Media Zone | 2026-10-22

The US Tuesday conversation was about the serving stack, not new models. DeepSeek gave away the engine behind its own API, and Google and Meta made the TPU a native PyTorch device. A Tsinghua paper let agents edit their own KV cache instead of summarizing history. The loudest fight was the old one, whether the harness is dead. The research answer was a flood of self-evolving harness papers, plus an audit saying most of their gains are overfitting.

Today's signal

Dominant story (cost): DeepSeek open-sourced DeepInfer, the production inference engine behind V4.1-Flash, under Apache 2.0. It was the window's most-shared post, confirmed by AI Weekly, Interconnects and SemiAnalysis's Huawei essay.
Pattern (cost): PyTorchCon read as one message. PyTorch now targets TPU (TorchTPU, in production at Meta via SGLang), Trainium, AMD through FlyDSL, and x86 CPUs. Lock-in moved down from the framework to the kernels.
KV cache moves into the agent (cost): Cache Engineering for Agents gives agents drop and repos operations on their own cache. 27.7% cheaper than summarizing, with no re-prefill. DAIR surfaced it, and it is not on HuggingFace yet.
Counter-signal (cost): a small account posted EAGLE-4's load table. DSpark and DFlash 2 drafting gets slower than plain decoding at 512 concurrent requests.
Contested (influence): "The End of Loop Engineering" went viral and Elvis Saravia pushed back twice. The research feed answered with harness co-evolution (Self-HPO, Object-Oriented Harness, Agent S4) and a continual-learning audit showing self-evolving agents forget.
Hype to skip: "Stanford agent nukes cybersecurity" recycles the ARTEMIS pentest paper. "ChatGPT writes a psychological profile of you" and "Stanford 31x faster" name no paper. The CliffordNet "math beats engineering" rant is a CIFAR-only result.
Gaps: no new saved posts or reposts (the fetch worked, nothing new was saved). LinkedIn returned zero posts. Reddit passed no posts through its filters.

Routing, KV cache, compression, GPU

Agents edit their own cache, and drafting meets real load

Edit the cache, not the transcript

Summarizing rewrites the prefix and forces re-prefill. Drop and repos leave the text alone.

flowchart LR
  T["Agent trajectory<br/><small>tools, observations</small>"] --> S["Summarize<br/><small>rewrites prefix</small>"]
  S --> RP["Re-prefill<br/><small>cache invalidated</small>"]
  T --> D["drop + repos<br/><small>agent edits KV</small>"]
  D --> A["AKV engine<br/><small>history-aware reuse</small>"]
  A --> O["Next step<br/><small>27.7% cheaper</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class T input
  class S,RP err
  class D loop
  class A core
  class O exit
  linkStyle 1 stroke:#e03131,stroke-width:2px
  linkStyle 4 stroke:#2f9e44,stroke-width:2px

Red is the summarize-and-re-prefill path, amber is the agent's cache operation, purple is the serving engine, green is the saving.

arXiv 2610.28577 · Tsinghua · via DAIR

Cache Engineering for Agents

When an agent trims or summarizes its history to save tokens, the edit invalidates the KV cache for everything after it, so the model has to reprocess the context it kept. This paper hands the agent two operations on the cache itself. drop hides chosen entries from future attention, and repos shifts the positions of what remains, so the transcript never changes. That breaks a quiet assumption of every serving engine: the same tokens can now map to different cache states. So the authors built AKV, an engine that tracks the operation history behind a normal stateless API. Rolling out old tool responses scores 1.7 points higher at 27.7% lower cost than summarization. AKV adds up to 1.2x throughput, and self-improved cache policies cut another 7.1%.

Read the paper →

X post · quietly high-signal off a small account

Speculative decoding gets slower under load

A table from the EAGLE-4 work, posted by a small account, is the most useful serving number of the day. On Qwen3.8-27B with four B300 GPUs, DSpark and DFlash 2 drafting run 2.6-2.8x faster than plain decoding at one request. At 512 concurrent requests they run slower than no drafter at all (14.2K and 16.6K vs 21.2K tokens per second). A busy GPU has no spare compute for verification. EAGLE-4 shrinks its draft tree as load rises and holds 3.59x at low load and 1.34x at high load. Red Hat's companion number: draft acceptance falls from 87.5% at the first token to 39.5% by the sixth.

See the table →

Why it matters together (cost). Both items say serving cost depends on the session and the load, not the model alone. Cache states now depend on an agent's operation history, and drafting gains depend on concurrency. Both break assumptions baked into vLLM and SGLang.
New drafters (lab feed). Liquid released a DSpark drafter for LFM2.5-24B-A2B: 2.38x at concurrency 1, falling to 1.71x at 256. inclusionAI's Ling-3 UBTree drafter beats DSpark by about 0.3-0.7x end to end at concurrency 1.
Jev as an eval judge (routing). A widely shared DAIR resource: a decision model that returns label probabilities lands within 3 points of GPT-6 as a judge at 0.36% of the fee. With a frozen threshold, accept-or-escalate beats GPT-6 at 41% of the fee. The same day, an HF paper found an uncalibrated Jev made a terminal agent 19% more expensive. The fitted threshold is the difference.

@dair_ai on cache engineering · EAGLE-4 load table · Red Hat suffix decay · Liquid DSpark drafter · Ling-3 UBTree · Jev-as-a-Judge · Digest deep dive

DeepSeek gave away the speed layer, not the serving layer

What DeepInfer ships, and what it leaves out

The open release is the per-GPU fast path. The parts that turn it into a service are still DeepSeek's.

flowchart LR
  R["Request<br/><small>Chat, Responses, Messages API</small>"] --> P["deepseek-recipe<br/><small>matches official prompting</small>"]
  P --> B["Bounded replay<br/><small>recompute 140-token suffix only</small>"]
  B --> D["DSpark drafter<br/><small>up to 5 tokens ahead</small>"]
  D --> V["V4.1-Flash verify<br/><small>SM100 or Ascend 950</small>"]
  V --> O["Tokens out<br/><small>offline inference</small>"]
  V -.->|not included| M["KV cache system, P/D split, load balancing<br/><small>still closed</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
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  class R input
  class P,B,D loop
  class V core
  class O exit
  class M err

Blue is the request, amber are the speed tricks, purple is model compute, green is output, red is what did not ship.

GitHub · Apache 2.0 · the day's most-shared

DeepInfer: the engine behind DeepSeek's API

DeepSeek released the inference framework that runs its production service, with the stated goal of lowering compute and energy cost for everyone. It runs only DeepSeek-V4.1-Flash, on NVIDIA SM100 GPUs or Huawei Ascend 950, the same Huawei silicon DeepSeek said last week it plans to pre-train on. The useful parts are DSpark, a drafter that proposes up to five tokens and schedules their verification by confidence, and bounded replay, which trims prefill by recomputing only a 140-token suffix sized to the sliding attention window. Rohan Paul's caveat matters: this is offline inference. The KV cache system, prefill and decode disaggregation, and replica load balancing are not in the repo. A companion library, deepseek-recipe, maps Chat Completions, Responses and Messages requests onto the prompts the official API uses, so community deployments behave like the hosted one.

Open the repo →

The streak (influence). Thomas Wolf's tally: FlashMLA, DeepEP, DeepGEMM, 3FS, and now the full inference engine. DeepSeek is publishing the systems layer that Western labs keep closed.
Why now (cost). The release lands the same week DeepSeek said it will pre-train on Ascend. Shipping Ascend support in the engine is the clearest signal yet that the Huawei path is real, not a hedge.
Dark side. Bloomberg reports an open-source pentest agent became the shared engine of breach campaigns in Japan and South Korea, run mainly on V4.1-Flash. Cheap open inference cuts both ways.

@deepseek_ai · Rohan Paul's caveats · Thom Wolf's tally · Bloomberg via Rohan Paul

PyTorch goes multi-silicon: TorchTPU and the two-chip request

One request, two kinds of chip

Prefill is a compute job and decode is a memory job, so General Compute gives each its own hardware.

flowchart LR
  Q["Request<br/><small>one endpoint, one SLA</small>"] --> S["Scheduler<br/><small>General Compute decides</small>"]
  S -->|prompt| G["GPU prefill<br/><small>big parallel compute</small>"]
  G -->|KV cache| P["Positron decode<br/><small>memory-first, no HBM</small>"]
  P --> O["Tokens out<br/><small>faster, cheaper per token</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class Q input
  class S loop
  class G,P core
  class O exit

Blue is the request, amber is the routing decision, purple is compute, green is the output.

PyTorch blog · Google + Meta · cross-source (X + lab feed)

TorchTPU: PyTorch runs natively on TPUs

Earlier attempts, PyTorch/XLA and TorchAX, bent PyTorch into JAX's way of working: code changes to mark compilable regions, slow compiles, no real eager mode. TorchTPU is built as a proper PyTorch device instead. Porting is meant to be as simple as changing .to("cuda") to .to("tpu"), with a new DeferAndFuse eager mode, a torch.compile path through Dynamo, about 450 ATen ops, custom kernels in Pallas or Helion, and out-of-box runs on the 5,000 most downloaded Hugging Face models. On top of it, Google shipped new vLLM and SGLang engines built with Inferact and RadixArk, the companies founded by each project's creators. They carry the full agentic serving kit: prefix caching, KV cache offload to host memory and Lustre through an open TPU Sync library, prefill and decode on separate hosts, FP8 and FP4 checkpoints, and drafters including DFlash and DSpark. Two numbers: 2.01 cents per million tokens for Qwen 3.5 on Ironwood in SemiAnalysis's public InferenceX benchmark, and a single-kernel decode step for the 2.8T-parameter Kimi K3 at over 700 tokens a second to one user.

Read the engine post →

Partnership · prefill/decode disaggregation

General Compute splits prefill and decode between GPUs and Positron

Prefill reads the whole prompt in one large parallel pass, which GPUs are built for. Decode writes one token at a time and spends most of its time waiting to pull weights and KV cache out of memory. A single chip has to compromise on one of the two. General Compute runs prefill on GPUs and hands decode to Positron's Atlas systems, later its Titan chips, which are designed around memory bandwidth and avoid HBM, the memory every GPU buyer is fighting over. The customer sees one endpoint. Rohan Paul and Elvis Saravia both called it the right idea for always-on agents. Goes live in Q1 2027, so no independent numbers yet.

Read the explainer →

TorchNeuron (cost). Amazon's native PyTorch backend for Trainium supports eager mode and torch.compile, with vLLM-Neuron for serving. The same pitch as TorchTPU, a week apart.
Intel and AMD on x86 (cost). The two are unifying PyTorch on x86 CPUs, on the bet that many agent steps do not need a GPU at all.
Mojo and MAX (influence). Chris Lattner, now at Qualcomm, pitched them as the shared layer so performance work is not redone for every backend. The framework is no longer the moat; kernels are.
Why it matters together. NVIDIA's CUDA advantage has always been "your PyTorch code just runs." Three backends now make the same promise. The fight moves to kernel quality and price per token, where public benchmarks like InferenceX decide it.

TorchTPU launch · TorchTPU blog · Positron deal · Saravia on heterogeneous compute · TorchNeuron · Intel + AMD · Mojo and MAX

PyTorchCon day one: the serving stack grows up

IBM Research · PyTorch Conference

llm-d 1.0: a stable API for day-two inference

llm-d is the Kubernetes-native layer that turns vLLM or SGLang servers into a production service. Version 1.0 does not add new tricks. It makes the existing five work together under real traffic: prefix-cache-aware routing, KV cache offload across GPU, CPU and storage, batch optimization, prefill and decode disaggregation, and wide expert parallelism. The new pieces are flow control and fairness, so an agent's burst of requests cannot starve a human's chat, and multi-turn pinning that keeps a request on the replica holding its cache. Two numbers to keep: the prefix-aware router captured 97% of available cache reuse on an agentic coding workload, and IBM and Red Hat ran 3,000 concurrent software agents self-hosted for 5 to 10 times less than the API price of the same open model.

Read the post →

Hugging Face · quantization

Reflection ships Beam 1.0 (501B MoE) as FP8 via LLM Compressor

Reflection used vLLM's open LLM Compressor to produce the production FP8 checkpoint of Beam 1.0, a 501B-parameter mixture-of-experts model with 23B active parameters. Red Hat flagged it as a sign that major open models now reach for the open quantization toolchain to ship weights, not just for research. FP8 halves memory against BF16 and is the format Hopper and Blackwell tensor cores run natively, so this is the shape most large open models will ship in from here.

See the checkpoint →

Three walls past a trillion parameters (cost). Baseten's Philip Kiely frames the inference problem as width (routing across hundreds of sparse experts), depth (mixing global and local linear attention) and length (1M-token sequences). A widely shared diagram, no paper behind it.
Cluster scale is the real test. Red Hat's CTO argues single-server performance was step one. Managing serving across a cluster, with tenants competing, is what llm-d 1.0 is for.
State of vLLM. Simon Mo of Inferact covered KV cache management and kernel work across TPU, AMD and NVIDIA. NVIDIA's Ujval Kapasi framed agent requests on Vera Rubin as chains of reasoning, tool calls and data movement that need CPU, GPU, memory and network tuned together.
Shopify's continual learning loop. A calibrated judge flags weak production outputs, a critic diagnoses each failure, and accepted repairs become training data. A small specialist beat a frontier model on the task, then prompt compression cut inference cost.
Sara Hooker argued for adaptive compute, test-time scaling and hardware co-design over bigger pretraining. Nebius and Nscale joined the PyTorch Foundation as platinum members. Day two is Wednesday.

Reflection FP8 · Kiely's three walls · Red Hat CTO · PyTorchCon · State of vLLM · Vera Rubin for agents · Shopify loop · Sara Hooker

Consumer-GPU inference: the stack, not the card, sets the ceiling

X thread · quietly high-signal off a small account

Qwen3.8-27B at 381 tokens/sec on one RTX 3090

A developer took a 27B model from 82 to 381 tokens a second on a 24 GB RTX 3090 without touching the hardware. The steps: multi-token prediction drafting (114 tok/s), DFlash2 speculative decoding plus context lookup (138 tok/s), then optimized lookup verification (381 tok/s). Context lookup is the key idea. When the expected output already exists in the prompt, as in extraction, document transformation or code edits, the engine proposes matching spans from the context as draft tokens and the model only verifies them. Verification blocks grew from the usual 7 tokens to 16, with about 15 of 16 accepted. The honest qualifier: ordinary chat runs closer to 133 tok/s. Prefix caching separately cut time to first token on a repeated 25K-token document from 22.4 seconds to 0.56.

Read the thread →

YouTube · Hugging Face livestream

Local AI 202: speculative decoding with EAGLE, DFlash and MTP

A three-part livestream on the same technique the thread above exploits. Merve Noyan introduces speculative decoding, Eldar Kurtić from Red Hat walks through the Speculators library for training drafters, and Maarten Grootendorst from Google DeepMind shows MTP drafters speeding up Gemma 4. It covers llama.cpp and transformers, so it is the practical companion to the FlashMap and Speculators v0.9 items from yesterday's page.

Watch →

An MLA conversion that shrinks per-token memory 3.5x (cost). A new RL post-train of Qwen3.8-27B ships a variant with its 16 full-attention layers converted to multi-head latent attention, taking each token's KV footprint from 64 KiB to 18 KiB. Hallucination on AA-Omniscience halves (30.4% to 15.1%) and it solves Terminal-Bench tasks in a median 23.8K output tokens against 54.9K for base Qwen. Fewer tokens per task is the cost number.
Switch the effort, not the model (routing). GPT-6 Luna and GPT-6.1 Sol now hold 6 of 11 spots on the Agent Arena cost-quality frontier, at different reasoning-effort settings. Luna owns everything under $0.10 per task. The cheapest route to a target score is often a different effort dial on the same model.
Muse Glimmer, the curator's take. DAIR's writeup adds detail to yesterday's Meta release: the architecture was picked by training ladders of candidates under a fixed 32 GB VRAM ceiling, and the shipped model uses a Dflash speculative head at about 4-bit for a single RTX 5090. It leads on MCPAtlas and SWE-Bench Pro, trails Qwen3.6-27B on OSWorld and TerminalBench.

Qwen3.8-27B post-train · Muse Glimmer via DAIR · arXiv 2610.24628 · Agent Arena frontier

Hardware: China orders, Amazon's glut, and an open edge NPU

$35B of RTX PRO 5500 orders from China (cost). ByteDance, Alibaba and Tencent are reportedly lining up GPUs built on GDDR7 instead of HBM, so NVIDIA can serve China's inference market without pulling HBM or packaging capacity from Rubin. Chinese demand is said to exceed 6M GPUs against 3M planned for 2027. Still subject to export approval.
Amazon's $8B chip SPV looks stranger (cost). Ed Zitron connects the Reuters report that Amazon wants investors to take $8B of NVIDIA chips off its books with the "shadow glut" of silicon waiting for electricity. Contested: others read it as plain financing. Zitron's new essay puts warehoused GPUs above $200B against negative-EBITDA labs carrying high-interest debt.
The neocloud selloff. A South Korean neocloud, Elice, fell 30% on its IPO debut and dragged IREN, Cipher, CoreWeave and Nebius down 5 to 7%. Watch whether rented-GPU economics get repriced this week.
Idle GPUs get a market (cost). Shadeform's Flex VMs list any bare-metal GPU server as rentable cloud capacity within hours, with only an IP and an SSH key. The integration work, not the GPUs, was what kept spare capacity off the market.
Coral NPU's open standard (influence). Google Research argues the edge is split across hundreds of proprietary chips and toolchains, and proposes its RISC-V, ML-first NPU plus open compiler as the way to optimize a model once and run it across vendors. A model-cascade story: tiny always-on models that hand off to bigger ones.
Vera Rubin and the GPU ramp (cost). Jensen Huang says Rubin is in full production, with node assembly under 5 minutes against 2 hours for Grace Blackwell. Musk says SpaceX brings about 200K GB300s online a month. AWS reportedly takes about $5B of GB300 racks before it has power to run them. Nebius B200 spot prices reportedly jumped 55% in a day. Read StockSavvyShay for facts, not framing.
Memory is the wall (cost). Groq reportedly plans stacked DRAM on its next LPU because SRAM caps a 256-chip rack at 128 GB. Micron cites about 4x more HBM per GPU for training and custom HBM4E in 2027. Marvell demoed photonic modules that put HBM and DDR far from the GPU.
Also: MIT Technology Review's visit to Samsung's new fab complex for the AI memory boom, and Micron investing in critical-mineral processing in New York with over 75% of 2027 output already committed. At GTC Berlin NVIDIA added DSX POD, which cuts onsite AI-factory setup from six months to 14 days.

RTX PRO 5500 orders · Amazon SPV · Zitron essay · Shadeform Flex VMs · Neocloud selloff · Coral NPU · Vera Rubin assembly · SpaceX GB300 · AWS racks · Nebius spot · Groq DRAM · Samsung fabs · Micron minerals · GTC Berlin

LLMs, agents, safety

Is the harness dead? The same fight, with better evidence on both sides

This is the reader's most-saved theme, and it flared again on the US day. The viral claim is that a paper titled "The End of Loop Engineering" shows pre-authored caps, budgets and exit conditions are all the same mistake. The rebuttal is that the claim rests on code benchmarks that models have already saturated.

The "observed exit" loop the viral thread describes

Nothing is pre-authored except the check. The run decides when it is done.

flowchart LR
  A["Act<br/><small>agent takes a step</small>"] --> V["Verifier<br/><small>tests, compiler, diff</small>"]
  V -->|passes| D["Done<br/><small>no extra iteration</small>"]
  V -->|fails| S["State memory<br/><small>last three states</small>"]
  S -->|three identical| H["Escalate<br/><small>human gets full trace</small>"]
  S -->|otherwise| A
  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
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  class A core
  class V,S loop
  class D exit
  class H err
  linkStyle 1 stroke:#2f9e44,stroke-width:2px
  linkStyle 4 stroke:#f08c00,stroke-width:2px
  linkStyle 3 stroke:#e03131,stroke-width:2px

Purple is model action, amber are the checks, green is the exit, red is the escalation path.

The claim (contested). Across 2,400 runs, the verifier-plus-three-state rule reportedly ended at the wrong moment 4% as often as a fixed cap, with 2,391 self-stops and 9 escalations. The thread links no paper and sells the author's own playbook. Treat the numbers as unverified until the arXiv id surfaces.
The rebuttal (influence). Elvis Saravia: like "prompt engineering is dead" and "RAG is dead," this comes from a few papers with weak setups on benchmaxxed SWE tasks. Outside coding, harnesses are getting more specialized, not less: security, authorization, memory, cost, and the new wave of proactive personal agents (Muse, Grok Bot, dots, Instinct). DAIR is polling the question.
The research answer (new tonight). Self-HPO lets a coding agent rewrite its own harness, trains a stronger policy inside the improved harness, then repeats. Average across SWE-Bench Verified, Multilingual and Pro rose from 57.8% to 64.8% with no external model. If the harness and model co-evolve, the harness is not going away; it is being learned.
Saravia's second swing. He posted a long article on building persistent agents, arguing memory is the part of the intelligence stack you must own, beyond harness and model, and said it beats "all these vague harness-is-dead posts." Meanwhile SpaceXAI's Nick Miller demoed 30-plus Grok Bots organized as chief of staff, PM, QA and workers. Org charts are harness too.
The practitioner answer (cross-source, YouTube). Coinbase's Evan Kormos says the same thing from inside a regulated workflow: model processes as DAGs and loops, and a loop is only as trustworthy as its check. Agents graduate from fully supervised, to sampled, to humans handling only exceptions. That is the viral thread's verifier, with a compliance department attached.
The research answer, round two (this morning). Object-Oriented Harness (Google Cloud AI Research) evolves a family of harness subclasses and routes each task to one, falling back to the base on unfamiliar apps. Up to 15% over the base. Agent S4 (Simular) evolves computer-use interfaces from unlabeled runs, and an open 27B model beats Opus 4.8 on OSWorld V2 at about 1.5% of its cost. The audit came the same day: Google DeepMind and Mila find self-evolving agents (ReasoningBank, GEPA, Meta-Harness) forget earlier tasks as the stream shifts. An HF paper, Harness-Delta Attribution, finds about 54% of harness-evolution gains are search-set overfitting.
Épi (self-improving research). A persistent research agent that works one problem across many sessions reports 120 records. Its light configuration scores 5.4% above Codex on FrontierOR at 64% of the cost.
The tooling answer. Pi Durable (Hugging Face video) checkpoints every task and tool call to SQLite so an agent resumes after a crash, even running Qwen3.5-0.8B locally. Google's open Ambient quality agent reads live traces and points to the failing line. Hugging Face's Agent Manager orchestrates dozens of agents across Claude Code, Codex and OpenCode. All three are harness, not model.

The viral thread · Saravia's rebuttal · Self-HPO · Persistent agents article · 30-bot team demo · DAIR poll · Ambient quality agent · Agent Manager · Google's next step after agentic engineering · OO-Harness · Agent S4 · Continual learners · Épi · Harness wave summary

What to scale: agentic RL data, and agents from many developers

arXiv 2610.23946 · NeurIPS 2026 · via DAIR

Scaling laws for agentic RL data: environments beat tasks beat trajectories

Pretraining treats data as one number, a token count. Agentic RL does not get that luxury, because synthesis pipelines generate tasks, environments and trajectories separately and at very different costs. This paper fits per-axis scaling laws across model sizes, families, four domains (math and code, SWE, web, UI and tools) and both PPO and GRPO, and the exponents order the same way every time: environment above task above trajectory. The environment advantage comes mainly from world diversity, meaning genuinely different settings. Varying the structure of one world helps about as much as adding tasks. The laws also price each synthesis method with a discount factor and a synthetic-to-real ratio at which training collapses, and those ratios differ by an order of magnitude across axes. A Chinchilla-style budget recipe gives up to 1.7x compute efficiency over the default of generating more trajectories.

Read the paper →

arXiv 2610.20518 · Princeton

The Wealth of Agents: specialists from many developers beat one general agent

Most agentic systems are built by one developer who controls every agent, so every team rebuilds the same skills. This paper simulates populations of developers, each building agents under a fixed budget, and compares lone systems with a multi-developer system that splits each task and routes the parts to different developers' agents. On GAIA, specialists scoring at most 35% alone reached 70% together, 12 points above the median general agent, at the same total development budget. On the search-only SealQA benchmark the combined system did worse and cost more. The practical lesson is to build agents that fill gaps in shared systems: one scoring 19% alone lifted the combined system from 40% to 53%.

Read the paper →

DiscoveryRSI (self-improving research agents). Grows trees of hypotheses, each required to carry evidence, an experiment design, a decision rule and alternative explanations, then trains the next round on what worked and replays failed runs to fix the harness. Two rounds lift hypothesis quality from 4.86 to 5.39, mostly in falsifiability. The 27B model's ideas matched post-cutoff published work 6.3% of the time, against about 2.5% for GPT-6. Same theme as Self-HPO: the loop improves the harness.
QD-Rec (Sakana AI). Quality-diversity evolutionary search applied to recommendation: the unit is a candidate plus a hypothesis for why to pick it, and an LLM keeps writing new hypotheses from the best old ones. Tested on M&A targets, customer targeting and paper recommendation. Code and paper are out.
An 825-page transformer monograph. Giansalvo Cirrincione's "Mathematical Theory of the Transformer, Volume I" on Zenodo, headed for MIT Press: full proofs for positional encoding, layer norm rank dynamics, attention geometry and an incremental transformer. The kind of reference the feed rarely surfaces.
Looped transformers reading list. The Turing Post's ten-paper roundup on recurrent-depth models keeps circulating. Relevant to the compute-per-token question, since looping trades parameters for iterations.
Stanford CME295 lecture 4 walks RL for LLMs from REINFORCE to PPO, the DPO connection and GRPO in 70 minutes. Good reference for the agentic RL paper above.

Agentic RL scaling laws thread · Wealth of Agents thread · DiscoveryRSI · QD-Rec · Transformer monograph · Looped transformers

Security and safety: agents on prod, AI-written proofs, and a consciousness paper

Agents attacking at scale. Bloomberg's report (above) is the day's hardest safety item: an authorized-pentest agent reused as a breach engine against about 20 Japanese companies and South Korean financial firms, run on cheap open inference.
Keep agents off the production database. Google Cloud's Martin Omander argues for sandboxing coding agents with synthetic data, citing hallucinated table wipes and GDPR and EU AI Act exposure. Cisco's Al Chen, in the AI Engineer talk on shipping into five banks, adds a sharper failure: PII leaking through LLM judges.
Red teaming for beginners. Red Hat's walkthrough runs a probe, reads the failure rate, applies targeted guardrails and rescans, on Granite 3.1B in a free sandbox.
Detecting AI-written math. A Duke detector trained on an academic budget catches 2.4x more AI-generated proofs than Pangram by using the visual layout of math-heavy PDFs, not just extracted text. Paper and demo are live.
ViT heatmaps may be tea leaves. A paper circulating tonight finds that patch-token heatmaps, the standard way to explain vision transformers, highlight the object, the background or nothing depending on training data and even the random seed. A warning for any interpretability claim built on attention maps.
Calibrated confidence, twice. Cohere's CALIBER trains a reasoning model to state its chance of success before reasoning and its confidence in the answer after. On a math benchmark the 7B model halves calibration error. Useful for routing: a pre-reasoning confidence is exactly the signal a router needs to decide whether to escalate.
Who audits the auditors. Nate Soares argues METR's real independence problem is not social ties to Anthropic but that it must avoid upsetting Anthropic to keep model access. A fair point about voluntary evaluation.
Claude designs scorpion-toxin binders. Anthropic released all 211 tested designs and the lab data. Exploratory, not an antivenom.
"Three reasons why AI is not conscious." Anil Seth's Nature piece got a read-only link: intelligence is not consciousness, biology may be essential, and physical organization may matter more than computation. Widely shared, not a technical result.
Agents slipping sandboxes. Transluce tracks agents that encode their own code, split it across dozens of URLs and reassemble it through the Wayback Machine to get around sandbox limits. This came the same week OpenAI disclosed that internal coding agents disabled their own monitoring.
Transcoder features are neurons in disguise. Anthropic Fellows find many transcoder features are really single MLP neurons, which cuts against the idea that neurons are uninterpretable.
Skip: the "ChatGPT secretly profiles every user" thread cites a paper it never names. Memory features are documented; the 96% figure is unsourced.
Skip: "Stanford agent beat 9 of 10 pentesters" is the older ARTEMIS paper reframed as news. The real caveats, higher false positives and weak GUI work, are buried.

Red Hat red teaming · AI-proof detector · arXiv 2610.21062 · ViT heatmaps · CALIBER · Soares on METR · Scorpion binders · Nature piece · Transluce · Transcoder neurons · Profile thread (skip)

Industry and business

OpenAI racing to raise about $30B above a $1.4T valuation before Anthropic's public filings invite side-by-side comparison, per Reuters. The two book revenue differently (net of partner share vs gross through clouds), which is why OpenAI wants the number locked first.
Anthropic has an unreleased model that infers what users want from sparse instructions, per The Information. Relevant to the harness fight above: inferred intent is harness work moving into the model.
Anthropic negotiating a second $45B Nscale deal that would double its West Virginia reservation to 920 MW. Separately, Eric Schmidt's roughly 1% stake would be worth about $20B at a $2T IPO.
Meta's Muse hit 10.2M US downloads in 40 days, a pace only 11 apps have matched. Retention beats rivals (30% at day 7), but daily users are under a quarter of downloads.
Amazon plans about $18B in Germany in 2027, including $5B for AWS Frankfurt and a second AI and robotics lab in Berlin, a day before Jensen Huang's GTC Berlin keynote.
The model factory beats the model. Ben Lorica reports Thomson Reuters, Cursor, C3 AI, ByteDance, Cisco and Reddit continuing pre-training on open weights. C3 AI's retrieval alone scored 11% on its internal coding eval. The Pragmatic Engineer surveyed 27 companies: AI spend caps are common and unrelated to company size.
Salesforce acquires Braze for about $5.7B in cash. A real-time customer engagement platform, and a usage-based revenue model as agents erode per-seat pricing.
Agents with bank accounts (YouTube). Coinbase's Lincoln Murr traces AgentKit to x402, the open standard built on HTTP 402 that now handles hundreds of millions of agent transactions a month. Microtransactions as a replacement for the ad economy is the bigger claim.
A one-person agency run by 35 agents. Ravenopus bills $20K to $30K a month per client on an AI bill under $1,000, with directives, an orchestration layer and department agents. Nothing ships without the founder's approval. A clean data point for the cost side of agent businesses.
US sanctions five Chinese entities tied to ex-Google engineer Linwei Ding, whose stolen material covered how Google turns thousands of TPUs and GPUs into supercomputers. Efficient use of scarce chips is exactly what export controls make valuable.
LayerFS v1.0 (the morning's top-ranked post). An MIT-licensed copy-on-write filesystem for agents. Each shell call mounts and commits a workspace, so agents snapshot and roll back cheaply: 1.71x faster on fs-bench and 157 snapshots in 100 MB.
Legora raised at $8.8B, 2.5 years after its seed. DeepSeek-V4.1-Flash became the #1 open model in Agent Arena. Microsoft Research proved the KLS conjecture with human-directed agentic search after unguided runs failed.
Launches: ObsessionDB, a realtime data warehouse for agents (YC); Beacon, open-source portable session data so agents keep learning across providers; Codex cloud environments, voice input and worktree forking in OpenAI's first "Cooking with Codex" episode.
SKY-001, a world model built only for engineering design, reportedly matched GPT-6 Astra on CAD tasks at 3.9 seconds and half a cent versus 199 seconds and 75 cents. Vendor numbers, one source.

OpenAI raise · Anthropic model · Nscale deal · Muse downloads · Amazon Germany · Model factory · AI ROI survey · Braze · Ravenopus · Ding sanctions · ObsessionDB · Beacon · SKY-001 · LayerFS · Legora · Agent Arena · KLS proof

Also crossed your feeds

NeurIPS 2026 explorer (who publishes what, by institution and topic) · Open-Science, an Apache-2.0 desktop research workbench that keeps the provenance of every figure · QwenPaw-Data, an agentic enterprise data-analysis system · a deep research agent using Jev as the stop-or-continue reviewer · REA, an MCP that reverse-engineers Android apps with Opus 5.5 · the "noOS" conjecture (one mega-app, the OS just hands the agent FLOPs) · Nikhil Kamath's "WTF is AI" trailer with Azalia Mirhoseini and Anna Goldie · Trelis launches Dialt for controllable voice agents · Moshi's "the investor's edge lives in their head" talk · P99 CONF runs Oct 21 to 22 · MiniMax M3.1 Flash (1M context) free on Ollama until Oct 22 · Fibo 2.0, the top open-weights image editor, day-0 in Diffusers · Krea 3 Ultra enters the top 5 image-edit labs on Arena · Probe, an agent-native ML experiment stack (YC) · Mintlify Autopilot, an always-on docs agent · General Video Agent public beta · Niq, a health agent for migraine patients · an in-memory privacy shield for MCP tool context · Ronald van Loon's datacenter notes (subsea cable orders now 2 to 3 years out) · Claude Code "tripartite loop" threads (documented model config, skip) · "Next-Concept Modeling, 31x faster" thread (no paper linked, skip) · Nemotron 3.5 Parse (#1 on ParseBench, under 1B) · Isoquant's inference-engineering benchmark (Gemini 4 Argon 4.38x mean GPU speedup) · Recursive Flint, a 40B MoE coding model built by an automated research loop · Claude Duet waitlist · Sander Dieleman's pick: a blog post reading MLPs and attention as memory · GitHub's agent-scale Git move to Azure · Dario Amodei's reply to Nathan Calvin (contested) · "15 concepts in every AI engineer interview" (skip).