Transcription
Multiverse Computing out of San Sebastian, Spain, has developed a way to compress LLMs using Quantum software. I spoke with Enrique Lisazo, the CEO of Multiverse Computing, in an online interview before AWS re:Invent.
We compressed 95% of a model. That was a model 2.7 R, maybe at that point. Uh, very, very small loss of accuracy, given that we were just throwing away, throwing away 95% of the model. The model was quite nimble. The cost of running this model, the on-inference time, which is super small as well. And we said, "Okay, maybe we have, we have something here." Because LLMs are expensive, and GPU hours are expensive, and there is no solution. If you buy the Nvidias, you ruin yourself. If you buy the GPU hours, you, you make reach the hyperscalers. In any case, you are not in, in a, in a very good position. So a lot of customers just stop deploying the system.
Quantum Computing is just of another dimension. I think that's what we're hearing from Lisazo. He's saying, basically, if you can really use quantum computing, maybe we will see the point where robots actually can fly. But right now, we're just spending so much energy and consumption of resources into building something that really can't mimic what nature can do. Why is it, why is it going to be better than everything at some point?
Let's talk about Quantum Inspired Computing. We can talk later about the future of quantum computing. You want to. Yeah, but the point is the Quantum Inspired algorithms. The way that a quantum computer deals with this information focus, as I mentioned before, in the correlation. So it's not, for example, all the relation is between the neurons and the weights in an LLM. It's about the correlations, how the model responds itself to some particular input, which is the input that you get to them, and how some parts of the model correlate with some other. This is what quantum computers are super good at. And Quantum Inspired methods, by mimicking the quantum computer, is not as good as a quantum computer, but very good at the other. So that means that you can select which part of the LLM is the, the, the, the one that is taking the heavy, the heavy weight of the, of the work to be done. So you can take out in a reasonable way and put out the part that doesn't do anything. Okay. And there are, there are a lot of ways in, uh, LLM that, uh, what they are just providing to the final answerer or to the final usage is very few, if any.
Okay. I can put some example from biology. Look, the fruit fly. The fruit fly is a fly that from, okay, they, I mean, one month or one month and a half ago, in Nature, I think. I don't know where it is from here or so on. Uh, uh, yes, this one. Look. Okay, they put here the, the, okay, see. Yeah, yeah, yeah, yeah, yeah. This is the, they put the brain map of the fruit fly, the complete map of the brain. Okay. That was all the connections between the different neurons, etc., etc. So, as you already know, the, the fly can fly, okay, can walk, can mate, can fight, can communicate. And if you try to kill the, the fly, it's quite difficult unless you just use something so super nice. I think if I remember correctly, 120,000 neurons, I think, and 50 million of connections. 50, 50. Oh, wow. Now, you compare that with an LLM, and which are on the billion parameters, and you notice the difference. So nature does the thing differently. Okay. So there is a huge, a huge space to improve the, the way that we are just developing the AI systems. Okay. We already know that because, I mean, if you look at the, at a dog, or if you have a child, and how the, your, your child communicates and goes through the, I mean, nearly can walk without just hitting with anything. Okay. Can communicate everything, everything. And in the process, we are not killing the planet by overheating the or using all the electricity and using now nuclear reactors just to train the model. Okay. It's quite simple. So there are some other ways we are proposing, one of them which is based on Quantum, which is super efficient.