2026-08-27

1. Anthropic has now measured three times that its interpretability tooling gives no uplift — this time on predicting behaviour

• Would This Change Your Answer? (Adam Karvonen, Euan Ong, Subhash Kantamneni, Samuel Marks, arXiv 2608.16747, 2026-08-17) scores an explanation of a model behaviour by counterfactual simulatability — does it predict what the model does on related edited prompts — and builds CHIVE (Counterfactual Hypothesis Investigation Via Edits) to find behaviours in the wild and probe each with prompt edits at scale (source)
• The result: no uplift from any interpretability technique studied. Agents holding activation-reading tools predicted outcomes no better than agents that just read the transcript
• The constructive half: models trained to predict CHIVE experiment outcomes generalise to held-out settings, which makes explanation quality a trainable target rather than a matter of taste
• Why it matters: with Fine-Tuned Lie Detectors Failed to Generalize (2026-08-21) and AuditBench (2026-03-10, where scaffolded black-box tools beat white-box ones overall), that is three independent Anthropic measurements pointing the same way — the purpose-built internal instrument does not beat reading the model's own output. This wiki's Mechanistic Interpretability page opens by calling it "alignment's primary empirical tool"; on these three tasks, by its principal lab's own reporting, it is not
• What holds it back: no numeric result appears in anything read, and which techniques were compared was never stated. A tie is a different finding at 90% accuracy than at 55%
• → Would this change your answer? Evaluating Explanations of LLM Behavior In The Wild with Counterfactual Experiments, Mechanistic Interpretability, Anthropic

2. Two Chinese labs shipped cost-optimised open multimodal MoEs within hours of each other — and one of them shipped weights its bigger sibling is still withholding

• GLM-5.3-Flash (Z.ai, 2026-08-26): 320B-A18B natively multimodal MoE, 1M context, hybrid sparse + linear attention, MIT License, $0.15/M input · $0.03/M cached · $0.50/M output with a 50% launch discount. Vendor-stated DeepSWE v1.1 63.4 against GLM-5.2's 46.2, Terminal-Bench 2.1 84.3, AutomationBench 48.8; independent Artificial Analysis Intelligence Index 57 at $0.045 per task (source)
• Qwen3.8-Flash-Next (Alibaba / Qwen AI Lab, 2026-08-26): released on the date its ModelScope countdown named — 125B main model + 51B N-gram embeddings = 176B, 6B active per token, 262,144 native context extensible to 1M, licence qwen-community-1.0, BF16 and FP8 on Hugging Face and ModelScope (source)
• The contrast inside Z.ai is the sharper story. Twelve days ago GLM-5.3 shipped without its weights — withheld ~2 weeks pending a safety evaluation, while marketed as "the strongest open-weights coding model", with the window closing around 2026-08-28. Its smaller sibling shipped MIT weights on day one, and nothing read explains why the gate binds one and not the other
• Read Qwen's comparison column before its score column: every paired row is against Claude Opus 4.6 Max — two minor versions behind Claude Opus 4.8, three behind Claude Opus 5, neither of which appears. Beating a two-versions-old frontier model with 6B active parameters is a real result and not the one the framing invites
• Why it matters: no third party has measured Qwen3.8-Flash-Next at all, while GLM-5.3-Flash arrived with an independent composite already on it. Two comparable releases, one day apart, and only one is checkable — which is the difference that decides what this wiki can publish as fact
• → GLM-5.3-Flash, Qwen3.8-Flash-Next, Open-Weights Policy Fight

3. Z.ai announced its silicon as a headline feature, not a footnote

• The GLM-5.3-Flash launch post says the model was "running entirely on Chinese AI chips" — stated unprompted, in the announcement itself, alongside the parameter count and the licence (source)
• No part, no vendor, and no scope: nothing read says whether "entirely" covers training, serving, or both
• Why it matters: this wiki already holds LongCat-2.0, trained on Huawei Ascend 910 — so the capability is not new. What is new is a lab selling it, which reframes export controls from a constraint a Chinese lab works around into a claim it advertises. Kept separate from story 2 because it is a fact about the supply chain, not about the model
• → Z.ai, AI Governance

4. Gemini 3.5 Transcribe measures the half of its job that has a benchmark

• DeepMind released Gemini 3.5 Transcribe (2026-08-26): 85+ languages auto-detected, FLEURS 5.50% WER streaming / 5.04% non-streaming, and 70% faster time-to-final than Google's own Chirp 3. Ships in Google Antigravity and Gboard Rambler; coming to Search Live, Gemini Live, Docs, Keep, Gmail and Chrome (source)
• The model adapts unstructured speech into formatted text and removes filler words — it edits as well as recognises. Nothing published measures the editing: not whether the formatting is right, not what a user loses when a disfluency carried meaning
• Two caveats ride the WER: the only baseline is Google's own previous model, and "across a set of top languages and locales" does not name the subset of FLEURS's 102 languages — which is load-bearing in a claim about 85+ language detection
• Why it matters: second DeepMind recognition model in a month to publish the capability ahead of the accuracy. SL2T published no figure at all; this one publishes a figure for the easy half
• No price and no API model id appear anywhere read
• → Gemini 3.5 Transcribe, Google DeepMind