NVIDIA bolts BioNeMo onto Claude Science

Good morning 👋 NVIDIA just showed its hand on where it actually wants to sit in the AI stack, and it's not next to the chatbots.

In today's issue:

• 🔭 NVIDIA bolts BioNeMo onto Claude Science
• 🧠 A Lean 4 prover that already caught a real bug
• 🔬 A dumb safety monitor that holds up, and where unlearning actually breaks
• 🛠️ Give your agent a tool to call instead of a guess to make
• 🔗 Robots, a presidential library, and a Bedrock procurement fix

🔭 THE ONE THING

🧬 NVIDIA bolts BioNeMo onto Claude Science

Four days after Anthropic's Claude Science workbench launched, NVIDIA shipped the part that makes it worth a scientist's time: the BioNeMo Agent Toolkit, packaging its GPU-accelerated stack (Parabricks, RAPIDS-singlecell, nvMolKit) as skills a Claude agent can call in plain language. The numbers do the talking: RAPIDS-singlecell takes a 52-minute cell-preprocessing job down to 25 seconds, and nvMolKit claims up to 3,000x on cheminformatics runs, per both companies' posts. NVIDIA says 18 of the top 20 pharma companies already run BioNeMo, so this agent layer is landing on infrastructure that's already doing real drug discovery, not a research toy, and it's open source and harness-agnostic besides, meaning those same skills show up wired into other agent stacks soon enough. NVIDIA isn't trying to win the chatbot war. It's making itself the compute layer every agent has to rent.

🧠 MODELS & RELEASES

• 🧬 NVIDIA built its BioNeMo Agent Toolkit into Claude Science, giving agents direct calls to Parabricks, RAPIDS-singlecell and nvMolKit (that last one claims up to 3,000x on cheminformatics; RAPIDS turns a 52-minute cell analysis into 25 seconds). Eighteen of the top 20 pharma companies already run BioNeMo, so this is Claude landing inside labs that already own the GPUs. NVIDIA

• 🧮 Mistral released Leanstral 1.5, a Lean 4 theorem prover that saturates miniF2F and solves 587 of 672 PutnamBench problems, and it already caught a live integer overflow bug in an open-source zigzag decoder. Apache-2.0, weights on Hugging Face, free API if you'd rather not host it. Mistral

🔬 RESEARCH HIGHLIGHTS

• Online safety monitoring A new paper tests the dumbest possible real-time guardrail: take a verifier model's confidence signal, threshold it, calibrate the threshold with risk-control math, raise an alarm when it crosses. On math-reasoning and red-teaming benchmarks, that simple setup matched fancier sequential-hypothesis-testing monitors. Watch, don't act: it's one paper's benchmarks, but if a threshold this dumb holds up, most teams already running a verifier get a safety monitor for free. Online Safety Monitoring for LLMs

• LACUNA Researchers built a testbed that injects fake PII into specific weights of 1B and 7B OLMo models, then checks whether unlearning actually erased it or just buried it. Most SOTA unlearning methods pass every output-level test and still resurface the data under the right probe. When the fact's location in the weights is pinned down precisely, though, a plain gradient-based method erases it cleanly. The bottleneck isn't the unlearning algorithm. It's finding where the fact lives. LACUNA: A Testbed for Evaluating Localization Precision for LLM Unlearning

• Program-as-Weights Compile a fuzzy spec (alert on important log lines, fix malformed JSON, rank by intent) into a small adapter for a frozen local model instead of paying an API per call. The paper's number: a 0.6B Qwen3 running the compiled program matches Qwen3-32B's direct-prompting quality at roughly a fiftieth of the memory, at 30 tokens/sec on a MacBook M3. Builder read: if that generalizes past their own FuzzyBench eval, it's a real argument for compiling stable logic once instead of renting a big model for it forever. Program-as-Weights: A Programming Paradigm for Fuzzy Functions

🛠️ TRY THIS

Give your agent a real tool to call, not a guess to make

NVIDIA didn't teach Claude biology. It wired BioNeMo's actual functions into it, so the model looks answers up instead of inventing them. The same move works for any narrow, repeated task your team keeps re-explaining to a chatbot.

1. Pick one recurring question where the model currently guesses at something it should be looking up (an internal API, a scoring table, a lookup call).

2. Wrap that function as a tool schema and register it with Claude, so it calls out instead of hallucinating a number.

3. Lock the system prompt to route matching questions straight to the tool, with no "just compute it" fallback.

4. Test with one query that has a verifiably correct answer. If the model's output matches the tool's output exactly, the wiring works. If it doesn't, that's a routing bug, not a model problem.

Prompt: You have access to {{tool_name}}. For any question about {{domain}}, call it and return its output verbatim. Do not calculate or estimate the answer yourself.

Worth a look

• AWS Model Profiler pulls model specs, pricing by consumption type, context windows, quota limits, and lifecycle status from seven separate sources into one searchable index. Useful the moment you're picking which Bedrock model sits behind that tool call. AWS
• Amazon Bedrock's phishing detector skips keyword filters entirely. It scores emails on word choice, style drift, and whether the request fits the sender's baseline, which is the same behavioral-anomaly logic you'll want once your own agents start sending mail on someone's behalf. AWS
• SageMaker's multi-turn RL playbook is the checklist to read before you reward-tune anything that calls tools across more than one turn: trustworthy training environment, external eval, reward tied to the end task, and metrics that catch what changes once the agent runs long. AWS

🔗 QUICK LINKS

• NVIDIA profiled Jaiveer Singh, the engineer behind Isaac ROS, the open-source stack letting robots actually see through their own cameras instead of stumbling around blind. NVIDIA
• Microsoft built an AI exhibit at the Theodore Roosevelt Presidential Library so visitors can "talk" to TR. Cute, and a preview of where museum budgets are headed. Microsoft
• AWS now lets orgs buy Bedrock model access once from a central account and hand it out across sub-accounts, no more Marketplace permissions per workload. Cuts a real procurement headache for anyone running Bedrock at scale. AWS

That's the signal worth your time today.

Pradeep Perugu