

Signal: Two lanes are splitting on the board: the models you run (LTX-2.5, Qwen-Image quants, coders) and the models you admire (laya, still at zero downloads). The delta between them is the best measure of which capabilities are actually deployable this week. Sources: HF trending

04 · Money Flow

Prime Intellect’s inference bet: 600B tokens a day

Prime Intellect launched Prime Inference, serving 600 billion tokens daily with NVFP4 KV compression, prefill/decode disaggregation, and sub-second time-to-first-token. The inference-supply business keeps scaling up: whoever serves the cheapest token at acceptable latency wins the agent-workload contracts, and the technical levers (KV compression, disaggregation) are now the marketing copy.

Ideogram’s 27x pricing spread

Ideogram 4.5 fights edit drift with multi-turn precision — and a 27x pricing spread across its tiers. Price discrimination has arrived in image generation: the casual edit and the production edit are now different products with different prices. It debuted at #23 on the image-editing leaderboard, leading on text edits but trailing the frontier on raw generation. Read it next to FLUX 3 Image’s expiring 50% discount: the image-model price war is being fought in launch promotions and tier spreads, not headline rates.

Signal: Inference is the commodity; the margin is in the tiers. Watch who copies the 27x spread — precision editing is the first image-gen feature priced like enterprise software.

05 · Leaderboard

The reshuffle: Claude Code takes back #3

The coding-agent board just had its choppiest day in weeks. Claude Code jumped ~16% to ~1.23T tokens, retaking the #3 coding spot from Kilo Code (~1.14T, flat) — and passing Cline (~1.06T, slight slip) for good measure. Codex cooled to ~1.01T from ~1.11T, giving back part of its two-day spike. Hermes Agent (~2.01T) and Framer (~1.29T) both slipped, which reads like rolling-window churn. Pi ticked up to ~456B. The read: Codex’s grab was real but not yet sticky; Claude Code’s jump smells like renewed enterprise flow, not vibes.

Artificial Analysis: Grok 4.7, Solar Mini 4, and local-agent benchmarking

Grok 4.7 (Low) got its AA evaluation on October 3 (model page). Korea’s Upstage released Solar Mini 4 on October 1 (AA’s write-up) — the Korean labs keep shipping compact models into the gap between flagship and edge. And AA launched AA-AgentPerf-Local: benchmarking local AI agents on laptops and workstations (article) — the benchmark layer finally follows the models onto the device.

Signal: The leaderboard that matters next quarter may be the laptop one. AA-AgentPerf-Local is the first serious attempt to measure what agents feel like on hardware people own — and it’s where the open-weights story gets tested, not just told. Sources: OpenRouter rankings

06 · The Argument

Court agrees with EFF: Utah’s VPN law demands a technical impossibility

697 points, 333 comments — the day’s biggest HN thread, and barely about AI at all. A court agreed with the EFF that Utah’s VPN law demands a technical impossibility: legislators wrote a mandate the technology cannot satisfy, and the thread is 333 comments of engineers explaining, again, that you cannot legislate around math. The honest argument underneath: when the law requires the impossible, compliance becomes theater — vendors perform it, lawyers bill it, and the security properties nobody asked about stay exactly as broken as before. The counter-argument: impossibility rulings like this one are the system working, slowly, expensively, correctly.

Signal: This is the template fight for the AI-regulation decade. Every impossible mandate — age verification, watermarking, “just detect the AI” — will get its own version of this thread. The winners are whoever can explain the math to a judge.

07 · Papers

ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New Research

A benchmark for the unglamorous superpower: finding the papers that inspire new research, not the ones that match keywords. Retrieval is being graded on generativity now — the right citation is the one that changes what you build next.

The Missing Primitive: Diagnosing and Repairing Mathematical Reasoning in Large Language Models

Diagnosing and repairing mathematical reasoning — the paper treats math failure as a missing primitive, not a scale problem. Lands the same week Meta’s Muse Spark co-authored six math papers: the labs are attacking mathematical reasoning from both ends at once.

One Basis to Animate Them All: Gaussian Blendshape Distillation for Real-Time Avatars

Gaussian blendshape distillation for real-time avatars — the distillation theme isn’t just corporate warfare this week (see The One Thing), it’s also a research technique making avatars cheap enough to run live. Same word, two meanings, one newsletter.

VISTA: A Visual Harness for Reasoning in an Interactive World

A visual harness for reasoning in an interactive world — agents that think by looking, in environments that push back. The harness layer keeps climbing the stack: from text to tools to vision to worlds.

Signal: The paper layer keeps rhyming with the product layer within days — reasoning repair behind the math co-authorships, distillation behind the takedown, visual harnesses behind the Sites-as-PaaS debate. arXiv is the leading indicator; the trending page is the lagging one.

08 · GitHub Trending

The skills stack keeps compounding; Cloudflare ships an agent OS

DietrichGebert/ponytail still leads the daily board (152,683 stars, “the best code is the code you never wrote”), pbakaus/impeccable holds #2 (74,843, and on the weekly board too). The new entries show the stack filling in: JuliusBrussee/caveman (109,366 stars — a viral skill and proxy that cuts 65% of tokens by making your agent talk like a caveman; the bit of the week), thedotmack/claude-mem (persistent context across sessions), Panniantong/Agent-Reach (89,477 — “give your AI agent eyes to see the entire internet,” one CLI, zero API fees), affaan-m/ECC (harness performance optimization — skills, instincts, memory, security — for Claude Code, Codex, Opencode, and Cursor), alirezarezvani/claude-skills (380 skills on the weekly board). And the enterprise tell: cloudflare/cloudflare-os (10,451 stars) — an agent workspace built on Cloudflare Workers for documents, apps, and agents with company context. Plus TencentCloud/Octop (6,521 — self-hosted multi-user, multi-agent assistant) on the weekly board alongside paperclipai/paperclip (96,586), vectorize-io/hindsight (44,942), and heygen-com/hyperframes (56,166).

Signal: Skills were the land grab; now come the runtimes and operating layers — Cloudflare and Tencent both shipping agent infrastructure in the same week is the platform tell. And “caveman” at 109K stars is the market pricing token frugality as a feature, not an insult.

09 · Product Hunt

Yesterday’s top 10 launches

1. Monospace from Directus — The governed API layer for every app, person, and agent (420 votes)

2. Yedric.ai — Let users control your SaaS with natural language (306 votes)

3. Dots by OpenAI — Always on agents built to handle everything (278 votes)

4. Omnia Agent — The AI Agent that does 95% of your GEO work (207 votes)

5. Polylane — AI agents that fix production before you wake up (183 votes)

6. DSH Desktop — Official app for DeepSeek’s open-source agent harness (170 votes)

7. statusbar — A statusbar for any terminal (131 votes)

8. Chat.sh — The help center I built after Intercom’s search broke (124 votes)

9. Bracket — The memory layer for your business (108 votes)

10. America.gov — Whatever you need from government, start here. (105 votes)

10 · Hacker News

Court agrees with EFF: Utah’s VPN law demands a technical impossibility

697 points, 333 comments — the day’s biggest thread. Covered in The Argument.

Sites in ChatGPT

311 points, 297 comments of prototype-vs-production. Covered in Lab Watch.

FLUX 3 Image

383 points on the bounding-box image model. Covered in Open Weights.

Muse Gadgets

216 points on Meta’s open-source hardware play. Covered in Lab Watch.

Also notable

Mike Tomlin spent 12 years building a Minecraft city (554 points, 127 comments): the most wholesome thread of the week, and the best argument that long-term creative projects survive the attention economy. Apple Pass Designer (484 points, 298 comments): Apple’s wallet-pass tooling gets its moment. Extra Big Ass Intelligence (377 points, 83 comments): the parody the newsletter industrial complex deserved. Show HN: Giving Opus 5.5 a simulated paint canvas (314 points, 101 comments): the model as a painter, and the comments as art critics. From the creator of Redis: run LLM locally with ds4 (301 points, 137 comments): dwarfstar.sh — local inference with a famous name attached. With most information hidden, the game Stratego had stumped AI until now (253 points, 119 comments): beaten on a budget, which is the interesting part. One month coding with GLM 5.3 Flash (193 points, 152 comments): the long-term field report format HN does best. Show HN: Germany’s new sovereign AI model Kolibri (143 points, 86 comments): covered in Open Weights. An Update on Orion for Linux and Windows (132 points, 59 comments): Kagi’s browser goes cross-platform.

11 · OpenRouter Apps

Claude Code +16% retakes #3; Codex gives back the spike

The full board this morning: Hermes Agent (~2.01T, slipping), Framer (~1.29T, slipping), Claude Code (~1.23T, up ~16%), Kilo Code (~1.14T, flat), Cline (~1.06T, slight slip), Codex (~1.01T, cooling from ~1.11T). Behind the coding six: Freebuff (~656B), omp (~621B), Pi (~456B, ticking up), Command Code (~435B), DeepSeek Harness (~282B), Zed (~277B). The coding-agent tier is now a six-way knife fight separated by ~200B tokens — and the only one still accelerating is Claude Code.

Signal: Rolling-window volatility is the story of the week: Codex’s two-day grab is already giving back, Claude Code’s enterprise flow is back. Nobody’s moat is wider than a weekend. Sources: OpenRouter rankings

12 · Watch

Did a 50 year old military secret just solve agent prompt injection? — Fireship

Fireship · Sep 30 — the prompt-injection episode, with a Cold War hook. Watch the week OpenAI is publishing takedowns of adversarial campaigns: the attack surface is the story, and Fireship is the fastest filter for whether the defense is real.

GEMINI 4 is nuts… — WesRoth

WesRoth · Oct 1 — the capability-maximalist take on Gemini 4, a week after Argon put Google back in the top three. Pairs with the Solar Mini 4 and Grok 4.7 evaluations in Leaderboard: the frontier is crowded again.

OpenAI Just Revealed Dots… This Changes ChatGPT Forever — TheAIGRID

TheAIGRID · Sep 30 — the video version of the Dots story, which also placed #3 on Product Hunt this week. The always-on agent, decompressed into ten minutes.

13 · Listen

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

Latent Space · Oct 2 — fresh. The reasoning-model research conversation, landing the same week the labs are policing reasoning traces as trade secrets (see The One Thing): the research frontier wants reasoning open, the product frontier wants it encrypted. That tension is the episode’s unspoken backdrop.

Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week — Latent Space

Latent Space · Sep 30 — the DevDay 2026 debrief: the computer-use debate and how fast OpenAI can now ship. Listen after the Sites card in Lab Watch — same company, same week, same “the chat app is becoming the platform” thesis.

14 · Read

Google sent Trillium TPUs to orbit

Project Suncatcher just put four Trillium TPUs into low Earth orbit on a Planet-built satellite, testing whether AI compute can survive and run beyond the grid. The wildest infrastructure story of the week, and a real question underneath the stunt: if power and cooling are the binding constraints on Earth, does orbit become a datacenter location with a straight face?

Meta’s Muse Spark helped mathematicians solve five open research problems

Muse Spark worked alongside mathematicians through the standard chat interface to co-author six research papers, five answering previously open questions. Not a benchmark, not a demo — published mathematics. The capability story that isn’t about leaderboards: frontier models as research instruments, with the human still holding the pen.

livenerf catches Claude Opus 5.5 drifting with a daily 78-question trap

An open-source benchmark runs a frozen test panel every day for 30 days to detect whether a frontier model quietly degrades after launch. The premise everyone suspects and nobody measures: your model changes under you. Silent updates, distillation swaps, routing changes — livenerf is the tripwire. See today’s Build on This.

Google moves Gboard’s private AI training into secure server enclaves

Google rebuilt federated learning around Trusted Execution Environments, moving gradient computation server-side with externally verifiable differentiation. The privacy story underneath the AI story: the keyboard learns from you, but the learning happens where you can check it. The enclave is becoming the unit of trust.

Signal: The week’s reads share a shape: measurement. Drift traps, verifiable enclaves, orbit as a test harness. The industry is building the instruments it needs to believe its own claims.

15 · Safe to Ignore

Permission to skip

Skip: Mike Tomlin’s 12-year Minecraft city. 554 points of wholesome. The city is real; the FOMO is not. Skip: Apple Pass Designer. 484 points, 298 comments of wallet-pass tooling. A fine developer release. Not this newsletter. Skip: Newgrounds nostalgia. 338 points. The games were great; the news is not. Skip: Extra Big Ass Intelligence. 377 points of parody. Delightful. Not signal. Skip: Updates to Full Disk Access in macOS. 242 points of platform plumbing. Admire from afar. Skip: Scientists invent underwater umbrellas to protect coral reefs. Real science, wrong newsletter. Skip: What is going on with ceiling fans. McMansion Hell remains undefeated. Not this newsletter. Skip: Show HN: Offrun — manage every coding agent from one workspace. 35 points. The agent-manager category is crowded; come back with users.

16 · Build on This

DriftWatch: catch your model drifting before your users do

Your production model changes under you — silent updates, distillation swaps, routing changes — and you find out from a customer, a week later, in a tone you don’t like. livenerf (in Read) is the proof of concept: a frozen 78-question panel, run daily, that caught drift in a frontier model.

Product: point DriftWatch at your prompts and watch it run a daily probe battery against your live model. Frozen golden questions plus your own regression set, diffed against baseline, with an alert the hour answers drift — not the week. Dashboard shows the drift curve per model version, so “did Tuesday’s deploy change the model or did the model change itself?” becomes an answerable question. Sell it pay-per-probe, no subscription — drift monitoring is insurance, and insurance bills per check, not per month. The wedge: start with the loudest “the model got dumber” threads on HN and GitHub, run the probes publicly, and publish anonymized drift reports. Nothing sells monitoring like a leaderboard of who’s degrading.