353 karma · joined September 8, 2022
Yes that's sort of the point of the approach. Imagine when the board is any position, you have a list of all possible moves which are ranked based on how good they are (i.e., how likely they are to lead to you winning). In order to make this ranking you've "flattened" the future outcomes of that move into a single number which is used to rank the list. Then, you can play optimally without search: simply pick the best move at every board position. (Richard Bellman showed that, for the right definition of "best", this gives the optimal strategy for playing entire games, not just single moves.)
Now the question is, how does one actually calculate this sorted list of the best moves? The straightforward answer is essentially search and various techniques to make the search computationally tractable. What Deepmind's approach is use such a list computed by Stockfish (which explicitly searches) to compile a dataset which is then trained on in order to approximate the function that produces this ranking (called the Q-function). But the point is that search is just an algorithmic strategy to estimate the Q-function, mathematically speaking it is merely a function of the current board state and therefore in theory could be determined without any search. So, maybe a deep neural network could learn how to do this?
So you're estimating from data a function which is itself not necessarily optimal. Moreover, the point is more like how far can we get using a really generic transformer architecture that is not tuned to domain-specific details of our problem, which Stockfish is.
>performing some function of the similarity of the input to those previously seen.
This is indeed what transformers do. But obviously it learns some sort of interpolation/extrapolation which lets it do well on board states/games outside the training set.
Random comments:
>when the states evolve in time and the observables don’t we are using Liouville’s picture; when the observables evolve in time and the states don’t we are using Hamilton’s picture.
I have never heard this terminology, I have only heard Schrodinger's picture vs. Heisenberg's picture.
>This means that, very unlike on a Riemannian manifold, a symplectic manifold has no local geometry, so there’s no symplectic analogue of anything like curvature.
Perhaps the only enlightening comment I have ever heard about the tautological 1-form/symplectic approach to Hamiltonian mechanics.
It's essentially the same as the relation between covariance and contravariance in category theory.
Obviously, a smart watch is strictly better than a regular watch in the sense that it has more features, but I think the downside of having to charge it so regularly makes the value-add not worth it. I am always sad about the downfall of Pebble watches, because the e-ink screens were excellent and gave you a watch that rarely needed to be charged and still gave the benefits of a smart watch (however small they may be).
The only thing that is scary is the hype, because this will make people sloppily use deep learning architectures for problems that do not need that level of expressive power, and because deep learning is challenging and not theoretically well understood, there will be little to no attempts made to ensure safe operation/quality assurance of the implemented solution.
Interesting to me how the top comment has talked about constructing logic gates out of biological circuits. I wonder if anyone has done the opposite, i.e., write a probabilistic programming language whose operations are under the same amount of noise as a cell?
In any case your threat model for container workflow is already 100x more complicated than some one-server blog with an open SSH port, I would give up on fail2ban immediately
Doing anything which involves more than one pipe or xargs is just painful. Bash as a language is not ergonomic at all.
I don't know if there's a nice alternative that keeps the same spirit. My feeling is something like emacs + eshell (a shell which can execute elisp expressions)? Elisp is not a very ergonomic language either, of course... but something like this.
VSCode was a revolution in giving an IDE with a good UI, every feature one could want, and fairly good performance. In 2023, I feel like everyone's on containerized/distributable work flows so there is a big need to actually return to "UNIX as IDE" type work, and now would be a great time for some new CLI tools that improve on the past.