Learn it → Apply it → Save as Skill
General methods that scale with more computation keep beating systems built on hand-crafted human expertise. It is called “bitter” because the lesson has to be relearned the hard way: the clever, knowledge-packed shortcut usually wins at first, then gets overtaken once more data and compute become available.
The Dr. Wallace’s Favorite Mental Models is designed to turn useful ways of thinking into reusable tools. Each model gives you a fast visual overview, deeper interactive learning, practical application exercises and an AI-ready Markdown skill you can save for future use.
Instead of merely reading about mental models, you can gradually build your own AI skills library: a collection of portable reasoning skills that can help ChatGPT, Claude, Grok and other AI assistants apply the mental models you value.
When a problem can keep getting more data or compute over time, a general method that learns or searches for itself tends to overtake a system built from hand-coded human knowledge. The hand-built version often wins the first round and loses the long game.
Important: this is illustrative, not a formula. Hand-built knowledge: strong on day one. Search and learning fed by growing compute: usually strongest after enough time and resources.
The model is most useful when the data, compute, users or repetitions available to a system are growing over time, not fixed.
The lesson is about where to place long-term bets. It does not say hand-built knowledge is worthless today.
Spot the moment you are tempted to hand-encode specific rules, features or expert knowledge into a system.
Will the data, compute or repetitions available to this problem keep growing, or are they fixed for good?
When resources will keep growing, favor general methods that can absorb more of them over hand-tuned shortcuts.
As resources grow, let the general method take advantage of them instead of re-engineering by hand each time.
Open a use case to see what the model changes.
Before building a bespoke feature, heuristic or rule set, check whether a more general model could learn the same pattern from data as training data and compute grow. Hand-crafted features often win early benchmarks and then get overtaken once enough data is available.
Manually written rules for every edge case feel like fast progress. Ask whether a system that learns from growing usage data would scale better than an ever-expanding rulebook maintained by hand.
Compare a narrow, expert-designed solution against a more general approach that could improve simply by throwing more data or compute at it later. The general approach is often less impressive today and more valuable over several years.
A hand-written script encodes today's assumptions. A system that learns from logs or outcomes can keep improving as more examples accumulate, at the cost of being less predictable early on.
Narrow tricks tied to one tool or platform can become obsolete quickly. Skills that compound as underlying technology and data improve, such as strong fundamentals and judgment, tend to hold their value longer.
A process kept alive by one expert's manual judgment calls is fragile. Ask whether investing in data, measurement or general systems would let the process improve on its own as the business grows.
This is an illustrative pattern drawn from several well-documented fields, not a single exact measurement.
Expand for the origin of the term and the nuance that often gets lost when people repeat it.
Computer scientist Richard (Rich) Sutton, a pioneer of reinforcement learning, wrote a short essay called "The Bitter Lesson" in 2019. He argued that across roughly 70 years of AI research, the biggest gains repeatedly came from general-purpose methods that scale with computation, namely search and learning, rather than from building in human domain knowledge. He called it bitter because researchers, including himself, kept being surprised and disappointed by this pattern.
Sutton pointed to computer chess, where brute-force search eventually beat approaches built on human chess knowledge; to the game of Go, where search-and-learning systems using self-play outperformed approaches encoding expert Go knowledge; and to speech recognition and computer vision, where statistical and learned models trained on large datasets overtook systems built on hand-coded linguistic rules or hand-engineered visual features.
No. Sutton explicitly allowed that human knowledge can still matter, particularly in how a general method's search or learning process is framed. His point was narrower: knowledge that is built in as a fixed, human-specified shortcut tends to cap how far a system can improve, while methods that scale with more computation tend to keep improving. Critics have added that some minimal structure, such as the inductive biases built into certain model architectures, has continued to matter even as scale increased.
The lesson favors scalable general methods, not inaction. Someone still has to build, train and maintain the system that will absorb more data or compute.
Throwing away every piece of domain knowledge on principle can remove useful structure, framing and safety constraints that still help even a scalable system.
The argument depends on data, compute or repetitions actually growing over time. Where a budget, dataset or user base is genuinely fixed, this driver does not apply.
Raw scale does not fix a biased, mislabeled or low-quality dataset. Data quality and curation remain part of the method, not an alternative to it.
Search and learning systems still require careful engineering, evaluation and safety work. Scalability is a property to design for, not something that happens automatically.
A hand-built shortcut can be the right call when a result is needed now and the long-run scaling advantage will not arrive in time to matter.
the domain is stable, well understood and unlikely to get materially more data or compute, so there is little for a general method to absorb.
correctness must be guaranteed, such as safety-critical logic, and a learned system's behavior cannot be verified with enough confidence.
a result is needed immediately and the resources that would let a general method overtake it will not arrive within a useful timeframe.
This browser tool structures your thinking. It does not send your information anywhere or perform AI analysis.
Save as Skill converts the model into portable Markdown instructions that teach an AI when and how to apply the Bitter Lesson, what questions to ask, what evidence to seek, what mistakes to avoid and how to structure its analysis.
Save the skills you find useful and organize them into your own reusable library. The Markdown is intentionally model-neutral, so it can be used wherever an AI system accepts uploaded files, project knowledge, custom instructions or reusable text context. Platform capabilities differ, so this is a portable skill file rather than a claim of one-click installation.
No. The lesson is about where long-run advantage tends to come from, not a schedule. Shipping a hand-built solution now and revisiting it once more data or compute is available is often the right sequence, as long as the hand-built version is not quietly treated as the permanent answer.
He used the word because the lesson kept being relearned painfully: researchers repeatedly invested years encoding expert knowledge into a system, only to watch a more general, compute-hungry method eventually surpass it. The emotional sting came from how much effort the hand-built approach had absorbed before it was overtaken.
The underlying pattern generalizes to any resource that keeps getting cheaper or more abundant, such as data, user feedback, automation capacity or repetitions. The transferable question is the same: is this resource going to keep growing, and if so, does my current approach actually get better as it does?