Hassabis steps back, Jeff Dean walks out the door

Good morning 👋 Demis Hassabis just traded the corner office for a lab coat, and Google's shares didn't like it.

In today's issue:

• 🔭 Hassabis steps back, Jeff Dean walks out the door
• 🧠 NVIDIA open-sources a reasoning model built for robotaxis
• 🔬 Ditch the JSON schema: code-style tool calls just won
• 🚀 A 20-year-old's second startup lands $28.5M for agent infra
• 🛠️ Turn a compliance doc into a Bedrock guardrail
• 🔗 OpenAI teams up with the APA on teen mental health, plus two more worth a skim

🔭 THE ONE THING

🧠 Hassabis steps back, Jeff Dean walks out the door

Sundar Pichai's memo reshuffles the top of Google's AI org: Demis Hassabis moves from CEO to Chair of Google DeepMind and the newly created Chief Scientist of Alphabet, ceding day-to-day control of Gemini and frontier research to Koray Kavukcuoglu, now SVP reporting straight to Pichai. Hassabis keeps Isomorphic Labs, so read this as a promotion into strategy, not a demotion. The more interesting exit is Jeff Dean's: 27 years in, he's out to co-found Discovery Loop with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le, a public-benefit corp betting AI can run the hypothesize-experiment-evaluate loop of science on its own, with Google staying on as investor and cloud partner. For anyone building on Gemini, the org chart matters less than the fact that the guy who invented Bigtable, Spanner, and MapReduce just bet his next decade on automating science itself.

🧠 MODELS & RELEASES

• 🤖 NVIDIA shipped Alpamayo 2 Super, a 34B-parameter open reasoning model for robotaxis, licensed under OpenMDW-1.1 so anyone can fine-tune it or redistribute it commercially. On LingoQA, the benchmark built for the rare long-tail scenarios that actually total self-driving cars, it beats GPT-4o by 23.2 points and Gemini 2.5 Pro by 15.1. Details
• 🌀 Google DeepMind's WeatherNext now reads a cyclone's track, intensity, and wind structure a full day further out, its three-day calls as accurate as last generation's two-day ones. It caught Hurricane Melissa's rapid intensification correctly in 2025, and the weights are open on GitHub and Weather Lab. Details

🔬 RESEARCH HIGHLIGHTS

• Skip the JSON schema. A new benchmark run across 14 models on BFCL v4 found that letting agentic models call tools as typed Python stubs, not rigid JSON blobs, matches or beats the JSON approach in 11 of 14 models. GPT-5.6 picked up a 10.6% gain just from the switch, and under parallel tool calls the code approach won or tied on 13 of 14. The bigger tell: when context gets noisy, JSON calling degrades 2.3% on average while code-based calling barely moves. If your agent stack still forces every tool through a schema, you're leaving accuracy on the table.

• TrajDebug goes after a problem anyone running multi-step agents knows too well: a trajectory fails after forty steps, and good luck finding which one actually broke it. The authors built TrajErrBench, 486 hand-annotated failed runs across tool-use and coding tasks, and their method traces each error's fate through the rest of the trajectory to isolate the one that mattered. It beat existing baselines at root-cause localization. Less eyeballing logs, more actually knowing where things went wrong.

• Selective trust, not blanket suspicion. Train a model to ignore misleading context and you also get a model that ignores correct context, which defeats the point. SCOPE, tested on a new benchmark called MIST that runs each question through clean, misleading, correct, and irrelevant versions, trains on preference pairs balanced across all four instead of just the adversarial case. Result: fewer flips from correct to wrong under misleading input, without tanking accuracy when the context was actually right. Worth watching if you're building anything RAG-shaped, since "resist bad context" was always the wrong target.

🚀 AI STARTUPS

• Naïve closed a $28.5M Series A led by Nexus Venture Partners, with YC, Zetta, Liquid 2 and a bench of operator angels (ex-HubSpot CFO JD Sherman, Amazon's Gert Lanckriet) filling out the round. The pitch: one API that gives an AI agent everything a founder needs to run a company solo, incorporation, payments, email, phone, compute, memory, and coordination across other agents. Founders Sean Dorje and Dennis Zax are 20 and already on their second startup after ezML. Naïve's own numbers, 2.3ms agent cold starts and memory queries at roughly 11x fewer tokens than the usual setup, are worth watching, not yet worth believing until someone outside the company reproduces them.

🛠️ TRY THIS

Bake a verifier into your agent's output, not a bigger prompt

Hassabis just traded CEO for chief scientist. Read into that what you want, but the signal is rigor over speed. AWS shipped something in that spirit this week: open source Agent Skills that let a coding agent build, test, and deploy a formal verification policy for your own agent's outputs.

1. Point a coding agent (Claude Code, Q Developer, whatever you run) at AWS's new Agent Skills for Automated Reasoning and hand it a compliance doc or internal SOP.

2. Ask it to draft a Bedrock Automated Reasoning policy that encodes those rules.

3. Have it generate test cases, some that should pass, some that should violate the policy, and run them against the draft.

4. Read the counterexamples it surfaces. They usually mean your original doc was vaguer than you thought. Fix the doc, then deploy the policy as a guardrail.

Prompt: Read the attached compliance policy. Draft an Amazon Bedrock Automated Reasoning policy that encodes its rules, then generate 10 test scenarios, 5 that should pass and 5 that should violate the policy, so I can validate it before deploying.

🔗 QUICK LINKS

• OpenAI teamed up with the American Psychological Association on youth mental health guidance, the kind of partnership that reads more like liability management than product news. link
• OpenAI published country-level ChatGPT usage data showing adoption patterns shifting from asking questions to getting tasks done. Worth a skim if you sell into a specific geography. link
• Baseten joined Hugging Face as an inference provider, bringing DeepSeek V4 Flash, Kimi K3, and GLM-5.2 to the Hub with no markup if you bill through HF. link

See you tomorrow.

Pradeep Perugu