⏳ The Bitter Lesson

In 2019, the researcher Rich Sutton wrote a short essay called "The Bitter Lesson" (the Bitter Lesson). The thesis: over decades of machine learning (ML), whenever researchers tried to embed “clever human knowledge” into the system, in the long run they lost to simple approaches that just threw more compute at the problem. Chess, Go, speech recognition, vision: in every case, brute-force computation outperformed handcrafted optimizations.

Why “bitter”? Because it’s humiliating for someone who’s smart. We wants that human ingenuity should prevail. But the engine behind the lesson is simple and relentless: compute grows very quickly and gets cheaper year after year. Your manual optimizations are fixed; compute is a tide that keeps rising. As Pocock sums it up in the video, the lesson is “trust that the underlying thing will improve"—the model will get stronger on its own, so it’s not as worthwhile to polish things by hand. Common mistake: thinking that the Bitter Lesson is a magic law that applies to everything, always, with no limits. It isn't — it's about a dimension (the model’s brain). Keep this caveat in mind: the rest of the module is precisely about where it no applies.

From the Bitter Lesson comes a dangerous temptation: "if the model is going to improve on its own, why should I bother? Just wait for the AGI come in and solve everything in my place." Pocock is direct about this in the video: "standing around waiting for AGI without doing anything was a really dumb idea." The point is deliberately strong—he admits he flirted with this attitude himself and realized he was wasting time.

The waiting problem has three layers. First: AGI may not arrive when promised — AI promises are always delayed. Second: even if a much better model arrives, it doesn't fix your messy codebase or your vague prompt — these parts (the harness) are still your responsibility. Third, and the cruelest: while you wait, you doesn't develop any skills. Whoever takes action while waiting reaches the “future” with sharp fundamentals; whoever just waits arrives with nothing. Waiting seems like the smart choice (let the machine do it), but it costs the most.

Where the Bitter Lesson is true: in the model’s brain. Training a larger model with more data and more GPUs really does deliver gains that no manual adjustment can match. That’s why David, in the interview, asks the obvious question: "why not both? Changing the engine improves everything instantly." He is right—a better model raises the whole tide: every task gets a little easier without you changing anything. Denying that would be foolish.

Where it has a limit: compute improves what is inside of the model. It no improves what’s outside — your vague prompt, your labyrinthine codebase, the missing skill, the wrong permissions. These parts of the harness don’t get GPUs; they only improve when you improves them. And there's one detail that changes everything, which Pocock uses all the time: a well-built harness lets a model cheaper deliver the same work— "have a codebase that's easier to change and you can use a dumber model to do the same job". In other words: improving the harness is a way to gather the compute gains without paying for the most expensive model. The two sides add up—they don’t compete.

The way out of the trap has a name: fundamentals. Pocock insists: "people focus on the wrong thing—the shiny new toy—when they should focus on what has been working for 30-40 years." Organizing code, writing a clear request, writing tests, and documenting well: none of that expires when a new model comes out. On the contrary—the better the model, the more it benefits from a well-maintained harness. The fundamentals are an asset that appreciates; the hype around this week’s model is what evaporates.

There's one more practical and powerful reason to act now: the setup agent-agnostic. If you build your prompts, skills, and codebase in a way that doesn't tie you to a single model, so—when the model improves—your work today still applies and it gets even better. It’s the best of both worlds: you get the benefit of the Bitter Lesson (a stronger model becoming available) without having sat idle waiting for it. Pocock ends with humility: "I'm not a guru/commentator — I'm doing the best I can with what I have right now." That’s the spirit: work with what exists today instead of betting everything on an uncertain tomorrow.

Here's the heart of the module, and where many people misunderstand Pocock. He no says "ignore the model, just work on the harness." It says it’s 50/50 (and not 90% model / 10% optimization, as most people think). When David challenges him — "use both: the best model AND the best harness" — Pocock agrees in the end: "it's 50/50; the problem is thinking through the model first." The order matters. If you start with the model, you lose focus on the fundamentals: "if I over-optimize around a model, I lose focus on the fundamentals."

Notice the nuance: Pocock admits that itself you may be falling into the Bitter Lesson by spending so much energy on the harness instead of simply trusting that the model will improve. It doesn't hide this risk—and that honesty is what makes the balance reliable. The practical synthesis that follows: improve your harness now (you control it; that's half the game) and use the best model available (it’s the other half, and you get it as a free gain when you switch). David describes exactly this mindset: actively improves the whole setup every day e uses the strongest model available. It’s not “or”—it’s “and.” Falling into the trap means treating it as “or.”

Wrapping up the module with a practical approach: the Bitter Lesson is real, but as a principle of AI training — not as an excuse for you to sit still. Whenever the temptation to "wait" hits, run the decision through the filter below. It turns the fear of "am I falling into the Bitter Lesson?" into an objective check. Copy and paste this when you catch yourself waiting:

Peguei-me "esperando o modelo melhorar"? Rode o filtro:
[ ] É treino de modelo (interno)? -> ok, isso melhora sozinho; nao e minha tarefa.
[ ] E harness (prompt/skill/codebase/permissao)? -> NAO espera; e comigo, agir HOJE.
[ ] Meu setup esta agent-agnostic? -> se sim, o ganho do modelo futuro ja esta garantido.
[ ] Estou usando o MELHOR modelo disponivel agora? -> use; e a outra metade (50/50).
Regra: pense pelo HARNESS primeiro. E "E" (modelo + harness), nunca "so esperar".