Drug workflows, optimization budgets and simulated worlds

Enveda’s review study grounds Anthropic’s drug-development report; two preprints examine optimization costs, and DeepMind introduces the Genie 4 preview.

3 stories · Yesterday in AI

TL;DR

The short version
  1. Enveda’s assisted review took 166 minutes against 1,553 unaided in a four-record study; known-issue coverage was consolidated across runs and human decisions remained central.
  2. GEPA batching and EvalAlloc address different optimization costs: elapsed time and evaluation allocation. Their separate benchmark gains do not establish combined savings.
  3. Genie 4 adds native sound and scene steering in a trusted-tester preview; continuity limits constrain what a realistic simulation can establish.

Big Moves

1 story

Claude-assisted batch review shows gains and a narrow evidence base

Enveda’s study behind Anthropic’s drug-development report records faster document review, while preserving human decisions and exposing limits in the comparison.

As a leader: Ask for a timed comparison on representative records that measures missed issues, unnecessary comments and the human decision work.

As a practitioner: Preserve the record version, page coverage, model and prompt configuration beside every accepted or rejected finding.

Enveda ↗Vendor documentation

Enterprise Ops

1 story

Two studies separate faster agent optimization from cheaper evaluation

Parallel GEPA proposals and EvalAlloc target different costs of improving agent workflows: elapsed time and the number of evaluations needed to choose a candidate.

As a leader: Separate elapsed time, evaluation executions and dollars when budgeting an agent-optimization experiment.

As a practitioner: Keep a held-out test set and record worker limits, candidate versions and final-selection evaluations.

Sichun Luo and colleagues / arXiv ↗Research preprint

Vendors

1 story

Genie 4 brings native audio to a limited simulation preview

DeepMind’s new world model generates sound and steers interactive scenes, while preview access and continuity limits bound what enterprises can evaluate.

As a leader: Define a bounded simulation task and measurable continuity failures before committing a workflow to a preview.

As a practitioner: Check whether objects, instructions and scene state persist through navigation and edits in any trial you obtain.

Google DeepMind ↗Vendor documentation