LIMITLESS A.I. IS COMING

 

PLASTICITY AND THE FUTURE OF ARTIFICIAL INTELLIGENCE

 
Plasticity unlocks the future of AI
 

A CONVERSATION WITH PROF RICH SUTTON, TURING LAUREATE

21 March 2025—Today I had a fascinating conversation with the great Rich Sutton, Professor of Computer Science at the University of Alberta. I was especially grateful because he had just won the Turing Award, the computing equivalent of the Nobel, so it had to be one of the busiest weeks of his career.* The Turing was awarded jointly to Rich and Prof Andrew Barto at the University of Massachusetts Amherst for their work on reinforcement learning, which laid the foundations for all modern AI, but the subject of our call was not the past or the present, but the future, specifically Rich’s efforts to build plasticity into neural networks to unlock continual AI learning.

LEARN LIKE BABIES, LEARN FOREVER

Some background:

Leading computer scientists have previously articulated to me their grand goal of making AIs that ‘learn like babies and learn forever.’ They’ve made great strides on the babies part, via multiple modes of learning and so on; not so much the forever part.

In animals and humans, the ability to keep learning and re-learning from new and changing experiences comes naturally. Our brains have plasticity. In machines, learning is still mostly a one-shot exercise: present an algorithm with a ton of up-front training data to get an amazingly capable neural network; ask it to learn from additional data and it might, if you are lucky, get a wee bit smarter for a wee bit longer; keep asking and …. it gets dumber.

Today’s AIs do not have plasticity.

Rich and his colleagues are building new software algorithms, such as their Continual Backprop algorithm, to fix this.

And while continual learning has been locked into my predicted future of AI since at least 2007, this wonderful conversation gave me a far better grounding on recent progress, renewed my confidence in the outcome, and the importance, and – best of all – helped me think even bigger about future implications.

 

PLASTICITY CHANGES EVERYTHING

Why is plasticity so important?

  1. Plasticity enables individual AIs to keep learning from experiences forever, which means EVERY AI increases in utility. Think: AI traffic systems that learn daily from live traffic data. AI weather forecasts that adapt as climate change alters local weather patterns. AI maintenance systems that adapt to machine wear and tear. Plasticity accelerates all the developments I described in Exponential AI.

  2. Continual learning slashes pre-training overheads, which punches a hole through the looming AI “resource wall.” The current training approach for Large Language Models (LLMs) is essentially ‘one-shot’ and requires exponentially more hardware, data and electricity for each improvement. 2018 A.M. Turing Award winner Yoshua Bengio of the Université de Montréal and 74 other scientists and AI experts predict that, on current trajectory, LLM training compute requirements in 2030 will be 10,000 times those in 2023. The entire tech industry is hurtling towards a point where no one can afford to make a smarter AI. Plasticity neutralizes this. Did you think DeepSeek was impactful, when it wiped $1 trillion off US stocks (principally NVIDIA) by demonstrating a step-reduction in the number of GPUs required for pre-training? Plasticity will be seismic.

  3. If we define artificial general intelligence (AGI) as “better-than-human performance in all cognitive tasks” then continual learning has to be the most obvious pathway to AGI.

 

STEADY PROGRESS, HIGH optimism

Rich described his Continual Backprop algorithm as “about 20% of the way” and emphasized “this is a marathon, not a sprint - there is a degree of struggle. It will take time to make this transformation.” At the same time, Rich is notably optimistic about getting there. Algorithms will refine. They will unlock learning from experience, and learn how to learn. You can hear and see his optimism for yourself in his talks, such as this one with colleague Shibhansh Dohare.

Rich estimated that we were “three to four years away” from continual learning impacting industry.

Predicting the pace of research is, of course, fraught with difficulty – one research breakthrough can change everything – and his estimate must be taken in the generous spirit in which it was offered, but note: he is the father of reinforcement learning and the winner of the Turing, and “three to four years” is not “ten years” nor “One day, if we get lucky” – it’s soon!

Let’s do the futurist thing and throw dates at it:

2029 – continual learning AIs start impacting industry in a meaningful way

2031 – full-blown continual learning revolution underway

2033 – all leading AIs are continual learning AIs  

That’s a starting point. Am I wrong? Are there other high-integrity viewpoints or datapoints I should consider? Everything good in life comes through other people, and I positively welcome constructive, dissenting voices, so let me know!

Oh yes, and Rich is pretty bullish about getting to AGI, too, giving a 50% probability of reaching human level intelligence by 2040.

 

“EXTENDED LEARNING” OR “UNLIMITED LEARNING?”

Recently I’ve been visualizing plasticity unlocking more extended learning periods: AIs learning for weeks, then better algorithms stretching this to months, then years, but in each case reaching some unspecified limit after which imperfections or perturbations kill performance, so I asked Rich if plasticity will allow AI’s to learn for an extended but bounded period, or learn essentially ‘forever.’ He said once we learn how to learn there should be no limits. He expects to unlock AI learning in the fullest sense, with AIs capable of learning for as long as we let them.

So, unlimited.

At this point he shared another way of looking at today’s AI: “The real problem is not the algorithms. The problem is the data.”

Current methods are limited by the volume of training data that must can be collected (volumes can be truly enormous; everything on the Internet in the case of LLMs) and quality (which is being diluted by more and more AI-generated “pollution”). As an example of how desperately technology companies are searching for additional training data, Meta was recently caught secretly feeding a vast store of pirated books and science papers into its LLM - including one of my novels - without permission and without compensation. Yuck.

Continual learning, which lets the experiences keep coming, removes that limit.

 

DEATH OF LLMs (AGAIN)

A corollary from Rich was the same message Yann LeCun and others have been repeating for some years now; LLMs are NOT the future of AI. Quite apart from the scaling issue, the path to truly powerful reasoning and problem-solving cannot logically come through one-shot training on a gigantic bucket of unstructured data.

Rich used the word strange to describe how the big industry players have gone all-in on LLM's and coaxing ‘conversational’ interfaces out of them. He said it without judgement, just genuine bafflement; LLMs are so obviously not the best way of doing things to him. Strange. Wonderful adjective! I’m going to borrow that!

 

NEW TRUST PARADIGM

My biggest surprise takeaway was this: when plasticity unlocks continual learning, we will need to make a major cultural adjustment on trust. Stepping it out:

(A) With a fixed model you can benchmark. You can test your AI and verify that it fits your measures of trust.

(B) AIs that learn continually will NEVER be benchmarkable, because what your AI “knows” keeps evolving.

Aaaah …

Think of the global efforts currently applied to testing and fine-tuning and confidence-building. Think of all the AI “guardrails” conversations around the world. Think of all the hesitation over opacity and hallucinations and trained-in biases.**

I put it to Rich that the way we will evaluate our trust in future AIs must become analogous to how we decide to trust humans. That is, as with a human we are thinking of hiring or taking advice from, we will have to make considerations about background and experiences, where the AI learned and what subjects and from which teachers, and we will have to ask questions like, ‘have they performed well so far’ and ‘have they demonstrated good intentions so far.’ He agreed. The example Rich added was, “when you're assessing whether to put a human in charge of your factory operations, you take a good look at the operator and assess whether you're happy to trust them.”

As I say, BIG.

I’m still thinking about it.

SUPER-SPECIALISTS & SUPER-GENERALISTS

Rich’s insights inspired me to revisit some of my predictions about the future of AI, in particular my predicted landscape of many thousands of ‘specialist’ modular AI’s collaborating with one another, what I call ‘AI-to-AI’ in my keynotes.***

Does unlimited learning alter that landscape? Does the transition to a new trust paradigm?

Here’s where I’m at:

  • Everywhere I look I still see the most exciting, impactful and highest return AIs in specialist models, or what some scientists like to call “narrow AI.” A key reason is quality: AIs trained on clean data to do one thing well, do it well.

  • In a world where AIs learn continuously and perform ever better, but can no longer be benchmarked, higher trustworthiness remains a valued asset. i.e. people and organisations will willingly pay for higher trustworthiness for certain classes of applications, such as autonomous vehicles, medical and legal applications, and applications providing significant competitive advantage.

  • I think custodians of continuously learning narrow AIs (let’s call them super-specialists) will (a) impose more quality control by closely curating exposure to context-relevant, high-quality learning experiences, and (b) test continuously in the sense of querying and validating and monitoring customer experiences, and therefore be well positioned to (c) attest to higher trustworthiness for their customers.

  • I’m thinking, for example, of how Prof Paul Newman at Oxford is building route-by-route assurance and certification for Oxa’s autonomous vehicles. And in medicine, how high-quality tumour data might in future be fed into a continuously learning best-of-breed diagnostics AI curated by an MD Anderson or Memorial Sloan Kettering.

  • In which case, super-specialists remain a preferred go-to over continuously learning general-purpose AIs (let’s call these super-generalists) for many, if not most applications.

  • BUT super-generalist AIs are going to be pretty damned smart and getting smarter, and we can be confident many millions will happily use them for a VERY wide range of tasks, with almost no attention to trustworthiness whatsoever, because that’s exactly what many millions are doing with today’s LLMs. Quality of course will be far superior, so people will feel more satisfied and comfortable.

  • On the other hand, higher AI infatuation plus higher complacency with super-generalist AIs will accelerate all the dangers I described in The Real Threat From AI.

NEUROMORPHIC JOURNEY

All modern AI is “neuromorphic,” in the sense that it is all neurologically inspired to an extent, and all developments fall on a neuromorphic continuum. This applies to software and hardware. Neural networks were inspired by the way layers of neurons work in nature, and GPU chips are more neuro-analogous than CPUs because they do so much more processing in parallel, as we see in nature. The future of AI can thus be seen as a series of intertwined software and hardware innovations that get better and better at emulating and augmenting the biological architectures of intelligence.

Once you start seeing it this way, you start noticing relevant developments all over the place. Like what neuromorphic pioneer Steve Furber and colleagues are doing with new versions of SpiNNaker at Manchester University, like Intel’s Loihi 2 processor and its Hala Point System, like the neuromorphic sensors Alexandre Marcireau demonstrated to me at University of Western Sydney’s International Centre For Neuromorphic Systems. Like the pure research into the inner workings of biological neurons I saw at TU Dortmund University that I believe will one day be relevant to computer scientists. Like how the incomparable Daniela Rus and her MIT colleagues are emulating neural architectures of worms at Liquid AI including the plasticity elements, of which more here. And of course, like what Rich Sutton and his colleagues are doing at the University of Alberta.

There is a LOT of headroom in the computer science, a LOT of good people are working on new AI architectures and neuromorphic computing at scale, and they are progressing on multiple fronts. Which is another reason I’m confident plasticity is inevitable and coming soon, and another reason I’m confident we’ll find both software and hardware pathways around the looming AI resource wall (although we may have to hit that wall first).

I asked Rich whether plasticity was dependent on progress in neuromorphic hardware, or whether it could all be accomplished in software. He believes the latter, but of course "everything will work better with neuromorphic hardware."

And since I’ve previously called neuromorphic chip research the most important computer engineering research in the world because of its promise to vastly improve AI capabilities while simultaneously reducing resource (especially energy) requirements, then the same logic applies and plasticity is the most important computing software research in the world.

 

SUMMING UP

Continual learning is coming.

Continual learning will utterly transform AI, and our future.

While predicted timeframes can change dramatically with one research breakthrough, “industry impacts by 2030” gives us an indicator: continual learning is not far away.

After which, a full-blown re-write of the AI landscape.

Professor Richard Sutton, congratulations again on your Turing Award, and my sincere and profound thanks your time and generous insights.

FURTHER READING

I highly recommend visiting Rich’s personal webpage, which features a bunch of useful links and resources.

For the technically minded, his top 10 reading list will get you up to speed on his AI views and research goals, and this talk — Maintaining Plasticity in Deep Continual Learning - CoLLAs 2022 — by Rich Sutton and Shabhansh Dohare provides an excellent overview of the plasticity challenge and the principles behind the Continual BackProp algorithm.

This Jan 2025 lecture by Rich — A Perspective On Intelligence — gives a fulsome overview of his thinking on the nature of intelligence, key challenges goals and pathways for machine intelligence going forward, and expectations for the coming decade or so. He discusses plasticity and touches on many of the points he made in our conversation.

This interview is a more informal window into Rich’s purpose and personality, and his expectations about understanding the mind, and making goal-directed machines that can reason and plan. I especially like his advice to researchers at the end about writing down ideas, especially nebulous ideas, as a way to assist thinking :-)

  

ENDNOTES

* Incidentally, this has happened to me before:- I was at UPenn trying to organize an interview with Katalin Karikó and Drew Weissman on the day they won their Nobel, and Rich won his Turing the day after I requested this conversation, so from now on I'm billing myself as a lucky charm for scientists!

** Researchers are also working to unlock transparency, principally by making more stages visible in the “flowchart” of work, from parsing the initial query to building the final response. See, for example, Google’s Gemma Scope. I’m confident that improved methods coupled with more modular AI approaches to problem solving (ie breaking work into components) will see AIs get much better at communicating their “reasoning.”

*** AI-to-AI is another game-changer heavily featured in my keynotes. AIs collaborating with one another, addressing component parts of a problem, supervising other AIs, training other AIs, improving the outputs of other AIs, etc, etc. All these things are happening today, and will get ever more sophisticated. Scientists such as Andrew Ng at Stanford use the term “agentic AI,” but agentic is also commonly used to describe single AIs empowered to take on more complex transactions, purchases and payments on behalf of users (i.e. having more “agency”). Hence, I like AI-to-AI when I’m emphasizing the collaborative element. Learn like babies, learn forever, and learn together!

WHAT ABOUT YOU?

Now is the perfect time to be opening our minds to new AI-driven opportunities and possibilities.

One of my greatest joys is imagining all the big new AI opportunities for a specific industry. All my keynotes are positive and opportunity-focused because this is the best way I know to activate people to make a positive difference to their own future. I’ve had the privilege of helping audiences of hospital administrators, doctors, nurses, bankers, insurers, investors, lawyers, realtors, public servants, brokers, food manufacturers, teachers, transport-operators, police, energy personnel, and more, connect with hundreds of game-changing AI opportunities.

Reach out, and let’s explore the opportunities for your industry in a keynote at your next event!

 
Next
Next

ELECTRIFICATION OF ROAD FREIGHT – WITH DAVID CEBON