[ExI] The present and future of AI

[The Avantguardian]

My pleasure, Jason. I must apologize for top posting but Yahoo seems to bounce my emails back to me as attempted spam if I don't top post on this account.

I did not factor in the time-complexity of the human brain at all. I simply extrapolated the average memory cost for individual neurons and synapse weights in tensorflow for AlphaGo and scaled up the number of neurons and synapse weights to that of the human brain in a linear fashion. So my analysis was more based on the Kolmogorov complexity than computational complexity.

I do realize that this results in oversimplification, but it was more of a back of the envelope first approximation rather than an in-depth analysis. 

If you factor in time complexity, then AGI will already be superhuman before space-complexity parity with the human brain is reached due to the speed differences.

How effectively one can trade space for time in AI is a fascinating question. I think it would necessarily be limited by the required non-linearity of the neural activation function. Let me think more on this.

Stuart LaForge

On Friday, June 12, 2020, 06:27:17 PM PDT, Jason Resch via extropy-chat <extropy-chat at lists.extropy.org> wrote: 
On Fri, Jun 12, 2020 at 5:22 PM The Avantguardian via extropy-chat <extropy-chat at lists.extropy.org> wrote:
> I would say that we are close enough that we need to have a better definition of AGI. For example does a honeybee have "general intelligence"? If so, we are already really close because we can currently compute neural networks with the complexity of a honeybee brain. Assuming neuron number is a correlate of intelligence, it will be about 44 years until Moore's law reaches human brain parity. But we might reach rat or raven parity in about 14 years.
> 
> See old thread "Benchmarking the SIngularity" for more info:
> 
>  http://lists.extropy.org/pipermail/extropy-chat/2019-July/097196.html
> 

Stuart,

Thanks for sharing that, it was very interesting.

Regarding parity of the human brain, how are you estimating the computational complexity of the brain? The Summit supercomputer can perform as many Ops/second as a generous estimate of total synaptic firings the human brain can have per second.

Jason 
_______________________________________________
extropy-chat mailing list
extropy-chat at lists.extropy.org
http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat