~ just post it in the comments! ~ in response to the essay “AGI soon, but Narrow Works Better” Please note! The argument that “RETRO and other models use memory, which could be used for following a multi-step plan” is erroneous; for RETRO or other memory-equipped networks to utilize their memory on the *task* of “remember which step of the Destroy-All-Humans-Plan that you’re on, now” they would have needed to *re-assign* the *meaning* of those memory locations, which would require plasticity of the weights. Yet, the original constraint was that weights be frozen; hence, no memory’s *meaning* can be re-assigned to “keep track of which step they’re on”… nor could frozen weights generate a new macro-task such as “Destroy all humans” unless that macro-level task was included in its training data.
