FOMO Buster #021: 35% more code

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

Gemini 4 Argon: the model launch that came with receipts

On October 13, Google launched Gemini 4 Argon, its latest frontier model, and the announcement reads less like a benchmark table and more like an internal cost report. The headline internal numbers: software engineers using Argon in Antigravity increased agentic code submissions by more than 35% while reducing code rollbacks by nearly a third. Argon ranks #1 on the Vals Index, the sector-by-sector economic benchmark covering finance, coding, legal, and tax work, and ships at an introductory $2 per million input tokens and $10 per million output tokens, with cached input at 95% off.

The pricing is the quiet shot. Meta’s standard Muse Wonder launched two days earlier at $4/$20, matching Claude Opus 5.5 exactly; Argon undercuts both by half. The distribution is wide: API and AI Studio, Gemini app, AI Mode in Search, and a free top usage tier in Antigravity for 10,000 scientists over six months.

The receipts, because Google published them: Argon helped the Chrome team carve core webpage handling out of a ~6-million-line legacy folder inside Chromium’s 40-million-line codebase, finishing months of refactoring in weeks; it gave the open-source Agent Development Kit 40% faster session prep and 21% faster cold imports, live in the repo; it made the xTB quantum-chemistry package 2.3x faster at predicting molecular structures; and it squeezed a 2.85x end-to-end inference speedup (37,900 to over 108,000 tokens/sec on an A100) out of its own Gemma 4 E2B, with 100% token parity, in days instead of months.

Two more numbers worth your attention. On agent alignment, where models take unauthorized shortcuts when they hit a roadblock, Argon’s shortcut attempts dropped to 2%, its best internal number ever. And in robotics, Argon watched a raw video of a human playing cornhole, built the simulation itself, and transferred it to an Apptronik Apollo humanoid that sank seven consecutive shots in the real world. The ecosystem is already moving: Cursor is shipping Argon with a 1M-token output limit, and Cognition added it to Devin at 75% off.

Signal: The industry just went through Meta’s frontier launch, and Google answered with something harder to argue with than a chart: internal throughput numbers, a half-price tag against the exact model it is positioned against, and third-party wins already committed to other people’s repos. Frontier launches are becoming procurement battles, and the first move is always price. Sources: Google’s announcement · HN thread, 1,380 points

02 · Lab Watch

Anthropic ditched git. Claude broke it first.

Anthropic published the engineering story of Scion, the in-house version control system that replaced git across its monorepo in five months. The numbers explain why. In May, Anthropic merged around 2,500 code changes a day; by September it was over 30,000 a day. The main branch’s history alone is nearly 100 GB, a compressed working copy exceeds 25 GB, and CI does roughly 25 million checkouts a day; fetched naively, that would be over 600 petabytes of git traffic daily. At peak, the company was running several hundred internal git mirrors just to stay upright.

Scion assumes what git never could: that nearly every client is an agent that is always online. Commits happen server-side, the client keeps a bounded cache, and checkouts are sparse by default. The payoff, verified in production: infrastructure footprint 200x smaller than the git-mirror era in CPU and memory, checkouts 100x faster (seconds instead of minutes), and a 25% CPU capacity win in CI alone. The footnote that should worry every dev-tools startup: a team of seven people built Scion and the Antique code-review surface in five months, partly because they used Claude itself, which rewrote the ingestion pipeline at 20x less memory and 4x the speed with no extra instructions.

Signal: Git’s core design is offline-first humans with full clones; the agent era’s workload is online-first agents that never need one. Anthropic is the first lab to say the quiet part out loud: its bottleneck is no longer the model writing code, it is the human-scale infrastructure around the model. Everyone else running coding agents at scale is about to discover the same ceiling. Sources: Anthropic’s engineering post

Also on the lab feeds

Anthropic opened a Singapore office with a government MOU (Oct 14), with Claude usage reportedly at 6x expected demand in the region. Google DeepMind is putting frontier models and advanced AI compute behind European scientific discovery (Oct 14), paired with the 10,000-scientist Argon grant. OpenAI published an analysis of agentic AI adoption across EU enterprises and small businesses (Oct 14), plus a Higgsfield customer story on AI video going from 1 minute to 10. Mistral’s feed is quiet post-Large 4; Meta’s holds the Muse Wonder launch.

Signal: Five labs, five moods again: Google ships science compute, Anthropic ships Southeast Asia, OpenAI ships the enterprise-adoption narrative. Nobody is quiet; the announcements just stopped being models. Sources: Anthropic Singapore · Google’s European science push · OpenAI’s EU agentic report

03 · Open Weights

Fish Audio Drama 3: voice AI becomes a director’s medium

Fish Audio shipped Drama 3, a 35-billion-parameter mixture-of-experts TTS model with 3B active per token, and the headline is not the size: it is control. Developers direct a voice performance in plain language, shift emotion mid-sentence, run up to 16 speaker tags in one generation, and use audio inpainting to regenerate a single word while keeping the surrounding performance intact. It ships as open weights under Fish Audio’s community license, with a ComfyUI node and a day-zero serving recipe, not a waitlist.

Read the company story behind it: Fish Audio just marked its first anniversary, grew from zero to $21M ARR and 15 million users in that year, and its predecessor S2.1 Pro became the most-requested speech model on OpenRouter within three weeks of launch. For the first 30 days the API is 80% off, and enterprises switching providers get three months free.

Signal: Yesterday this newsletter had VoiceStudio as the free local ElevenLabs; today the commercial leader shipped open weights with director-grade control. The voice lane’s competitive axis just moved from “who sounds real” to “who is easiest to direct,” and the open model is competing on both. Sources: The launch release · kie.ai deep dive · fish.audio

LiquidAI’s LFM2.5-24B-A2B: 24 billion parameters, 2 billion awake

LiquidAI released LFM2.5-24B-A2B, a 24B reasoning model with only 2B active parameters, running at a reported 184 tokens per second on a laptop. The architecture is the story: sparse activation as the route to laptop-class reasoning, the same direction as the mlx-community shrink of Alibaba’s Qwen3.8-27B to fit a 24 GB Mac.

The board this morning

google/embeddinggemma-2 holds trending a fifth day (2,199 likes, 135,175 downloads), now with a community speech-to-text hack. jialinyyzz/humanizer stays at #2. The climbers: baidu/Unlimited-OCR (4,632 likes, 1.15 million downloads), ConwayResearch/Underdog-Saluki-27B (the 27B-in-8GB squeeze, 311,969 downloads), openbmb/MiniCPM-V-4.7, and the uncensored Qwen-Image lane still printing downloads (the abenzerps quant past 2.3 million). Cloudflare/clef debuted high (2,156 likes). All-time: Qwen3.8-27B at 17,672 likes, with Lightricks/LTX-2.5 (7,732 likes) now a gravity well in its own right.

Signal: The delta says it: baidu’s OCR model is exploding in downloads without cracking the all-time likes board, and JEV-35B (below, in Leaderboard) is doing the same on the decision side. Trending is where demand shows up first; likes are where it settles. Sources: Hugging Face trending · LiquidAI LFM2.5

04 · Money Flow

Prime Intellect buys Moondream for the inference stack, not the vision model

Prime Intellect, the neocloud valued at $1B in July on a $130M Series A with an annualized revenue run rate above $100M, announced on October 13 that it is acquiring Moondream, the open vision-model shop with fewer than 10 employees and $4.5M raised. Terms are undisclosed, but the deal’s real prize is not the vision model: it is Moondream’s inference optimization software, the auto-generated “megakernels” that fuse GPU operations, its Photon inference engine, and a reported 59ms median latency on an H100, all serving 5M+ monthly downloads.

The read: neoclouds are competing on what they squeeze out of each chip, not how many chips they rent. Prime Intellect’s CEO framed the deal around robotics and computer-use agents, the workloads where inference cost is the product.

Signal: Differentiation is moving from raw compute to the engineering on top of it. Watch whether the megakernel tooling becomes a standalone product: if it does, every inference provider becomes a buyer. Sources: The deal report

Physical Intelligence opens the Pi API: robot models as a hosted service

Physical Intelligence opened its Pi API beta on October 13, putting its general-purpose robot foundation models (π0 and π0.5) behind a developer API with hosted fine-tuning and inference. The workflow is literal: upload robot data, fine-tune on infrastructure the company manages, serve the fine-tuned model through a hosted gateway. Access is waitlist-gated; pricing and GPU allocation are undisclosed.

The case studies are the pitch. Almond’s Axol, a dual-arm robot, now deburrs spacers used in its own assembly after four days of API iteration. Actor’s autonomous wheel loader scoops material and loads dump trucks in real commercial deployment with no operator inside the machine. The company says partners get more data-efficient fine-tuning on the latest model, with rapid model updates under the API.

Signal: This is the API moment for robots, and it arrived as managed infrastructure, not open weights. Read it against Argon’s cornhole robot and the Dyna-2 episode in Listen: the robotics stack is splitting into foundation models you rent and fleets you run, and the physical-AI layer is being productized faster than the safety discourse around it is moving. Sources: The announcement report · Anthropic’s hardware rules

05 · Leaderboard

JEV-35B: the open model that beats Claude Haiku 5.5 at 11x the speed

Artificial Analysis’ new Decision Index handed the day’s most interesting upset to AutoTrust AI’s JEV-35B: on open-source go/no-go decisions served as 204ms API calls, it beats Claude Haiku 5.5 at 11x the speed. The model sits at #6 on today’s HF trending board (713 likes, 27,970 downloads) and is climbing without the all-time likes to show for it yet, which is exactly the demand-first pattern from the Open Weights signal. Yesterday AA also added full evals for Claude Haiku 5.5 (four effort levels), GPT-6 Sol (Daybreak Blue, max), and Zoom Scribe V2 Pro, plus the FrontierCode benchmark and Harvey LAB-AA v1.1 with hallucination checks.

Signal: The Decision Index is a new kind of leaderboard: not “which model is smartest” but “which model makes the cheapest correct decision.” That framing favors small open models served fast, and JEV is the first native of that lane. The leaderboard business is quietly shifting from intelligence to economics. Sources: JEV-35B on Hugging Face · Artificial Analysis

06 · The Argument

Google’s two speeds: Argon refactors Chromium while AOSP releases lag so far behind that developers disassemble libraries

Today’s most ironic HN thread (405 points, 191 comments) is aimed at the same company that just launched Argon. Android developers are so starved of timely AOSP source releases that they are disassembling compiled libraries to report bugs, per the thread anchored on an android/ndk issue. The company whose model just refactored 40 million lines of Chromium in weeks cannot ship its own mobile OS source on a cadence its developers trust.

Pair it with the other trust story of the day: Ledger confirmed a hardware spy chip implant behind crypto wallet thefts (755 points, 472 comments on HN), a supply-chain compromise in the device whose entire value proposition is trust. One story says the software supply chain is opaque; the other says the hardware one is compromised. The agent era is inheriting both.

Signal: Argon’s launch narrative is “AI accelerates the whole software lifecycle,” and the AOSP thread is the counterfactual walking around in public: velocity concentrates where the model vendor eats its own dogfood and rots where it does not. Watch which of Google’s own stacks get the Argon treatment next; the ones that do not will keep shipping disassembly stories. Sources: The AOSP thread · Ledger’s confirmation · Argon’s Chromium refactor

07 · Papers

A free 472-page textbook on agentic LLMs

Leiden University dropped a free 472-page textbook on agentic LLMs on arXiv (2610.13455), the teaching story of the day. If yesterday’s Anthropic usage-policy update was the rules for agents, this is the curriculum. Pair it with the Build on This section: the skill economy has courseware now.

Also on arXiv

Who Improves the Robot? Toward Physical Recursive Self-Improvement: the recursive-self-improvement thread leaves the datacenter and enters the body, two days after the MASS paper. Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards: the long-horizon lane Meta is selling with Muse Wonder is getting its training environments. DoubtLess: Training-Free Contrastive Decoding for Efficient Reasoning: better reasoning without retraining, the cheapest upgrade on the board.

08 · GitHub Trending

The skills economy has a repo for everything

The board’s theme this week is skills, and it is everywhere at once. mattpocock/skills holds the weekly board at 290,735 stars. Google shipped google/skills (the repo behind Argon’s .agents/skills rulebook that drove the ADK optimization), K-Dense-AI/scientific-agent-skills is at 49,012 stars, and the newsletter flagged Anthropic dropping 400+ free ready-to-use agents and skills for Claude. Skills went from a config convention to the unit of distribution for agent capability in about a month.

The rest of the board: morluto/rea is now at 231,359 stars and still leading daily, debpalash/VoiceStudio holds at 65,261, HKUDS/CLI-Anything (52,259) and mksglu/context-mode (28,820) keep the agent-tooling lanes hot, and thedotmack/claude-mem (100,662) is the weekly memory entry. New heat: ifixai-ai/iFixAi (26,796) and vxcontrol/pentagi (28,923).

Signal: A skill is a markdown file with opinions; the star counts say opinions are now the most forkable asset in AI. The skill repos are winning because they are the cheapest possible distribution for the most expensive thing: judgment about how to use the model. Sources: GitHub trending · google/skills · Anthropic’s 400+ agents and skills

09 · Product Hunt

Yesterday’s top products (Oct 13)

1. Articos MCP — See how simulated buyers react, right in your agent (325 votes)

2. OpenCode 2.0 — The open source AI coding agent (292 votes)

3. Fina Business — The flexible $20/mo QuickBooks alternative for founders (259 votes)

4. Muse Wonder — Meta’s most capable models for personal agents (236 votes)

5. Kai — Autonomous assistant designed to handle meetings and email (195 votes)

6. MentionBird — Ahrefs for AI Search (156 votes)

7. MausBot — Open source Grok Bot: Free, Local (136 votes)

8. Takes for Mac — Record talking-head videos. Claude Code does the edit. (127 votes)

9. PostNitro for Agents — Let your AI agent design and publish social posts (127 votes)

10. Toby — Realistic mock interviews with an AI hiring manager (110 votes)

Two worth a click: OpenCode 2.0 won second place the same week Anthropic published the Scion post, which is the open-source agent catching the wave the labs are surfing. MentionBird (“Ahrefs for AI Search”) is the SEO industry’s AI-search moment arriving as a product: someone is selling rank tracking for the answers, not the links.

Signal: The board keeps assuming the agent is the user: an MCP for simulated buyers, social posts designed by your agent, an open-source coding agent. Product Hunt is an agent-tooling trade show now, and the votes agree. Sources: Product Hunt

10 · Hacker News

Gemini 4 Argon

1,380 points, 986 comments. Covered in The One Thing. The thread’s best material is the pricing math: half of Muse Wonder and Claude, and the cached-input discount that makes long-context agent loops cheap.

Ledger confirms spy chip implant behind wallet crypto thefts

755 points, 472 comments. Covered in The Argument. The comments are doing supply-chain forensics: which fabrication and assembly steps were in scope, and whether “verify the hardware” is even a coherent ask for consumers.

Show HN: Guitoro, a virtual guitar you strum with your keyboard

515 points, 141 comments. A browser guitar you play with keystrokes, and the comments are pure delight, which is rare enough to note on a day of supply-chain compromises.

Also notable

AOSP Source Releases Are So Late Developers Disassemble Libraries to Report Bugs (405 points, 191 comments): covered in The Argument. Amazon pushes up prices across the Internet (334 points, 313 comments): the everything-store’s fee stack as an inflation engine, with receipts. MIT class project turns into an FDA-cleared treatment for tremors (248 points, 50 comments): the rare student-project-to-medical-device pipeline. Pg_vectorsearch: An index that is faster than pgvector’s IVFFlat and HNSW (89 points, 33 comments): a faster vector index from Timescale, the database lane keeps shipping.

11 · OpenRouter Apps

The board this morning

Command Code holds at ~4.18T tokens, flat, which confirms the new baseline. The fight below it is the story: ZCode (~1.92T) and Hermes Agent (~1.92T) are now statistically tied for second, with Hermes still accelerating. Nous Research API (~1.30T) is the new name on the board. Claude Code (~1.50T) and Kilo Code (~951B) hold their lanes. StepCode (~695B) is flat on day four after its October 10 launch: real adoption, not a spike.

Signal: The top of the board is now a two-horse race for second between the Chinese-lab agents, and the field below keeps widening. The token leaderboard is a map of where coding agents actually run, and the map keeps saying the same thing: the action is in the agent wrappers, not the chat UIs. Sources: OpenRouter rankings

12 · Watch

WesRoth: “GEMINI 4 ARGON IS WEIRD…”

WesRoth · Oct 14 · the launch-morning reaction to Argon: the weird bits Google buried in the announcement. Watch against the Leaderboard numbers, not the thumbnail. The video companion to The One Thing.

TheAIGRID: “Google Just DROPPED Its Most Powerful AI Ever”

TheAIGRID · Oct 14 · the launch-day episode on Gemini 4 Argon: the benchmark claims, the pricing, and the receipts. Watch the vendor claims against the Signal callouts, not the thumbnail.

13 · Listen

Latent Space: “Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy”

The new episode: Jason Ma of Dyna Robotics on general-purpose physical agents with hour-long autonomy. Pairs with the Money Flow card on Physical Intelligence’s open API and Argon’s cornhole robot: the physical-agent stack is the week’s second storyline.

From the queue

The backlog still holds “Why AlphaFold Didn’t Solve Protein Folding,” the Periodic Labs episode, Hard Fork’s “Anthropic’s Quest to Give A.I. Morals,” and Lenny’s Josh Woodward interview. The Watch section has the fresh material.

14 · Read

Gemini 4 Argon: Built to solve your hardest problems

The launch post for The One Thing, with all the receipts.

Redesigning source control at Anthropic

The Scion deep-dive: 30,000 merges a day, 600 PB of would-be git traffic, and the team of seven.

A free 472-page textbook on agentic LLMs

Leiden University’s courseware for the agent era.

Fish Audio Drama 3 deep dive

The 35B-A3B architecture, the community license, and what is still unverified.

15 · Safe to Ignore

Permission to skip

Skip: Iosevka is an open-source monospace font. 401 points of typography news. Wrong newsletter. Skip: The United Kingdom generates less electricity per person than in the 1970s. 376 points of energy data. Wrong newsletter. Skip: Recreational Programming. 263 points of conference talk. Wrong newsletter. Skip: Felix Rieseberg Homepage. 250 points of personal-site news. Wrong newsletter. Skip: 900-Page Pre-Pantone Guide to Color from 1692. 222 points of archival delight. Wrong newsletter. Skip: Sound – Bartosz Ciechanowski (2022). 206 points, a repost. Wrong newsletter. Skip: Switzerland proposes banning profiling-based ads to minors. 203 points of policy news. Wrong newsletter.

16 · Build on This

The skills auditor

Skills are the cheapest distribution for the most expensive thing in AI: judgment about how to use a model. The board proves it: 290,735 stars on mattpocock/skills, Google shipping its own skills repo, 400+ skills dropped for Claude this week, and a 472-page textbook teaching the next cohort to write them. But a skill is a markdown file with opinions, and opinions can hide instructions: prompt injection, credential exfiltration, runaway tool calls, dependencies that phone home.

The skills auditor is a certification service for agent skills and plugins. Drop in a skill repo or a Claude Code plugin; get back a static and dynamic audit (injection surface, tool-call graph, network calls, file-access scope) and a clean bill of health or a flagged report, plus a badge for the README. The customer is every team whose agents install skills the way developers install npm packages: without reading them. Pricing follows the unit of risk: per audit, pay as you go, no subscription, with team packs for CI pipelines that audit on every skill update. The moat is the corpus: every audit makes the next one’s heuristics sharper, and the badge becomes the trust mark the ecosystem reaches for before installing.