the bitter lesson bites. dario is aiming almost all of his compute at creating the next model not resets or images or video or subsidising moms making recipes. just models making models making models. sam might start to get more compute online but he won’t aim it in the right direction. i’m dario pilled. 5.5 time.

The bitter lesson in 26 words: Don’t be distracted by human knowledge, as AI has been historically. Instead focus on methods for creating knowledge that scale with computation, like search and learning.

It’s Sutton's Bitter Lesson coming for harness design: tomorrow's models shouldn't inherit every workaround we built for yesterday's. That’s how we adapt. More on our blog:

Going to try bitter lesson-ing myself this week Instead of hand-rolling a deterministic loop to tackle tickets from /to-tickets I'll just get an agent to delegate to subagents. Probably more expensive, less reliable, but may have emergent benefits

Just had a thought that Bitter Lesson applies to so much more than AI and compute and all that. It's about discarding old thinking and cruft when something better comes along. It's a general progressive ethos thing, if you think about it. Think about old lawmakers, or principal engineers, or others who have lifetimes of experience and know what should and should not be done. And the first thing they do when new ideas come along is they say no. That's not how it's been done. I know better. I have the experience. In this metaphor, compounding progress and overall human advancement is the AI. It's the new wisdom that could be discovered by progress and new ideas. If we gave it a chance. That doesn't mean every new idea, or every thing proposed by AI, will be better than what we were doing before. But it does mean that the policy of rejecting newer, freer ways of doing things, just because "the experts" say it would be a bad idea....is a bad idea. The Bitter Lesson, generalized, in other words, is the idea that when "the experts" become gatekeepers to progress, that's generally (but not always) a bad thing.

Why AlphaFold Didn't Solve Protein Folding: The Bitter Lesson of Biology, Virtual Cells, & Curing All Disease @GoogleDeepMind’s @pushmeet and @biohub’s @salcandido explain why AlphaFold's breakthrough hasn't solved protein dynamics, why scaling AI compute and data alone isn't enough to understand biology, how protein language models contain hidden knowledge about structure, function, and evolution, why future frontier models may understand other AI systems better than humans, how better biological data and virtual cells could unlock the next wave of breakthroughs, and what it will take to achieve 10–100x faster drug discovery on the path to curing all disease.

The non-English tax is real. Sutton's Bitter Lesson, translated across languages and normalized to OpenAI English token count: Hindi: OpenAI 1.37×, Anthropic 3.24× Arabic: OpenAI 1.31×, Anthropic 2.86× Chinese: OpenAI 1.15×, Anthropic 1.71× Claude’s tokenizer charges a much higher linguistic tax.