A growing number of people building AI for science describe the same future. AI becomes as fundamental to the laboratory as the microscope or the computer: not a tool scientists reach for occasionally, but an intelligence layer connecting observations, experiments, instruments, and decisions. In that future, science itself becomes a continuously learning system.
I share that vision, and continuous learning is the part I care about most. But it raises a question I think deserves more airtime: who is doing the learning?
The loop closes around the scientist
In most pictures of the AI-enabled lab, the learning loop closes around the scientist. The AI watches the experiment, the scientist gets better guidance, and the next run improves. That is real progress, and I want it to happen.
It also leaves someone out. Every experiment runs on tools that somebody else designed: the kit on the bench, the reagent in the tube, the instrument on the counter. If the insight from each run stays with the scientist or the lab, those tools never get better. The next lab starts from the same instructions and hits the same wall.
What I saw from the manufacturer side
I spent more than a decade building life science tools, first at QIAGEN and Bio-Rad and then as a co-founder of Fluent BioSciences, which Illumina acquired in 2024. The pattern I saw again and again was that the hardest part was not the invention. It was helping customers execute correctly once the product was in their hands.
We built tools we were proud of, handed them to talented scientists, and then watched them struggle to use them the way we intended. At Fluent, we could send our own scientists to stand next to customers. At Illumina scale, that simply doesn't work. Once a kit ships, it disappears into a black box.
What comes back is thin. Support tickets capture the loud failures. Surveys arrive months later and reflect what people remember, not what they did. Field application scientists are the gold standard, but they can't be at every bench. The everyday runs that only partly work, the quiet workarounds, and the lab that tries a kit twice and then switches vendors never reach the people who could fix them.
A learning system has two loops
If science is going to become a continuously learning system, I think it needs two loops running at once.
Two learning loops
- The scientist's loop. Guidance through each step, answers without stopping work, and an accurate record of what actually happened, which methods sections have never captured.
- The tool maker's loop. Which steps raise questions, where people pause, where they deviate, and which runs succeed, fed back into a clearer protocol, a more forgiving reagent, or a better next-generation instrument.
Most of the current excitement is about the first loop. The second matters just as much. A better-guided scientist gets one experiment right. A better tool gets thousands of experiments right, in labs its maker will never visit.
Science moves at the speed of its tools
None of this takes anything away from that vision. It extends it. The bench has always been where science happens. With the right records, it can also be where the tools learn, and the distance between a problem at the bench and a fix in the box gets shorter with every run.
Science advances faster when the tools get better, faster. That is the half of the loop I want to help build.
Full disclosure: I founded Adept Scientific, which makes Kavia, a voice copilot that guides scientists through protocols, and Spotlight, which shows tool makers how their products are used at the bench.