Does the Bitter Lesson Have Limits?
dbreunig.com
dbreunig.com
[0] https://www.tudelft.nl/en/2025/lr/autonomous-drone-from-tu-d...
https://www.nature.com/articles/s41586-023-06419-4
It's similar with options pricing. The most sophisticated models like multivariate stochastic volatility are computationally expensive to approximate with classical approaches (and have no closed form solution), so just training a small NN on the output of a vast number of simulations of the underlying processes ends up producing a more efficient model than traditional approaches. Same with stuff like trinomial trees.
Yes, it's called distillation.
Throwing a deep network on a problem without some physical insight into the problem has also its disadvantages it seems.
The specific scenario was estimating the orientation of a stationary semi trailer. An objectively measurable number and it was consistently off by 30 deg, yet I was the jerk for suggesting we move from end to end DL to trad Bar Shalom techniques.
That scene isn't for me anymore.
They will learn. At least when the competition beats their solution with a hybrid approach they can't begin to understand.
Doesn’t any such claim come with huge caveats — pre specified track/course, no random objects flying between, etc…? ie. train & test distributions are ensured same by ensuring test time can never be more complicated than training data.
Also presumably better sensing than raw visual input.
This well-known critical paper shows examples of AI articles/techniques applied to popular datasets with good-looking results. But, it also demonstrates that, literally, a single line of MATLAB code can outperform some of these techniques: https://arxiv.org/pdf/2009.13807
^ that's MPC. (MCP = Model Context Protocol)
The Bitter Lesson is from the perspective of how to spend your entire career. It is correct over the course of a very long time, and bakes in Moore's Law.
The Bitter Lesson is true because general methods capture these assumed hardware gains that specific methods may not. It was never meant for contrasting methods at a specific moment in time. At a specific moment in time you're just describing Explore vs Exploit.
Except in the last year or two, which is why people are citing it a lot :)
Thirty-five years ago they gave me a Ph.D. basically for pointing out that the controversy du jour -- reactive vs deliberative control for autonomous robots -- was not a dichotomy. You could have the best of both worlds by combining a reactive system with a deliberative one. The reactive system interfaced directly to the hardware on one end and provided essentially a high-level API on the other end that provided primitives like "go that way". It's a little bit more complicated than that because it turns out you need a glue layer in the middle, but the point is: you don't have to choose. The Bitter Lesson is simply a corollary of Ron's First Law: all extreme positions are wrong. So reactive control by itself has limits, and deliberative control by itself has limits. But put the two together (and add some pretty snazzy image processing) and the result is Waymo.
So it was no surprise to me that Stockfish, with its similar approach of combining deliberative search with a small NN computing its quality metric blows everything else out of the water. It has been obvious (at least to me) that this is the right approach for decades now.
I'm actually pretty terrified of the results when the mainstream AI companies finally rediscover this. The capabilities of LLMs are already pretty impressive on their own. If they can get a Stockfish-level boost by combining them with a simple search algorithm the result may very well be the GAI that the rationalist community has been sounding the alarm over for the last 20 years.
>In retrospect, in the story of the three-layer architecture there may be more to be learned about research methodology than about robot control architectures. For many years the field was bogged down in the assumption that planning was sufficient for generating intelligent behavior in situated agents. That it is not sufficient clearly does not justify the conclusion that planning is therefore unnecessary. A lot of effort has been spent defending both of these extreme positions. Some of this passion may be the result of a hidden conviction on the part of AI researchers that at the root of intelligence lies a single, simple, elegant mechanism. But if, as seems likely, there is no One True Architecture, and intelligence relies on a hodgepodge of techniques, then the three-layer architecture offers itself as a way to help organize the mess.
The rules / dynamics / objectives of chess ( and Go ) are trivial to encode in a search formulation. I personally don't really get what that tells us about AGI.
It is a very wide term, IME, that means anything besides "one-shot through the network".
I think the thing about the search formulation, which is amenable to domains like chess and go, but not other domains is critical. If LLMs are coming up with effective search formulation for "open-ended" problems, that would be a big deal. Maybe this is what you're alluding to.
That's like saying that Darwinian evolution is simple. It's not entirely wrong, but it misses the point rather badly. The thing that makes search useful is not the search per se, it's the heuristics that reduce an exponential search space to make it tractable. In the case of evolution (which is a search process) the heuristic is that at every iteration you select the best solution on the search frontier, and you never backtrack. That heuristic produces a certain kind of interesting result (life) but it also has certain drawbacks (it's limited to a single quality metric: reproductive fitness).
> Beam search is an example of TTC in this modern era.
That's an interesting analogy. I'll have to ponder that.
But my knee-jerk reaction is that it's not enough to say "put reactivity and deliberation together". The manner in which you put them together matters, and in particular, it turns out that putting them together with a third component that manages both the deliberation and the search is highly effective. I can't say definitively that it's the best way -- AFAIK no one has ever actually done the research necessary to establish that. But empirically it produced good results with very little computing power (by today's standards).
My gut tells me that the right way to combine LLMs and search is not to have the search manage the LLM, but to provide search as a resource for the LLM to use, kind of like humans use a pocket calculator to help them do arithmetic.
> If LLMs are coming up with effective search formulation for "open-ended" problems, that would be a big deal.
AFAICT, at the moment LLMs aren't "coming up" with anything, they are just a more effective compression algorithm for vast quantities of data. That's not nothing. You can view the scientific method itself as a compression algorithm. But to come up with original ideas you need something else, something analogous to the random variation and selection in Darwinian evolution. Yes, I know that there is a random element in LLM algorithms, and again I don't really understand the details, but the way in which the randomness is deployed just feels wrong to me somehow.
I wish I had more time to think deeply about these things.
Sometimes the solution is in the tree, but it is too deep, and one runs out of time before it is found.
Statistical based learning could act as a branch predictor. Sometimes guiding the search to go very deep in the right place and find the hidden solution. Sometimes guiding the search to go very deep in the wrong place; one runs out of time as usual.
Notice the strength of hybrid approach. One isn't accepting the probably correct answer of the statistical part. It is only a guide, and if the answer is found, and the symbolic part of the software is correct, the answer will be reliable.
I think this is already being done with maths problems. The LLM is writing proof attempts in Lean. But Lean is traditional symbolic AI. If the LLM can come up with a proof that Lean approves, then it really has a proof. From your comment I learn that Stockfish has already got something like this to work very well.
I think at the moment the best source of data is the chat log, with 1B users and over 1T daily tokens over all LLMs. These chat logs are at the intersection of human interests and LLM execution errors, they are on-policy for the model, right what they need to improve the next iteration.
General-purpose-algorithms-that-scale will beat algorithms that aren't those
The most simple general purpose, scaling algorithm will win, at least over time
Neural networks will win
LLMs will reach AGI with just more resources
This is your reading of Sutton. When I read his original post, I don't extract this level of nuance. The very fact that he calls it a "lesson" rather than something else, such as a "tendency", suggests Sutton may not hold the idea lightly*. In other words, it might have become more than a testable theory; it might have become a narrative.
* Holding an idea lightly is usually good thing in my book. Very few ideas are foundational.
Interestingly, this hasn't happened for wafer fabs. A modern wafer fab costs US$1bn to US$3bn, and there is talk of US$20bn wafer fabs. Around the year 2000, those would have been un-financeable. It was expected that fab cost was going to be a constraint on feature size. That didn't happen.
For years, it was thought that the ASML approach to extreme UV was going to cost too much. It's a horrible hack, shooting off droplets of tin to be vaporized by lasers just to generate soft X-rays. Industry people were hoping for small synchrotrons or X-ray lasers or E-beam machines or something sane. But none of those worked out. Progress went on by making a fundamentally awful process work commercially, at insane cost.
Perhaps we will find something better in the future, but for now awful is the best we got for the cutting edge.
Also, when is cutting edge not the worst it's ever been?
If you compare it as % of gdp, or relative to m2 or % of S&P 500, or even size of electronics industry it's maybe 2x-4x or something. Which is still an increase and still a lot but doesn't seem as crazy to me.
Another lens on it: the most valuable car company in 2000 was $50bn and in 2025 it is $1000bn, which I think says more about dollars than cars.
Be careful when anyone, even a giant in the field such as Sutton, posits a sweeping claim like this.
My take? Sutton's "bitter lesson" is rather vague and unspecified (i.e. hard to pin down and test) for at least two reasons:
1. The word "ultimately" is squishy, when you think about it. When has enough time passed to make the assessment? At what point can we say e.g. "Problem X has a most effective solution"?
2. What do we mean by "most effective"? There is a lot of variation, including but not limited to (a) some performance metric; (b) data efficiency; (c) flexibility / adaptability across different domains; and (d) energy efficiency.
I'm a big fan of Sutton's work. I've read his RL book cover-to-cover and got to meet him briefly. But, to me, the bitter lesson (as articulated in Sutton's original post) is not even wrong. It is sufficiently open-ended that many of us will disagree about what the lesson is, even before we can get to the empirical questions of "First, has it happened in domain D at time T? Second, is it 'settled' now, or might things change?"
https://en.wikipedia.org/wiki/Fifth_Generation_Computer_Syst...
In the comments, zero_k posted a link to the SAT competition's parallel track. The 2025 results page is here: https://satcompetition.github.io/2025/results.html Parallel solvers consistently score lower (take less time) than single-threaded solvers, and solve more instances within the time limit. Probably the speedup is nowhere near proportional to the amount of parallelism, but if you just want to get results a little bit faster, throwing more cores at the problem does seem like it generally works.
For comparison, an H100 has 14,592 CUDA cores, with GPU clusters measured in the exaflops. The scaling exponents are clearly favorable for LLM training and inference, but whether the same algorithms used for parallel SAT would benefit from compute scaling is unclear. I maintain that either (1) SAT researchers have not yet learned the bitter lesson, or (2) it is not applicable across all of AI as Sutton claims.
Then, as the article mentions, some new fundamental shift happens, and practitioners need to jump over to a completely new way of working. Monkeypatching to make it all work. Rinse repeat.
How much of the recent bitter lesson peddling is done by compute salesmen?
How much of it is done by people who can buy a lot of compute?
Deepseek was scandalous for a reason.
However, retrieval is not just google search. Primary key lookups in my db are also retrieval. As are vector index queries or BM25 free text search queries. It's not a general purpose area like compute/search. In summary, i don't think that RAG is dead. Context engineering is just like feature engineering - transform the swamp of data into a structured signal that is easy for in-context learning to learn.
The corollory of all this is it's not just about scaling up agents - giving them more LLMs and more data via MCP. The bitter lesson doesn't apply to agents yet.
Looking at the original claim, we can take from birds a number of optimization regarding air flows that are far beyond what any plane can do. But, the impact that could be transfer to planes would be minimal compared to a boost in engine technology. Which is not surprising since the way both systems achieve "flight" are completely different.
I don't believe such discourse would happen at all if it was just considered to be a number of techniques, of different categories with their own strength and weaknesses, used to tackle problems.
Like all fake "laws", it is based on a general idea that is devoid of any time-frame prediction that would make it falsifiable. In "the short term" is beaten by "in the long run". How far is "the long run"? This is like the "mean reversion law", saying that prices will "eventually" go back to their equilibrium price; will you survive bankruptcy by the time of "eventually"?
But modend neural nets ARE based on biological neurons. It's not a perfect match by any means, but synaptic "weights" are very much equivalent to model weights. Model structure and size are bigger differences.
But its become a lazy crutch for a bunch of people who meet none of those criteria and perverted into a statement more along the lines of "LLMs trained on NVIDIA cards by one of a handful of US companies are guaranteed to outperform every other approach from here to the Singularity".
Nope. Not at all guaranteed, and at the moment? Not even looking likely.
It will have other stuff in it. Maybe that's prediction in representation space like JEPA, maybe its MCTS like Alpha*, maybe its some totally new thing.
And maybe it happens in Hangzhou.
Stockfish is a culmination of a lot of computer science research, chess knowledge and clever, meticulous design.
>Any one open to that world...
The "world" in question being a brand of Marxism that's super-explicitly anti-human. No, I'm not kidding or exaggerating.
I think it's quite a bit more likely for HRM to scale embarrassingly far and outstrip the tons of RLHF and distillation that's been invested in for transformers, more of a bitter lesson 2.0 than anything else.
This article cites Leela, the chess program, as an example of the Bitter Lesson, as it learns chess using a general method. The article then goes on to cite Stockfish as a counterexample, because it uses human-written heuristics to perform search. However, as you add compute to Stockfish's search, or spend time optimizing compute-expenditure-per-position, Stockfish gets better. Stockfish isn't a counterexample, search is still a part of The Bitter Lesson!
Tldr; ML using neural networks is not really replacing human knowledge with computation; human work goes into encoding data, building datasets, and designing hyperparameters.
Great, we're safe!
When AlphaZero came along it blew stockfish out of the water.
Stockfish is a top engine now because besides that initial proof of concept there's no money to be made by throwing compute at Chess.
43% of American workers have used AI at work, they are mostly doing it in informal ways, solving their own work problems. Scaling AI across the enterprise is hard
In my experience LLMs are incredibly useful from a simple text interface (I only work with text, mainly computer code). I am still reeling from how disruptive they are, in that context.
But IMO there is not a lot of money to be made for start ups in that context (I expect there is not enough to justify the high valuations of outfits like Open AI). There should be a name for the curse - revolutionary technology that makes many people vastly more productive, but there is no real way to capture that value. Unless "Scaling AI across the enterprise" can succeed.
I have my doubts. I am sure there will be niches, and in a decade or so, with hindsight, it will be clear what they are. But there is no reliable way to tell now
The "Bitter Lesson" seems like a distraction to me. The fundamental problem is related: this technology is generally useful, much more than it is specifically useful.
The "killer app" is a browser window open to https://chat.deepseek.com. There is not much beyond that. Not nothing, just not much.
But so long as you have not bet your farm on "scaling AI across the enterprise" nor been fired by someone else who is trying, we should be very happy. We are in a "steam engine" moment. Nothing will ever be the same.
And if Open AI and the like all go belly up and demote a swathe of billionaires to be normally rich, that is the cherry on the top
Arguably, so is the alternative: explicitly embedding knowledge!
Nothing is immune to GIGO.
Trying to train it to do what humans do, working with each other, when they navigate a company they designed (by plan and happenstance), to achieve a spaghetti plate of objectives, is just another attempt at putting humans between data and performance.
Only broad goals, i.e. revenues up, costs down, profits rising sustainably, no lost legal cases ... needed.
Using the hodgepodge of measures traditionally used to (try to) track intermediate human performances? ...would be just another attempt at putting humans between data and performance.
Where I furrow my brow is the casual mixing of philosophical conjecture with technical observations or statements. Mixing the two all too often feels like a crutch around defending either singular perspective in an argument by stating the other half of the argument defends the first half. I know I'm not articulating my point well here, but it just comes off as a little...insincere, I guess? I'm sure someone here will find the appropriate words to communicate my point better, if I'm being understood.
One nitpick on the philosophical side of things I'd point out is that a lot of the resistance to AI replacing human labor is less to do with the self-styled importance of humanity, and more the bleak future of a species where a handful of Capitalists will destroy civilization for the remainder to benefit themselves. That is what sticks in our collective craw, and a large reason for the pushback against AI - and nobody in a position of power is taking that threat remotely seriously, largely because the owners of AI have a vested interest in preventing that from being addressed (since it would inevitably curb the very power they're investing in building for themselves).
This is about AI, the title is ambiguous.
Despite was used unambiguously wrong.
But... The title being ambiguous despite the content being unambiguously about AI.
What?
Is the content of the complain that the article is insufficiently ambiguous despite the ambiguity of the title? No
Only tangentially related, but this has to be one of the worst metaphors I’ve ever heard. Garbage cans are not typically hotbeds of chaotic activity, unless a raccoon gets in or something.