With the way that AI models are trained is that they will never be broadly smarter than the collective sum of humanity. The will be AGI-- general expert human level intelligence across all domains. But they will never pick tokens outside of that distribution-- they materially can't. With the current regime of mixture-of-experts, they may not be able to deliver on the promised cross-domain synthesis, either. However, many problems are solvable by exhaustive search. They are not solved because the quantity of human experts available to solve them is not saturated-- LLMs can scale expert knowledge production 10000x and we will see an enormous amount of math get proven. The LLMs really cannot generate new mathematical theorems well yet. Although they can translate the idea from a great mathematician into a Lean proof. To some extent, LLMs are the product of exhaustive search across the domain of possible architectures. The transformer was not obvious, in fact many people saw dense attention as too costly to ever work, and that they worked came certainly from intuition and unbelievable amounts of trial and error. But there are infinite architectures. And this one has obvious flaws: the weights are frozen, it can't learn new information without forgetting old information or collapsing, the training process is extremely sample inefficient. And we have existence proof that there is a better way: evolution. We're orders of magnitude from the capability and power efficiency that evolution brings to the table in every living creature. Superintelligence is a kind of intelligence that is smarter than all humans. By its very nature, it cannot be trained on human data, or not alone at least. It has to be able to learn, grow, experiment and explore. It has to exhibit adaption to environment. It has to be able to create new tokens to represent concepts that have never been conceived of before. Compared to our current systems, it would look like a complete alien, being able to communicate with us through human channels only as an artifact of being able to learn and do everything. Superintelligence is a set of algorithms that meet these constraints. Rich Sutton's Alberta Plan points toward this. If you have an algorithm that can identify the important features and adapt at runtime, learn new information without losing old information and forget anything that isn't important to keep capacity for information that is relevant, surprising, traumatic or delightful. I think this is buildable today. I think you can pull Fable or GPT-5.6 and have it take on something like the Alberta Plan. Continuous learners can beat MLPs. Attention can be plastic. JEPA and world modeling can be extended to continuous domains like utterances and we can leave the shortcut paradigm of tokens entirely. ASI might take a while to scale, and might underperform the human level high scale AGI systems for sometime, but the seeds of it are here now, the algorithms are just waiting to be discovered, or rediscovered with new compute and seen for what they are.(Shaw (spirit/acc))