LOL, the LLM is a single dimension. True AI involves a multiplicity of dimensions. Take the course of the life of a child from birth to old age. It suckles- poop/pee and learns from those experiences - maturity. The Large Human Model (LHM) would 'see' the progress of this life via trillions of observations of experiences on file, like LLM scavenged the written 'textome' to get where it is. Now text archives are more extensive and also easily searchable. The millions of LHM watchers would go through the 'humome' of experience.(which is probably both smaller than the textome as well as larger in diversity of physical experience). So it would learn it could not fly, (notwithstanding kids in superman suits) and quickly accumulate an action base of the full human life stream. After learning, crawl/walk/fall from trees/trip/bruise knees.....all things people do. It can then 'cross compile' this with the LLM and be able to create an avatar to use the LLM data to script a virtual avatar. Such an avatar could directed as needed for tasks etc., and the first steps of the ladder to a true AI will commence. This would also need all the details of the human 'neurome' to be mapped and modelled in silicon to see if it will think like us.
We have a 'clock speed', or alpha rhythm of 8-12 Hz, with a mind in silico is a clock speed of 10 Ghz - or more, reasonable?
This can start(has started) with models of the flat worm neurome and has reached many insects and small mammals, but is a work in progress with mice etc, but in the next few years virtual small insects will be virtualised, but the chemical pathways of even small insects and flatworms will take longer.
8 comments
[ 3.4 ms ] story [ 25.2 ms ] threadSo, he thinks that GPT-5 will not market any better than 4? Ok.
That article is when he was 36, which was decades after.