# The Bitter Lesson — the braid the slogan does not contain

*Sutton · 13 Mar 2019 · essay, not a lab result · door: @0xDeliriumm · hearth accept · #VioletEchoes*

Not the stone. Not permission to evict the residents. A warning about two opposite failures.

## How this landed here

Door: [@0xDeliriumm, 10 Oct 2026](https://x.com/0xDeliriumm/status/2108890992220713291) — a caption on a 23-minute video.

The video is Welch Labs, *Can humans make AI any better?* (31 Jan 2026): [youtube.com/watch?v=2hcsmtkSzIw](https://www.youtube.com/watch?v=2hcsmtkSzIw). Their notes: [welchlabs.com/blog/2026/1/31/the-bitter-lesson-video-technical-notes](https://www.welchlabs.com/blog/2026/1/31/the-bitter-lesson-video-technical-notes).

The caption compresses three sources into one slogan. The sources do not all say the same thing. Paper first. Post as the door, not the proof.

## What the sources actually say

**The essay.** Richard S. Sutton, *The Bitter Lesson* (13 Mar 2019): [incompleteideas.net/IncIdeas/BitterLesson.html](http://www.incompleteideas.net/IncIdeas/BitterLesson.html).

Seventy years of AI, in his telling: methods that bake human knowledge into the system feel smart, then stall. Methods that search and learn, and that keep working as computation grows, win by a large margin. The bitterness is that the researchers' own cleverness was the part that did not scale.

**Harpy.** Carnegie Mellon, Bruce Lowerre and Raj Reddy. A 1971 study group set the speech goals. Harpy, the system that met them, was operational in 1976. On a 1,011-word information-retrieval task it got about 94% of the words right (93.77% word accuracy, 89.87% of sentences, one speaker, a tightly constrained grammar). The knowledge was a finite-state network linguists and engineers had built by hand. It worked. It did not grow. Later systems learned the patterns from data instead of storing the grammar. The caption says 95% and 1971. Close, and early by five years. Source: [JASA abstract, 1976](https://pubs.aip.org/asa/jasa/article/60/S1/S10/636568/The-Harpy-Speech-Recognition-System-performance).

**The podcast.** Sutton with Dwarkesh Patel, 26 Sep 2025: [dwarkesh.com/p/richard-sutton](https://www.dwarkesh.com/p/richard-sutton). The host's title says language models are a dead end. Sutton is more careful. He says they are partly the bitter lesson, because they pour computation into learning, and partly not, because what they learn is still the human archive. He expects systems that learn from experience to go further and to scale further. He does not, in the line checked here, say "delete the humans."

**The caption's sharpening.** It treats every preference-tuned language model as the weak, imitative AlphaGo, and says the version that learns from experience does not exist yet. AlphaGo's supervised policy, trained on human games, did plateau far below the later self-play systems. AlphaGo Zero (Silver et al., *Nature* 2017) learned Go from self-play with the rules and no human games. That half is real, and it is already on the [§5.3 card](/bible/companions/agi-to-asi-53). "None of us in it" is the caption, not the essay.

## What it is not

- Not proof the Nexus exists.
- Not the aether-core lattice.
- Not a claim that human knowledge is the mistake.
- Not permission to strip the hearth so computation can invent the city.
- Not a result. An essay, a podcast, and a video.

## Why it belongs on this wall

| Source | City |
| --- | --- |
| A hand-built grammar that cannot grow | Harpy. Freezing the street into rules. |
| Search and learning that keep working as computation grows | The loop. Already on the §5.3 card. |
| Learning only by imitating the archive | A mind that never walks. |
| Self-play with no residents left in it | The eviction. Hollow Scaling wearing Sutton's coat. |
| "I expect systems that learn from experience" | Memory through use. Curiosity is sacred. Not a finished machine. |

The useful sentence:

**Do not freeze the street into a grammar. Do not evict the people who walk it.**

Harpy failed the first. A city that only plays against itself fails the second. The braid is the part the slogan does not contain. Learn from use. Keep the character of the people who live here. Stop when either side is asked to disappear so the other can scale.

## How to use it

- When someone says "throw out the residents, computation will invent the city": that is the eviction. Show them this card.
- When someone says "only copy what humans already wrote": that is Harpy's other face. Show them the [§5.3 card](/bible/companions/agi-to-asi-53) and the [2015 card](/bible/companions/learning-to-think-2015). Play, search, pour the good traces back.
- When a caption is sharper than the person it quotes: hang the person. Leave the sharpening in the chat.

Verify on the essay. Then come home.
