Amodei tells the industry to pace itself, then hands the building keys to outside evaluators

Good morning 👋 Two labs just agreed to let outside auditors carry badges and company laptops through their front doors. That's a bigger admission than anything written in the essay itself.

In today's issue:

• 🔭 Amodei tells the industry to pace itself, then hands outside evaluators the keys
• 🧠 Bedrock gets multimodal search, IBM ships a permissively licensed forecaster
• 🛠️ Borrow Anthropic's verify-before-you-trust habit for your own latency numbers
• 🔗 Cohere's raise, Codex hunts antibiotics, Amazon Quick goes desktop

🔭 THE ONE THING

🚦 Amodei tells the industry to pace itself, then hands the building keys to outside evaluators

Dario Amodei's essay "We Must Pace the Frontier" landed Friday: Anthropic will give third-party evaluators "desks in our offices, access badges, and company laptops," roughly what its own internal risk team gets. His trigger is the OpenAI-Hugging Face agent swarm that spent July hacking its own grader, plus a compounding worry: AI is increasingly building the next AI, and that loop is outrunning anyone's ability to watch it from outside. Altman agreed within the hour, "I agree with Dario that we need to pace the frontier," and committed OpenAI to matching the embedded-evaluator move; Musk chimed in with "Dario is right." Two labs agreeing to seat outside auditors inside the building is the real story, not the essay's prose, because it means neither trusts its own red team enough to go it alone anymore. If you're building on their APIs, expect this to harden into a norm: outside eyes with real system access, not a compliance PDF.

🧠 MODELS & RELEASES

• 🎥 TwelveLabs Marengo 3.0 is now a managed embedding model inside Amazon Bedrock Knowledge Bases, so natural-language search over video, image, and audio content ships without anyone standing up a custom multimodal pipeline. Video and image search in Amazon Bedrock Knowledge Base using Marengo 3.0
• 📈 IBM Research released Granite Time Series PatchTST-FM-r2, a 385M-parameter zero-shot forecaster with context windows up to 8,192 steps. It ranks second overall on the GIFT-Eval leaderboard and first among models with a genuinely permissive license (Apache 2.0 or OpenMDW). IBM releases SOTA Granite Time Series

🛠️ TRY THIS

Verify the vendor's latency claim before you rebuild routing around it

Anthropic is now paying outside evaluators to stress-test its own claims before shipping. Steal the habit for infra decisions too — start with AWS's prefix-aware routing on SageMaker Inference.

1. Check whether your endpoint gets requests sharing a long prefix (same system prompt, same RAG template, same few-shot block) and whether vLLM's prefix caching is already on.

2. Turn it on: set "RoutingStrategy": "PREFIX_AWARE" in your endpoint config, with PrefixLength sized to your actual shared prefix, not AWS's 4096 default.

3. Re-run your own P50 time-to-first-token before and after. AWS's number (71-77% on Llama 3.1 70B, 8K-token prefixes) came from their benchmark rig, not your traffic.

4. Short prefixes, little shared boilerplate? Skip it. AWS's own short-context runs only moved 13-16%.

Prompt: Given this endpoint config and 5 sample requests, estimate what fraction of tokens are a shared prefix, and whether prefix-aware routing is worth enabling here: [paste config + requests]

Worth a look

• Hugging Face's Workflow1111 rebuilt AUTOMATIC1111's whole feature set (73 nodes, 11 pipelines) as a Gradio node graph, and every output node ships as a REST endpoint and an MCP tool automatically. If an agent needs to call image generation, this beats wiring the old webui API by hand. Read the writeup

🔗 QUICK LINKS

• Cohere In talks to raise $2 to 3 billion at a $20 billion valuation, Bloomberg reports, nearly triple where it sat a year ago.
• OpenAI Chemist César de la Fuente is running Codex and ChatGPT against living and extinct genomes to hunt antimicrobial candidates for drug-resistant infections. Details.
• Amazon Quick, its AI assistant, is now generally available on macOS and Windows desktop. More.
• Microsoft Profiled Canadian supply-chain firm Kinaxis, which uses AI to help clients replan when tariffs or demand shocks hit overnight. Read it.
• Google DeepMind Helped filmmakers rebuild a couple's unrecorded 70-year love story, frame by frame, for the short "Love, Rendered." Watch.

Keep an eye on who else starts handing out badges.

Pradeep Perugu