[HN Gopher] Does the Bitter Lesson Have Limits?
       ___________________________________________________________________
        
       Does the Bitter Lesson Have Limits?
        
       Author : dbreunig
       Score  : 175 points
       Date   : 2025-08-01 20:21 UTC (1 days ago)
        
 (HTM) web link (www.dbreunig.com)
 (TXT) w3m dump (www.dbreunig.com)
        
       | o11c wrote:
       | This is about AI, despite the title being ambiguous.
        
         | andy99 wrote:
         | Is there more than one bitter lesson?
        
           | terminalshort wrote:
           | I've learned many
        
           | bigstrat2003 wrote:
           | I don't know, but considering I had never heard of "the
           | bitter lesson" before reading this article I agree the title
           | is not especially clear. Not everyone lives and breathes AI
           | and has heard of stuff like that, after all.
        
         | thrawa8387336 wrote:
         | If we're going to be pedantic:
         | 
         | This is about AI, the title is ambiguous.
         | 
         | Despite was used unambiguously wrong.
        
           | o11c wrote:
           | "Despite" is absolutely correct when you realize that
           | cheating in the title is a way to make people look at
           | articles when they would rather ignore AI in favor of
           | actually-useful/interesting subjects.
        
         | myhf wrote:
         | The original "Bitter Lesson" essay is about machine learning,
         | but the linked article appears to be trying to apply it to
         | LLMs.
        
       | criemen wrote:
       | All links render as blue strike-through line in Firefox
       | (underline in Chrome), hurting legibility :(
        
         | thwg wrote:
         | interesting. i see underlines in firefox. but the width of the
         | line is 2px in chrome, 1px in firefox.
        
         | penneyd wrote:
         | Fine over here using Firefox
        
         | Cerium wrote:
         | I'm getting the same effect, seems to be the css property
         | "text-underline-position: under;"
        
       | grubbypaw wrote:
       | I was not at all a fan of "The Bitter Lesson versus The Garbage
       | Can", but this misses the same thing that it missed.
       | 
       | 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.
        
         | schmidtleonard wrote:
         | Right, and if you spot a job that needs doing and can be done
         | by a specialized model, waving your hands about general purpose
         | scale-leveraging models eventually overtaking specialized
         | models has not historically been a winning approach.
         | 
         | Except in the last year or two, which is why people are citing
         | it a lot :)
        
           | hinkley wrote:
           | Probably because this is how bubbles happen.
        
         | anp wrote:
         | I think there might be interesting time scales in between "now"
         | and "my entire career" to which the bitter lesson may or may
         | not apply. As an outsider to ML I have questions about the
         | longevity of any given "context engineering" approach in light
         | of the bitter lesson.
        
         | energy123 wrote:
         | The bitter lesson becomes more true over time, because
         | inductive bias becomes less useful over time. Case in point:
         | PCA/hand engineering -> CNN -> ViT.
        
       | Animats wrote:
       | The question is when price/performance hits financial limits.
       | That point may be close, if not already passed.
       | 
       | 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.
        
         | schmidtleonard wrote:
         | Fundamentally awful but spiritually delightful.
        
         | azeirah wrote:
         | Sometimes awful is the best we have, we don't have anything
         | that performs at a similar level to EUV machines by ASML but
         | are much simpler or tenable than what we have right now, right?
         | 
         | 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?
        
         | xyzzy123 wrote:
         | It's hard to value $20bn in 2025 vs say $2bn for cutting edge
         | fab in 2000. Personally I don't think CPI captures the relative
         | value of _investments_ well.
         | 
         | 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.
        
       | thrawa8387336 wrote:
       | This brings about a good point:
       | 
       | 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.
        
       | fdav wrote:
       | A better lesson: https://rodneybrooks.com/a-better-lesson/
        
         | ripe wrote:
         | Rodney Brooks's analysis, as usual, is spot-on.
         | 
         | Tldr; ML using neural networks is not really replacing human
         | knowledge with computation; human work goes into encoding data,
         | building datasets, and designing hyperparameters.
        
       | benlivengood wrote:
       | I think it's a little early (even in these AI times) to call HRM
       | a counterexample of the bitter lesson.
       | 
       | 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.
        
         | turnsout wrote:
         | I didn't know about HRM until this article, but it looks
         | incredibly impressive, even in its current state. A 27M
         | parameter model that can be trained to perform difficult
         | domain-specific tasks with better results than flagship
         | reasoning models? That is compelling.
        
       | PaulHoule wrote:
       | One odd thing is that progress in SAT/SMT solvers has been almost
       | as good as progress in neural networks from the 1970s to the
       | present. There was a time I was really interested in production
       | rules and expert system shells and systems in the early 1980s
       | often didn't even use RETE and didn't have hash indexes so of
       | course a rule base of 10,000 looked unmanageable, by 2015 you
       | could have a million rules in Drools and it worked just fine.
        
         | bmc7505 wrote:
         | The difference is that SAT/SMT solvers have primarily relied on
         | single-threaded algorithmic improvements [1] and unlike neural
         | networks, we have not [yet] discovered a uniformly effective
         | strategy for leveraging additional computation to accelerate
         | wall-clock runtime. [2]
         | 
         | [1]: https://arxiv.org/pdf/2008.02215
         | 
         | [2]: https://news.ycombinator.com/item?id=36081350
        
           | PaulHoule wrote:
           | RETE family algorithms did turn out to be somewhat
           | parallelizable, enough to get a real speed-up on ordinary
           | multicore CPUs. There was an idea in the 1980s that symbolic
           | AI would be massively parallelizable that turned out to be a
           | disappointment.
           | 
           | https://en.wikipedia.org/wiki/Fifth_Generation_Computer_Syst.
           | ..
        
             | bmc7505 wrote:
             | You could argue that since automatic differentiation and
             | symbolic differentiation are equivalent, [1] symbolic AI
             | did succeed by becoming massively parallelizable, we just
             | needed to scale up the data and hardware in kind.
             | 
             | [1]: https://arxiv.org/pdf/1904.02990
        
           | yorwba wrote:
           | > [2]
           | 
           | 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.
        
             | bmc7505 wrote:
             | > The solvers participating in this track will be executed
             | with a wall-clock time limit of 1000 seconds. Each solver
             | will be run an a single AWS machine of the type
             | m6i.16xlarge, which has 64 virtual cores and 256GB of
             | memory.
             | 
             | 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.
        
       | jamesblonde wrote:
       | I see elements of the bitter lesson in arguments about context
       | window size and RAG. The argument is about retrieval being the
       | equivalent of compute/search. Just improve them, to hell with all
       | else.
       | 
       | 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.
        
       | sorenjan wrote:
       | The last time I was reminded of the bitter lesson was when I read
       | about Guidance & Control Networks, after seeing them used in an
       | autonomous drone that beat the best human FPV pilots [0].
       | Basically it's using a small MLP (Multi Layer Perceptron) on the
       | order of 200 parameters, and using the drone's state as input and
       | controlling the motors directly with the output. We have all
       | kinds of fancy control theory like MCP (Model Predictive
       | Control), but it turns out that the best solution might be to
       | train a relatively tiny NN using a mix of simulation and
       | collected sensor data instead. It's not better because of huge
       | computation resources, it's actually more computationally
       | efficient than some classic alternatives, but it is more general.
       | 
       | [0] https://www.tudelft.nl/en/2025/lr/autonomous-drone-from-
       | tu-d...
       | 
       | https://www.nature.com/articles/s41586-023-06419-4
       | 
       | https://arxiv.org/abs/2305.13078
       | 
       | https://arxiv.org/abs/2305.02705
        
         | logicchains wrote:
         | >It's not better because of huge computation resources, it's
         | actually more computationally efficient than some classic
         | alternatives
         | 
         | 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.
        
           | William_BB wrote:
           | Interesting. Are these models the SOTA in the options trading
           | industry (e.g. MM) nowadays?
        
           | cactusfrog wrote:
           | This is really interesting. I think force fields in molecular
           | dynamics have underwent a similar NN revolution. You train
           | your NN on the output of expensive calculations to replace
           | the expensive function with a cheap one. Could you train a
           | small language model with a big one?
        
             | lossolo wrote:
             | > Could you train a small language model with a big one?
             | 
             | Yes, it's called distillation.
        
           | jwjohnson314 wrote:
           | Can you elaborate on this? I am curious what processes are
           | being simulated to feed the network.
        
         | bsoles wrote:
         | But I have also seen people trying to use deep networks to
         | identify rotating machinery faults, like bearings, from raw
         | accelerometer data collected a high frequencies like 40 kHz.
         | Whereas the spectrum of the data from running FFT on the signal
         | contains fault information much more obviously and clearly.
         | 
         | Throwing a deep network on a problem without some physical
         | insight into the problem has also its disadvantages it seems.
        
           | jvanderbot wrote:
           | Yeah, we're shouting into the wind here. I have had people
           | tell me directly that my ideas from old school state
           | estimation were irrelevant in the era of deep learning. They
           | may produce (in this case) worse results, but the long game
           | I'm assured is superior.
           | 
           | 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.
        
             | AnonymousPlanet wrote:
             | > 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.
        
           | physicsguy wrote:
           | I'm working on this sort of thing right now in a SaaS product
           | that previously didn't have support for vibration data. One
           | competitor is all ML-d up to the hilt but customers don't
           | like the black box and keep finding it gives false positives
           | with no explanation. I think one problem is those buying not
           | understanding problem; they just want to plug in a sensor and
           | insights to happen, but without information about the machine
           | that's never going to be able to provide useful insights
        
         | ssivark wrote:
         | > an autonomous drone that beat the best human FPV pilots
         | 
         | 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.
        
         | bsoles wrote:
         | There are also extremely misleading research articles out there
         | promising good results with deep networks in the area of
         | anomaly detection, without adequate comparison with more
         | classical techniques.
         | 
         | 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
        
         | chrisweekly wrote:
         | > "MCP (Model Predictive Control)"
         | 
         | ^ that's MPC. (MCP = Model Context Protocol)
        
       | mentalgear wrote:
       | The Neuro-Symbolic approach is what the article describes,
       | without actually naming it.
        
         | jgalt212 wrote:
         | Perhaps, but I see it more as an endorsement of careful feature
         | selection. Subject matter experts can do this, and once done,
         | you can get a away with a much smaller model and better price /
         | performance.
        
       | pu_pe wrote:
       | I'm not so sure Stockfish is a good example. The fact it can run
       | on an Iphone is due to Moore's law, which follows the same
       | pattern. And Deepmind briefly taking its throne was a very good
       | example of the Bitter Lesson.
        
         | mtlmtlmtlmtl wrote:
         | Stockfish being so strong is not merely a result of scaling of
         | computation with search and learning. Basic alpha-beta search
         | doesn't really scale all that well with compute. The number of
         | nodes visited grows exponentionally with the number of plies
         | you look ahead. Additionally alpha-beta search is not
         | embarassingly parallel. The reason Stockfish is so strong is
         | that it includes pretty much every heuristic improvement to
         | alpha-beta that's been thought of in the history of computer
         | chess, somehow combining all of them while avoiding bugs and
         | performance regressions. Many of these heuristics are based on
         | chess knowledge. As well as a lot of very clever optimisation
         | of data structures(transposition tables, bitboards) to
         | facilitate parallel search and shave off every bit of overhead.
         | 
         | Stockfish is a culmination of a lot of computer science
         | research, chess knowledge and clever, meticulous design.
        
           | plumthreads wrote:
           | While what you mention is true, I'm not sure how it
           | undermines the bitter lesson. Optimizing the use of hardware
           | (which is what NNUE essentially does) is one way of
           | "increasing compute." Also, NNUE was not a chess specific
           | technique, it was originally developed for Shogi.
        
       | aydyn wrote:
       | Does anyone else see the big flaw with the chess engine analogy?
       | 
       | 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_.
        
         | Mistletoe wrote:
         | I'm wondering how prophetic this is for LLMs. My hunch is a
         | lot.
        
       | itkovian_ wrote:
       | I don't think people understand the point sutton was making; he's
       | saying that general, simple systems that get better with scale
       | tend to outperform hand engineered systems that don't. It's a
       | kind of subtle point that's implicitly saying hand engineering
       | inhibits scale because it inhibits generality. He is not saying
       | anything about the rate, doesn't claim llms/gd are the best
       | system, in fact I'd guess he thinks there's likely an even more
       | general approach that would be better. It's comparing two classes
       | of approaches not commenting on the merits of particular systems.
        
         | eldenring wrote:
         | Yep this article is self centered and perfectly represents the
         | type of ego Sutton was referencing. Maybe in a year or two
         | general methods will improve the author's workflow
         | significantly once again (eg. better models) and they would
         | still add a bit of human logic on top and claim victory.
        
           | visarga wrote:
           | The point about training data stands. We usually only think
           | of scaling compute, but we need to scale data as well, maybe
           | even faster than compute. But we exhausted the source of high
           | quality organic text, and it doesn't grow exponentially fast.
           | 
           | 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.
        
         | joe_the_user wrote:
         | It occurs to me that the bitter lesson is so often repeated
         | because it involves a slippery slope or moot-and-castle
         | argument. IE, the meaning people assign to the bitter lesson
         | ranges between all the following:
         | 
         |  _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_
        
           | arrowsmith wrote:
           | motte and bailey*
        
         | xpe wrote:
         | > I don't think people understand the point sutton was making;
         | he's saying that general, simple systems that get better with
         | scale tend to outperform hand engineered systems that don't
         | 
         | 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.
        
       | throw1289312 wrote:
       | This article focuses on the learning aspect of The Bitter Lesson.
       | But The Bitter Lesson is about both search _and_ learning.
       | 
       | 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!
        
         | jgalt212 wrote:
         | Stockfish is neither. It's a hybrid approach.
        
       | beepbooptheory wrote:
       | I should know better than to speak anything too enthusiastically
       | about the humanities or feminism on this particular forum, but I
       | just want to say the connection here to Donna Haraway was a
       | surprise and delight. Any one open to that world at all would
       | behoove themselves to check her out. "The Cyborg Manifesto" is
       | the one everyone knows, but I recently finished "Living with the
       | Trouble" and can't recommend it enough!
        
         | frozenseven wrote:
         | >Donna Haraway
         | 
         | >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.
        
       | benreesman wrote:
       | When The Bitter Lesson essay came out it was a bunch of important
       | things: addressing an audience of serious practitioners,
       | contrarian and challenging entrenched dogma, written without any
       | serious reputational or (especially) financial stake in the
       | outcome. It needed saying and it was important.
       | 
       | 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.
        
       | stego-tech wrote:
       | Not familiar with the cited essay (added to reading list for the
       | weekend), but the post does make some generally good points on
       | generalization (it me) vs specialization, and the benefits of an
       | optimized and scalable generalist approach vs a niche,
       | specialized approach, specifically with regards to current LLMs
       | (and to a lesser degree, ML as a whole).
       | 
       | 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).
        
       | kazinator wrote:
       | > The bitter lesson is dependent on high-quality data.
       | 
       | Arguably, so is the alternative: explicitly embedding knowledge!
       | 
       | Nothing is immune to GIGO.
        
       | strangescript wrote:
       | The problem with the Bitter Lesson is that it doesn't clearly
       | define what is a computational "hack" and what is a genuine
       | architecturally breakthrough. We would be no where without
       | transformers for example.
        
       | titanomachy wrote:
       | > This views organizations as chaotic "garbage cans" where
       | problems, solutions, and decision-makers are dumped in together,
       | and decisions often happen when these elements collide randomly,
       | rather than through a fully rational process
       | 
       | 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.
        
       | quantum_state wrote:
       | It would not be surprising if a bitter lesson 2.0 comes about as
       | a bitter lesson to the bitter lesson.
        
       | bravesoul2 wrote:
       | Well hardware has limits so I guess so. Humans evolved with
       | faculties not a massive organic generic compute engine.
        
       | worik wrote:
       | 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
       | 
       | A lot of firms starting into this business are "betting the farm"
       | on "scaling AI across the enterprise"
       | 
       | 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
        
       | softwaredoug wrote:
       | In my experience, there's an opposing "bitter lesson" when trying
       | to make incremental, tactical progress in user-facing AI / ML
       | applications: _you're not a researcher_. Stay to tried-and-true,
       | boring ML methods that have been proven at scale and then add
       | human knowledge and rules to make it all work.
       | 
       | 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.
        
       | xpe wrote:
       | > The biggest lesson that can be read from 70 years of AI
       | research is that general methods that leverage computation are
       | ultimately the most effective, and by a large margin.
       | 
       | 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?"
        
       | jgalt212 wrote:
       | The Bitter Lesson assumes Moore's law is alive and well. It may
       | still be alive, but not as full of vim and vigor as it once was.
        
       | lisper wrote:
       | Por que no los dos?
       | 
       | 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.
        
         | incognito124 wrote:
         | Did you name that law by yourself?
        
           | DonHopkins wrote:
           | https://flownet.com/ron/papers/tla.pdf
           | 
           | >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.
        
         | gugagore wrote:
         | What you're describing -- search + NN -- is presently under the
         | term "test-time compute".
         | 
         | 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.
        
           | lisper wrote:
           | No, I don't think that test-time compute is the same thing at
           | all. It's a little challenging to find a definitive
           | definition of TTC, but AFAICT it is just a fairly simple
           | control loop around an LLM. What I'm describing is a merging
           | of components with fundamentally different architectures,
           | each of which is a significant engineering effort in its own
           | right, to produce a whole that is greater than the sum of its
           | parts. Those seem different to me, but to be fair I have not
           | been keeping up with the latest tech so I could be wrong.
        
             | gugagore wrote:
             | I think search is a fairly simple control loop. Beam search
             | is an example of TTC in this modern era.
             | 
             | 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.
        
         | alan-crowe wrote:
         | I've been thinking of the issue like this: a traditional
         | approach to symbolic artificial intelligence frames problem
         | solving as a search of a tree. Do you go depth first, dodging
         | infinite branches? Or breadth first (do you have enough
         | memory?). What about iterative deepening?
         | 
         | 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.
        
       | user_of_the_wek wrote:
       | > If AI agents can train on outputs alone, any organization that
       | can define quality and provide enough examples might achieve
       | similar results
       | 
       | Great, we're safe!
        
       | woolion wrote:
       | I believe the main problem could be reframed as an improper use
       | of analogies. People are pushing the analogy of "artificial
       | intelligence" and "brain", etc. creating a confusion that leads
       | to such "laws". What we have is a situation that is similar to
       | birds and planes, they do not operate under the same principles
       | at all.
       | 
       | 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"?
        
         | phkahler wrote:
         | >> People are pushing the analogy of "artificial intelligence"
         | and "brain", etc. creating a confusion that leads to such
         | "laws". What we have is a situation that is similar to birds
         | and planes, they do not operate under the same principles at
         | all.
         | 
         | 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.
        
       ___________________________________________________________________
       (page generated 2025-08-02 23:02 UTC)