12 · AI & Machine Learning

This chapter is the tower's other stairwell — it touches every floor, and hides the oldest assumption in the tower: that somebody, somewhere, wrote the rule down.

Can behavior be grown from data instead of written by hand?

In July 1959 the IBM Journal of Research and Development published Arthur Samuel's account of a checkers program that had started beating him (Samuel, 1959). This should have been impossible in the terms the rest of this book has used. Every floor below is built from rules a person stated: a gate's truth table, a protocol's state machine, a query optimizer's cost model. Samuel had stated the rules of checkers, not the rules of playing checkers well — that part the program accumulated by playing, adjusting the weight it gave to board features according to which ones had preceded winning. Its skill existed nowhere in his source code. It existed in numbers no human chose. Everything since has been an argument about how far that idea goes.

Learning means generalizing, and generalizing is the only hard part. A system that reproduces its training data perfectly has learned nothing; a lookup table does that. The measure that matters is performance on data the system has never seen, and the gap between the two — the model fitting noise it should have ignored, "overfitting" — is what every technique in the field is ultimately fighting. Wolpert and Macready's No Free Lunch theorem shows that averaged over all possible objective functions, every learning or optimization algorithm performs identically (1997). Every bit of real learning ability comes from assumptions baked in about which problems are likely — that nearby pixels are related, that word meaning depends on context, that the world is smoother than it could be. A learner with no prejudices cannot learn. Choosing the prejudices is the engineering.

If every step is differentiable, the machine can debug itself. Frank Rosenblatt's perceptron (1958) could adjust its own weights, and the US Navy's press conference promptly told the New York Times it expected a device that would eventually walk, talk, see, write, and be conscious of its existence. It could not compute exclusive-or, as Minsky and Papert demonstrated in 1969 for the single-layer case — a limit historians now think did far less damage to the field than the surrounding funding politics (Perceptrons). Stacking layers fixes the expressiveness but creates the credit-assignment problem: with millions of weights and one number for how wrong the answer was, which weight was at fault? Backpropagation answers it by running the chain rule of calculus backwards through the network, computing for every weight at once how much a small nudge would have reduced the error, at roughly the cost of one forward pass. Rumelhart, Hinton and Williams brought it to the field's attention in Nature in October 1986 (323: 533–536); Seppo Linnainmaa had published the underlying method as reverse-mode automatic differentiation in 1970 and Paul Werbos had applied it to neural networks in a 1974 thesis. This is the floor's real trick, and it is smaller than its reputation: a program you can take the derivative of is a program improved by measurement rather than by argument.

What gets learned is the representation, not just the answer. For decades the labour of machine learning was feature engineering — humans deciding which measurable properties of a photograph or a sentence the statistics should chew on. The shift was to let those intermediate descriptions be learned too (Bengio, Courville & Vincent, 2013). The public demonstration came in 2012, when Krizhevsky, Sutskever and Hinton's convolutional network cut the ImageNet top-5 error rate to 15.3% against the runner-up's 26.2% (NeurIPS 2012) — a margin large enough that a research community changed direction within a year. Nobody had told it what an edge or a texture was. Those appeared in its early layers because they were useful.

Scale beat cleverness, repeatedly, and the people who noticed said so plainly. Rich Sutton's essay of 13 March 2019 states the pattern without hedging: "The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin" (The Bitter Lesson). Bitter, because the human insight painstakingly encoded into a system tends to be the part that has to be thrown away. But scale is not a slogan; it is a quantity with a recipe, and the recipe has already been corrected once. Kaplan and colleagues charted how loss falls with model size, data and compute (2020); two years later the Chinchilla result showed the prevailing practice had the ratio badly wrong — for a fixed compute budget, models were being built too large and fed too little, with roughly 20 training tokens per parameter closer to optimal (Hoffmann et al., 2022). The lesson survived. The arithmetic under it did not.

Attention made scale usable. Sequence models used to force a whole sentence through a fixed-size bottleneck, one step at a time, remembering worse the further back it went. The 2017 architecture proposed by eight authors at Google replaced the recurrence entirely: let every position in a sequence look directly at every other and weight what it finds relevant (Attention Is All You Need). What mattered was less memory than hardware — with the sequential dependency gone, training parallelizes across the machines of chapter 04, and the bitter lesson's compute could actually be spent. (Readers who want that one idea traced from 2014 onward will find it is the subject of its own book in this library.)

The limits are structural, not embarrassing. These systems produce fluent, confident, false statements — "hallucination," surveyed rigorously by Ji and colleagues (ACM Computing Surveys, 2023). One prominent argument holds this is not a bug to be patched but the predictable result of training and benchmark regimes that reward a confident guess over an admission of ignorance (Kalai et al., 2025) — one live account among several, not a settled verdict. The deeper issue is that these systems optimize the objective you gave them, which is never quite the thing you wanted. That gap is why the field built machinery to train against human judgments of quality instead (Christiano et al., 2017; Ouyang et al., 2022) — which relocates the problem into the question of whose judgments, rather than dissolving it. And because the behavior is grown from data, the data is itself an attack surface: poison what a system learns from and you have programmed it — chapter 11's adversary, one floor up.

The field's own institutions have now conceded the point. Hopfield and Hinton took the 2024 Nobel Prize in Physics for the foundational work enabling machine learning with neural networks; the Chemistry prize that year went to Baker, Hassabis and Jumper for computational protein design and structure prediction (Nobel). A method for growing behavior from data had become an instrument of natural science.

Connections

Below: chapter 02, because a gradient step is an algorithm with a cost — and training is dominated by exactly the matrix multiplication whose exponent that floor is still chasing — and chapter 04, because this floor's history is the history of what hardware made affordable — 2012 was not a new idea but an old idea a graphics processor could finally afford. Chapter 03 supplies the framing of a model as compression, and chapter 01 draws the outer boundary: learning does not repeal undecidability. Sideways: like security, learning from data is not really a floor with a single tenant. It has been quietly installed on nearly every floor — branch predictors in 04, query optimizers in 08, congestion control in 09 — which is why chapter 00's map gave it a stairwell rather than a floor.

Open questions

Does the bitter lesson keep holding, or was it a description of one era with unusually cheap compute — and how would the field tell the difference before the fact? Is generalization from data a route to understanding, or a good imitation of it that fails in ways we notice only afterwards? And the question chapter 00 left standing: if people increasingly command machines in ordinary language, is natural language becoming the newest floor of the tower — a surface under which every abstraction in this book is hidden, including the leaks?

Go deeper

• Ian Goodfellow, Yoshua Bengio & Aaron Courville, Deep Learning, MIT Press, 2016 — the standard graduate text, full text free at deeplearningbook.org.
• David Rumelhart, Geoffrey Hinton & Ronald Williams, "Learning representations by back-propagating errors," Nature 323 (1986): 533–536 — nature.com.
• Jordan Hoffmann et al., "Training Compute-Optimal Large Language Models," 2022 — arXiv:2203.15556.