Comment by chriswarbo
2 hours ago
Tests (and type-checkers, linters, formal specs, etc.) ground the model in reality: they show it that it got something wrong (without needing a human in the loop). It's empiricism, "nullius in verba"; the scientific approach, which lead to remarkable advances in a few hundred years; that over a thousand years of ungrounded philosophy couldn't achieve.
The scientific approach is not only or primarily empiricism. We didn't test our way to understanding. The scientific approach starts with a theory that does it's best to explain some phenomenon. Then the theory is criticized by experts. Finally, if it seems to be a promising theory tests are constructed. The tests can help verify the theory but it is the theory that provides the explanation which is the important part. Once we have explanation then we have understanding which allows us to play around with the model to come up with new things, diagnose problems etc.
The scientific approach is theory driven, not test driven. Understanding (and the power that gives us) is the goal.
> The scientific approach starts with a theory that does it's best to explain some phenomenon
At the risk of stretching the analogy, the LLM's internal representation is that theory: gradient-descent has tried to "explain" its input corpus (+ RL fine-tuning), which will likely contain relevant source code, documentation, papers, etc. to our problem.
I'd also say that a piece of software is a theory too (quite literally, if we follow Curry-Howard). A piece of software generated by an LLM is a more-specific, more-explicit subset of its internal NN model.
Tests, and other real CLI interactions, allow the model to find out that it's wrong (~empiricism); compared to going round and round in chain-of-thought (~philosophy).
Of course, test failures don't tell us how to make it actually pass; the same way that unexpected experimental/observational results don't tell us what an appropriate explanation/theory should be (see: Dark matter, dark energy, etc.!)
The theory still emanated from actual observations, didn't it ?
This only holds if you understand what's in the tests, and the tests are realistic. The moment you let the LLM write the tests without understanding them, you may as well just let it write the code directly.
It most certainly is not. All your tests are doing is seeding the context with tokens that increase the probability of tokens related to solving the problem being selected next. One small problem: if the dataset doesn't have sufficiently well-represented answers to the specific problem, no amount of finessing the probability of token selection is going to lead to LLMs solving the problem. The scientific method is grounded in the ability to reason, not probabilistically retrieve random words that are statistically highly correlated with appearing near other words.
This assumes that tests are realistic, which for the most part they are not.