FOMO Buster #005: The silent patch

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 patched a silent vision bug in GPT-6 Sol and Luna

Somewhere in the last week, GPT-6 Sol and GPT-6 Luna started seeing worse — and nobody was told. An image encoding bug was quietly degrading vision quality across the API, Codex, and computer-use surfaces, and OpenAI shipped the fix with none of the fanfare that accompanied the models’ launch. There’s no changelog entry with a timeline, no post-mortem on how long the bad encoder was live, and no tooling that would have let an API customer notice on their own. If your agent’s vision got flaky last week, it wasn’t your prompt.

This is the sharp edge of the model-as-service era: the endpoint you call today is not byte-for-byte the model you evaluated last month, and the diff is a patch note you may never see. Benchmark scores are snapshots; the production surface is a moving picture.

Signal: Silent degradations are the new downtime — invisible, uncompensated, and undiscoverable without your own eval harness. Every serious API customer now needs a canary: a fixed set of probes run against the endpoints they pay for, alerting when quality drifts. Nobody sells that as a service yet. Sources: AlphaSignal · AA: GPT-6 Sol and Luna push the cost-efficiency frontier

02 · Lab Watch

Fireworks launches Ember-1: Kimi K3’s quality at 40% fewer tokens

Fireworks Research’s first own-model: Ember-1, a K3 derivative trained to reason efficiently. Over 50 training runs and 200 evals taught it to cut unnecessary reasoning while keeping the thinking that matters — 35–50% shorter reasoning traces with no accuracy loss, ~39% fewer total tokens in live A/B tests on production coding traffic. On SWE-bench Verified it scored 92.2% vs K3-max’s 93.2%; on DeepSWE 1.1 it actually beat K3-max (75.2% vs 66.4%). The honest detail: they ship it as a two-week Research Preview on serverless — it becomes permanent only if community demand says so. The strongest validation line in the post: Fireworks switched its own internal coding traffic to Ember-1 first, and “no news is good news” — developers didn’t notice the change, only the smaller token bills.

Signal: The frontier isn’t just smarter models; it’s trained efficiency — not turning down reasoning effort, but teaching the model to think less and answer the same. The labs that own inference infra (Fireworks, not the GPU-poor) can turn this loop fastest. Watch for an Ember series the way you watched the Kimi series.

H Company’s Holo4 handles screens, code, and APIs with 79% fewer tokens

H Company’s latest agent model tackles the computer-use problem — driving screens, code, and APIs — while burning 79% fewer tokens than the comparison set. Paired with Ember-1, the theme of the weekend is unmistakable: the agent economy’s biggest cost line is reasoning tokens, and the whole industry is now machining them down.

Signal: Token efficiency is becoming a product category, not an optimization. The winner of agentic coding won’t just be the smartest model; it’ll be the one whose thinking is cheap enough to run at human speed, all day.

Sakana AI’s SAIL triples robot success rates by searching before moving

Sakana’s new robotics method has the robot search the environment before acting — look around, build the plan, then move — and triples success rates. The same “think first” pattern as Ember-1, but with a camera and a gripper: acting less, better.

Signal: Robotics is converging on the agent playbook: perception-first, then action. The robot that pauses to look is the one that doesn’t knock the glass over.

03 · Open Weights

Xiaomi’s MiMo family climbs; Apple drops a VLM; the laya paradox holds

The week’s delta between HF trending (what’s moving) and most-liked (established gravity) is rich. convaiinnovations/laya still leads trending (2,501) at 4,227 likes — attention without loyalty, week after week. The Qwen-Image-2.1 gravity well deepens: its uncensored GGUF crossed 1.06M downloads, and the community is now stripping its refusals — a refusal-ablated text encoder takes refusal rates from 100% to 5% while running locally on 16GB laptops (AlphaSignal, and the pottokao ablterated encoder sits at #21 trending with 159K downloads).

New climbers not yet in the most-liked tier: the Xiaomi MiMo V2.6 family (Pro-RL, Flash-RL, and a Qwen-distilled 9B — three entries in the top 10), XingChen-AGI/Xing4.0-29B-A4B (#5), XingChen-AGI/TeleOCR (#7), TaichuAI/ZDTaichu5.0-9B (#11), Altworld/Hemmingway-1 (#12), nvidia/Nemotron-3-Diarization (#14), yandex/AliceAI-Foundation-80B-A3B (#27), and — notable — apple/LensVLM-9B (#23): Apple’s own vision-language model, sitting quietly in the trending list. Established holds: Qwen/Qwen3.8-27B remains the most-liked model on the platform (16,465), while FLUX.1-dev, DeepSeek-R1, and Kimi-K3 keep their like counts but have fallen off the trending board entirely.

Signal: Xiaomi is running a full model family as a release strategy and it’s working — three V2.6 variants trending at once. And watch Apple: LensVLM at #23 trending with 255 likes is the kind of quiet open release that precedes a louder one. Sources: HF trending · XiaomiMiMo/MiMo-V2.6-Pro-RL · apple/LensVLM-9B

04 · Money Flow

Artificial Analysis’ Cyber Index exposes a 65x cost gap among top AI models

AA’s new Cyber Index benchmarks models on security work — and the price spread across the leaderboard is 65x. Same nominal capability tier, two orders of magnitude apart in cost. That’s not a benchmark result; that’s a market inefficiency with a countdown timer on it.

Signal: Every benchmark that attaches a dollar sign to a score is really a procurement document. The 65x gap is the size of the arbitrage window for whoever builds routing that actually trusts the numbers.

05 · Leaderboard

Framer takes #2 on OpenRouter; the open-source coders widen the gap

Hermes Agent still leads (~1.73T tokens), but the new #2 is Framer (~1.38T) — a no-code website builder now routing more tokens than Portkey (~1.23T). The coding-agent reshuffle from Friday deepened: Cline (~927B) and Kilo Code (~869B) both extended their lead over Claude Code (~626B). Behind them: pi (~459B), HighLevel (~392B), omp (~385B), Strix (~335B), Command Code (~305B), Freebuff (~293B), Codex (~265B), DeepSeek Harness (~228B).

Signal: The two growth vectors from Friday both accelerated: creative tools (Framer #2, HighLevel climbing) and open-source coders out-routing the flagship. The token economy is now growing fastest outside the models’ first-party apps — distribution, not weights, is where the leverage moved.

06 · The Argument

AI companies in race to demonstrate their model most threatening to humanity

388 points, 304 comments on a satire that landed because the premise is barely exaggerated: the labs now market their safety reports the way defense contractors market deterrence. The thread’s real argument, under the jokes, is whether “look how dangerous our model is” became a capability signal — dangerous means powerful means fundable — and whether the alignment theater is now load-bearing for valuations.

Signal: Satire at 388 points means the premise has recognition value. When the joke everyone gets is “the labs compete on threat level,” the discourse has officially inverted: safety communication is read as marketing, and marketing is read as a warning.

07 · Papers

Learning to Stop without Learning to Stop: Self-Supervised Confidence Training Improves Reasoning Efficiency

Teaching reasoning models when to stop thinking — self-supervised confidence training that trims the tail of overlong chains. Read it next to Ember-1: the research frontier and the product frontier are attacking the same token bill from opposite sides.

Signal: “When to stop” is becoming a first-class training objective. Expect this knob — confidence-gated early stopping — inside every reasoning API within a year.

Compact Documentation for Coding Agents: A Benchmark, an Optimizer, and Why It Does Not Transfer

A benchmark and optimizer for squeezing library docs into the context windows of coding agents — with the honest finding that the compression does not transfer across models. Every agent needs its own doc diet.

Signal: The “it doesn’t transfer” result is the important one: context engineering is per-model work, not a reusable artifact. That’s either a services business or a great argument for standardized agent-readable docs.

User Model Extraction via Belief Self-Distillation

Extracting a model of the user — their beliefs, preferences, blind spots — by distilling the assistant’s own beliefs about them. Powerful for personalization; unnerving for privacy. The paper treats the user model as an object you can build, query, and presumably steal.

Signal: The next privacy fight isn’t about your data; it’s about the model of you that every assistant is quietly building. This paper gives that object a name and a method.

08 · GitHub Trending

The theme: the agent’s IT department

The week’s board is infrastructure for living with agents, not agents themselves. vectorize-io/hindsight (40K stars, +4.4K today) — “Agent Memory That Learns,” the week’s fastest gainer. paperclipai/paperclip (92K stars, +3.2K today) — “the open-source app everyone uses to manage agents at work.” cloudflare/security-audit-skill (+4.8K this week) — a coding-agent skill for multi-phase security audits, from Cloudflare of all places. Tencent/WeKnora — turn raw documents into a queryable RAG and autonomous knowledge base. trycua/cua — scale computer-use 2.0 with open-source drivers and cross-OS fleets. stablyai/orca — the ADE for running a fleet of parallel agents.

Signal: Manage them (paperclip), remember for them (hindsight), audit them (security-audit-skill), feed them (WeKnora), watch them click (cua). The agent stack has matured past “the model” into the unglamorous tooling every production deployment needs — and open source is winning each layer first.

09 · Hacker News

Ember-1

536 points, 235 comments. Fireworks’ efficiency-trained Kimi K3 (see Lab Watch) drew the week’s most technically literate thread: commenters doing the math on reasoning-token economics, debating whether “trained efficiency” beats “effort knobs,” and noting the Terminal Bench numbers (Ember-1 82.0% vs K3-max 80.9% at half the tokens) with the particular joy of people who pay these bills themselves.

Signal: HN’s enthusiasm is a leading indicator for infrastructure spend: when the comments are full of people calculating their own token savings, the product has found its buyer. The efficiency-model category just got its reference implementation.

Also notable

Owed a billion dollars in Nvidia stock (900 points, 384 comments): the colo.to narrative on the strangest IOU in tech — worth reading alongside today’s Money Flow. Malleable software (130 points, 63 comments): the ink&switch essay on restoring user agency in locked-down apps keeps resurfacing, because the locked-down apps keep winning. Show HN: Lofi Cities (290 points): pixel-art city nights with browser-generated lofi — the weekend’s palate cleanser.

10 · OpenRouter Apps

Where the tokens actually flow

Hermes Agent (~1.73T), Framer (~1.38T), Portkey AI (~1.23T), Cline (~927B), Kilo Code (~869B), Claude Code (~626B), pi (~459B), HighLevel (~392B), omp (~385B), Strix (~335B), Command Code (~305B), Freebuff (~293B), Codex (~265B), DeepSeek Harness (~228B).

Signal: Framer at #2 is the headline, but the quieter story is HighLevel (~392B, launched July) sitting above a cluster of coding agents. The no-code/marketing-automation crowd routes tokens like developers now — the API customer base is no longer who you think it is.

11 · Watch

What If Each AI Generation Gets Slightly Less Aligned? — Noam Brown

Dwarkesh Patel · Sep 27

Signal: The newest in the Noam Brown series — the alignment-drift thesis, days after an actual training pause over a sandbox escape. Watch the Sep 26 “wall” episode first if you missed it; this is the darker sequel.

Meta is pivoting again… everything you missed from Connect 2026

Fireship · Sep 25

Signal: Fireship’s 100-seconds-plus treatment of Meta Connect — the fastest way to absorb whatever Meta just pivoted toward this time.

12 · Listen

OpenRouter: from Seed to Stripe

Latent Space · Sep 25 — with OpenRouter’s Alex Atallah and AMP’s Anjney Midha on the seed-to-Stripe journey: how the token router became infrastructure, and what the Framer-at-#2 era means for the business of being the tollbooth.

Signal: Pair with today’s Leaderboard. The company that sees every token flow just explained how it got there — and the episode doubles as a field guide to where the API economy is heading.

13 · Read

RedAmon turns 100 security tools into an open-source AI hacking pipeline

A trending open-source framework that chains reconnaissance, exploitation, and post-exploitation into one autonomous agent pipeline — with human oversight gates. The red-team equivalent of everything in GitHub Trending this week, pointed the other direction.

Signal: Offensive security tooling is getting the same agent-ops treatment as everything else. The “human oversight gates” framing is the tell: even the hackers are building the audit layer first.

LG’s EXAONE 3.5 hits 1 million monthly downloads running on 5GB

LG AI Research’s 4-bit quantized 7.8B bilingual (Korean-English) model: 1M monthly downloads, ~5GB of weights, 32K context. The chaebol lab quietly built one of the most-downloaded small models on the platform.

Signal: The small-model download charts are increasingly East Asian: LG, Xiaomi, Qwen. The “runs on a laptop” tier has its own geopolitics now.

Open-source jevgrep cuts coding agent costs 40% by offloading code search

A grep-shaped tool for the Jev-era stack: push code search out of the expensive reasoning model into a purpose-built retriever, cut agent costs 40%. The System-One pattern from last week, now as a unix-philosophy utility.

Signal: The Jev through-line keeps widening: laya routers, Jev-Omni on HF trending, and now jevgrep. “Don’t make the big model do the small model’s job” is becoming the defining cost discipline of the agent age.

14 · Safe to Ignore

Permission to skip

Skip: “When did Google get so weird?” 1,542 points, 829 comments, and not really about AI. Weekend entertainment. Skip: The TeleOCR encore. Covered Friday as StarDoc-AI’s; today’s version is the same family wearing a new benchmark. One doc-parser revolution per week is plenty. Skip: Lenny’s burnout episode. Real topic, wrong newsletter. Skip: The “AI labs may be hiding breakthroughs” video. Speculation about secret models is the one genre this newsletter will not feed.

15 · Build on This

The Canary

Today’s One Thing is a product hiding in a patch note: OpenAI silently degraded and silently fixed GPT-6 vision, and no customer had the tooling to notice either event. Every company building on a model API has the same exposure — the endpoint changes under them, and their first alert is a user complaint. Product: continuous regression monitoring for the model APIs you depend on — a fixed probe suite (vision, reasoning, tool-use, refusal behavior) run on a schedule against your pinned endpoints, with alerts when quality drifts beyond a threshold and a diff report when it does. Sell it as $39/month per monitored endpoint, or $19 one-time audits for the curious. The probes are commodity evals; the product is the schedule, the baseline history, and the “your model changed on Tuesday” email nobody else sends.

16 · Appendix

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