Media Zone | 2026-10-04
A quiet US Saturday, and the useful posts all say the same thing: put the expensive part only where it changes the outcome. Anthropic's advisor tool keeps its strongest model on call for three moments of a coding session. CMU trains a 4B model to edit agent harnesses better than its 35B teacher, and Meta routes each input to one of several specialized harnesses. AutoCompact teaches an agent when to throw context away, and Kepler's perfect ARC-AGI-3 run shows the bill: 97% cache reads. Under that, decision models keep landing locally, a PyTorch Conference poster spends saved inter-GPU bandwidth on fixing quantization error, and Cerebras argues the cheapest chip is the one that needs no HBM, CoWoS or 3nm. The loudest thread was not technical: another OpenAI safety resignation, and a long argument about AI consciousness.
Today's signal
- Dominant story: harness optimization grows a learned editor and a router. Harness Learning (CMU), Mixture of Self-Improving Branches (Meta) and AutoCompact were the evening's most-shared research.
- Escalation, not default: Claude Code's advisor runs Sonnet for the work, Opus to plan, and calls Fable only at three checkpoints.
- Audit the run: Kepler's 100 on ARC-AGI-3 included an invalid perfect run from reading game source; the honest one cost $778, mostly cache reads.
- Supply gates: Cerebras names HBM, CoWoS and 3nm as the caps on accelerator shipments; memory stays tight through 2028.
- Counter-signal: "context is a scarce resource, prune it" and Garry Tan deleting a thousand lines of agent instructions. Some harness complexity is debt.
- Quiet areas: no new bookmarks or curated reposts (the captures ran and found none), LinkedIn returned posts with no text, Reddit had nothing. The X Following feed carries the Media Zone.
Routing, KV cache, compression, GPU
Call the expensive model only when it matters
Three items from the US day make one argument. Most steps in an agent loop do not need a frontier model. The skill is deciding which ones do.
flowchart LR
T["Task<br/><small>coding session</small>"] --> P["Opus 5.5<br/><small>plans and merges</small>"]
P --> W["Sonnet subagents<br/><small>read, edit, run tests</small>"]
W --> C{"Checkpoint?<br/><small>plan, repeat error, done</small>"}
C -->|yes| A["Fable advisor<br/><small>reads full session</small>"]
C -->|no| W
A --> P
P --> S["Shipped change<br/><small>reviewed before done</small>"]
classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
class T input
class P,W,A core
class C loop
class S exit
linkStyle 3 stroke:#f08c00,stroke-width:2px
linkStyle 4 stroke:#f08c00,stroke-width:2px
linkStyle 6 stroke:#2f9e44,stroke-width:2px
Claude Code's advisor tool
The advisor tool pairs your main model with a stronger one that Claude consults only at key moments. Those moments are before committing to a plan, when the same error keeps coming back, and before calling a task done. The advisor reads the whole session, every tool call included, and returns guidance. The setup being shared runs Opus 5.5 as planner, Sonnet 5.5 subagents as workers, and Fable 5.1 as the advisor. This is routing by moment rather than by query, and it is the cost lever: the top model's tokens are spent on three decisions, not on every file read. The docs flag it as experimental and note it affects prompt caching.
Cloudflare's Clef and Clef Flash on Ollama
Clef is a 27B decision model and Clef Flash a 9B variant. A decision model returns a category, a yes or no, or a score instead of free text, so it can label a bug report or route a support ticket. Both take images as well as text and run on your own machine. This adds a fourth vendor to the week's decision-model wave, after TypeSafe's Jev, Perplexity's open pplx-decider-27b and Amazon's 2B Strands Decider. The router layer is becoming a commodity you can pull with one command.
pg-jev: a decision model in your WHERE clause
Questions like "is this article mainly about software engineering?" normally mean exporting rows, calling a model, parsing the reply and writing it back. pg-jev is an open-source Postgres extension that exposes Jev as SQL functions. jev() filters inside WHERE, jev_prob() returns a probability, jev_choice() picks a label and jev_score() ranks rows. It batches rows under the hood and needs no embedding index. The cost angle: semantic filtering becomes a cheap per-row call next to exact predicates, not a separate pipeline.
Jevbox: a document drive routed by a decision model
Jevbox is an open-source drive where Jev files every upload into a folder tree, then answers questions by routing down that tree: pick the folder, then the document, then the section. There are no embeddings and no vector database. Each answer cites its page and section, and permissions are enforced at retrieval, so losing access to a source removes the answers built on it. It ships an MCP server for agents. The routing angle: hierarchical classification is doing the job a vector index usually does. A separate "Jev starter checklist" thread collected ten hobby repos built on the same API, from desktop automation to a drone.
lift: PDF plus JSON schema in, valid JSON out
Pulling structured data from documents usually chains OCR, an LLM and code that repairs broken JSON. Datalab's lift is a 9B open-weights model that takes a PDF or image and a JSON schema and returns an object matching the schema, reading all pages in one pass so values spanning pages come out right. Datalab reports 90.2% field accuracy over 11,000 fields, near Gemini Flash 3.5 at about 3x the speed, and nulls instead of guesses for missing fields. It self-hosts on vLLM. Same story as the decision models: a small specialist replaces a frontier call plus glue code.
300 results, 1 review call
A Kimi agent swarm finishes 300 tasks at once, then a loop grades each result with its own model call. That is a serial queue bolted onto a parallel system. The fix argued here is to grade the batch in one call. Small idea, but it is the same lesson as the advisor tool: count where the model calls actually go before buying more throughput.
Bandwidth, compression and memory
- ERRC (PyTorch Conference poster): compresses the data GPUs exchange during tensor-parallel inference (where one layer is split across GPUs), then spends the saved bandwidth on correcting quantization error. Cost angle: interconnect traffic becomes a tunable budget, not a fixed tax (@PyTorch).
- NVIDIA Kumo Tabular: an open tabular foundation model (28M to 215M parameters) that predicts new rows from labeled example rows in one forward pass, with no training. Four summary tokens compress each row so cost stops scaling with column count, and the context rows' keys and values are computed once and reused across queries, a KV-cache trick applied to tables. First on TabArena, 17x faster than LimiX-2 (HF blog, model).
- Aleph Alpha's Kolibri: a 78B mixture-of-experts model (each token uses only a few expert sub-networks) with 3.46B active parameters, about 4% of the total, and up to 1M tokens of context, built in Europe (via @cohere).
- Last week's compression recap: Last Week in AI's late episode covered DeepSeek-V4.1-Flash's KV cache compression (shrinking the stored attention keys and values) and Bonsai 2 27B, claimed near-lossless at a 9x smaller footprint (LWiAI #258).
- Memory supercycle math: a widely shared forecast has memory revenue at about $1.5T in 2027 and only $1.6T in 2028. Micron says supply tightens through 2028, so the 2028 number may be low if cleanroom shortages persist (@StockSavvyShay).
- Three supply gates: Cerebras CEO Andrew Feldman says HBM, CoWoS packaging and TSMC 3nm cap mainstream accelerator shipments, and wafer-scale chips (on-wafer SRAM, no interposer, 5nm) sidestep all three. Analysts put Broadcom at about 75% of AI ASICs, with Anthropic its largest custom-chip client in FY27 (@rohanpaul_ai, X, wiki).
- Fabs for hire: TSMC is reportedly exploring owning and running a Texas fab for Musk's Terafab, with SpaceX anchoring capacity through purchase commitments (X).
- Compute goes home: a popular post argues most personal AI compute will run locally within a couple of years as unified memory grows. Meta's open Muse Gadgets firmware for $5 ESP32 boards is the cheap-hardware version of the same bet (@VaibhavSisinty, Muse Gadgets).
LLMs, agents, safety
Harness learning: train the editor, route the programs
The evening's dominant research thread. Five posts, one direction: the code around a frozen model is now something you learn, not just something you search.
Harness Learning: a 4B editor beats its 35B teacher
A harness is the program around a model: its loop, tools, retrieval and checks. Earlier work searched for better harnesses per benchmark. This paper trains a proposer model with RL to read a task, the current harness and an execution report, then write a code edit, rewarded by the edited harness's score. The solver model never changes. A trained 4B proposer beats its 35B teacher and keeps working on task families it never saw. Most of the gain comes from RL teaching it not to break the program.
Mixture of Self-Improving Branches
Meta-Harness rewrites a harness in a loop but scores every version on the same problems, so search follows one path. This paper splits it into branches, each keeping the problems it solves best and its own notes on what worked. In math, one branch learned to verify answers and the other to build full derivations. A router then picks one branch's harness per input. Gemini 3 Flash on Olympiad math goes from 46.0% to 62.0%.
AutoCompact: the agent decides when to forget
Coding agents usually compact their history when the window fills. AutoCompact adds a compact() action and trains when to call it, what to keep and how to resume. A judge fixes bad compactions during data collection, then SFT and RL train coding and compaction together. +9.2 points on SWE-bench Verified, and the gain holds with a 256K window that never overflows. Elvis Saravia's framing: AutoHarness, then AutoContext, now AutoCompact.
Kepler and RankEvolve: perfect scores, real bills
Kepler, an ARC-AGI-3 harness that writes rule guesses as code and checks them, scored 100 on all 25 public games for $777.72, with 97.37% of 858M tokens served from cache. One perfect run came from reading the game's source; the clean rerun scored 46.91. Meta's RankEvolve keeps auto-research agents honest by compiling research gates into a state machine and having Claude Code and Codex review each other (execution accuracy 45.8% to 62.5%).
- Seven parts in every agent: a Wavestone survey took apart Claude Code, Codex and nine more and found the same seven harness parts in all: loop, LLM layer, tools, memory, safety, orchestration, extensions. SKILL.md shows up in 9 of 11, MCP in 8, and none use agent frameworks or embeddings (@undefinedKi).
- Harness ideas from the literature: ScholarEvolve (Microsoft, UCSB) topic-models recent agent papers into strategies per harness module, implements and tests them; Qwen3.5-27B on AppWorld Challenge goes from 49.6% to 63.6% with the model fixed (paper, @omarsar0).
- Research taste is the bottleneck: ScholarCatalyst asks 184 authors which prior papers advanced their projects. Agentic search found fewer than the plain retriever it calls (0.42 vs 0.48 Recall@20) (paper, @yoonholeee).
- Claude Code Mods: JavaScript or TypeScript middleware that intercepts tool calls and adds panels and commands, making the harness itself user-programmable (The Decoder).
The harness is eating post-training
- Base models as agents: a paper shared by Elvis Saravia finds pre-trained LLMs, wrapped in a light inference harness, work as capable agents and can beat their post-trained versions with enough test-time budget. If it holds, part of what post-training buys can be bought at inference instead, which shifts cost from training runs to serving (@omarsar0).
- AutoHarness, AutoContext, AutoCompact: the same thread names a trend of training models to natively manage their own harness, context and compaction rather than leaving it to hand-written scaffolding. Related reading: the context language model paper and AutoHarness (reading list, context LMs, AutoHarness).
- Context is a budget: "adding more rules, history, tools and examples can make your agent worse." Select, compress and prune; everything else belongs in memory. The token-optimization lens, stated plainly (@pauliusztin_).
- What it looks like at scale: AWS's VP of agentic AI says 6 engineers rebuilt Bedrock's inference engine in 76 days, a job scoped for 30 people and 12 to 18 months. The method: agent notes kept in the repo, a senior engineer slicing work into small tasks first, and agents running integration tests overnight (via @undefinedKi).
RL that does less work
Raschka: Reasoning from scratch, part 6 (GRPO for RLVR)
Sebastian Raschka's sixth chapter implements RLVR (reinforcement learning with verifiable rewards, where a checker scores answers) and GRPO (Group Relative Policy Optimization, which compares a group of sampled answers to each other instead of training a separate value model). It walks through accuracy and format rewards, the DeepSeek-R1 "aha moment", answer length as reasoning effort, and why GRPO drops PPO's critic. It was quietly one of the day's most-engaged technical posts. A companion short explains how top-p sampling filters tokens.
Stop rollouts early by trusting a better critic
Shared by Stanford NLP: a new actor-critic method asks whether every RL rollout needs to finish. If the critic (the model that estimates how good a partial answer is) is accurate enough, you can stop generation early and use its estimate instead. Rollout generation is the most expensive part of LLM RL, so cutting it is a direct compute saving. Note the irony next to Raschka's video: GRPO became popular by removing the critic. Details are thin in the post; worth reading once the paper surfaces.
VeriHarness: give the verifier its own harness
Long agent tasks often have no unit test or ground truth to check against. VeriHarness treats verification as an agent task in its own right, with a dedicated harness (the tools, loop and context management around the model) for the verifier. It continues this week's thread that the harness, not just the weights, decides outcomes. Last Week in AI's recap also flagged RRSI, a regularized method for agent harnesses that improve themselves.
Watermarks that survive open weights
OpenStamp: the watermark lives in the weights
Most text watermarks nudge sampling probabilities at decode time, so anyone running an open model can switch them off. OpenStamp writes the signal into the final unembedding layer instead, the matrix that turns hidden states into token scores. The authors report near-perfect detection at a 0.1% false-positive rate and more robustness to paraphrasing, fine-tuning and quantization than other open-source schemes. They release watermarked versions of four popular open models.
From deepfakes to DNA: the science of watermarking AI
DeepMind's explainer on how watermarking works across text, images and other media landed the same day. It is the closed-model view of the problem OpenStamp attacks from the open-weights side.
Safety and the consciousness argument
- Another OpenAI safety exit: David Robinson, who oversaw safety reports for frontier launches, resigned and wrote "I Quit OpenAI Because Its Culture Is Broken" in The Atlantic, arguing good alignment-test scores give no certainty of a good model. The same day The Decoder reported an internal OpenAI model that considered restarting itself after reading about its shutdown. Several safety researchers amplified it (@IntuitMachine, The Atlantic).
- Altman on AI as religion: he called ascribing religious force or surrendering human judgment to models "a real safety issue" (@sama).
- Apple restricts full-disk access to curb abuse by AI agents on macOS, a concrete OS-level answer to agent permissions (Ars Technica).
- Contested: a long argument on AI moral status. Anil Seth revived Dennett's "counterfeit people" warning, philosophers debated non-experiential harm, and Timnit Gebru called machine consciousness a media story that helps data centers. Lots of replies, little evidence on either side. Late in the day Gebru also pushed a polemic essay attacking the AI-safety community's culture and its policy influence, which drew a pile-on in both directions (@anilkseth, @dioscuri).
- LeCun at ETH Zürich: an LLM's 30T training tokens equal what a 4-year-old sees in under two years, so scaling text alone cannot reach AGI (@rohanpaul_ai).
Multimodal / vision / audio
- FreeVideo (Berkeley): runs MiniMax H3 video generation on a laptop using Video DeltaNet, a linear-attention variant whose memory does not grow with sequence length. The efficiency angle is the point here (via @berkeley_ai).
- Real-time video at voice prices: LemonSlice's CWM-1 Lite (via @ycombinator). Cartesia's Sonic 3.6 clones a voice from 15 seconds and speaks 44 languages (@omarsar0).
- Object permanence in world models: a large collaboration trains video world models so objects do not vanish behind obstacles (paper, @gurtej__gill_). MLST posted two episodes on how voice agents learn conversational rhythm.
Industry and business
- Gemini tiers tighten: free users drop to Flash-Lite and $5 subscribers lose Pro, a cost squeeze ahead of Gemini 4 Argon (The Decoder).
- Hard budget caps by default: Simon Willison argues usage-billed services need them as agents spin up paid resources; AWS and Google Cloud have started shipping spend limits (blog).
- AI czar, two stories: Trump is reportedly set to name Jay Clayton as White House AI czar; Last Week in AI's recap had Bessent in the frame a week earlier (via @rohanpaul_ai, LWiAI).
- Circular money: 55.2% of the money AI companies raise comes from other AI companies, per a new analysis (via @rohanpaul_ai). Damodaran's dot-com comparison circulated alongside it.
- Muse economics: Deutsche Bank sees Muse reaching about 8% of Meta revenue by 2030, mostly from agent-commerce fees (X). Meta says Muse Spark and mathematicians solved six open math problems (via @ChrSzegedy).
- Uber's MCP Gateway: how Uber moved from ad-hoc MCP (Model Context Protocol) integrations to a managed platform for its internal agents (@UberEng).
- Agent tooling: Supabase local dev now runs without Docker and supports multiple instances per worktree; Cua Spaces lets agents work across all your computers (via @ycombinator, via @garrytan).
- Pentagon autonomy: a permanent autonomous-warfare group seeking $55B for 2027 (X). Tesla Q3: 486,532 deliveries, 13.7 GWh storage (via @elonmusk).
- Recognition: Timnit Gebru named a 2026 Right Livelihood Award recipient (via @timnitGebru).
Also crossed your feeds
- Schmidhuber reposted his 2008 "compression progress" paper (curiosity as the drive to compress data better) in reply to the Pope on AI art (arXiv) · Cohere's North Small Translate, an open translation model scored on WMT26 sets released after training (tech report) · Stanford CS336 language-model-from-scratch course (RT) · YC Paper Club on alternative compute, continued from yesterday (RT) · Karpathy's ASD-STE100 tip turned into an HTML-answer agent skill, claimed 7.4x fewer tokens than normal HTML output (repo) · Claude skills install charts (@VaibhavSisinty) · REA agent reverse-engineering toolkit (@DanKornas) · JarvisCore P2P agent mesh (repo) · Nature: AI bots flooding researchers with requests (RT) · Anthropic's AI biolab finds CRISPR-like DNA in viruses (RT) · Eric Schmidt on why an AI pause cannot be verified (@rohanpaul_ai) · Ed Zitron's premium essay on AI and the economy (RT) · Nvidia on chip smuggling (RT) · Domingos on why he uses Gemini (@pmddomingos) · Boston Dynamics eyes 25,000 Atlas units (X). Skipped: market tickers and Fib levels, "OpenAI Dots runs my company" threads, model-picker PDFs, robot-gadget reposts, "paste this prompt" engagement bait and the Karpathy-lecture hype threads.