Ali Asif

The AI log, frontier · agents and coding · science

Claude's black hole proofs, Gemini Tool Retrieval, Theorizer

Anthropic reports two physics results from Claude, Gemini loads tool schemas on demand, Ai2 opens Theorizer, and NVIDIA targets RL rollouts.

By , published

What shipped

  1. Anthropic

    Claude solves two open problems in the physics of black holes

    Two quantum gravity problems, one with a proof formally checked in Lean and reviewed by a physicist, after about six weeks of background agent work.

  2. Google

    Introducing Tool Retrieval in the Gemini API

    A preview feature that holds back tool parameter schemas until the model asks for them, with prefix caching intact and MCP servers supported.

  3. Ai2

    Theorizer comes to Asta

    An agent that reads papers and proposes explanations linked to their sources, available on the web and as an open-source library.

  4. NVIDIA

    Reducing Rollout Bottlenecks in Reinforcement Learning with NVIDIA Dynamo

    Dynamo routing, weight delta transfers and worker recovery for RL rollouts, with adopter-reported gains of up to 15x on weight updates.

Four items today, all from pages I opened and read. I have not run any of the tools, so the numbers below are what each source reports.

Claude on two open black hole problems

Anthropic describes two open problems in quantum gravity that Claude worked on. A researcher, Hongbin Chen, gave Claude a list of open problems. Claude proposed 22 candidates, and separate agents worked on them in the background for about six weeks.

The first is the Hawking-Page transition in weakly coupled N=4 super Yang-Mills. A 2003 calculation was never finished because the answer hinged on one number three loops in, which needed many thousands of terms. Almost every piece was done twice by independent agents and parts were checked against published results. The page reports that the transition is abrupt.

The second is a bound on the mass of the lightest black hole in 3D gravity using the modular bootstrap. Claude proved an exact ceiling with Λ of about 9.1266 and showed it is optimal for this version of the method. The page says the result was verified in Lean and that a physicist checked every main result. It also says an earlier agent failed on this problem in 61 parallel attempts.

Tool Retrieval in the Gemini API

Tool Retrieval is a preview in the Interactions API. You add a tool_search tool and set defer_loading: true on tools to hold back. The model sees names and descriptions, requests a tool by name, and the API loads the schema in the same request. Because deferred schemas stay out of the start of the prompt, prefix caching still works across turns. Remote MCP servers are supported, and you can run the search on the client if you want your own registry or permission filter, at the cost of one round trip per search.

There is no separate charge, but each loaded schema is billed as input for the rest of the conversation. The page gives no benchmark figures.

Theorizer in Asta

Ai2 added Theorizer to Asta. It searches the literature for a question, extracts findings from each paper, and proposes theories linked to the papers behind them. Ai2 says it can miss relevant studies and misjudge originality, and that researchers still decide what is worth testing. One example is a neuroscience query that returned 50 candidate theories where the researcher expected one or two. It runs at asta.allen.ai and the code is open source at github.com/allenai/asta-theorizer.

NVIDIA Dynamo for RL rollouts

NVIDIA describes Dynamo as a serving layer for rollout generation in reinforcement learning. The router balances KV cache reuse against worker load, ModelExpress copies weights between GPUs and can send only deltas, and replacement workers can catch up from a checkpoint. The numbers are adopter reports: CoreWeave cites up to 15x faster weight updates, and H Company cites up to 20% shorter training steps. The post gives no benchmarks of its own.

What I'd watch next

Whether Google publishes token savings for Tool Retrieval, and whether Anthropic's multi-agent checking pattern shows up in other fields. I also want to see independent numbers for Dynamo outside the adopters.

Sources

Topics: Anthropic, NVIDIA, Gemini, Google, Claude, Ai2, Dynamo, Reinforcement Learning