Partnership with Georgia Tech on Project SPEED DIAL
Machines and AI can now design algorithms, and tools like Agda and Lean can be used for assistance with proofs (especially for maths related to Homotopy Type Theory). Over the past two years the evidence has accumulated steadily. Researchers have evolved Kalman filter variants that outperform the standard filter in conditions where its assumptions no longer hold. Others have discovered iterative methods for linear algebra rather than deriving them by hand. Meta solvers have been found that outrun conventional iteration on nonlinear partial differential equations. One system identified a way to multiply four by four complex matrices in forty eight scalar operations, improving on a benchmark that had stood since 1969.
Very little of this work has reached production use.
The common explanation is that the field is young and the tools need another development cycle. I believe that reading is mistaken. The tools function. What is absent is any reason for a practicing engineer to trust what comes out of them.
Consider the position of that engineer. You are responsible for a computational fluid dynamics workflow supporting a vehicle that must survive reentry. Someone presents you with a preconditioner that a search process produced overnight. Would you place it in the loop? Three conditions would have to be met first. You would need to read the method and understand its behavior. You would need to know where it fails, not merely that it performed well on the cases it was scored against. And you would need to see it measured against the method it replaces, using data the search process never encountered.
Each of those is a verification problem. None is a discovery problem.
That distinction determines what is worth building. If discovery is the difficult part, investment belongs in search. If verification is the difficult part, search becomes a component that can be replaced as the field advances, and the lasting assets lie elsewhere. They lie in the schema that converts an imprecise engineering complaint into a well posed search, the harness that evaluates a candidate honestly, the provenance record establishing where a piece of code originated, and the index that surfaces the right prior solution when a new problem arrives. Those outlast any particular search method. A platform committed to a single paradigm will be obsolete before it ships.
There is a more demanding version of this standard. The measure that matters is not whether a discovered algorithm runs quickly. It is whether someone who did not build the system can obtain a real improvement from it. That is a considerably higher bar, and it is the correct one, because a tool that only its author can operate is a research artifact rather than a capability.
The failure modes are specific and deserve to be named. A discovered solver can prove numerically unstable in a region no one sampled, and a speedup figure will never expose it. A search guided by a language model can reproduce a solution memorized during training and score it as novel. Background search has no natural stopping point and will consume a compute budget without necessarily producing anything admissible. Each of these requires a policy established before the first run rather than a caveat appended at the end.
None of this argues against automated discovery. It is an argument about sequencing. The discovery half is close to settled and growing cheaper. The trust half has barely been addressed, and it is what stands between a strong body of research and anything that reaches an engineer's desk. This is what we intend to do with Project SPEED DIAL.
Comments