Self-improving modular accelerators

The idea behind self-accelerating agents for science is that you train and deploy AI that can learn the scientific method of hypothesizing, testing, and learning. It can then build enough prior and methodology knowledge to iterate fast and explore vast design spaces in a short time. The goal is for this to reach new discoveries, products, or novel practices owing to its ability to A) reason in high dimension systems and B) exploit them fast.

There are two ways people are approaching this:

1. The self-accelerating stack with broad industry focus. It learns on the job, trains its own machine learning model, understands the workflows and adapts with time to become the company's brain. Examples include Mirendil and recently DiscoveryLoop.

2. The focused approach where companies build novel datasets and harnesses for specific industries targeting certain workflows and embedding themselves deeply into the design process. Examples include LILA (materials), Periodic (semiconductors), or Cradle (protein design).

There has been much more funding per company and frontier lab and big tech veterans flowing into the first.

I am much more bullish on the second. Taking a bet on the first presumes that the autonomously accelerating AI stack will know what the endpoints are in an organization and solve for them correctly. It will become aware of its role and the tasks it has to take to converge to the right solution. I believe that this AI is "un-post-trainable" because, in technical terms, the number and breadth of factors, uncertainty, and noise playing into the long-horizon context are too ambiguous and too complicated to reason about, compress, and use as a form of input to generalize from.

The second approach proposes a well demarcated scope with the human-in-command. It decides the exact role that the AI should play in the engineering and design problem, and the fairly tractable context it should exist in. The endpoints here are much cleaner. Cradle-1, for example, helps scientists design proteins and optimize leads. The user knows exactly what they want from it. Continual learning lives on a very clear feedback paradigm with a defined verification environment where the output is "directly" measurable.

I like to think about AI as a modular accelerator that helps you converge to a reachable endpoint, then make the output accessible to other AI modules hitting at other endpoints. The silos share context and move together, orchestrated by a human vs. an autonomous supermodel that we expect to ICL-itself into broad intelligence.