Quantum computing (QC) and AI have one thing in common: they make mistakes.
There are two keys to dealing with mistakes in QC: We have made tremendous progress in error correction in the last year. And QC focuses on problems where the solution is extremely difficult to formulate, but easy to verify. Think about factoring 2048-bit prime numbers (about 600 decimal digits). This is a problem that would take years on a classical computer, but a quantum computer could solve it instantly – with a significant probability of the wrong answer. So you have to test the result by multiplying the factors to see if you get the original number. Multiply two 1024-bit numbers? Easy, very easy for modern classical computers. And if the answer is wrong, the quantum computer tries again.
One problem with AI is that we often incorporate it into applications where verification is difficult. Tim Bray recently Reading His AI-generated biography on Grow Wikipedia. There were some major errors, but there were also many subtle errors that no one could identify except him. We’ve all done the same with one chat service or another, and all had similar results. Worse, some of the sources referenced in the biography to verify claims actually “fail completely to support the text” – a well-known problem with LLMs.
Andrej Karpathy Recently Proposed A definition of Software 2.0 (AI) that puts verification at the center. He writes: “In this new programming paradigm, the newest predictive characteristic to look for is verifiability. If a task/job is verifiable, it is adaptable either directly or through reinforcement learning, and a neural network can be trained to do the job very well.” This formulation is conceptually similar to quantum computing, although in most cases verification for AI will be much more difficult than verification for quantum computers. The little facts from Tim Bray’s life are verifiable, but what do they mean? Does the verification system require Tim to contact Tim to verify details before authorizing a biography? Or does it mean that this kind of work should not be done by AI? Although the EU’s AI Act Lays the foundation for what AI applications should and shouldn’t doWe’ve never had anything that was easily, well, “calculable.” Also: In quantum computing it is clear that if a machine fails to produce the correct output, it is okay to try again. The same will be true for AI; If you ask the question again, we already know that all interesting models give different outputs. We should not underestimate the difficulty of verification, which may prove to be more difficult than even the LLM training.
Karpathy’s focus on verifiability is a huge step forward, despite the difficulty of verification. Then from Karpathy: “The more verifiable a task/task is, the more amenable it is to automation… This is what is driving the ‘complex’ limit of progress in LLM.”
What differentiates it from Software 1.0 is simple:
Software 1.0 easily automates what you can specify.
Software 2.0 easily automates what you can verify.
This is the challenge Karpathy poses for AI developers: determine what is verifiable and how to verify it. Quantum computing is easily overkill because we only have a few algorithms that solve straightforward problems like factoring large numbers. Validation for AI will not be easy, but it will be necessary as we move into the future.
