I am sharing Citrini's post on X because realistic expectations about AI in biology are too easily mistaken for a dismissal of its value.
Citrini's purpose is to temper expectations for how quickly AI can deliver medical breakthroughs. Experiments and human observation take time, even when models improve rapidly. They remain optimistic that AI will help gather useful data and shorten the path to treatments.
My reason for highlighting this is that often people dismiss AI as a research accelerator because it will take years before we have robust, accurate, high fidelity robotic wet labs automating investigation, testing, discovery and creation in the physical world. However, today we are able to automate and accelerate planning, analysis, results filtering, and coordination that will increase the pace of laboratory work and discovery.
Biology has its own clock
In software, an AI can often propose a change, run tests, and revise it within minutes. In biology, feedback may require making a molecule, running an assay, or following patients. Faster proposals do not automatically make those steps equally fast.
If a question requires six months of human follow-up, three months does not provide the same evidence. The FDA's clinical research overview explains why human studies remain necessary after preclinical work: researchers need to establish safety, dosage, effectiveness, and adverse effects.
Organizational efficiency still matters
Even if making biology itself 100 times faster proves unrealistic, I would still expect substantial gains from standard AI automation, LLM assistance, and generative AI. Much of that opportunity involves helping people organize and carry out research more efficiently.
Consider the work around an experiment: reviewing literature, drafting plans, coordinating tasks, preparing documentation, writing analysis code, filtering results, and identifying findings that deserve closer attention. These are practical places to apply tools that assist with language, information, and routine workflows.
Making some of those tasks five or ten times more efficient seems a reasonable ambition. Those multipliers are my expectation for selected workflows, rather than a measured claim about an entire laboratory or the time needed to develop a treatment. Researchers would still need to check the outputs.
That distinction does not diminish the opportunity. Less time preparing reports, searching through results, or coordinating handoffs could leave more time for scientific judgment. Better planning and filtering could also help teams choose more informative experiments and avoid repeating work. An experiment may still take weeks, but the planning before it and decisions after it need not be slowed by administrative work or fragmented information.
Improve the whole research cycle
Citrini's follow-up is optimistic about making useful biological data easier to gather. I would extend that optimism to the organization around gathering and using it. Better models, better measurements, and better everyday workflows can reinforce one another.
We should be realistic about the time needed to observe biological outcomes while remaining ambitious about removing avoidable delays. AI does not need to deliver 100x faster biology for improvements in planning, execution, and analysis to make a substantial difference.