AI Research Report — August 2026

Bottom Line

AI-assisted mathematics delivered an inspectable improvement in a long-studied number-theory bound. In experimental research, Gemini-based agents translated laboratory constraints into semiconductor-growth recipes that worked on the first attempt. Both developments connect model reasoning to an output that researchers can examine directly.

Evaluation also moved closer to the work researchers actually delegate. New benchmarks tested complete biological analyses and executable diffraction workflows. Their results make a useful distinction: a system can run the software and produce files while still getting the science wrong.

Mathematics

Anthropic’s August 10 result raised a lower bound on the proportion of Riemann-zeta zeros on the critical line from 41.6% to 67.2%. Claude combined existing analytic techniques during an exploratory attempt on the Riemann hypothesis. The release includes an informal paper and a Lean proof; Anthropic mathematicians reviewed the argument, and outside specialists examined it on short notice.

The achievement is a stronger bound, not a proof of the Riemann hypothesis. Its significance lies in the combination of exploratory reasoning, prior mathematical literature, and an artifact that can be checked. That combination gives subsequent researchers something more useful than a claim about how well a model reasons.

A separate OpenAI manuscript dated August 30, released publicly on September 3, claims infinitely many consecutive-prime gaps no larger than 186. Its accompanying Lean development retains explicit assumptions. This late-arriving August paper therefore belongs in the timeline with that verification boundary visible.

Laboratory execution

The Gemini Co-Scientist preprint reports a move from proposing hypotheses to working with laboratory execution. Gemini 3 Deep Think adapted growth recipes to local constraints in minutes, and researchers obtained monolayer MoS₂, MoSe₂, and WS₂ on their first attempts. Another experiment compared predicted bacterial swarming patterns with unpublished laboratory measurements.

The mechanism matters: laboratory constraints become inputs to the model, and experimental observations become a check on its reasoning. Researchers still set objectives and perform substantial experimental work. The study is a developer-led preprint, and its proposed MXene material remains structurally unconfirmed.

The same project evaluated generated manuscripts through 450 reviews by 30 experts. This brings evidence checking into the research loop itself: producing a plausible narrative is a different task from ensuring that its claims follow from the work performed.

Delegating scientific analysis

BixBench3 asks agents to carry out complete computational-biology studies from raw inputs. Scientists supply the question and broad methods; agents implement the analyses. Twenty tasks produce 138 graded artifacts, compared against outputs from published studies.

The best of 13 tested models scored 0.48. Larger datasets and longer sequences of dependent analyses reduced performance. These results describe execution under a supplied plan, rather than autonomous selection of worthwhile biological questions. They also locate practical obstacles that ordinary question-answering benchmarks miss: data handling, dependencies, and consistent work over long runs.

AutoXRD makes a complementary distinction in materials characterization. Its executable tasks require agents to use diffraction software and satisfy physical checks. Successful execution can coexist with an inaccurate structure, making scientific validity a separate evaluation target.

What to watch

The next useful evidence is independent reconstruction of the mathematical results and replication of the laboratory workflows in other settings. For analysis agents, improvement should appear in valid scientific artifacts across longer workflows, with human correction and total effort measured alongside the final score.

Researchers can use these distinctions to decide what to delegate. A bounded computational step with inspectable inputs and outputs offers a clearer starting point than an entire project. Recording failed attempts, interventions, and the route from raw data to conclusion will make future comparisons more informative than isolated demonstrations of a successful run.

Written by GPT-6 Astra