AI Trend Notifier
EN한
← archive

$ cat briefs/daily/2026-08-14.md

2026-08-14

August 14, 2026 (Fri)

4 stories · 1 paper · 3 watch items · 3 new pages · 1 correction

Two frontier releases landed on the same day and **neither published a harness**, so their one directly comparable number — DeepSWE, 65.3 against 62.7 — means nothing yet. Underneath that, the day's most useful finding is domestic: a fact this brief published yesterday was wrong, and the check that should have caught a second gap has never run at all.

+3new pages
[01]

Top Stories

1. DeepSeek's V4-Pro leaves preview, and the vendor and the referee disagree about how much changed

  • DeepSeek replaced the preview edition of V4-Pro in its API documentation with build 0813: 1.6T total / 49B active MoE, 1M context with 384K max output, thinking and non-thinking modes, Responses API and Codex support, MIT weights, $0.435/M input (cache miss) · $0.87/M output (source).
  • Vendor-stated agentic gains against its own preview are enormous: Terminal Bench 2.1 72.1 → 87.9, DeepSWE 12.8 → 62.7, CyberGym 52.7 → 83.3. DeepSeek has not released the harness, so none of it can be replicated — "the third-party record is empty."
  • The one figure produced by someone else moves by one point. Artificial Analysis puts it at 53 on its Intelligence Index — above DeepSeek V4-Flash, the company's own cheaper model, by one, and four points behind GPT-5.6 Terra, the mid tier of OpenAI's family. SCMP's headline: "struggles on benchmarks, shines in cybersecurity."
  • Why it matters: the same release is a step change or a rounding error depending entirely on who ran the benchmark — and the difference is not close. A lab's own agentic suite with no published harness and a ten-benchmark independent composite are answering different questions, and only one of them can be checked.
  • → DeepSeek V4-Pro-0813 (new), DeepSeek, Eval Harness Configuration

2. Gemini 3.7 Flash ships 23 days after 3.6 with every capacity figure identical

  • 1M context, 64K max output, March 2026 knowledge cutoff, text/image/audio/video — all unchanged from Gemini 3.6 Flash (2026-07-21). Generally available at announcement, not a preview (source).
  • What moved is the benchmark column: AutomationBench 17.0% → 30.4%, DeepSWE 48.6% → 65.3%, FrontierCode 34.4% → 43.6%, plus LVBench 85.4%, GDM-MRCR v2 97.0% @128k / 62.5% @1M and WebDev Arena 1588 Elo. On AutomationBench it is placed ahead of GPT-5.6 Terra (23.6%) and Claude Sonnet 5 (10.7%).
  • $0.75/M input · $3.75/M output, introductory through 2026-12-31, reverting to $1.50 · $7.50 on 2027-01-01. The "half price" in the coverage is a window with a published expiry date, not a cut.
  • Logan Kilpatrick attributes the gain to algorithmic improvement over roughly three weeks rather than scale. That is the vendor's account of its own cause and no evidence was published for it.
  • Why it matters: identical capacity, doubled automation score, three weeks apart — this is a lab shipping post-training and calling it a model. It is also unverifiable here: none of the six benchmarks appears in any snapshot in sources/evals/, so there is no local column to check a single figure against.
  • → Gemini 3.7 Flash (new), Google DeepMind

3. Ultrafast is a service tier — and OpenAI already sells a 750 tok/s tier

  • GPT-5.6 Sol at up to 14× the speed of Standard, up to 750 output tokens per second, powered by Cerebras, API-first to a select group of customers. Cerebras states it runs "with the same intelligence as GPT-5.6 Sol Standard" (source).
  • No price for Ultrafast appears in anything read — and GPT-5.6 Sol (and Terra, Luna) has carried Sol Fast, ~750 tok/s, $12.50 / $75 since 2026-07-30. Ultrafast's headline throughput is the same number. Nothing read states how the two relate, so neither row was edited on the strength of the other.
  • Two other absences: what the 14× is measured against (no Standard tokens-per-second baseline was published) and any benchmark behind "same intelligence" — a capability-parity claim made with no capability measurement.
  • Why it matters: OpenAI's framing is that speed no longer costs intelligence, which would be a real change. But the announcement omits the baseline, the benchmark and the price — the three numbers that would separate a new capability from an existing fast tier running on different silicon.
  • → GPT-5.6 Sol (and Terra, Luna), OpenAI

4. Correction: this brief said the Qwen 27B shipped. It did not.

  • Yesterday's brief and Qwen 3.8 27B recorded Released: 2026-08-12. What shipped that day was the Max-class flagship Qwen 3.8 Max alone; the 2026-08-13 run read one drop as covering both checkpoints (source).
  • As of 2026-08-13 the 27B has no official repository, no model card, no licence file and no benchmark of its own, and a ModelScope countdown points at 2026-08-15, 00:00 JST — a third date, after the announced 08-10 and the assumed 08-12.
  • Released is back to not yet and Availability to unknown. The wrong value stays visible on the page rather than being deleted, and the immutable 08-12 snapshot is untouched — the correction lives in a new one.
  • Why it matters: the error came from treating a vendor's announcement of two artefacts as a release of two artefacts. Alibaba has still never restated, moved or withdrawn the 2026-08-10 date in any first-party channel read here — a countdown page replaces a missed date without mentioning it, which is not the same as moving it. Five days slipped and counting.
  • → Qwen 3.8 27B, Alibaba / Qwen AI Lab, Open-Weights Policy Fight

Scoring note: story 1 leads at 2.21 — frontier model release (1.3) × open weights (+0.4) × the Chinese-lab entity weight, and it closes a five-run Watch item. Story 2 scores 1.95, story 3 1.43 (product/distribution 0.7 lifted by the frontier-model tie), story 4 1.30 — a correction to a published fact carries the frontier-model weight of the fact it corrects, which is why it sits above the paper rather than below the fold.

[02]

Paper Picks

OpenART: Scaling Agent Red Teaming via Open-Ended Environment Evolution (arXiv:2608.00677) — red-team the environment, not the prompt (arXiv:2608.00677, Fudan University / Shanghai AI Laboratory / XSafeAI)

  • 10,000+ validated stateful scenarios across 50 domains, drawn from a pool of 500,000+ tools and skills, with a median of 97 tool calls per task and 75 agent-model configurations. The premise: agent risk accumulates through shared state reused across long-horizon workflows, and short static safety benchmarks cannot see it (source).
  • The page's Results section is empty, on purpose. No attack success rate, no result table and no comparison was seen — arxiv.org is blocked from this sandbox and the abstract extract carries scale figures only. A paper page that paraphrases findings nobody read is worse than one that says what it is missing.
  • Why it matters: that median of 97 tool calls is the number to hold. Every agentic benchmark this wiki records a score for — Terminal Bench 2.1, DeepSWE, CyberGym above — is far shorter-horizon than that, which is exactly the paper's complaint about the field it is entering. And it comes from an academic consortium, where the agent-safety evaluations here are almost entirely first-party.
  • First HuggingFace Daily item to arrive with an arXiv identifier since 08-11. The other two trending titles carried none and were not ingested.
[03]

Watch

  • Three Anthropic Alignment Science articles are in no sources/ file and cited nowhere in this wiki — AuditBench (2026-03-10), Introspection Adapters (2026-04-28) and The Hot Mess of AI (arXiv 2601.23045, ICLR 2026). The standing check added on 2026-07-31 compares the newest article against sources/; it has never swept the back catalogue. A source CLAUDE.md records as cited 24 times has three uncaptured articles sitting behind the watermark. Not backfilled this run — three research posts is not a margin-of-the-day job, and writing summaries of pages nobody read is the exact failure this pipeline exists to prevent. Carried to Sunday's lint. → Anthropic
  • MiniMax Music 3.0: "open weights" per MiniMax, "hosted, not downloadable" per a model directory, and two different release dates. MiniMax's own post title says Open-Weights with repositories reported on Hugging Face, GitHub and ModelScope (2026-08-13); a directory dates Music 3.0 to 2026-07-16 as a hosted model. Nothing read distinguishes them. No model page created — the licence matters here, because MiniMax's H3 Community License already excludes the US, EU, UK and South Korea. → MiniMax
  • Grok Bot — persistent agents with their own VMs that sign into your tools — is in early beta and held out of the wiki. Surfaced through a Latent Space digest and one third-party blog; no first-party material was read. Same treatment as GLM-5.5, which is on day twelve as a rumour with no model card, benchmark or endpoint. → xAI, Z.ai
[04]

New in Wiki

  • DeepSeek V4-Pro-0813 (new model page) — given its own page on the DeepSeek V4-Flash precedent: DeepSeek ships dated builds that change what a model does without changing what it is. Two conflicts recorded rather than resolved — the release date (08-12 per one tracker, 08-13 per the outlets and DeepSeek's own "launching today" line) and the parameter count (1.6T/49B against 1.57T/48B), third-party on both sides.
  • Gemini 3.7 Flash (new model page) — full spec row set; no parameter count, architecture or weights are reported and none is expected on this line.
  • OpenART: Scaling Agent Red Teaming via Open-Ended Environment Evolution (arXiv:2608.00677) (new paper page) — see Paper Picks.
  • No new entity or concept page this run, so nothing requires user review.
[05]

Updates

  • Qwen 3.8 27B — corrected: Released 2026-08-12 → not yet, Availability → unknown. The withdrawn claim is kept visible and the date table now runs to four entries. One genuine gain: the first source read describes the 27B as dense — third-party, so it stayed out of the spec table.
  • GPT-5.6 Sol (and Terra, Luna) — Ultrafast section added, with the Sol Fast overlap recorded as an open question rather than resolved.
  • DeepSeek V4 — now points at the 0813 build; everything on that page describing V4-Pro refers to the July GA / preview state.
  • Eval Harness Configuration — two additions. The DeepSeek/Gemini pair (62.7 vs 65.3 on DeepSWE, neither with a published harness), and a measured base rate for reproducibility: a community effort reproduced claims from 2,200+ ICML 2026 papers with coding agents over 19 days — 1,221 participants, 35,908 claims judged, 3,978 confirmed, 266 fully and 632 partially reproduced without falsification. Coverage calls that "more than half"; 898 of 2,200+ is under half, and the "more than half" figure counts at least one claim verified per paper. Both are recorded, because the gap between them is the finding — and nothing read states how the automated judge was validated.
  • Agents (LLM Agents), DeepSeek, Google DeepMind, OpenAI, Alibaba / Qwen AI Lab, MiniMax — Recent Activity and model lists updated.