DeepMind's AI Beat Its Human Research Partner 2-to-0 on Drug Candidates. DeepMind Says That's the Problem.

Co-Scientist proposed 3 drug candidates to slow liver fibrosis. 2 worked in live-cell assays. Its human research partner proposed 2 of his own. 0 worked. DeepMind's public-policy team used that exact result this week as evidence for an essay warning that AI hypothesis generation has outrun the world's capacity to test it — a problem the authors name outright: "proof indigestion." A win becomes a warning once you can't verify the next thousand of them fast enough. This week's other releases put hard numbers on how fast that gap is opening.

DeepMind's own admission: the hypothesis side is winning

"Conjecture Machines: AI Agents and the New Validation Bottleneck in Science" (Google DeepMind, published this week, by Don Wallace, Conor Griffin, Sean O'Neill, Thang Luong, and Owen Larter) argues that AI agents have gotten dramatically better at generating hypotheses and candidate solutions, while the experimental capacity to test them — wet labs, peer review, physical assays — hasn't moved. "Refutations remain physical and institutional," the authors write, "and so, costly and slow. We are moving toward a future of serious 'proof indigestion.'" Their sharpest line: "the validation gap in most disciplines is widening, not closing."

The essay's own supporting evidence makes the point. Co-Scientist matched a decade-long superbug-research hypothesis in 2 days; on drug repurposing for liver fibrosis, researcher Gary Peltz's own 2 candidate picks showed no benefit in assays, while 2 of Co-Scientist's 3 candidates did. DeepMind's Aletheia math agent is credited with solving 6 of 10 unpublished problems in the February 2026 First Proof challenge. None of that is new this week — what's new is DeepMind citing its own results as evidence that the bottleneck moved. The essay points to existing national programs as a sense of scale for what closing that gap costs: $100 million from the US NSF for a distributed automated-lab network, and £81 million for the UK's Materials Innovation Factory. Both are rounding errors next to what frontier labs spend on compute.

NVIDIA: 71 points in a research loop that ran for days, not months

NVIDIA's "How to Run an Autoresearch Workflow with RL Agent Skills and NVIDIA NeMo" (July 14, Vinh Nguyen) walks through a Codex (GPT-5.5) agent that autonomously designed and ran reinforcement-learning experiments using NeMo RL and NeMo Gym, no human in the training loop. On a Qwen3-VL-2B-Instruct star-counting eval, the agent-run loop pushed accuracy from 25.0% to 96.875% in a matter of days. NVIDIA also reports its reimplementation of OAPL (an off-policy RL method) beat a DAPO baseline on Qwen3-1.7B with fewer training steps, though it doesn't publish the head-to-head numbers. To its credit, the post names its own failure modes — context drift, steering drift, and "low-signal loops" where the agent gets stuck optimizing against its own eval — and states plainly that "agents are not replacing researchers." It's a good-faith writeup of exactly the kind of number DeepMind's essay is warning about: a 71-point jump, agent-generated and agent-graded, that nobody outside NVIDIA has independently checked for whether it generalizes past the eval it was scored on.

What a "158 out of 158" replication run actually hides

"Coding-agents can replicate scientific machine learning papers" (arXiv 2607.02134, Atharva Hans and Ilias Bilionis, Purdue, posted July 2 and revised July 10) ran a coding-agent workflow across twelve independent attempts to replicate the computational claims of four SciML papers. All twelve runs passed the completion gate, and all 158 recorded targets were matched with evidence in the replication report. But the authors are explicit about what that pass rate doesn't tell you: "repeated runs differ in how papers are divided into targets, in numerical fidelity to the source papers, in elapsed replication time, in the number of intermediate executions replaced before final evidence is accepted, and in the rules used to accept evidence." Their fix isn't a better score — it's making completion depend on "workspace evidence and validation checks rather than on the agent's final message." A perfect aggregate number, kept without the underlying run, is exactly the kind of conjecture DeepMind says nobody can afford to keep re-verifying by hand.

Also new this week

Elicit shipped an API and MCP server (July 15) exposing semantic search across 138 million papers and ClinicalTrials.gov records, plus a Systematic Review endpoint Elicit reports hitting 95% search recall, 97% abstract-screening accuracy, 99% full-text-screening accuracy, and 96% extraction accuracy across 994 Cochrane reviews. The MCP server works from Claude, ChatGPT, and other MCP-compatible clients, on Pro and higher plans. Separately, llama.cpp release b10032 (July 15) shipped a CUDA kernel for the "lightning indexer" op that powers DeepSeek-style sparse-attention long-context models, and relaxed the GPU requirement down to Turing-generation hardware — meaning that model family's long-context mode now runs locally on a wider range of machines, including older ones.

What this means for reproducible, local-first science

DeepMind is arguing, in public, that its own field's bottleneck has moved from generating hypotheses to checking them. NVIDIA's post is a live demonstration of how fast an agent-graded score can move without anyone outside the loop confirming it holds up. Purdue's paper shows that even a "100% pass" replication run hides real run-to-run variance unless someone keeps the underlying evidence, not just the final tally. None of that gets fixed by a bigger model or a longer training run. It gets fixed by keeping the receipts every run leaves behind — exportable, and checkable by something other than the agent that produced them.