To solve AI's energy crisis, 'rethink the entire stack from electrons to algorithms,' says Stanford prof The Stanford Institute for Human-Centered Artificial Intelligence (HAI) on Wednesday celebrated five years of cat herding, which is to say shepherding the responsible development of machine learning. Following optimistic introductory remarks from HAI leadership about the plausibility of designing systems that augment people …
The current LLM training model and depending entirely on scale, is the energy crisis. Past AI work has not triggered any energy crisis. Future work will presumably not, either. And I like some of the chatter here about getting good-enough answers from sloppy mechanisms. So I think at least some of these HAI topics are on target.
More like 1929 all over again.. Or worse. It seems like we are headed for a combination of a Great Depression event (breakdown of international trade.. Shipping, Brexit, Trump, the new Cold War) AND a Dotcom Bubble event caused by the AI hype.
Just about the only thing AI is "good" for is swindling people, poisoning democracy etc. We could have another Civil War in the US at the same time as WWIII heats up in Europe.
So I'd rather have a DotCom bubble now (i.e. let the AI bubble burst as soon as possible please) rather than the complete armageddon that I think will happen if it is allowed to carry on along its current growth curve.
Numerical (digital) linear algebra in massively parallel GPUs are an incredibly inefficient way to simulate a system that is inherently analog. I imagine the ideal neural processor chip to be mostly analog. Perhaps transistor amplifiers for the neurons, with digital potentiometers set at load time for the synapses. DAC in, big mess of amps and resistors, then ADC out. You keep the digital representation of the weights, with the speed and efficiency of analog computation. Dunno about training, but it'll speed up inferencing. Icon because obviously.
There's a nice interview (podcast, transcript, linked papers) of "Next-Gen Neuromorphic Researchers" over at EEtimes, where four young'uns talk about analog stuff, spiking neural nets (snn), surrogate gradients for horror-backpropagation, energy consumption vs digital nets, leaky integrate-and-fire models, brain inspiration, and so forth. Its a nice foray into related machine "madness" and ties in nicely to Tom's article IMHO (Ganguli's and Hawkins' major points, equally sedatedly; much unlike the more characteristic AI wrecking ball of the ultra-hype-orama). Worth a gander.
For sentient AI is duplicating the mechanism(s) of consciousness itself.
It seems that quantum computation may be a potential solution, in the case of an AI the quantum neural network
would be based on 28Si based isolated qubits and a 3-D lattice broadly similar to a positronic network.
It would need conventional components as well but a 'Positronic Brain' might be relatively compact at least once
the whole cooling and fabrication issues are worked out.
Idea here is to use a relatively low end PC to test this idea, build a superconducting lattice that uses the oxygen
vacancies in cuprates as "synapses" as they can be changed and moved around with relative simplicity.
As Tc increases Jc decreases so it may not be required that the lattice operate at very low temperature, a sufficiently
finely engineered material using the right isotopes can work at about the temperature of a domestic freezer
with the pattern locked inside like the floating gates on a conventional Flash chip.
I don't know if there is any prior art for this.
int main(enter the void)... The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
