I recently sat down with Jeremy Wertheimer in Davos to talk about some of the biggest issues surrounding generative AI and today’s computing environment.
One of his most intriguing points concerned some of the differences between engineering and science, or rather the scientific process.
“When we build things,” Jeremys said, “we know what we’re doing…we have to do it right, we have to make sure our process is good, we have to avoid mistakes, but we don’t have to make errors. new discoveries. We just have to do the engineering…and we should get what we want.
On the other hand, he says, in science, we are faced with uncertainties. You have to be, in his words, “lucky”.
Sometimes, he suggests, we don’t know whether something is more of an engineering problem or a scientific problem.
“People might think it’s one, but it’s the other,” he said.
Jeremy gave an example of hearing Jeff Bezos talk about great language patterns. Bezos, Jeremy explained, suggested in a podcast that we should say we “discover” big language patterns instead of “inventing” them.
Jeremy said he liked it. He gave the example of a plant: we plant it and feed it, but we don’t know exactly what it will do. We didn’t build the factory or design the seeds!
Jeremy talked about the smartphone, which he pointed out would have been science fiction just a few decades ago. It understands what you say, for example, and can tell you the weather, etc. It’s engineering.
But then, he says, there is the LLM: and it’s science!
Unlike the smartphone, we did not build the LLM. We find out what he can do. The nature of machine learning and AI means that some of these technologies will NOT be designed, in the classical sense, but rather studied, like biology, as a force of nature. We’ll study them to see what they do!
This, I think, is the main takeaway. From there, during my conversation with Jeremy, he brought up a prediction he personally makes: in the future, everything will have the same three lines of code:
“Build a model, train a model, and apply a model.” »
To illustrate, he gave the emblematic example of so many technological advances over the years: the toaster!
When you think about what a toaster does, you can consider things like moisture content of the bread, heat, and other factors – but ultimately, if you don’t start from a purely deterministic point of view, you won’t. I don’t know how the model works – not entirely.
Regardless, Jeremy, who majored in neuroscience while researching AI, also made the analogy to the human brain.
“Brains are very complicated,” he said, adding that today’s LLMs are also becoming more complicated, with many more neurons, and will eventually defy easy dissection. He talks about the phenomenon of training versus construction, and our expectations of LLMs, calling realizing our limitations a “bitter pill” – as we will eventually learn that we cannot always understand how or why a model does something. .
It seems to me that this is a very instructive way of looking at technology. Either we will know how a given system works or we will not. We’ve talked a lot about explainable AI – at conferences and in the media. And there’s this general idea that we need to be able to continue to exploit AI while keeping it explainable. But Jeremy’s claim sort of contradicts this, or at least points to a certain kind of different perspective: that we will have to settle for a far from complete understanding of how complex models work.