One idea gets quoted in AI more than any other: the "bitter lesson." Its claim is simple and uncomfortable — over the long run, general methods that leverage more computation beat clever, hand-crafted approaches that encode human knowledge. It's bitter because researchers keep learning it the hard way.
The core argument
Again and again, AI history shows the same pattern. Experts build systems full of hand-coded human insight (chess strategies, grammar rules, vision features). Those systems work for a while, then get crushed by simpler, general methods that just learn from data and scale with compute. The lesson: don't build in what the system can learn; build systems that improve as computation grows.
Human cleverness feels satisfying and doesn't scale. Brute general learning feels unsatisfying and does. Compute keeps winning.
Does it still hold in 2026?
Largely, yes — the transformer is a general architecture that scaled, and the biggest wins keep coming from more compute and data plus general methods, not hand-engineered features. But there's nuance. The pure "just scale it" era is showing diminishing returns; recent gains come from how we spend compute (reasoning, better data, post-training) — still general methods, but more targeted. And efficiency and data quality now matter alongside raw scale.
The refined lesson
The 2026 version might be: general methods that scale still win, but which axis you scale matters more than before. Raw size alone is hitting limits; scaling reasoning, data quality, and inference compute is where the leverage moved. The spirit holds — bet on general, learnable, compute-hungry methods over hand-crafted cleverness — but the naive "bigger is all you need" reading needs updating.
Why it matters
The bitter lesson is a strategic compass: when choosing between encoding human expertise and building something that learns and scales, history favors the latter. It's a warning against over-investing in cleverness that computation will eventually render obsolete — and a reminder to keep asking which axis is worth scaling next.
