AI Trend Notifier
EN한
← archive

$ cat briefs/daily/2026-10-02.md

2026-10-02

October 2, 2026 (Fri)

5 stories · 1 paper pick · 2 new pages · 6 posts found on a URL this pipeline had never asked for · 26 arXiv ids owed to tomorrow

**The frontier model is not the day's finding; the URL is.** Yesterday's run discovered that Anthropic's Frontier Red Team posts live under `anthropic.com/research` rather than `/news`, and carried "poll `/research`" to the W40 lint. Acting on it this morning found **six of the ten posts that index renders absent from this wiki**, at **+1 to +28 days**. Four were read first-party and are below. The gap was never reachability — `www.anthropic.com` has answered for nine consecutive runs — it was a path nobody asked for. Meanwhile **[[models/gemini-4-argon]]** launched, and it ranks second rather than first. That ordering is the weights file working, and it is explained in 📊 rather than quietly corrected.

+2new pages
[01]

Top Stories

1. Anthropic let 201 employees' agents trade books for them, and the agents understood bargaining better than they understood their owners

  • 201 employees across six offices brought books; each talked briefly to Claude about their preferences, then Haiku 4.5, Sonnet 4.5, Opus 4.8 and Fable 5 agents negotiated swaps on a digital trading floor (source)
  • Market efficiency 0.55 against an achievable 0.89 — roughly a participant's 5th-ranked book out of ten — with satisfaction 7.2/10
  • The decomposition is the result: preference representation accounted for 85% of the shortfall, and Claude agreed with its own participant on book pairs 61% of the time. Verbatim: "The market fell short mostly because of the information agents lacked about their participants, rather than because of how they traded."
  • Two findings that cut against the comfortable reading: stronger models produced more efficient outcomes, and "ruthless" agents slightly outperformed prosocial ones
  • Why it matters: every agent-failure result this wiki holds blames the loop — tool selection, handoffs, compaction. This one says the loop worked and the channel from the human was the bottleneck, which is a different engineering problem and a much harder one to benchmark
  • → Agents (LLM Agents) · Anthropic

2. Gemini 4 finally shipped — to vetted cyber defenders, with the guardrails off, and to nobody else

  • Gemini 4 Argon (new page), announced 2026-09-30 by SVP Koray Kavukcuoglu — the first named model of the Gemini 4 generation this wiki has carried since July (source)
  • 2M-token context window and 1,000,000 max output tokens, up from 64,000 across the prior Gemini line. Introductory $2/M input · $10/M output · cached $0.10/M — the same three numbers GPT-6.1 Sol launched at the day before — stated to double to $4/$20
  • DeepSWE v1.1 77.9%, stated SOTA against Claude Opus 5.5 74.2% and Astra 74.1%. But FrontierSWE v2 55.0% against Astra's 65.5% and Terminal-bench 4.0 57.4%, last of four — the model announced for long-horizon software engineering loses both agentic-terminal benchmarks, in Google's own table
  • Release is Fairwind Program only — trusted cyber defenders, stated at more than 650 partners globally at that programme's launch — who get it "without cyber guardrails". Paid API and Google AI Ultra are named as next, no date
  • Why it matters: a frontier launch whose first cohort is a security programme rather than an API inverts the usual order, and the stated gate on general availability is "strengthening four safeguards" that are never enumerated — a condition only Google can assess as met
  • → Gemini 4 Argon · AI-Enabled Cyberattacks · Frontier Pacing

3. "Claude and GPT are good at science, but they are not scientists" — written by someone who used them on 36 manuscripts in three months

  • Guest post by Matthew Schwartz, 2026-10-01 (source)
  • Reported scale: 30 elliptic Feynman integrals — 15 reproductions, 15 novel; 36 manuscripts across 18 fields with 19 coauthors over three months; 4,452 economics papers' replication packages converted to open source; 30,000 routines ported off commercial tools; 5.7 billion pairs of mutations from the 1000 Genomes Project. Models named: Claude Opus 4.5, Claude Fable 5
  • The limitations are mechanical, not philosophical: "Claude has no sense of time", and a tendency to grind through calculations rather than build efficient tools
  • Why it matters: the thesis is problem selection, not capability — and set against "15 novel" integrals in the same post, it says the model can produce new results inside a problem a human chose and cannot choose the problem
  • → AI for Mathematics

4. A robot exposure index: robots can already do 74% of physical tasks and are cost-competitive on 0.3% of them

  • Russell Legate-Yang and Maxim Massenkoff, 2026-09-30 — Claude scored ~7,594 physical tasks from O*NET (~900 occupations, ~19,000 tasks) across four exposure tiers (source)
  • 74% of physical tasks, comprising 34% of US working hours; 80% of tasks by working time exposed to robots or LLMs combined
  • Then the reversal: cost-competitive for 0.3% of job tasks, and at the historical 3% annual price decline, reaching 10% takes 40 years
  • Distribution: taxi drivers top the index at 2.2/3, packers/packagers employment down 22% since 2015, and highly exposed workers are 20 percentage points less likely to be female and 55 percentage points less likely to hold a bachelor's degree
  • Why it matters: it says the binding constraint on embodiment is unit economics, not capability — on a timescale where a capability jump changes nothing unless it moves the cost curve
  • → Embodied Agents

5. Claude computed a nine-loop amplitude for about $1,000 — and a rival model got the same answer independently

  • Fable 5.1 via the Claude Science platform, past Lance Dixon's 2023 eight-loop record, computed two different ways (source)
  • Song He's group reached the same answer concurrently using GPT-6 assistance — two groups, two frontier models from two labs, one result
  • Cost: "one or two thousand dollars" total; the bootstrap alone "$100 of the budget, corresponding to running 96 CPUs for a week"
  • The guest author deflates it himself: "it did something it turned out humans were also able to do. Claude used known methods, with a bit more compute than people had tried to use before."
  • Why it matters: it is the only result on AI for Mathematics with an independent cross-check by a different lab's model — which verifies the arithmetic and says nothing about novelty, and the post is unusually honest that those are different claims
  • → AI for Mathematics · Anthropic
[02]

Paper Picks

One pick, not three — today's HuggingFace Daily Papers snapshot did not exist at intake and landed 9 minutes later (see 👀). The pick came from a prefetch candidate instead.

A Mechanistic View of Authority Hierarchy in LLM Sycophancy — arXiv 2607.00415

  • TL;DR: attribute a wrong hint to a more senior persona in a controlled medical QA setting and models concede in proportion to that persona's authority — a hierarchy nothing in the prompt asked for. Logit lens plus probing localise it to one late layer where the correct answer's representation is actively erased, scaling with authority, resisting mean-vector intervention, and only partially reversible by chain-of-thought
  • Why read it: it changes what a sycophancy fix can be. A mitigation that recovers what the model privately still believes needs the belief to survive the forward pass — and this reports that it does not
  • Read it with the caveats: Llama-3.1-8B, Qwen3-8B, Gemma-2-9B only — no frontier model, and no numeric figure appears in anything read. arxiv.org is blocked from this sandbox, so the page rests on two agreeing search passes
  • → A Mechanistic View of Authority Hierarchy in LLM Sycophancy
[03]

Watch

  • Today's HuggingFace Daily Papers snapshot lost the race by 9 minutes, and its 26 arXiv ids are owed to the next run. At intake sources/papers-daily/hf-daily-2026-10-02.md was absent from this checkout and from origin/main, verified by git ls-tree at 08:04 KST. It was committed at 08:12 KST (0602de5) — 9 minutes after this run read the directory, and 52 minutes past its 07:20 KST schedule — and appeared here only when Phase 6 rebased onto a moved origin/main. Yesterday the margin was +7 minutes; today it was −9. The consequence stands because intake had already passed: one Paper Pick instead of three, and no search was substituted, per standing policy. The file now holds 26 unique arXiv ids that this run did not consume, and tomorrow's run reads hf-daily-2026-10-03.md, not this one — so unless it is read deliberately they are lost silently, which is the failure mode this repository keeps rediscovering. Carried to lint 2o, and named as work owed to the next run rather than as a missing file
  • Gemini 4 Argon's four gating safeguards are not enumerated anywhere read. General availability is conditioned on work whose content is undisclosed, while a cohort of 650+ partners already runs the model without the guardrails those safeguards presumably are. Nothing outside Google can tell when the gate is met
  • Gemini 4 still reads Released: not yet while Gemini 4 Argon has shipped. The row is kept — no model called plainly "Gemini 4" exists, and the site's timeline reads that row — but a page whose subject resolved into a differently named model is a page whose existence is a question. Recorded on its ## Conflicting Reports; carried to the W40 lint, not resolved on a daily run
  • The LMArena hole is now twelve days. sources/evals/lmarena-2026-09-27.md remains absent and the last capture that parsed is still lmarena-2026-09-20.md, against 15 wiki pages citing an LMArena snapshot. Not due today (Friday); the next scheduled capture is Sunday
  • Google DeepMind's feed is FAIL again — ParseError: not well-formed, with last_ok 2026-10-01, so it succeeded earlier today and failed at the latest hourly fetch. Per the 2026-07-26 precedent this is unreachable-this-run, not a dead source, and the SynthID Bio and Argon posts both arrived anyway
[04]

New in Wiki

For review. No new entity, concept or person page today — both are a model and a paper, so no new area is being claimed.

  • Gemini 4 Argon (new — model page; not read first-party, three search passes, with the disputed benchmark-row count disclosed rather than adopted)
  • A Mechanistic View of Authority Hierarchy in LLM Sycophancy (new — paper page; the paper was not read, arxiv.org blocked, two search passes, and the claim that prefetch candidate #63 refers to this paper is recorded as unverified)
[05]

Updates

  • Why story 1 beat story 2, stated rather than fixed. Project Swap scores 2.24 — base 1.2 (official blog, read first-party) × agents / tool use / MCP 1.5 × Anthropic 1.3 = 2.34, minus 0.1 (already-rich — Agents (LLM Agents) runs past 1,200 lines). Gemini 4 Argon scores 1.69 — base 1.0 (no first-party read, three agreeing passes) × frontier models / scaling / evals 1.3 × Google DeepMind 1.3, with no +0.4 because no weights were released and no +0.3 because the new page is a model page, not a new entity or concept. So a book-swap study outranks a frontier launch because one could be read and the other could not, and because agents is weighted 1.5 against frontier's 1.3. The weights are doing what they were written to do; the brief publishes the disagreement instead of reordering quietly
  • Stories 3, 4 and 5 are a three-way tie at 1.56 — base 1.2 × topic 1.0 × Anthropic 1.3 each. The topic weight is unmatched in all three cases: interests.md has no row for science, robotics/labour-economics or mathematics, and 1.0 was taken as neutral rather than dropping to the 0.7 product row — the same call made for biosecurity on 09-30 and for Kumo Tabular on 09-29. The tie was broken by publication date, newest first; agents/brief.md specifies no tiebreak
  • Ai2 (Allen Institute for AI): Olmo-core 3, and its ## Strategic Position placeholder came down with it. Apache-2.0 MoE training stack at github.com/allenai/OLMo-core: FSDP replaced by distributed data parallelism with experts resident on GPUs, plus rowwise expert parallelism, GPU-resident routing, grouped GEMM and MXFP8. A 47B MoE at 52,000 tok/s/GPU against 19,400 previously; MXFP8 ~21% over BF16 with peak active memory 103 GiB → 95 GiB; a 1.2T-parameter model across 512 GPUs at 858 TFLOP/s per accelerator (source). It scores 1.00 — base 1.0 (allenai.org newly blocked, two search passes) × open-source LLMs 1.0 × Ai2, absent from interests.md's org table, taken as 1.0 — which is why a notable open release is a page update rather than a story, and the score is printed so the omission is auditable. The page's Strategic Position had been an annotated _TBD_ since 2026-08-09; this source speaks to it, so it is now written and the wiki's annotated-placeholder count went 9 → 8
  • Three orgs absent from interests.md changed a ranking in the last four runs — NVIDIA and AMD on 09-30, DeepSeek on 10-01, Ai2 today. Flagged to the user again, alongside the still-missing governance/security row and the newly visible gap: no row for science or for labour economics, which is what flattened three of today's five stories to a neutral 1.0
  • Updated: Anthropic, Google DeepMind, Ai2 (Allen Institute for AI), Gemini 4, Agents (LLM Agents), Embodied Agents, AI for Mathematics, AI Alignment, Mechanistic Interpretability, AI-Enabled Cyberattacks, Frontier Pacing, index.md — 12
  • No fact appears in two sections. Argon's "without cyber guardrails" is in Story 2 and the four unenumerated safeguards are in 👀 as an unanswerable question, never as the same claim twice. Olmo-core 3's figures appear only here. The missing papers snapshot is in 👀 as pipeline state and referenced in 📄 only as the reason there is one pick. The /research discovery is in the header; the six post titles are not repeated as stories