In AGI, "general" does not mean large, and it does not mean trained on everything. It means the system can transfer. Give it a task outside its experience and it should recognise the gap, reason from what it does know, and act, rather than fail because the pattern was not in the data.
Three ways to read generality
Task generality is breadth of domains. Transfer generality is the ability to carry skill from one domain to another. Learning generality is the ability to acquire a new skill after deployment, without a retraining run. The third is the hardest and the one most often skipped in marketing.
Why scale is not generality
A model trained on more text sees more patterns but still only interpolates within them. Genuine generality shows up when the input is unlike anything in the training distribution, which is precisely where current systems are weakest.
The test that follows
If a system only performs well on benchmarks that resemble its training data, its generality is an illusion produced by coverage. The useful question is what it does on the first problem nobody anticipated.