every jump found a new cheap verifier
The usual story of AI progress is a list of architectures: perceptron, backpropagation, convolution, LSTM, transformer. Walk the dependency graph on the ai timeline and a different spine appears. Each large jump came when someone found a new source of training signal that could not be fooled, and pointed optimisation at it.
Next-token prediction is the first one. The corpus verifies itself: every position in every document is a labelled example, the supervision is free and unlimited, and the binding constraint moves from labels to compute. That is the precondition for scaling laws to mean anything, and GPT-1's real claim was not the architecture but that this one objective is universal.
Human preference is the second. Christiano's 2017 insight was that comparisons are cheaper and more reliable than scores or demonstrations, and that a learned reward model amortises a few thousand of them into an unlimited oracle. InstructGPT's 1.3-billion-parameter model preferred over the 175-billion base is the measure of what that bought. The failure mode arrived with it: optimise hard enough against the proxy and you find its errors, which is why the KL penalty is mandatory, and why sycophancy is learned from the preferences themselves rather than being a labelling mistake.
Exact match is the third. Chain-of-thought showed that letting the model write intermediate tokens is literally buying it serial compute, and STaR showed that a model plus an answer key is a data pump: sample rationales, keep the ones that reach the gold answer, fine-tune, repeat. DeepSeek-R1 stated the principle from the other side when it deliberately avoided neural reward models at scale: a verifier that cannot be fooled is worth more than one that is smart. The 2025 pass@k results are the caveat. Much of what RLVR buys is sharpening, moving probability onto paths the base model could already produce, rather than expanding what it can do, and the exploration-preserving fixes and the move to richer environments are responses to exactly that.
The Lean kernel is the fourth, and the one that matters for mathematics. AlphaProof's reward is one bit, "does the proof compile", which is perfectly honest and infinitely queryable, so RL can run as long as compute allows. Everything on the math timeline from DeepSeek-Prover to Aristotle to Astra's ten certificates inherits that recipe. Its residual weakness is not correctness but fidelity: the kernel checks the Lean statement, not the English one.
A programmatic evaluator is the fifth. FunSearch and AlphaEvolve let a language model be a mutation operator inside an evolutionary loop where a human-written scorer is the only ground truth. Tao's account of driving AlphaEvolve across 67 problems locates the human skill precisely: designing evaluators the system cannot exploit, because it will otherwise find degenerate solutions that game floating point.
Task completion in a real environment is the sixth, and the current one. Agentic RL scores a whole trajectory of tool calls and edits by whether the tests pass or the artifact is correct. It is the only known source of fresh, non-scraped signal at scale, and it is why METR's time-horizon curve is doubling every few months. The 2026 survey of self-improvement loops reached the same conclusion from a thousand papers: nearly every loop that works grounds its critique in something external, and self-critique without grounding does not compound.
The efficiency work sits underneath all of this and is easy to miss. FlashAttention, grouped and latent attention, quantisation and paged KV caches did not just lower cost; they moved context from two thousand tokens to a million and made agent loops that emit millions of tokens per task affordable. A tenfold cost reduction turns an infeasible technique into a default one. The zeta result was two Claude Code sessions and 31 million output tokens across sixty subagents. In 2022 that was not a budget anyone had.
If the pattern holds, the next jump is not an architecture. It is whatever the next thing is that cannot be fooled.