The strongest mechanistic case for AGI from next-token prediction deserves more respect than it gets, but it also deserves sharper scrutiny than the usual Reddit thread allows. The question posed here is the right one: not whether progress will continue, but whether the learning mechanisms we already understand can plausibly scale into open-ended intelligence. The answer is not obvious, and pretending otherwise does nobody any favors.
What makes the case compelling is compression. A next-token predictor that learns to anticipate the next word in a vast corpus is not memorizing strings; it is building a model of the world that generates those strings. To predict text about physics, economics, or human behavior with any accuracy, the system must internalize regularities that go far beyond syntax. That is a causal claim, not a vibes claim. The model learns latent variables that correspond to real structures, because those structures are what make prediction possible. This is the same logic that drives a practical guide to diffusion models that balances deep math with clear insight, where generative systems are forced to represent the underlying distribution rather than surface patterns. If compression is intelligence, then scaling compression is scaling intelligence.
But the gap between compression and open-ended agency is where the argument gets honest. Predicting the next token is not the same as deciding what to do next. A system that excels at forecasting the next word in a conversation is not necessarily a system that can set goals, navigate novel environments, or improve its own capabilities. The mechanisms that drive prediction are passive in a way that agency is not. This is where the practical engineering view matters. We can test the compression hypothesis directly by building systems that learn from active feedback, not just static text. Work on teaching AI to learn from screen video and improve over time points in that direction, and it is a far more useful experiment than debating definitions of consciousness. The question is not whether next-token prediction is sufficient in theory. It is whether we can extend the same learning principles to interactive, consequential environments.
That extension is the real frontier. The mechanistic case for AGI does not rest on current models alone; it rests on whether the principles that make next-token prediction work, gradient descent over predictive error, can also produce systems that act, observe, and learn from the consequences of their actions. This is an engineering problem, not a philosophical one. We can build systems that predict the outcome of an action before taking it, then use that prediction to choose better actions. That is a concrete path from passive prediction to active intelligence. The path is not guaranteed. There may be fundamental limits to what gradient descent can discover, or the computational cost may become prohibitive. But the argument is testable, which is more than can be said for most AGI speculation.
What we should watch is not the next benchmark score but the shift from prediction to intervention. If AI systems start learning from the outcomes of their own choices, rather than only from the text humans have already written, that is the moment the mechanistic case becomes credible. Until then, next-token prediction remains a powerful compression engine with an open question at its core. The people asking that question on Reddit are not being naive. They are doing the work that the field needs, and the rest of us should pay attention to the answers.