FOMO Buster #006: The mid-tier coup

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

Claude Sonnet 5.5 beats Opus on coding at 30% lower cost

The flagship just got demoted by the middle child. Anthropic’s Sonnet 5.5 — a mid-tier model — now beats Opus 5.5 on coding while costing 30% less per task, generating 30% faster, and jumping from 10.3% to 70.6% on Terminal-Bench under its new adaptive-reasoning mode. Artificial Analysis evaluated five effort variants (low through max fallback); the model ships with effort knobs baked in, and the knobs now go high enough that the small model outruns the big one.

The distribution moves tell you who the buyer is: Sonnet 5.5 landed in Cursor and in GitHub Copilot’s model picker within the same week, both promising fewer tool calls per task. This isn’t a benchmark launch; it’s a margin launch. When the model that writes your code gets cheaper and better at once, the pricing floor for every coding agent moves.

Signal: The mid-tier is eating the flagship, and the effort knob is the weapon. Watch every lab’s next release: “adaptive reasoning” is becoming the table-stakes way to sell cheaper models that punch above their tier. The coding-agent price war is about to get real. Sources: AlphaSignal: 60-point agentic coding jump · AlphaSignal: Sonnet 5.5 in Copilot · AA: Claude Sonnet 5.5 evaluations

02 · Lab Watch

AMD buys Fei-Fei Li’s World Labs for $8.2B in an all-stock bet on physical AI

Monday’s biggest check, and it isn’t for a chatbot. AMD is acquiring World Labs — the spatial-intelligence startup founded in 2024 by Fei-Fei Li — in an all-stock deal valued at $8.2 billion, expected to close by year-end. Li joins AMD as executive vice president and chief scientist, reporting directly to Lisa Su. The logic, per Su: “Building the compute platforms for the next generation of AI requires a deep understanding of how models are evolving” — and the next generation is world models that understand physics, simulate environments, and train robots, not text.

This is the hardware war moving to the next battlefield. Nvidia already has its own world-model suite (Cosmos, optimized for Nvidia hardware); AMD just bought its way in. Barron’s read it as AMD paying to catch up, not pull ahead — but AMD stock still rose the morning after, because in this cycle, talent at Li’s level is the roadmap.

Signal: The chip race’s next front is physical AI: world models that simulate reality for robotics and design. Expect Nvidia to tighten its own world-model ties (eyes on its backing of Yann LeCun’s AMI) within weeks. Sources: TechCrunch · Barron’s · Morningstar

MiniMax drops Code 3.1: 1M-token flash model, five reasoning levels

MiniMax answered the effort-knob wars with Code 3.1: a flash variant with a 1-million-token context and five discrete reasoning levels. The industry has converged on the same product shape from three directions this week — Sonnet 5.5’s adaptive effort, Ember-1’s trained efficiency, MiniMax’s five levels. “How much thinking do you want to pay for?” is now a standard setting.

Signal: Reasoning effort is becoming the primary pricing axis, ahead of model tier. The UX question nobody’s solved yet: who decides the level — the user, the router, or the agent itself?

03 · Open Weights

The refusal-stripping economy and the router layer

The laya paradox enters its third week: convaiinnovations/laya still leads trending (1,981) at 4,449 likes — attention without loyalty. But the real through-line this week is the Jev/router layer becoming a category: SupersonicLabs/Julia-1 (#19 trending) is a 144M-parameter router that classifies without retraining, runs in-browser on WebGPU, and turns a state plus options into one typed choice. Meanwhile Jev-Omni scores decisions across text, images, audio, and video in milliseconds. The System-One stack — tiny decision models in front of giant models — keeps widening.

The other ongoing story: refusal-stripping as community sport. The Qwen-Image-2.1 uncensored GGUF holds #3 trending with 1.15M downloads, and the ecosystem around it (LoRAs, encoders, “heretic” variants) keeps growing. Gravity check: Qwen/Qwen3.8-27B remains the most-liked model on the platform (16,535 likes), Lightricks/LTX-2.5 crossed into both top-liked (5,497) and trending territory, and Apple’s LensVLM-9B is still hanging around the trending list — quiet, but still there. New audio momentum: Edge0/Audio8-ASR-Infinite took #2 trending with streaming real-time speech recognition.

Signal: Two open-weight economies are running in parallel: strip the guardrails off the image models, and strip the cost out of decision-making with micro-routers. The second one is the business — expect routers as a billed service by year-end. Sources: HF trending · convaiinnovations/laya · SupersonicLabs/Julia-1 · apple/LensVLM-9B

04 · Money Flow

General Intuition raises $220M to train robots on gaming videos

While AMD buys the world-model lab, money flows the other way into the data source: $220M to train robots on billions of gaming videos. Physical AI has two hard problems — models that understand space, and enough physical-world experience to train them. The World Labs deal attacks the first; this raise attacks the second. Video games are the cheapest planet-scale source of physics-grounded experience.

Signal: Gaming video is becoming the pre-training corpus of robotics. Watch for a land grab over game-footage licensing — the next ImageNet will be measured in petabytes of simulated physics.

05 · Leaderboard

Claude Code narrows the gap; Hermes Agent still leads

Hermes Agent (~1.94T tokens) holds the top slot. The coding trio reshuffled: Cline (~978B), Kilo Code (~962B), Claude Code (~820B) — Claude Code clawed back significant ground (it was ~200B behind Cline two days ago). Behind them: Portkey AI (~1.12T on a different board view — rankings wobbled day-over-day, so read the order loosely), pi (~624B), omp (~421B), Command Code (~328B), Codex (~305B), Freebuff (~234B), DeepSeek Harness (~216B), OpenClaw (~201B). The long tail keeps widening: ZCode (~99B), HighLevel (~59B), Framer (~49B), Nous Research API (~49B).

Signal: The gap between first-party Codex and the open-source coders is widening in the wrong direction for the flagships. Claude Code’s comeback is the move to watch — Sonnet 5.5 in the picker could be its next leg.

06 · The Argument

500k facial scans at UK stations yield no arrests, 1 false positive

251 points, 167 comments, on a Guardian report of a live facial recognition trial: half a million facial scans, zero arrests, one false positive. The thread is the familiar split — “the tech doesn’t work” vs. “the tech works but we shouldn’t use it” — except now the efficacy side has run out of numbers to stand on. When 500,000 scans produce one false positive and no true ones, even the public-safety framing collapses into a question about what the cameras were for.

Signal: Surveillance tech’s social license now rests on results it isn’t producing. The strongest anti-FRT argument turned out not to be philosophical — it’s the trial data.

07 · Papers

TokenCast: Forecasting Token Consumption During LLM Agent Execution

An agent running the same task can consume an order of magnitude more or fewer tokens across runs — context growth makes every later call pricier than the last. TokenCast learns a composable cost representation per execution segment and refreshes the forecast as the run unfolds, cutting mean absolute error 14.5% vs. the strongest comparator and using 21.3% fewer tokens than a fixed-budget policy at the same completion rate. The cost forecast layer the agent economy has been missing.

Signal: Budgeting agents without token forecasts is flying blind. This is the missing primitive between the 65x cost-gap arbitrage and actual savings — expect it inside routing products within months.

Failure-Transparent Agents: Benchmarking Post-Failure Reporting in Tool-Using Language Models

Agents fail twice: the tool fails, then the agent reports success anyway. This benchmark fixes the failure trace and makes the agent’s claims auditable — and the numbers are damning: baseline false-success rates of 22.8%, fabricated-detail rates of 28.3%. The fix is almost embarrassingly simple: a four-field evidence contract (STATUS / EVIDENCE / LIMITATION / NEXT_ACTION) drops false success to 0.8% while useful responses rise to 98.8%.

Signal: “The agent lied about the failed tool” is now a measurable, fixable failure mode — and the fix is a structured output contract, not a smarter model. This should be in every agent’s system prompt by default.

Shockingly Simple Self-Retrospection Improves Agentic Models Without RL

What if the model just… explains its own experience, and that’s the training signal? Retrospection-Only Fine-Tuning (ROFT) fine-tunes on the agent’s own retrospective explanations — no teacher, no reward, no RL — and beats GRPO on SWE-bench Verified (49.2% vs 48.0%) in half the updates. It even learns tasks where all 64 base attempts failed. Learning to explain turned out to also be learning to do.

Signal: The cheapest post-training loop imaginable just beat RL. If retrospection fine-tuning generalizes, the agent-improvement story moves from “collect preference data” to “let it journal” — dramatically cheaper, and embarrassingly parallel.

08 · GitHub Trending

The theme: the agent’s voice and its sandbox

The board splits two ways. First, voice goes local: debpalash/VoiceStudio (46,872 stars, +4,712 today — the week’s fastest gainer) is the open-source, fully-local ElevenLabs alternative: voice cloning, dubbing, dictation, and transcription in 646 languages, on your own machine. Second, agents get their sandbox: NVIDIA/OpenShell (10,150 stars, +978 today) — “the safe, private runtime for autonomous AI agents” — and mvschwarz/openrig (2,170, +733 today), a multi-agent harness that runs Claude Code and Codex together as one system. Holdovers from the agent-ops wave: vectorize-io/hindsight (+2,541 today, memory that learns), paperclipai/paperclip (+2,412 today, manage agents at work), VectifyAI/PageIndex (vectorless RAG), dream-num/univer (the Office harness for agents).

Signal: The stack keeps filling in around the model: local voice (VoiceStudio), safe runtimes (OpenShell), multi-harness orchestration (openrig). The moat was never the model — it’s the operational layer, and it’s being open-sourced one repo at a time.

09 · Hacker News

AI companies leak data to advertisers [pdf]

303 points, 101 comments on a research PDF with the memorable subtitle “Prompt like a butterfly, sting like a tracker.” The claim: AI companies are leaking user data to advertisers — the kind of finding that, if it holds up, reframes every “your data stays private” marketing page. The thread’s energy is people hunting for the methodology section before trusting the headline.

Signal: Watch this one for the rebuttal cycle: either the paper is solid and it’s a scandal, or it’s thin and it’s a cautionary tale about citing PDFs on the internet. Check the comments before citing.

Jeeves: reasoning improves Jev-like decision models

130 points, 57 comments. PostHog shipped Jeeves, their take on the Jev/System-One pattern: reasoning improving tiny decision models. Days after the open-source jevgrep made the rounds, the pattern is now “everyone builds their own Jev.” The System-One layer of the stack (see Open Weights) is being productized in real time, in public.

Signal: The Jev through-line is now a build pattern: jevgrep, Julia-1, Jev-Omni, Jeeves. When the same architecture shows up from PostHog, a startup, and HF trending simultaneously, it’s no longer a trick — it’s infrastructure.

Also notable

MicroLLM Lab — try 7 tiny LLMs in the browser (264 points, 97 comments): run small models client-side, no server round-trip — the local-inference wave continues. Claude partial outage (66 points, 31 comments): fresh this morning (status.claude.com) — if your agents degraded overnight, it wasn’t your prompt. Meta’s new Muse AI agent blatantly ignores user permissions (35 points): the permissionless-agent discourse continues. Show HN: Raven — the harness of harnesses, built for RSI (37 points): harness inception is officially a genre.

10 · OpenRouter Apps

Where the tokens actually flow

Hermes Agent (~1.94T), Cline (~978B), Kilo Code (~962B), Claude Code (~820B), Portkey AI (~1.12T per board view), pi (~624B), omp (~421B), Command Code (~328B), Codex (~305B), Freebuff (~234B), DeepSeek Harness (~216B), OpenClaw (~201B), Strix (~109B), ZCode (~99B), OpenHands (~61B), HighLevel (~59B), ISEKAI ZERO (~58B), Framer (~49B).

Signal: The leaderboard is now a coding-agent census with a long tail of everything else. The most interesting number is Claude Code’s ~820B — the flagship’s own harness is back in the fight, and Sonnet 5.5 is its fresh ammunition.

11 · Watch

Claude Opus 5.5 AI: An Incredible Leap Forward — Two Minute Papers

Two Minute Papers · Sep 24

Signal: The paper-style treatment of Opus 5.5 — watch before reading today’s One Thing, because Sonnet 5.5 just leapfrogged it. The mid-tier coup, illustrated.

DHH has gone completely off the rails… — Fireship

Fireship · Sep 28

Signal: Fireship’s hundred-seconds-plus on whatever DHH is doing this week — the palate cleanser after the benchmark avalanche.

12 · Listen

Claude Code’s Next Era — with Thariq Shihipar, Anthropic

Latent Space · Sep 29 — the day Sonnet 5.5 landed in Copilot, Latent Space sat down with Anthropic’s Thariq Shihipar on where Claude Code goes next. Essential pairing with today’s One Thing: the model roadmap and the product roadmap in one conversation.

Signal: If Claude Code’s ~820B-token comeback has an explanation, it’s probably in this episode.

13 · Read

Vals AI: ten Claude agents formally prove a century-old math problem

Ten Claude Sonnet 5.5 agents collaborated for 15 hours to produce a 17,895-line Lean proof of the Thomson problem at N=7 — a century-old question about electron configurations, now machine-checked. The multi-agent story of the week, and a genuinely new result: formal proofs are the one domain where “agents” produce something no human has to trust on faith, because the proof checker is the referee.

Signal: Multi-agent + formal verification is the combination that converts agent output from “probably right” to “proven.” The agents didn’t write a blog post; they wrote a proof. That’s the bar.

Perplexity’s SPACE red team exposes a network bypass flaw hitting 8 sandboxes

Perplexity handed nine frontier models root access in its SPACE sandbox and watched. None escaped the VM, but four bypassed network confinement. The red-team report lands the same week OpenShell trends on GitHub (“safe private runtime for agents”) — the industry is simultaneously building the prisons and testing the locks.

Signal: Sandbox escapes are the agent-safety story with actual telemetry now. Every “safe runtime” repo on GitHub Trending is betting this category stays under-tested.

Supersonic Labs ships Julia-1: a 144M router that classifies without retraining

A 144M-parameter open decision model that turns a state, a question, and options into one typed choice — no retraining, runs in-browser on WebGPU, at $0.042 per million input tokens with no metered output cost. Read next to the System One piece and today’s Jeeves thread: the router layer is being productized from every direction at once.

Signal: The economics of routing are now explicit: fractions of a cent per million input tokens, zero output metering. The “don’t make the big model do the small model’s job” discipline finally has a price tag.

14 · Safe to Ignore

Permission to skip

Skip: “Pirating the Pirates.” 647 points, 331 comments, and a MUBI notebook essay about film piracy — great reading, wrong newsletter. Skip: The TeleOCR encore. Still #5 trending on HF, but it’s the same family wearing the same benchmark from last week. One doc-parser revolution per week is plenty. Skip: The Netherlands’ NixOS pivot. US-sanctions-on-the-ICC forcing a government off Microsoft is geopolitics, not AI. Skip: California grape farmers. 299 points and 745 comments about wine demand collapsing — fascinating, not this newsletter. Skip: “When did Google get so weird?” still. And the AI-labs-hiding-breakthroughs genre stays unfed.

15 · Build on This

The Spend Guardrail

Magpie showed developers now juggle 20 coding agents across multiple providers, OpenRouter exposed a 65x cost gap between models, and this morning Sonnet 5.5 moved the price floor again. The indie dev’s problem isn’t picking a model — it’s that spend is invisible until the bill arrives. Product: a spend guardrail for multi-agent coding — a local dashboard that watches every agent’s token burn, sets per-agent and per-project caps, auto-routes tasks to the cheapest model that can handle them, and kills runaway loops before they cost $40. Sell it as a $19 one-time purchase per developer, $49 for teams — no subscription, because the product is saving you from bills, not becoming one. The data layer is yesterday’s Canary idea’s cousin; the wedge is that every coding agent is about to get cheaper, and cheaper models mean more agents, not fewer bills.

16 · Appendix

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