My hope for AGI: not a bigger model, but a unified calculus of logic and probability

I don't hope AGI arrives as a bigger version of what we have. My hope is that it arrives when someone finds a way to unify logic and probability at the bottom of computation — the way physics hopes to unify its forces — so that reasoning and uncertainty become two expressions of one language, instead of two modules passing notes. Let me define what I mean, because the definition does most of the work. I mean a system that (1) can revise its own rules and stay reliable; (2) keeps learning instead of freezing after training; (3) models its own limits; (4) can judge an improvement against something it cannot edit; (5) has a causal, persistent world model. None of this requires consciousness. Without these five, "general" is doing no work — it is just a wider distribution. Today's systems are extraordinary statistical engines, but they fit rather than derive. The distinction sounds philosophical until you test it. Rename a variable, lengthen a chain of trivial steps, change the surface form of the same problem — performance drops in ways that have nothing to do with the difficulty of the problem. Training by prediction optimizes the next token, not rule-following. A network can represent logic; that was never the issue. The issue is that logic is not the organizing principle. And no amount of in-distribution data can separate two hypotheses that agree on all of it and disagree outside it: they are indistinguishable from the inside, so scaling changes coverage, not kind. The gap comes from a conflict I keep running into. Logic is rigid: a step is valid or it is not. Probability is graded: beliefs come in degrees. Gradients need continuity, so every trainable logic softens "must" into "mostly" — fuzzy operators, soft constraints, probabilistic logic programs. In-distribution, "mostly" is fine. Out of distribution, "mostly" is exactly where the failures live. That is why the current fixes — hybrid models, tool use, RLHF, long context — are real progress and still patches on the same floor. I would not be shocked if continued optimization eventually papers over much of this; predictions of a wall have a bad track record. But better patches do not give you a single language in which both are native; they give you two languages and a translator, and the translator is where the guarantees leak. So my bet is a mathematical one. I suspect AGI needs a constructive calculus with strata: continuous inside a stratum, discrete between strata, with checkable transitions, so that reasoning and committing to a branch are different operations rather than a compromise. Cheap certificates, so each step can be checked without re-deriving everything. And an anchor outside the editable set, because if the judge can be edited, the cheapest improvement is to edit the judge. Start where checking is cheap — formal math, typed program synthesis, constrained design — not because those are the most important domains, but because they are the only ones where "correct" is not a matter of taste. The hard part is that no such structural unification exists. Category theory, topos theory, and linear logic can describe both the discrete and the continuous, but none of them gives a compiler — a set of rules you can run on hardware, train where appropriate, and check where necessary. The architecture that would unify our mathematical theories may simply not exist yet. And three walls remain: continuity cannot enforce hard constraints; self-reference cannot certify its own soundness; and complexity does not disappear when you change the notation. Current AI can help: formalize, search, verify, propose, exhaust the space of small ideas. I don't think it can invent the missing language by scaling, because the missing thing is not a pattern in the data; it is a new calculus. That is the difficulty I mean. Verification is becoming the bottleneck; institutions are not built for a thirty-year program with no interim product; and we may be wrong about the direction. I think our generation probably won't see it finished. I'm not a researcher. This is a personal view; I may be wrong. But I don't think the distance to AGI is another trillion tokens.