No. 004

What AI Can (and Can't) Do

• Human scientists trounce the best AI agents on complex tasks

  Nature, April 13 2026

  The 2026 Stanford HAI AI Index finds that frontier AI agents score roughly half as well as PhD-level experts on complex, multi-step scientific workflows. Meanwhile, 6-9% of natural-science publications now mention AI, and AI-adopting researchers publish 3x more and receive 5x more citations. The tension between individual acceleration and the agents' actual ceiling on hard tasks is the report's most useful finding for working scientists.

• MirrorCode: preliminary results on AI software reimplementation

  Epoch AI, April 10 2026

  A new benchmark from Epoch AI and METR tests whether AI models can reimplement existing software from specifications alone, without seeing source code. Claude Opus 4.6 successfully recreated gotree, a 16,000-line bioinformatics toolkit that would take a skilled human 2-17 weeks. The catch: performance depends on having a detailed, checkable specification, which most real-world scientific software does not have.

• Harnessing AI to build virtual cells

  bioRxiv, April 14 2026

  VCHarness is an autonomous AI system that constructs perturbation-response models by pairing a coding agent with multimodal biological foundation models. Across benchmarks, it identifies architectures that outperform expert-designed approaches and reduces development time from months to days. The system also surfaces non-obvious architectural patterns, suggesting automated search can move beyond replicating known design strategies.

Tools & Infrastructure

• Article Galaxy MCP: your AI now talks to your literature library

  Research Solutions, April 7 2026

  Research Solutions shipped an MCP server that lets AI agents search scientific literature, check institutional access rights, and acquire full-text articles without leaving the AI environment. It plugs into Claude, ChatGPT, Copilot, and any MCP-compatible tool. The practical upside: a literature search that currently requires switching between three or four systems can happen in a single agent session.

• China builds its largest scientific AI computing cluster

  CGTN, April 14 2026

  A 60,000-chip AI computing cluster in Zhengzhou reached full operational status, hosting over 1,000 open-weight models. Researchers submit requests in natural language and the system allocates compute and invokes models end-to-end. Whether the "natural-language submission" interface works as described remains to be seen, but the scale of dedicated scientific AI compute is notable.

Scholarly Publishing

• The journal article is not the job

  The Scholarly Kitchen, April 15 2026

  Ashutosh Ghildiyal argues that as AI automates parts of writing and increases the volume of available knowledge, the constraint shifts from access to reliability and relevance. Publishers who stop at producing articles are missing the point; the real work is making knowledge trusted, discoverable, and connected enough to produce downstream impact.

• Exploring data spaces in scholarly communications

  The Scholarly Kitchen, April 17 2026

  A primer on data spaces for scholarly publishing: federated infrastructure for controlling and auditing how data flows between organizations. The immediate relevance for researchers is that as AI tools consume and reprocess scholarly content at scale, machine-actionable data governance is becoming the mechanism by which institutions decide what their content can be used for.