Are LLMs capable of reaching AGI?
"AGI" is defined here as a computer program that can do any intellectual task that a human can do, excluding unfair cases like asking for secret information that only the human knows or testing sub-second reaction time. It must be able to do the task at least as reliably and as quickly as the most skilled living human (at market creation) could do it. It also must be capable of running on all of the Earth's computing power as of market creation, to prevent trick answers like "use an LLM to implement brute force search on 10^10 galaxies worth of GPUs". For training, it must be trainable within 100 years. In the event of disagreement over whether something counts as AGI, I will err towards no; it needs to be pretty clear that computers have attained intellectual parity with humans. e.g. it must be the case that there is no longer any economic reason why a company would want to hire humans over AIs, excluding physical tasks and things like "our customers will pay more for human outputs" or "we don't trust the AIs to not defraud us". "LLM" is defined here as any neural network that's trained primarily on human text. It can be multimodal, but any non-text inputs like images or synthetic mathematical data must constitute less than 50% of the training data. Reasoning models and agentic frameworks do count as LLMs, provided the scaffolding is just some form of feeding its answers back into itself, giving it multiple scratchpads, or other simple things like that. It cannot add any external system that would itself seem "intelligent". In the event of disagreement over whether something counts as an LLM, I'll err towards seeing how other people - particularly those who developed the AGI and other experts in the field - are using the term and whether they think it applies. If there's disagreement among experts, I'll err towards no; it must be pretty unambiguously an LLM. This market resolves once there's a broad consensus as to the correct answer, which likely won't be until after AGI has been reached and humanity has a much better understanding of what "intelligence" is and how it works. This definition is open to minor modifications if a problem in it is pointed out to me that makes it not align with the spirit of the title. In the event of severe disagreement over what constitutes an LLM or AGI according to this description, I'll defer to a vote among Manifold users.
