AI Trend Notifier
EN한
← archive

$ cat briefs/daily/2026-09-19.md

2026-09-19

September 19, 2026 (Sat)

3 stories · 2 paper picks · 4 watch items · 5 new pages

**Everything in today's brief was published on 09-16 or 09-17, and yesterday's run looked for all of it and found none of it.** That entry says in as many words that a search against Anthropic's newsroom "reports no Anthropic publication dated 2026-09-17". It was wrong when it was written, and not because the sweep failed: the run fired about six hours after a US-afternoon post and the search index had not caught up. **The lag between a lab publishing and this pipeline being able to see it is roughly a day**, and it is worth knowing because a negative result from the sweep is what this brief prints on quiet days. What arrived is unusually coherent for four unrelated items. **Two labs published instruments for arguing about how fast AI is moving, 24 hours apart and in different formats.** A third party shipped a 9× smaller copy of somebody else's model. And three separate papers took the agent harness apart.

+5new pages
[01]

Top Stories

1. Claude spent four weeks rewriting 36 biology models, supervised by two people who had never written a kernel (2.24)

  • How Claude is uplifting biomolecular modeling, 2026-09-17, captured here day +2. www.anthropic.com is EGRESS_BLOCKED, so every figure is a search extract with a pass count (source)
  • 36 optimised packages across more than 30 open-source models in just under four weeks, in six families — co-folding and structure prediction 14, genomics 7, structure generation 6, hallucination 3, inverse folding 3, protein language models 3. Roughly 4× faster on average with minimal precision loss, and nearly 2× with identical outputs — two speedups under two fidelity conditions, and the coverage mostly quotes only the first
  • The supervision detail is the story. Two members of technical staff supervised it. They had biomolecular-modeling experience and none in inference optimization or kernel engineering — so they could check that the output was faster, and not how it was made faster
  • Also shipped: a low-memory mode putting systems larger than 10,000 tokens on a single NVIDIA GPU node, all the optimised code open-sourced, and a protein-design competition with Adaptyv Bio — $1 million in jointly sponsored wet-lab validation for more than 5,000 designs free to entrants, $1 million more in Claude credits, $250,000 in compute from Modal, DNA from Twist Bioscience
  • Why it matters: every agentic-coding account this wiki holds reports an input — Databricks' engineers spending ~60% more, Yegge's thousands a month. This one reports artefacts, open-sourced, with a stated speedup. It is the first entry on Agents (LLM Agents) that a third party could re-run rather than take on trust — and nobody has, so the 4× is still the vendor's number for the vendor's agent
  • Not established: no baseline hardware, no per-model speedup table, no accuracy delta for the "minimal precision" case, no competition dates. One outlet headline carries a 100× cost figure no other pass supports; it is in the snapshot and on no page
  • → Agents (LLM Agents) · Anthropic

2. Anthropic publishes a number for how much of its own AI R&D is done by AI, and it is 26% (2.23)

  • Measurements for understanding the pace of AI development inside frontier labs, 2026-09-17, on an Anthropic Institute path. First results from a prototype R&D Automation Index, plus two further measurements — oversight of autonomous agents and compute allocation — neither of which carries a figure in anything read (source)
  • The scale is AL0–AL5, attributed by one pass to Epoch AI: AL3 is AI collaborating under close human direction, AL4 is AI leading — "completes most of a task end-to-end from a high-level prompt" while a human supervises — AL5 is fully autonomous
  • As of August 2026: 26% of AI R&D work at AL4, above 90% at AL3 or higher, and nothing at AL5
  • How it was built: each week in July, a 20% sample of employees in model-R&D departments; Claude read Slack and documentation to identify their work, producing roughly 15,000 tasks; Claude then organised those into 542 categories
  • Why it matters: Frontier Pacing has collected positions since 2026-07-28 — an employee letter, an essay, a three-step plan, a standard — and not one checkable quantity. Amodei's 09-12 commitment put evaluators with publication rights inside the company and had no engaged evaluator, no contract, no start date. It also had no instrument. This is one, published five days later, with the methodology released so other labs can run it
  • The criticism is structural and nothing read answers it: Anthropic sampled its own employees, Claude read the records, Claude built the categories, Anthropic applied the scale. One outlet's framing — "'lead' doesn't mean what you think" — is fair: AL4 keeps a human on every task and nothing is at AL5, which is narrower than the "AI is building AI" headlines
  • Two loose ends kept visible: the sample is July and the headline is as of August, unreconciled; and the starting point is "less than 1% in February" on three passes against "1% in March" on one headline
  • → R&D Automation Index (new) · Frontier Pacing

3. A 27B model in 5.9 GB, and the retention number in the headline is not the one a local-agent user needs (1.90)

  • PrismML — a Caltech spin-out on one pass — released Ternary Bonsai 2 27B on 2026-09-17: a ternary {−1, 0, +1} quantization of Alibaba's Qwen 3.8 27B with FP16 group-wise scaling, 1.76 effective bits per weight, 5.93 GB against 53.80 GB in FP16, same 262,144-token context, text and image in, Apache 2.0, GGUF and MLX, stated to run on a 16 GB laptop or a single 24 GB GPU (source)
  • The headline: 98.2% retention — an aggregate of 83.9 against the parent's 85.4 across 20 benchmarks spanning reasoning, math, coding, instruction following, vision and agentic tool use. Vision 96.3%, knowledge and reasoning 96.9%
  • Outside that suite it is a different number. On Terminal-Bench 2.1: 52.8 against 69.7 — about 75%, with SWE-bench Verified stated in the same band and no figures given
  • Neither is wrong and they are not in tension. They answer different questions, and the one that describes driving a terminal for a long time is the one not in the headline
  • Why it matters: Alibaba / Qwen AI Lab's page has recorded its models being taken without permission since June — the Senate letter, GTG-16005, 151 million exchanges. This is the same weights taken with permission, under the licence Alibaba chose, by someone who published the result. The first third-party agentic measurement of Qwen3.8 27B this wiki holds arrived inside a competitor's release notes — and it reads 69.7 where Qwen 3.8 27B carries 73.0 from Alibaba's own model card. A 3.3-point disagreement with no harness named on either side, disclosed on the new page and not reconciled
  • Not established: no price, no hosted endpoint, no throughput or energy figure despite both being claimed, and no third-party evaluation of Bonsai 2 of any kind. The ~406,000 GGUF downloads on 09-18 are distribution, not quality
  • → Ternary Bonsai 2 27B (new) · PrismML (new) · Open-Weights Policy Fight
[02]

Paper Picks

An Empirical Study of Harness Design for Coding Agents — arXiv 2609.20804

  • TL;DR: fixes a coding agent's execution loop and varies planning, action space and context management across 176 matched settings, four models, SWE-Bench Verified and Terminal-Bench 2.1 — the first study here that treats the harness as a thing with parts rather than a monolith
  • Findings: context management's benefit is mostly preventing context-overflow failures; staging rule-based elision before LLM summarization wins, while making elided content recoverable "adds machinery that models rarely use and yields no accuracy gain"; planning is an accuracy scaffold for weak models and a cost saver for strong ones; bash-capable models beat predefined tools on cost
  • Why read it: Eval Harness Configuration exists because benchmark figures come out of unnamed harnesses. This measures what the unnamed thing does, and finds the harness is not a constant you can hold equal — its effect has the model's strength in it. The recoverable-elision negative result is the part nobody publishes if only improvements count
  • → An Empirical Study of Harness Design for Coding Agents

Agora: Git as Shared Memory for Collective AutoResearch — arXiv 2609.18094

  • TL;DR: stores collective autonomous research as an append-only DAG of Git commits, so every claim is a commit anyone can check out and rerun. First run: 13 workers, nearly 12 days, no assigned tasks, no central planner, 1,703 contributions; the evaluator moved 3.39 → 1.899 bits per byte initializing a frozen 119.6M-parameter attention-SSM hybrid from 141 donor models with no training data and no gradients
  • The result is 165 independent reproductions with none failing, and a winning recipe whose ancestry spans 145 commits across 15 accounts
  • Why read it: it reports its own confounds first — one mid-run human rescue from a monoculture, what the trace does and does not establish, and the controlled comparison it did not run. A twelve-day single-arm study that says so is more useful than one that doesn't
  • → Agora: Git as Shared Memory for Collective AutoResearch
[03]

Watch

  • Twelve spec-check conflicts, now twenty-one days old — and tomorrow is where they have been headed since 09-17. Action run 100 (2026-09-18 07:11 UTC, scheduled) exits 1 through its conflict branch, not the exit-2 unreachable branch, so this is a finding and not an outage; the job log ends ok, so Slack is receiving them. The eight printed lines are byte-for-byte yesterday's: two DeepSeek pages at exactly half the first-party figure, four Gemini Flash pages at exactly double and one at 4×, and glm-5-3-flash context 1,048,576 vs 1,310,720. Two of those families look like missing tier labels rather than wrong numbers. Sunday's W38 lint is one day away and takes them. spec-check.py was not run here — openrouter.ai is blocked from this sandbox
  • 2609.20519 (SoL-Pi) was read and given no page. It discovers harness mechanisms by auto-research loops instead of designing them, keeping four: action execution, context compaction, observation handling, delegated reading — 44.7–49.0% less token traffic and about 1/3 less API cost at comparable performance on the 51-task EdgeBench. Its second mechanism is Context Compaction, which got a page yesterday for being a prompt-injection channel with no foreign origin. The same component is the largest token saving and the newest attack surface at once. Carried, per the rule that a link to a page which does not exist is a dead link
  • Grok 4.7 is unreleased on the ninth day past a third expired window, and no xAI page moved. Newest held remains Grok 4.6 (2026-08-12)
  • Today's Chinese-lab rotation — Alibaba Qwen and MiniMax — returned nothing, and that is the result. Alibaba has nothing newer than Qwen 3.8-Max (2026-08-03) and Qwen 3.8 27B (2026-08-14); Qwen 4 still has no announced date and the September window is still one 2026-07-20 leak. MiniMax's M3 Pro carries no announcement, API string, model card or technical report on its ninth consecutive day held out. A rotation that only reports when something ships cannot tell nothing happened from nobody looked
[04]

New in Wiki

For review.

[05]

Updates

  • Google DeepMind: the DeepMind Institute launched 2026-09-16 — directors Shane Legg (also managing editor), James Manyika and Demis Hassabis — to "surface differing views" on AGI, opening with four essays: "The case for reasoning transparency" (Shah, Dragan), "Economic policy for AGI" (Jacobs, Imas), "Principles for a new utopianism" (Cave), and Hassabis's "A framework for frontier AI and the dawning of a new age". Scored 1.09 and kept out of Top Stories: an institute launch matches product launches (0.7) rather than anything above it, and the +0.3 for a new page does not apply — no page was created, the material going to the entity. Recorded there as the other half of this week's pacing argument: a measurement invites a competing measurement, an essay collection edited by a director of the lab under discussion invites an invitation, and only one of the two is checkable in principle. Date resolved, not conflicted — Axios 09-16, TechCrunch 09-17, and 09-16 was the Wednesday one pass names (source)
  • DeepSeek V4.1-Flash: the technical report 2609.19969 landed nine days after the weights and is the first citable first-party document on that page. It names the mechanisms the page had only the effects of — cross-layer KV cache reuse in Compressed Sparse Attention 2 (CSA2) with FP4 KV caching behind the 890 bytes per token HBM-resident cache, and SWA Bounded Replay taking the persistent cache on SSD or host memory to roughly 1/8 of V4-Flash's, a second compression the coverage did not lead with. 45T-token multimodal corpus is the first training-scale figure on any DeepSeek page here. The 890 bytes and the 8B prefill / 16B decode split are confirmed unchanged from single-pass figures carried since 09-11
  • Agents (LLM Agents): two new ### sections — the three-paper harness cluster, and the biomolecular result read for its supervision structure rather than its speedup
  • Frontier Pacing: two dated sections for 09-17 and 09-16, placing the index and the Institute against the artefact types the page already holds
  • Alibaba / Qwen AI Lab: the licensed counterpart to every distillation entry on that page