A Bitter AI Lesson - Compute Reigns Supreme!

📝 The article "The Bitter Lesson" is available here:

🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
313V, Alex Haro, Andrew Melnychuk, Angelos Evripiotis, Anthony Vdovitchenko, Brian Gilman, Bruno Brito, Bryan Learn, Christian Ahlin, Christoph Jadanowski, Claudio Fernandes, Dennis Abts, Eric Haddad, Eric Martel, Evan Breznyik, Geronimo Moralez, Javier Bustamante, John De Witt, Kaiesh Vohra, Kasia Hayden, Kjartan Olason, Levente Szabo, Lorin Atzberger, Marcin Dukaczewski, Marten Rauschenberg, Maurits van Mastrigt, Michael Albrecht, Michael Jensen, Morten Punnerud Engelstad, Nader Shakerin, Owen Campbell-Moore, Owen Skarpness, Raul Araújo da Silva, Richard Reis, Rob Rowe, Robin Graham, Ryan Monsurate, Shawn Azman, Steef, Steve Messina, Sunil Kim, Thomas Krcmar, Torsten Reil, Zach Boldyga, Zach Doty.

Image sources:

Károly Zsolnai-Fehér's links:

[sddfg]

I'm good with 10 minute papers too.

[shaun_rambaran]

Man, the subtlety of Károly's humour. I love it.
"Welcome to Two-Minute Papers!"

[erikharkonen]

This phrase, from the end of the article, would make a great AI mantra:
"We want AI agents that can discover like we can, not which contain what we have discovered"

[darabat]

Rather than a bitter lesson, my reaction to that at first was "well, that's something I've come to accept a long time ago", which is the fact that human thinking is not sacred, flawless nor always superior to other possibilities.

[davidwuhrer]

This is not only relevant to AI, but to programming in general.

At our university the mathematicians needed a more efficient matrix processor. They gave a computer scientist a library of optimisation tricks. The programmer implemented the most basic matrix operations, and it already outperformed everything the mathematicians had. Later analysis showed that looking for conditions that allowed to apply tricks made the system slow.

Our head of IT has taught me this saying: "Don't be clever."

[cherubinth]

All examples had an absurd amount of data too. Alpha creates its own infinite data and OpenAI example has the giant english corpus. What about situations where we have little data, but know very well how it works?

[Bonesters]

In my AI class in college, this is one of the biggest lessons that people had to learn. When making an AI play connect 4 using a min-max search with alpha-beta pruning, it worked better to only check if you won or lost, instead of trying to evaluate the current board in more detail. The more simple your evaluation of the board, the more turns ahead your AI could look, and the better it performed.

[ABaumstumpf]

AI-research ended up at Quantum-mechanics: "Shut up and calculate"

[oktaycomu]

I think I would mostly agree with Sutton's points. He isn't saying that we should just idly wait around for the next step in hardware improvements, but we should work towards optimizing our search and learning algorithms so that they can make use of the hardware we have.
His point is also analogous to human learning. I remember hearing this example somewhere: When you ask a tennis player to analyze their own serve and explain each muscle movement that takes place, you will notice a significant drop in their performance. Thinking about and paying conscious attention to an action often inhibits performance. In reality, that tennis player got probably got some good tips and tricks from their coach, but, in the end, what allowed them to master the serve was performing the serve over and over again. With each new serve the brain unconsciously makes tiny adjustments and can understand which adjustment contribute to the performance and which ones don't. The tips and trick might be an initial boost to performance, but real mastery comes from repetition .

[my_temporary_name]

The same problem exists in bringing up children. People used to force strict rules on them instead of relying on autonomous learning. It works great in the short term (a child terrorized with rules doesn't cause many problems) but can be devastating in the long term (passive people with no motivation, permanently frustrated people and even murderers are created this way).

[TwoMinutePapers]

You can also weigh in on this topic in this twitter thread:

[foxps]

In the spirit of favoring simple first principles: look to thermodynamics - you can't reduce complexity, you can only displace it. You can certainly build in an optimization, and it will give you a head start over brute force, but you better be right about it being the global minimum.

[NeWincpp]

He put the finger on my biggest problem on all neural network on this world (including humans) and why I don't want to spend any time on AI yet. I don't really care about how better the algorithm produced is better. What I do care is understanding why. Basically what he is saying is that we should put humans out of the equation but I'm not sure we (as humans) will gain anything if we just get an all mighty do everything tool. Especially me, I'm not just gonna sit and wait it work for me. We are all here because we love to understand and go deeper yet we are building a tool that generate black box algorithms. For me AI will because a valid tool when they will be able to publish their work and write tutorials about it or it's basically as useful as this hero-programmer in your studio that say nothing and just push on your code and don't want to explain what was the bug.
Sure the project goes faster and so what ? I'm not doing this project just to ship it I'm doing it because I wanna be a better human I wanna build the adventure of searching and building with all humans interaction with it. I want to put parts of me in it so diversity and singularity are still value that count just like efficiency.

[JasonShermanYouTube]

5:30 - This is my favorite parenting tips channel

[AexisRai]

Thoughts:
1) This seems to clearly echo the sense that people outside the field who call AI safety researchers alarmist simply don't understand how unbound by our human sense of what's possible AI really is, how unrelated its sources of effectiveness are to the ways people think their intelligence works.
2) I can't help but think that when he says "stop trying to design methods, just implement meta-methods that can learn the methods", that this is somehow pointing out that the meta-methods are actually more basic, and this whole class of research failure is only because we don't recognize this more basic principle as a primitive of our own cognition, even though it might be.
3) "Don't solve any concrete problem with special methods, solve everything with general methods" does not seem like the right takeaway here; if special methods solve the particular problem faster and are less unwieldy, use them. I think this is more about where the fruitful research directions are than what the fastest path to a working system is. ...maybe?

[iestynne]

I love these more reflective videos. You've reviewed enough papers at this point to draw some pretty insightful conclusions about broad trends.

[brantworks]

Makes perfect sense. I’m sure consciousness is built on massive compute to pattern search and predict. Compute, test, repeat.

[frankx]

"Smart kids don't need help from dumb dads"

[wolframstahl]

I had a good laugh at the AI-written article at 5:00. It looks like the AI has learned that _"The [incident] will have significant negative consequences on public and environmental health, our workforce and the economy of our nation"_ is just a great all-purpose sentence that someone will always say about some unfortunate event.

Overall I'm really impressed with the article and the amount of reasonable information that was not contained in but deduced from the human-written prompt, like which departments would be involved, where the nuclear material came from, and that _"the Nuclear Regulatory Commission did not immediately release any information"_ which would very likely happen in an event involving theft of nuclear material, but would also not be relevant information in most other types of incidents.
Amazing stuff!

[MobyMotion]

It's hard to disagree with the sentiment of this paper, given the number of times hand crafted algorithms have lost out to computational equivalents in the last few years. But that makes it all the more bitter - the entire field of machine vision used to be based on hand crafting these ways to extract meaning from pictures, and it's now almost obsolete. I wouldn't like to be one of those researchers who's whole career has been made redundant