AI models must adapt or die
The technology has already consumed almost all high-quality data — experience is now the dominant medium of improvement
John Thornhill
The nosebleed valuations in the US tech sector partly reflect the belief that artificial general intelligence is within sight.
Even though few agree on what AGI means exactly, investors seem convinced that a stronger form of generalisable AI will transform economic productivity and make mountainous fortunes for its creators.
To sustain that story, US tech firms have been pouring hundreds of billions of dollars into building more AI infrastructure to scale their computing power.
The trouble is that scaling is now producing diminishing returns and some researchers doubt whether the AI industry’s route map will ever lead to fully generalisable intelligence.
Arch-sceptic Gary Marcus wrote recently that generative AI models were still best viewed as “souped-up regurgitation machines” that struggled with truth, hallucinations and reasoning and would never bring us to the “holy grail of AGI”.
The debate about the limits of scaling has been raging for years and, up until now, the doubters have been proved wrong.
In 2019 the computer scientist Rich Sutton wrote The Bitter Lesson, arguing that the best way to solve AI problems was to keep throwing more data and computing power at them.
The bitter lesson was that human ingenuity was overrated and constantly outstripped by the power of scaling.
“The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin,” he wrote.
For the biggest AI companies, which have all pursued such deep learning approaches, that lesson still underpins their vast investments.
But the doubters have kept doubting and they were given extra ammunition by OpenAI’s launch of GPT-5, which led some to argue that generative AI is finally hitting a wall.
Interest is now picking up again in alternative approaches to advance AI, including some of the so-called “good old fashioned AI” techniques, long scorned by deep learning researchers.
One expert proposing a different path is Karl Friston, a professor of neuroscience at University College London and chief scientist at the Canadian cognitive computing company Verses, who thinks we still have much to learn from biological intelligence.
He acknowledges that the latest generative AI models are “absolutely astounding”.
But he argues they will never achieve AGI as they “have no agency under the hood”.
That can only be achieved through “active inference”, which means having a theory of the world and the ability to predict and adapt.
“For true AGI, you have to be active and embodied and situated physically in a world that you can act upon before you can even think about applying the notion of intelligence,” he tells me.
As an example, generative AI models might be good at analysing historic patterns of traffic flows in a city.
But what happens when Taylor Swift comes to town?
A truly intelligent system has to have a constantly evolving model of how the world works and predict the likely effect of novel factors.
“AI breaks when it hits the real world because the real world keeps changing,” says Gabriel René, Verses’ chief executive.
His company’s vision is to help build a network of billions of small, adaptive, domain-specific agents that do one thing well rather than rely on a massive system that tries to do billions of things well.
“Intelligence is about adaptation.
It’s not about compressing historical knowledge and memory.”
Intriguingly, even Sutton appears to be warming to this way of thinking.
In a recent paper called Welcome to the Era of Experience, Sutton and fellow researcher David Silver argue that almost all the high-quality data that could improve an AI agent’s performance has been consumed.
That means AI agents will now have to generate fresh inputs from interacting with the real world to gain experiences and data points.
“AI is at the cusp of a new period in which experience will become the dominant medium of improvement and ultimately dwarf the scale of human data used in today’s systems,” they write.
“The pace of progress driven solely by supervised learning from human data is demonstrably slowing, signally the need for a new approach.”
Investors may be reassured that there remain millions of lucrative use cases for existing generative AI models.
And the big AI companies can, of course, pivot and increasingly pursue hybrid approaches to build more efficient models in future.
Still, investors had better hope that the AI companies, as well as their agents, learn fast from experience and adapt to a changing world.
0 comments:
Publicar un comentario