Prediction vs. Construction

In the AI Story

The prediction paradigm is embedded deep in how contemporary AI systems are built and sold. Scaling laws are prediction claims: given this much compute and this much data, the model will achieve this much capability. Benchmark evaluations are prediction claims: given this test, the model will achieve this score. Investment theses are prediction claims: given current trends, this company will win. At every level of the AI industry, the fundamental orientation is toward forecasting what will happen on the assumption that the future is determinable from the present.

Smolin's framework denies that the future is determinable. Not because forecasts are impossible — patterns in the past do constrain what the near future can be — but because the universe is not finished. Configurations that have no precedent produce outcomes that are not determined by preexisting laws. Systems undergoing phase transitions produce states that cannot be extrapolated from their prior states. Genuine novelty enters the universe through the thick present, and it enters in ways that the prediction framework cannot accommodate because the prediction framework assumes the future is already there to be extracted.

The construction paradigm reorients AI development around a different question. Instead of asking 'what will happen next?', it asks 'what could we bring into being that does not currently exist?' The analogy Smolin offered at Toronto is instructive. A baby does not attempt to predict who she will meet next. She engages each new encounter with an open question: 'Who is that?' The encounter is genuine because the person encountered is treated as genuinely novel rather than as an instance of a preexisting category. AI systems built for construction would operate similarly — not as engines that extrapolate from past data to forecast future configurations, but as participants in the generation of futures that had no precedent.

The practical implications are substantial. An AI designed for prediction optimizes for accuracy against known benchmarks. An AI designed for construction would optimize for its contribution to the generation of genuinely new possibilities — a criterion much harder to measure but arguably more important. An AI designed for prediction treats the user as a consumer of forecasts. An AI designed for construction treats the user as a collaborator in the creation of futures. Segal's account of his collaboration with Claude in The Orange Pill — the laparoscopic surgery example, the punctuated equilibrium insight — describes construction rather than prediction. The valuable outputs were not forecasts; they were genuinely new connections that opened new possibilities.

The reframing does not require rejecting prediction entirely. Prediction is valuable within the domains where extrapolation is reliable — typing-assistance, spelling correction, weather forecasting at short timescales. But it becomes actively misleading when applied to domains where the future is genuinely open — strategic decisions, creative work, institutional design, the governance of transformative technology. Distinguishing these domains is itself a temporal-literacy problem. It requires asking: is this a situation where the past reliably constrains the future, or is this a situation where something genuinely new is emerging and extrapolation will miss the signal?

Origin

Smolin articulated the prediction-versus-construction distinction at a 2023 symposium at the University of Toronto and has developed it in subsequent essays and interviews. The framing builds on decades of his work on the reality of time and connects directly to his collaborative work with Kauffman on combinatorial innovation and with Lanier on the autodidactic universe.

Key Ideas

Two paradigms. AI can be designed to predict what will happen or to help construct what has never existed — these are structurally different orientations.

Paradigm tracks ontology. Prediction presupposes the block universe; construction presupposes temporal naturalism.

The baby's question. A genuinely open engagement with what is novel — 'Who is that?' — rather than a pattern-matching against prior experience.

Different optimization targets. Prediction optimizes for accuracy against benchmarks; construction optimizes for contribution to the generation of genuinely new possibilities.

Domain-sensitivity. Prediction is valuable where extrapolation is reliable; construction is necessary where genuine novelty is emerging.