Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

The Bitter Lesson (2019) [pdf]

cs.utexas.edu

Post reply on HN
Print view View original

31–40 of 62 posts

Re: The Bitter Lesson (2019) [pdf]

#31

Post by reacweb »

isaacfrond wrote: post #19
YeGoblynQueenne wrote: post #16

Earlier quoted context omitted.

>> Go has been described as the drosophila of AI. John McCarthy described chess as "the drosophila of AI": http://jmc.stanford.edu/articles/drosophila/drosophila.pdf Who was it that described Go as the drosophila of AI?

It was John McCarthy, in the very paper that you cite! See section 7: As a fourth Drosophila I would like to mention the research on Computer Go.

I finished my studies in 1993. Artificial intelligence was my major and I had done a study on programming the game of Go (a review of research state). At that time, the success of brute force for chess had caused much disillusionment in the AI world, which had hoped to see a drosophila. The transfer of this hope to the game of go was present in numerous publications. Even if it doesn't show up very well in a Google search, there are plenty of allusions in articles about programming the game of Go. Since the creation of go language, it is more difficult to find.

Re: The Bitter Lesson (2019) [pdf]

#32

Post by the8472 »

niemandhier wrote: post #8

I think the Bitter Lessons do not consider two important points with regards to neural nets : 1. Not all nets work for all problems, those that work tend to have the right inductive biases. We discovered the architectures partially by trial an error, nevertheless they work because of encoded prior information. 2. Data and computation are bounded. GPT4 was basically trained on all text, further advancements probably n…

> GPT4 was basically trained on all text

Well, there's multi-modal training. There's tons of untapped audiovisual data.

Re: The Bitter Lesson (2019) [pdf]

#33

Post by cubefox »

Sutton cites people who were not convinced by the success of Deep Blue in chess:

> They said that ``brute force" search may have won this time, but it was not a general strategy, and anyway it was not how people played chess.

But it is now clear that these people were completely right! The Deep Blue approach (apparently mostly brute force search) didn't scale. E.g. Go wasn't solved this way, and the AlphaGo approach (based on reinforcement learning) did in fact scale to a lot of other games with AlphaZero and MuZero.

Sutton in fact says that both search and learning produce the big success in AI. But now, with the dominance of ML, it seems that mainly learning is responsible the big successes in AI, not search. Machine learning means that the AI models aren't written by hand, GOFAI style, like Deep Blue, but instead by a learning algorithm, which in turn is written by hand. That seems to be the breakthrough, not brute force search. (Do GPTs use "search" in any conventional way? I don't think so.)

Re: The Bitter Lesson (2019) [pdf]

#34

Post by famouswaffles »

niemandhier wrote: post #8

I think the Bitter Lessons do not consider two important points with regards to neural nets : 1. Not all nets work for all problems, those that work tend to have the right inductive biases. We discovered the architectures partially by trial an error, nevertheless they work because of encoded prior information. 2. Data and computation are bounded. GPT4 was basically trained on all text, further advancements probably n…

The essence of the bitter lesson is that the less inductive bias we try to bake in, the better performing they tend to be.

Transformers are preferred to ConvNets these days in Computer Vision despite the latter having all sorts of vision based inductive biases.

GPT-4 has not been trained on all text lol

Re: The Bitter Lesson (2019) [pdf]

#35

Post by Xcelerate »

nologic01 wrote: post #23
Xcelerate wrote: post #12

> The second general point to be learned from the bitter lesson is that the actual contents of minds are tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents of minds, such as simple ways to think about space, objects, multiple agents, or symmetries. I agree and think “ML explainability” efforts are doomed to fail as ML becomes increasingly more effective. There is…

This (popular) opinion makes no sense. How are you going to improve something you dont understand? Throwing ramdom software pieces to see what sticks? The recent progress was possible because people understood intuitively enough of the limitations of earlier models to think and invent a remedy.

> This (popular) opinion makes no sense. How are you going to improve something you dont understand?

How does that opinion not make sense? There are numerous things humans have invented for which we have little understanding of how they work: medical drugs, anesthesia, certain quantum phenomena utilized in semiconductors, etc.

I would argue that for current state-of-the-art LLMs, the implementation is likewise ahead of the theory at the moment.

Re: The Bitter Lesson (2019) [pdf]

#36

Post by Xcelerate »

layer8 wrote: post #27
Xcelerate wrote: post #12

> The second general point to be learned from the bitter lesson is that the actual contents of minds are tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents of minds, such as simple ways to think about space, objects, multiple agents, or symmetries. I agree and think “ML explainability” efforts are doomed to fail as ML becomes increasingly more effective. There is…

There is also no a-priori reason why we shouldn’t be able to understand how the higher-level behavior emerges. And without such an understanding, trying to improve or control the behavior is like poking around in the dark.

> trying to improve or control the behavior is like poking around in the dark

Which is basically what we've been doing with AI so far, as the paper notes (among other fields, like medicine). Have you heard the joke "grad student descent"?

Re: The Bitter Lesson (2019) [pdf]

#37

Post by niemandhier »

famouswaffles wrote: post #34
niemandhier wrote: post #8

I think the Bitter Lessons do not consider two important points with regards to neural nets : 1. Not all nets work for all problems, those that work tend to have the right inductive biases. We discovered the architectures partially by trial an error, nevertheless they work because of encoded prior information. 2. Data and computation are bounded. GPT4 was basically trained on all text, further advancements probably n…

The essence of the bitter lesson is that the less inductive bias we try to bake in, the better performing they tend to be. Transformers are preferred to ConvNets these days in Computer Vision despite the latter having all sorts of vision based inductive biases. GPT-4 has not been trained on all text lol

1. Partially that just moves where we must search for implicitly included symmetries. Word embedding most famously create structure that allows to define shift operations in the embedding space “doctor + female = nurse” might be the most infamous. By evolution only those word embedding so that yield good results are used. Again just because we did not put the structure there, but discovered it by trial and error does not mean that the structure is not key to the success.

2. Gpt 4 was trained on 13T tokens, all books ever written would be about 6.5 T tokens by Fermi calculation.

Re: The Bitter Lesson (2019) [pdf]

#38

Post by cs702 »

cubefox wrote: post #33

Sutton cites people who were not convinced by the success of Deep Blue in chess: > They said that ``brute force" search may have won this time, but it was not a general strategy, and anyway it was not how people played chess. But it is now clear that these people were completely right! The Deep Blue approach (apparently mostly brute force search) didn't scale. E.g. Go wasn't solved this way, and the AlphaGo approach…

We train deep learning models with SGD, which is in fact a form of search -- in a high-dimensional parameter space.

The purpose of SGD is to find parameter values that minimize a training loss.

Re: The Bitter Lesson (2019) [pdf]

#39

Post by nologic01 »

Tuna-Fish wrote: post #25
nologic01 wrote: post #23

Earlier quoted context omitted.

This (popular) opinion makes no sense. How are you going to improve something you dont understand? Throwing ramdom software pieces to see what sticks? The recent progress was possible because people understood intuitively enough of the limitations of earlier models to think and invent a remedy.

Not that I necessarily agree with the above poster, but this: > How are you going to improve something you dont understand? Is just nonsense. Evolution understands nothing, yet produced a mind. Closer to us, the early people who produced all the crops that led to the shift to agriculture, and it's later improvements, absolutely did not understand how any of it worked. Evaluation and selection are sufficient to improv…

These are just absurdly remote analogies that have nothing to with how ML algorithms have developed thus far, nor how they will develop in the the immediate future.

So how would you "improve" here an now any algorithm. Create and insert random code and evaluate? Jeepers people are losing touch with reality.

Re: The Bitter Lesson (2019) [pdf]

#40

Post by reaperman »

the8472 wrote: post #32
niemandhier wrote: post #8

I think the Bitter Lessons do not consider two important points with regards to neural nets : 1. Not all nets work for all problems, those that work tend to have the right inductive biases. We discovered the architectures partially by trial an error, nevertheless they work because of encoded prior information. 2. Data and computation are bounded. GPT4 was basically trained on all text, further advancements probably n…

> GPT4 was basically trained on all text Well, there's multi-modal training. There's tons of untapped audiovisual data.

GPT4 is a multi-modal model. They haven't exposed a way to use the image embeddings, so consumers can't utilize it, but the model accepts image input and it was trained on images and text. Yes, there are other modals that can be incorporated, and reinforcement learning is still pretty nascent/very much unsolved.
Post reply on HN