2 ms·
That's not true even before mixture of experts because the systems are stochastic. They are only identical if we were to always initialize with the same startin
by godelski 3y ago
That's not true even before mixture of experts because the systems are stochastic. They are only identical if we were to always initialize with the same starting seed and rng_state.
- ben_w 3y agoWhile they do indeed have a random number generator, I think the main strengths and weaknesses of these models (and humans) are in the structures rather than at the token temperature stage — I expect a human to answer any given question slightly differently at different times, but asking the same doctor for a second opinion as the first is considered the setup for a joke rather than a sensible idea.
- godelski 3y agoThis is the assumption of a flat and smooth latent manifold. If either of those conditions are not met, then seeds can have a wild effect on the model output. There's only bounds on the latent structure being locally smooth or locally flat btw. This is the underpinning of such works as the Lottery Hypothesis but there's a whole rabbit hole to dig down here and it is why Alpha Go did so poorly in round 2. This is an important part of the discussion of generalizability btw. I do not think your analogy works out and it is inconsistent with your prior comment, because you're right that anyone asking GPT is akin to anyone asking the same doctor. The second doctor is a second model (or potentially a second GPT checkpoint depending on how we want to consider divergence).