123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214

No. 002 · 2026-10-14 · ESSAY

SOURCE Rich Sutton (March 13, 2019) · read the original ↗

The Bitter Lesson

My Understanding

I started reading ‘The Bitter Lesson.’ I read the first page, where he talks about how we keep ignoring and re-ignoring this lesson again and again. He gives the example of the computer chess champion, and then the gaming example, where relying on human knowledge did not work out over the long run. I read up to the speech-recognition example, where he says methods that learned from data and computation worked better over time. I stopped before computer vision and need to continue from there. I started reading the next two paragraphs, where he compares computer vision to a similar pattern and then says that the lesson has to be learned repeatedly. He says that hard-coding a neural network with our assumptions about edges, shapes, or how vision should work has limitations. If we build in too much of our own understanding, we may prevent the system from learning. We make it too rigid. It feels like a general point that this is not supposed to be so constrained or so heavily designed by people. Computation will keep growing. We cannot put so many restrictions on that growth, and if we do, it is not going to help us in the end. I thought the essay was written in 1990. Then I realized it was written in 2019, which is not that far back. The 1990s were part of the examples he was discussing, especially in speech recognition. What he is saying is that we have to keep learning this lesson. We ignore it again and again, so he gives several examples of what happened in the past. The early cases were the chess example with IBM’s Deep Blue and then Go, which I had already read about yesterday. Apparently, IBM’s team did not get much credit at the time. I worked at IBM, but that was many years after the Deep Blue match. The systems people initially designed did not always give the expected output. Eventually, they had to stop expecting the system to understand everything through human reasoning and let it work more through computation, search, and learning. At some point, people feel good about building human knowledge into a system. They think, ‘Oh, this is a good thing,’ and feel that it gives the system more strength. That can help at first. Later, though, the system may stop growing and become stagnant. This is a personal reflection in the essay too. Sutton says that people have seen this happen again and again but still forget it. He does not want AI researchers to fall into the same trap while creating what comes next. The gains from building in human knowledge can be worthwhile, but over the long run, they can be a plateau or even an obstacle. That line is brilliant because it says what we keep doing. AI researchers have often tried to build knowledge into agents. It can help initially and feel satisfying, but later it plateaus or gets in the way. Breakthrough progress can come from the opposite approach: computation that scales through search and learning. That, I think, is the bitter lesson. I finished it. In the last two paragraphs, he says that one important thing is that methods should keep scaling with computation, such as search and learning. He also says that we should let go of this need to make systems work through our own human reasoning about how the world should be understood. The world is complicated, and these systems may understand or navigate its complexities in ways that are not natural to us. We need some detachment from insisting that they reason as we do. That final thought is interesting: we want AI agents that can discover, as we can, rather than agents that only contain what we have already discovered. They should not be confined to our current discoveries. That was something. Rich Sutton’s name felt familiar, and then I remembered Reinforcement Learning: An Introduction. He co-authored that book with Andrew Barto. We were introduced to it during my reinforcement learning course last spring, so that is where I had seen his name before.

Agreement / Disagreement

I also like that this is not a pitch for one particular method. He has looked at the pattern and given examples. We have made a mistake, and he is asking us to recognize it. It feels almost like he is pleading, ‘This has already happened. Please, not again.’ I do not think there is anything here that I particularly disagree with. He is saying what happened, and then giving his perspective on it.