408 karma · joined November 27, 2022
https://www.lesswrong.com/posts/jGuXSZgv6qfdhMCuJ/refusal-in...
I uploaded a picture of a forest and it came back with visually similar images. So the first thing it might be doing is some kind of KNN, and if the pictures have location labels associated applying some sort of weighted average to determine GPS coordinates. This is pretty cool.
I also tried flipping the image horizontally, and it came back with the same images. So their embedding isn't based off of exact matches (good) and seems to be invariant to some basic translations (good). It also seems like it's directly extracting visual features from the image. This can be done with something like Blip[0].
Then I uploaded a screenshot from Magic School Bus. It still extracted information to guess the "location" of the cartoon (San Francisco, which is wrong). So that's probably how it works.
I also found the text output is similar in some ways with OpenGVLab InternViT [1]. So perhaps this or something like it is being used to extract features.
And of course there may be an LLM on top of these extracted features with some sort of prompt template. But I should add that the text explanation is the least useful part of the result, since it is unreliable and less informative than the "boring" similarity metrics above.
[0] https://huggingface.co/Salesforce/blip-image-captioning-larg... [1] https://internvl.opengvlab.com/
But if it's all local? It would be a great product. No internet unless I turn on a physical switch, no ads, no monthly fee. Just do the thing I want you to do. Maybe at the moment it will still need to talk to my GPU. But we're getting there.
> So what is a blunder for a strong player may merely be a mistake for a weaker player.
Statistically this intuition appears to be correct. Your winning probability is still more than 25% when down a queen against an 800 rated player, but under 10% if playing a 2200: https://web.chessdigits.com/articles/when-should-you-resign#...
So it would make sense for the definition to take into account the opponent's Elo rating.
What this means in practice would be something like an "anti-tactic". In normal tactics, we need to find the best move which will usually win the game or give a decisive advantage. But an anti-tactic is a fight with your intuition: you want to make a move but that move is actually a blunder. There are several correct answers but one or two attractive but very wrong answers.
Of course the "rest of the owl" is how to determine which moves are very attractive to certain players. That is something I am working on now.
Maybe you could compare it to playing chess against a much higher rated player. In that sense, minimizing your weaknesses is still useful. And looking at the "blunders" the author lists we can definitely say that making these errors is worse than not making them.
One place where this analogy breaks down is that chess is a zero-sum perfect information game. Business is neither zero-sum (competition can encourage innovation and better products) nor perfect information. In chess, we can calculate for each of the candidate moves and have theoretically perfect knowledge of the future state of the board. In business and many other situations, this kind of thing isn't possible. The upshot is that several of the "blunders" the author lists may only turn out to be blunders after the fact.
As far as the wiki link, it's a reference for the code sample I use in the demo.
Convert your code to an image from the command line. No internet connection needed, although it can also download and convert code from Github too. Made it because currently the best alternative (carbon-now-cli) sends your code over the wire and there's a lot of situations where that's not desirable.
I usually write about machine learning & statistics topics as well as chess AI. I'm most proud of the search bar, which is all performed client-side so it's lightning fast.
> positions where the engine may consider the position to be almost equal, but it could be objectively worse for one of the players
That is a good point, centipawns aren't the best metric for the subjective experience of playing. Engines are really good at defending, finding the "only move" defense where a human will struggle. And the mental work of finding "only moves" is pretty taxing. If you put someone under that kind of pressure for long enough, they will break.
The analysis is also flawed because it doesn't take into account the fact that the opponent is also playing a word. If I know that my opponent must play a "G" move on his next turn, I should play a move to take advantage of that fact. Stockfish's game tree is not eliminating impossible moves and will tend to overrate the opponent's position.
(As a concrete example, suppose I can make a move M1 where the opponent's only move to not lose immediately is an F move, but if the opponent plays it they are at an advantage. In this situation the opponent must play a G move, a fact which a human might be able to guess. Stockfish will not play M1, but instead play a different move M2 that at least maintains equality. So the stockfish evaluation is pretty flawed.)
> Researcher Rodney Brooks complains: "Every time we figure out a piece of it, it stops being magical; we say, 'Oh, that's just a computation.'"
Autopilot might be safer, but this statistic is just a textbook example of selection bias. In order to compare one to the other, crashes in places autopilot doesn't work need to be excluded.
I think if the purpose of the competition is to apply 21st century skills then they shouldn't ban 21st century tools.
My issue is that I am working on a project (I wouldn't call it a startup exactly) which does directly involve AI, so I attract attention from the exact types of people you mentioned, fair weather fans. If you have crypto in your profile and are now interested in AI, that tells me all I need to know.
I've been able to scare most of them away by not having a functioning landing page and directing them to a repo instead. Since I am looking for another technical person this has been a great filter.
60 seconds to sort 400 tabs is a hell of a price to pay. It needs to be worth it. Do you really need semantic search? Do you really need semantic search that queries a huge proprietary model?
(A really cool feature that Chrome doesn't have would be the ability to search the content of the webpage as well.)
What about something simpler, both architecturally and computationally. Start with something boring, like BM25. In my experience it works surprisingly well. If that's not good enough, then running a free model from Huggingface would be the next best option.
The benefit of this approach is that you can do it in the browser, in real-time (or at least locally with a downloaded model). Even 400 tabs will be a piece of cake. Plus you won't need to proxy precious API calls to a service which may become unavailable or very expensive with no warning.