“The Bitter Lesson” is wrong. Well sort of (opens in new tab)

The main problem with the “Bitter Lesson” is that there’s something even bitter-er behind it — the “Harsh Reality” that while we may scale models on compute and data, that simply broadly inserting tons of data without any sort of curation yields essentially garbage models.

The “Harsh Reality” is that while you may only need data, the current best models and companies behind them spend enormously on gathering high quality labeled data with extensive oversight and curation. This curation is of course being partially automated as well, but ultimately there’s billions or even tens of billions of dollars flowing into gathering, reviewing, and processing subjectively high quality data.

Interestingly, in the time that this paper was published, the harsh reality was not so harsh. For example in things like face detection, (actual) next word prediction, and other purely self supervised and not instruction tuned or “Chat” style models, data was truly all you needed. You didn’t need “good” faces. As long as it was indeed a face, the data itself was enough. Now, it’s not. In order to make these machines useful and not just function approximators, we need extremely large dataset curation industries.

If you learned the bitter lesson, you better accept the harsh reality, too.

The bitter lesson is becoming misunderstood as the world moves on. Unstated yet core to it is that AI researchers were historically attempting to build an understanding of human intelligence. They intended to, piece-by-piece, assemble a human brain and thus be able to explain (and fix) our own biological ones. Much like can be done with physical simulations of knee joints. Of course, you can also use that knowledge to create useful thinking machines, because you understand it well enough to be able to control it. Much like how we have many robotic joints.

So, the bitter lesson is based on a disappointment that you're building intelligence without understanding why it works.

The Leela Chess Zero vs Stockfish case also offers an interesting perspective on the bitter lesson.

Here's my (maybe a bit loose) recollection of what happened:

Step 1- Stockfish was the typical human-knowledge AI, with tons of actual chess knowledge injected in the process of building an efficient chess engine.

Step 2. Then came Leela Chess Zero, with its Alpha Zero-inspired training, a chess engine trained fully with RL with no prior chess knowledge added. And it has beaten Stockfish. This is a “bitter lesson” moment.

Step 3. The Stockfish devs added a neural network trained with RL to their chess engine, in addition to their existing heuristics. And Stockfish easily took back its crown.

Yes sending more compute at a problem is an efficient way to solve it, but if all you have is compute, you'll pretty certainly lose to somebody who has both compute and knowledge.

From TFA:

> No machine learning model was ever built using pure “human knowledge” — because then it wouldn’t be a learning model. It would be a hard coded algorithm.

I guess the author hasn't heard of expert systems? Systems like MYCIN (https://en.wikipedia.org/wiki/Mycin) were heralded as incredible leaps forward at the time, and they indeed consisted of pure “human knowledge.”

I am disturbed whenever a thinkpiece is written by someone who obviously didn't do their research.

Sutton was talking about progress in AI overall, whereas Pinhasi (OP) is talking about building one model for production right now. Of course adding some hand-coded knowledge is essential for the latter, but it has not provided much long-term progress. (Even CNNs and group-convolutional NNs, which seek to encode invariants to increase efficiency while still doing almost only learning, seem to be on the way out)

In my opinion the useful part of "the bitter lesson" has nothing to do with throwing more compute and more data at stuff, it has to do with actually using ML instead of trying to manually and cleverly tweak stuff, and with actually leveraging the data you have effectively as a part of that (again using more ML) rather than trying to manually label everything.

"The Bitter Lesson" certainly seems correct when applied to whatever the limit of the current state of the art is, but in practice solving day-to-day ML problems, outside of FAANG-style companies and cutting edge research, data is always much more constrained.

I have, multiple times in my career, solved a problem using simple, intelligible models that have empirically outperformed neural models ultimately because there was not enough data for the neural approach to learn anything. As a community we tend to obsess over architecture and then infrastructure, but data is often the real limiting factor.

When I was early in my career I used to always try to apply very general, data hungry, models to all my problems.. with very mixed success. As I became more skilled I started to be a staunch advocated of only using simple models you could understand, with much more successful results (which is what lead to this revised opinion). But, at this point in my career, I increasingly see that one's approach to modeling should basically be to approach the problem more information theoretically: try to figure out the model with a channel capacity that best matches your information rate.

As a Bayesian, I also think there's a very reasonable explanation for why "The Bitter Lesson" rings true over and over again. In ET Jaynes' writing he often talks about Bayes' Theorem in terms of P(D|H) (i.e. probably of the Data given the Hypothesis, or vice versa), but, especially in the earlier chapters, purposefully adds an X to that equation: P(D|H,X) where X is a stand in for all of our prior information about the world. Typically we think of prior data as being literal data, but Jaynes' points out that our entire world of understand is also part of our prior context.

In this view, models that "leverage human understanding" (i.e. are fully intelligible) are essentially throwing out information at the limit. But to my earlier point, if the data falls quite short of that limit, then those intelligible models are adding information in data constrained scenarios. I think the challenge in practical application is figuring out where the threshold is that you need to adopt a more general approach.

Currently I'm very much in love with Gaussian Processes that, for constrained data environments, offer a powerful combination of both of these methods. You can give the model prior hints at what things should look like in terms of the relative structure of the kernel and it's priors (e.g. there should be some roughly annual seasonal component, and one roughly weekly seasonal component) but otherwise let the data decide.

I'm not sure if the Bitter Lesson is wrong, I think we'd need clarification from Sutton (does someone have this?)

But I do know "Scale is All You Need" is wrong. And VERY wrong.

Scaling has done a lot. Without a doubt it is very useful. But this is a drastic oversimplification of all the work that has happened over the last 10-20 years. ConvNext and "ResNets Strike Back" didn't take off for reasons, despite being very impressive. There's been a lot of algorithmic changes, a lot of changes to training procedures, a lot of changes to how we collect data[0], and more.

We have to be very honest, you can't just buy your way to AGI. There's still innovation that needs be done. This is great for anyone still looking to get into the space. The game isn't close to being over. I'd argue that this is great for investors too, as there are a lot of techniques looking to try themselves at scale. Your unicorns are going to be over here. A dark horse isn't a horse that just looks like every other horse. Might be a "safer" bet, but that's like betting on amateur jockies and horses that just train similar to professional ones. They have to do a lot of catch-up, even if the results are fairly certain. At that point you're not investing in the tech, you're investing in the person or the market strategy.

[0] Okay, I'll buy this one as scale if we really want to argue that these changes are about scaling data effectively but we also look at smaller datasets differently because of these lessons.

The Bitter Lesson is saying, if you're going to use human knowledge, be aware that your solution is temporary. It's not wrong. And it's not wrong to use human knowledge to solve your "today" problem.

it should be noted that "The Bitter Lesson" is not a general principle, for example LLMs are not able to sum 1000 digits numbers, but a python program with a few lines can do it. Also we provide tools (algorithms) for LLMs, these examples show that "The Bitter Lesson" is not true beyond a certain context.

Perhaps the Bitter Lesson, more data and compute triumph over intelligent algorithms, is true only in a certain range, and a new intelligent algorithm is needed to go beyond. For example, nearest neighbors algorithm obey the Bitter Lesson depending of the relation between grid of data points and complexity of the problem.

I first read the "Blitter" lesson and got excited for a moment.

the bitter lesson is to _let_ data and compute do the majority of the work, it's not to remove humans completely, but wherever possible