Even perfect AI is limited by experimental timelines, argues physicist Connor Blake in an essay on accelerating science
Even perfect AI is limited by experimental timelines, argues physicist Connor Blake in an essay on accelerating science
In an essay published on November 14, Blake considers what would slow a laboratory down if AI could choose a solution instantly. A reader counters that AI used by suppliers could prevent failures that would take a long time to resolve.
Imagine a factory where the proportion of defective products suddenly rises. In Blake's thought experiment, AI uses instrument readings to link the defect to an impurity in the raw material. Tests must establish which supplier introduced the impurity. The factory then needs to choose a different raw material or a purification system, wait for delivery, install equipment if necessary, and check product quality. Even if the solution is chosen instantly, delivery, installation, and testing take time.
Blake applies Amdahl's law to this process. The 1967 paper addressed the speedup achievable with multiple processors: some operations still have to run sequentially. In Blake's model, the stages occur in sequence, so their durations add up. Once data analysis becomes faster, growing cells may account for a larger share of the cycle; once cell growth becomes faster, the next test may take up that share. In this model, achieving another severalfold speedup in the whole process requires a substantial reduction in the time taken by the slowest remaining stage.
When intelligence becomes free, instantaneous, and almost omniscient, which stage will limit the pace?
Biological experiments conducted under the same conditions produce different results, so further observations are needed to distinguish a consistent effect from random variation. Blake argues that even an ideally chosen experiment may require a long period of data collection because of its duration or the variability of its results. One example he gives is delayed drug side effects in humans: evidence of these emerges through long-term follow-up. When the result of one experiment determines the next step, the wait delays the entire sequence.
In commenter DaemonicSigil's scenario, the supplier also uses AI. The supplier's AI compares raw material requirements with information about a factory being built nearby and asks that factory's AI which particles it will release into the air. The supplier upgrades its filters before the neighboring factory starts operating. “You never even notice the problem,” the commenter writes. AI thus prevents the failure and removes the need to investigate its cause and carry out repairs.
In his reply, Blake offers other scenarios: a hard-to-reach machine part fails prematurely, a supplier goes out of business, or a raw material contains an impurity whose harmful effects were previously unknown. In his view, even an economy with AI in widespread use would still face a long list of failures requiring new information and physical action. Estimating how much a particular laboratory could speed up therefore requires considering how long the necessary experiments take and how much downtime can be prevented.
- https://bosoncutter.substack.com/p/on-the-brutality-of-amdahls-law
- https://www.cs.cmu.edu/~18742/papers/Amdahl1967.pdf
- https://www.lesswrong.com/posts/bMsQ44sKqrFWjuHPb/on-the-brutality-of-amdahl-s-law/comment/sFLy7Z3o3ThiPLkBb
- https://www.lesswrong.com/posts/bMsQ44sKqrFWjuHPb/on-the-brutality-of-amdahl-s-law/comment/4yCvkEpg75Koonejy