2026-08-12

1. NVIDIA shipped a cheap agent model and, the same day, the software that decides when to use one

• Nemotron 3.5 Lightning, 2026-08-11: a 30B hybrid MoE with 3B active parameters and a 1M-token context window, open weights in BF16 and NVFP4 under OpenMDW-1.1 on Hugging Face and NGC. Distilled from the frontier Nemotron 3 Ultra; interleaved Mamba-2 and MoE layers, multi-token prediction, DFlash speculative decoding (source).
• Alongside it, NeMo Switchyard — an open-source model routing library that picks a model per step of an agent workflow, with tuning-free routers (LLM classifier, stage router, escalation router) plus tunable ones. NVIDIA's internal claim: frontier-level accuracy at nearly one-third the task-completion cost of Opus 4.8 alone.
• NVIDIA's model figures: PinchBench 86.5%, up to 4x output speed, up to 30% faster agentic task completion. Artificial Analysis independently scores it 24 on its Intelligence Index. PinchBench appears in no eval snapshot this repo holds, so 86.5% has nothing local to check it against — and three of the four figures measure speed or cost, not capability.
• Why it matters: the pairing is the announcement. A router whose headline is a cost ratio against a competitor's flagship is an argument for putting small models inside someone else's workflow — and NVIDIA, which sells the compute either way, is the party with the least stake in which model wins. It also breaks something this wiki assumes: once a router is in the path, the price on a model page stops describing what a workflow costs.
• → Nemotron 3.5 Lightning, Model Routing, NVIDIA

2. Claude's output is now watermarked worldwide — and Anthropic publishes the reasons it will not always work

• Anthropic detailed how it marks AI-generated content: an imperceptible pattern inserted directly into generated text, expected to survive copy-and-paste, plus C2PA-signed provenance metadata on generated SVG, PNG and JPG files (source).
• Effective 2026-08-02 — the day EU AI Act Article 50 became applicable — at the model level for new Claude models from that date, with older models a stated work in progress. Applied worldwide, not EU-only, across Claude Platform, Claude, Claude Code, Cowork and Claude Tag.
• The caveats are Anthropic's own: a detected mark is not conclusive evidence Claude produced the content, absence of a mark guarantees nothing, and detection fails on paraphrase, translation, heavy editing or short text. Independently: a not-yet-peer-reviewed evaluation reports paraphrasing removed nearly all detectable marks, and The New Stack notes code is close to unmarkable — prettier, black and gofmt rewrite style deterministically. No public detector exists.
• Why it matters: this is the first governance requirement this wiki has recorded that reaches inside the generation loop rather than into what a lab must publish. The compliance shape is the part to keep: Article 50 is binding, the Code of Practice describing how to satisfy it is voluntary, and reporting says no technology yet meets all four criteria the Article imposes. A binding duty is being discharged by a mechanism its own vendor says proves processing, not authorship.
• → Content Provenance (AI output marking), Anthropic, AI Governance

3. Hidden chain-of-thought is locked with a key every model in the ecosystem holds

• arXiv:2608.09867 (submitted 2026-08-10, ELLIS Institute Tübingen / MPI-IS): frontier providers hide reasoning traces by returning them to the client as encrypted blocks the client passes back each turn, rather than storing them server-side (source).
• Those blocks are fully interchangeable across sessions, users and models within one provider. Inject a strong model's encrypted trace into a weaker, less-safeguarded sibling and it decrypts and prints the trace verbatim. Anthropic, OpenAI and Google are named.
• The strong model is never attacked — its safeguards are irrelevant, because decryption is a property of the ecosystem rather than of any one model. No vendor response was found in anything read, and no figures are carried onto the wiki page: none appeared in the abstract, and a third-party "182 credentials" headline appears in no other source, so it stays in the snapshot as unverified.
• Why it matters: hidden reasoning was a product decision — the traces are training data for a distillation attack, which is the premise under half the export-control argument on AI Governance. If they can be pulled through a provider's own API at scale, that premise moves. It is also the same cross-model substitutability that makes routing work, which means Story 1 and Story 3 are describing one property from two sides.
• → Stealing Reasoning Traces from Proprietary LLM APIs (arXiv:2608.09867), Reasoning Models