AGI is the saturation of the model until it is expert human level at everything. Generally intelligent by human standards. It is the ideal of the data production and training regime, and it is bounded by human data and preference. We are approaching AGI not as a threshold but as a limit. Superintelligence, or ASI, is something else. General intelligence is table stakes. ASI learns from its own experiences and has potentially unbounded capability. Nobody has yet discovered a learner capable of meeting all of the conditions for ASI. Current RL cannot correctly assign credit to individual tokens for reward, so it is distributed across massive numbers of rollouts— not exactly a single agent learning from experience. I do believe this is possible, and like the math conjectures currently being demolished may just be a matter of time and compute.(Shaw (spirit/acc))