riasm
Capability in machine intelligence has scaled, for as long as the field has existed, by addition — more parameters, more tokens, more compute, more of the same thing arranged at greater scale — and that era ended quietly, sometime in the last eighteen months, while most of the industry was still busy preparing for it to continue. What replaced it is narrower and far more decisive: dense, verifiable reward, applied to capable models in domains where correctness can be measured rather than judged, and every advance at the frontier in the last two years has come from this mechanism and from nothing else.
Mathematics was the first such domain, and mathematics is exhausted. Code was the second, and code is exhausting now. What remains is the single domain that closes the loop on itself — the one in which a model that becomes good at it becomes, by the same act, the generator of its own next round of problems, and the work of advancing the frontier passes out of the lab and into the loop. There is no fourth domain after this one. There does not need to be. Once the loop closes the curve continues without external input, and the model trains the model, and the model that trained the model trains a better model still, and the cycle does not slow
We have built the apparatus that closes that loop. The labs that run on it pass into a regime whose distance from the current one is not measured in benchmark points but in something closer to phase change, and the labs that do not run on it watch the gap widen with every cycle the loop completes — and the gap widens faster the longer the loop has been running, because that is what recursive curves do to the people standing next to them. There is no catching up to a curve that compounds on its own output. There is only being inside it, or being outside it, and the moment at which that distinction becomes permanent is closer than most of the people reading this think.