Lab Newsletter — July 24, 2026: Open by Design

A quieter but consequential fight is taking shape: not which model is smartest, but whether the AI stack scientists work in is open or owned. This week it got concrete.

🔓 An open science workbench arrives

OpenScience (Synthetic Sciences) is an Apache-2.0, model-agnostic AI workbench for biology, chemistry, physics and ML: a browser workspace on a local runtime that runs the full research loop — literature, hypothesis, code, experiment, analysis, write-up — with 250+ editable skills and 30+ scientific databases (UniProt, PDB, ChEMBL, arXiv) wired in as tools. The pitch is pointed: your keys stay on your machine, and you route each request to whatever model you like (Claude, GPT, Gemini, DeepSeek, or a local fine-tune). It’s explicitly framed as the open answer to Anthropic’s Claude Science. Why it matters for the lab: this is our thesis, shipped by someone else — open, agent-readable, model- swappable tooling that a lab can run on its own infrastructure. It’s exactly what BioEngine and Hypha are built to be.

🏛️ Open vs. proprietary — and open can be frontier

The comparison is the story. OpenScience and Claude Science both run the loop, render science inline, and prize reproducibility; the difference, as the writeup puts it, is “openness and model choice” — any provider vs. Claude-only, your infra vs. lab machines, 250+ editable skills vs. a curated set. And crucially, open doesn’t mean second-rate: the most capable general-purpose biomedical agent of the year, Biomni, shipped in Science fully open, and open image- analysis agents like Agentic-J are proliferating. Why it matters for the lab: the scientific AI stack is genuinely up for grabs right now — and the open path is both viable and winning contributors, which is where a small, open lab can punch above its weight.

🌍 Why the lab bets open

This isn’t only ideology. Model-agnostic, local-data, open-source tooling means data sovereignty (no records leaving your institute), no vendor lock-in, and reproducibility by default — and, as a recent perspective on open biology and the Global South argues, it means access for labs without big budgets. That’s the entire design of our own stack: ImJoy runs in a browser with no install, the BioImage Model Zoo is free and testable in-page, and BioEngine deploys at a facility’s own institute. Why it matters for the lab: openness isn’t a feature we add — it’s the differentiator and the mission. The more the field’s tooling opens up, the more our bet pays off.

The smartest model of the month will change. Whether scientists can own, inspect and run their tools is a more durable question — and this week the open side got a serious new entry.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.