A puzzle the last three lessons leave behind
The previous lesson, Verifiability, gave you a rule for where AI surges: wherever an answer can be automatically checked. But sit with a stranger fact. A model can score superhuman on a hard maths or programming competition — a genuinely checkable, genuinely difficult test — and still cannot be trusted to improve your codebase with any taste, or reason that walking to a car wash fifty metres away defeats the point of having a clean car. It aced the exam. It did not become generally smarter.
That is the crack this lesson is about. The first three lessons built one side of the argument: general methods plus scale win (The Bitter Lesson), scaling reliably improves a model (Scaling Laws), and it improves fastest where answers are checkable (Verifiability). Now meet the strongest objection to that whole picture — the one Ilya Sutskever, who co-authored the results that launched the scaling era, calls the most fundamental unsolved problem in the field.
'These models somehow just generalise dramatically worse than people. It's a very fundamental thing.' — Ilya Sutskever
What 'generalise worse' actually means
To generalise is to take what you learned in one place and use it somewhere you have never been. It is the difference between a student who memorised ten past papers and one who understood the subject: both pass the mock, only one handles a question phrased in a way they have not seen. The generalisation gap is the claim that today's models are far more like the first student — and it shows up in two measurable ways.
- Sample efficiency — how much they must see to learn. A frontier model is trained on trillions of words, more text than a person could read in a thousand lifetimes. A five-year-old learns to recognise cars from a handful of examples, and a teenager drives competently after roughly ten hours behind the wheel. The human gets there on a rounding error of the data.
- Robustness — how well the learning holds up off the beaten path. A person mentored into a craft picks up the way of thinking and carries it into situations no one drilled them on, with no grader marking each step. A model needs a bespoke, hand-built training environment for each new skill. As Sutskever puts it, 'the robustness of people is really staggering.'
Why you can't blame it on evolution
The tempting rebuttal: of course humans learn vision and movement from almost no data — hundreds of millions of years of evolution pre-wired us for them. Fair enough for seeing and walking. But it collapses on the cases that matter most. Humans are also fast, robust learners at maths and programming — skills far too recent for evolution to have hard-wired any head start. If we learn those efficiently too, the advantage cannot be a stack of inherited tricks. Something more general is going on. Sutskever's blunt reading is that this
'...is more an indication that people might have just better machine learning, period.'
In other words, the human brain may be running a fundamentally better learning algorithm than anything the models use — not more knowledge, a better way of acquiring it from little. That is a claim about the recipe itself, and it is why the gap is unsettling rather than merely inconvenient.
Why 'just add more environments' doesn't close it
Here is the move a scaling optimist reaches for, and why it doesn't quite work. If a model is superhuman at coding competitions but a poor everyday programmer, why not simply train it on more kinds of tasks — stack up more checkable environments until the gaps fill in? The catch is that this treats the symptom. You can keep adding the 'ten-thousand-hour competitive programmer', over and over, and still not get the hundred-hour student who merely gets it. Sutskever's own framing of the question is the sharp one: if being superhuman at coding competitions doesn't make you a more tasteful programmer more generally, maybe the answer is not to keep stacking environments —
...but to figure out an approach which lets you learn from one environment and improve your performance on something else.
That last phrase is generalisation. Which is why this lesson sits underneath the previous one: jagged intelligence — brilliant here, clueless there — is the visible surface of the gap, and stacking more verifiable rewards paints over the holes without fixing what causes them.
What might actually close it
Nobody has the answer, but Sutskever names one suspected missing ingredient worth understanding: a value function. Today's reinforcement learning often waits until the very end of a long attempt, then hands out one final grade and smears it back across every step — the 'supervision through a straw' from the last lesson. A value function is an internal sense, at every moment, of whether things are going well or badly — the teenage driver who feels the car drifting and corrects, with no instructor calling it out. Sutskever's striking analogy is that in humans this internal signal is roughly what emotions are: a simple, ancient, remarkably robust running estimate of how you're doing. A learner with that can improve mid-course instead of only from the final tick.
This is why Sutskever argues the field has left the age of scaling — where the recipe was known and the only question was how much compute to buy — and re-entered an age of research, 'just with big computers'. If the gap is real, the winning move is no longer to pour in more of the same; it is to find the better learner.
What it means — and the honest caveat
This is the fourth lesson in a set, and it turns the argument. The first three tell you the scaling machine is real and predict where it bites. This one tells you what that machine may not reach on its current recipe: reliable, sample-efficient, robust learning of the kind a child does without trying. It is the crux of the wider debate laid out in The Road to AGI — whether the distance left is quantitative (more compute, more environments, same recipe) or qualitative (a learning principle we do not yet have).
- Don't read a benchmark score as general competence. Superhuman on a checkable test means superhuman on that test. It does not promise good judgment one step outside the test's shape — the exact place real work lives.
- Watch how much a system had to see. A capability that took trillions of examples to instil may be brittle where a human's cheaply-learned version is robust. Sample-hungry often means beaten-path-only.
- Hold the caveat honestly. This is a live disagreement, not a settled verdict. The scaling optimists point to reasoning models and inference-time compute as evidence the curve has not bent; and Sutskever's own company is staked on the gap being real, which is a reason to weigh the claim, not to swallow it. The gap is easy to assert and hard to prove in advance.
Go deeper
The single best source for this idea is Ilya Sutskever's conversation with Dwarkesh Patel, 'The Age of Research, Continual Learning, and Alignment', where he lays out the generalisation gap, the value-function hunch, and why he thinks scaling's transformative returns are flattening — watch it here, or read the annotated episode. For the same gap seen from the engineering side — why current reinforcement learning is such a wasteful teacher — hear Andrej Karpathy on AGI timelines and RL's limits.