Ramez Naam linked the speed of AI progress in biology to how organisms respond to interventions
Ramez Naam linked the speed of AI progress in biology to how organisms respond to interventions
On 26 August, Ramez Naam discussed Dario Amodei’s prediction in a lecture: powerful AI could compress the next 100 years of progress in biology into 5–10 years. Naam argues that the pace depends on how quickly an experiment produces an unambiguous result that reflects the actual goal.
Naam proposes evaluating a task by four criteria: whether success can be distinguished reliably from failure; whether enough data are available; whether feedback arrives at computer speed; and whether the measured outcome matches the actual goal.
In Go, the outcome is available immediately: the game ends, and the winner is known. Verification in formal mathematics works in much the same way. In programming, a test shows how code behaves in a specified scenario, although the programmer’s intent may extend beyond what the test covers. A rapid result still makes it possible to adjust the next attempt immediately.
In the search for ways to slow aging, an experiment in cells is only the first stage. Researchers must then determine whether the effect also occurs in animals and humans. Naam describes the gap as follows:
“What we observe in cells is different from what we find in an organism.”
A cell that survives longer in a dish provides evidence only about cells. Determining whether an intervention extends human life requires testing its effect in the whole organism. The measured outcome itself changes as research moves from cells to animals and humans: a system may improve an early indicator, while the central question concerns an outcome in the whole organism.
In his essay “Machines of Loving Grace”, Amodei describes AI as a virtual biologist that plans and conducts experiments, proposes ways to measure their results, and directs laboratory robots or human researchers. Amodei also accounts for the time required to perform the experiments themselves:
“Experiments and equipment design involve an unavoidable fixed delay.”
Amodei expects careful experiment selection and large scale parallelism to help researchers continue searching for solutions while individual experiments run their course. Naam raises a further issue: how quickly a result in cells can yield an answer about humans. If this transition requires prolonged observation of an organism, that observation sets the pace of learning.
In a review of AI agents for biological laboratories, the cycle worked as follows: a model reviews the literature, proposes a hypothesis, and passes it through screening and experimental testing. Naam adds a condition that determines the speed of this cycle: the experiment must return a signal that is linked to the desired outcome in humans.