機械はいつか知能を持つのか？

Technical advances are moving at such a rapid pace that it can be challenging to define the tomorrow we’re working toward. In The Shape of Things to Come, Microsoft Research leader Doug Burger and experts from across disciplines tease out the thorniest AI issues facing technologists, policymakers, business decision-makers, and other stakeholders today. The goal: to amplify the shared understanding needed to build a future in which the AI transition is a net positive.

In this first episode of the series, Burger is joined by Nicolò Fusi of Microsoft Research and Subutai Ahmad (opens in new tab) of Numenta to examine whether today’s AI systems are truly intelligent. They compare transformer-based large language models (LLMs) with the human brain’s distributed, continuously learning architecture, exploring differences in efficiency, representation, and sensory-motor grounding. The discussion probes what intelligence really means, where current models excel or fall short, and what future AI systems might need to bridge the gap.

Learn more:

Matching Features, Not Tokens: Energy-Based Fine-Tuning of Language Models

Publication | March 2026

A Thousand Brains: A New Theory of Intelligence (opens in new tab)

Book | Jeff Hawkins | 2022

Thousand Brains Project (opens in new tab)

Homepage

A Framework for Intelligence and Cortical Function Based on Grid Cells in the Neocortex (opens in new tab)

Publication | January 2019

Why Neurons Have Thousands of Synapses, a Theory of Sequence Memory in Neocortex (opens in new tab)

Publication | March 2016

On Intelligence (opens in new tab)

Book | Jeff Hawkins with Sandra Blakeslee | 2005

Subscribe to the Microsoft Research Podcast:

Apple Podcasts

Email

Android

Spotify

RSS Feed

Transcript

[MUSIC]

DOUG BURGER: This is The Shape of Things to Come, a Microsoft Research Podcast. I’m your host, Doug Burger. In this series, we’re going to venture to the bleeding edge of AI capabilities, dig down into the fundamentals, really try to understand them, and think about how these capabilities are going to change the world—for better and worse.

In today’s podcast, I’m bringing on two AI researcher-experts: Nicolò Fusi, who is an expert in digital, transformer-based large language model architectures and learning, and Subutai Ahmad, who is an expert in biological architectures, specifically the human brain. And the question we’re going to discuss is, are machines intelligent?

And what I mean by that: are digital intelligence, large language models, on a path to surpass humans, or are the architectures just so fundamentally different that one will do one set of things well, the other will do something else very well? And so we’ll be debating the architecture of intelligence across digital implementations and biological implementations because the answer to that question, I think, really will determine the shape of things to come.

[MUSIC FADES]

I’d like to ask each of my guests to introduce themselves. Tell me a little bit about your background and what you’re currently working on—to the extent you can talk about it—in AI. So, Nicolò, would you please start?

NICOLÒ FUSI: Yeah, thank you, Doug, for having us and having me here. It’s so much fun. So I’m Nicolò Fusi. I’m a researcher at MSR [Microsoft Research]. So Doug is my boss, so I will be very, very, very good to Doug in this podcast.

No, but jokes aside, my own background is in Bayesian nonparametric. That’s what I started studying. So Gaussian processes and things like that. And then equally, I would say, in computational biology, because I found it, like, one of the most interesting use cases for AI techniques. And that, kind of, has been true throughout my career. And pretty much like everybody else, eventually, I moved away from the kernel methods and the Bayesian nonparametrics and I started working more on language models, transformer models, with a particular eye towards information theory and the connection between information theory and generative modeling. And that’s, kind of, one of the main things I do today other than, kind of, managing the research of people who do much more interesting work than I do. [LAUGHS]

BURGER: I have to interject there, Nicolò, because you dragged a piece of bait across my path.

FUSI: I figured.

BURGER: You know, at Microsoft Research, I have a management rule that I can’t tell anyone what to do because we hire some of the best people in the world. You have to trust them. And everyone is always completely free to call BS on me. And so Nicolò was joking there; [LAUGHTER] he does not have to toe the party line. In fact, I encourage him not to. So, so …

FUSI: I just have to be well-behaved. That’s the only thing I will say. [LAUGHS]

BURGER: Yeah. Thank you, thank you for baiting me. [LAUGHS] Because he knew exactly what he was doing. And I love him for it.

Subutai, can you tell us a little bit about yourself?

SUBUTAI AHMAD: Sure. Thank you so much, Doug, for having me. I’m really looking forward to the conversation between us all.

So I see myself fundamentally as a computer scientist. You know, I’ve been studying computer science for longer than I care to admit. But something changed for me during my undergrad years. I decided to minor in cognitive psychology, and I started to get really interested in how the brain works.

And to me, understanding intelligence and implementing intelligence was the hardest problem a computer scientists could ever solve. So I got very, very interested in that. You know, I couldn’t see how to really commercialize that. I was very interested in making products and stuff. So I stopped, you know, working on that for a while. I did a number of startups doing computer vision, you know, video processing, a lot of that stuff.

And then when Jeff Hawkins started Numenta back in 2005 with the idea of really deeply understanding how the brain works and figuring out how to apply that to AI, for me, it was like all my worlds coming together. This, like, this is what I had to do. None of us thought [LAUGHS] it would take as long as it did. We spent the last couple of decades really deeply trying to understand neuroscience from a computer scientist—from a programmer’s—standpoint, the underlying algorithms. And that’s really what I’m passionate about, just trying to translate what we understand about the neuroscience to today’s AI.

And in terms of what we’re working on today, it’s, you know, the human—maybe we’ll get into some of this—the brain is super efficient in how it works—power efficient, energy efficient—and we’re trying to embody those ideas and trying to make AI a lot more efficient than it is today.

BURGER: Great. I think we’ll get into efficiency a little bit later in the podcast because that’s a subject that’s near and dear to my heart, you know, being a computer architect originally by training.

I want to go back to, you know, one of the reasons I got involved with Numenta is, you know, Subutai and I have been exchanging emails, like, discussing collaborations, you know, visiting each other through the years, and the thing that really stuck with me was when I read one of the earlier books from Jeff On Intelligence (opens in new tab). And there was an example in the book that talked about how, you know, the human brain learns continuously. I think biological organisms in general learn continuously.

And the anecdote that I remember was this anecdote if you’re walking down your basement steps, you know, you’re walking down the stairs to your basement and there’s one step that’s always been a few inches off and you decide to fix it, and so you raise it so it’s even with the others, and then the next time you go down the stairs, you don’t remember and you’re wildly off and, you know, you hit that step, you hit it earlier or later than you anticipated, you go out of balance. You’re flailing around. You know, you get all this adrenaline. You think you’re going to pitch headfirst down the stairs. Hopefully you don’t. And then the second time you do it, you’re a little off balance, but it’s not crazy. And the third time you maybe notice a little bit, and the fourth time, it’s, like, it’s your basement stairs.

And so somewhere between that first time down and the third and fourth times down, there are molecular changes in your brain that have learned the new timing of your basement steps. And I remember just that example vividly from the book. And that got me thinking, wow, this is so different from the way our digital AI works. I’ll turn it over to you to comment for that and then I think we’ll go into the digital.

AHMAD: Yeah, no, that’s a great example. I think it’s remarkable how our brain is constantly modeling our entire world at such a granular level, and we’re not even aware of it perceptually. Like, you know, that example of the steps is probably not … you wouldn’t consciously be aware of it, yet if something is different about anything in your world that you’re very familiar with, you’ll instantly notice it. And then you’ll, you know, you’ll update your world model, you’ll adjust, and you’ll continue on. It’s really remarkable how the brain’s able to do that so seamlessly.

BURGER: And a lot of that is based on neurotransmitters, right? Because there’s just a … you know, when you have that physical reaction to “I’m about to pitch down the stairs,” you get a flood of transmitters that actually changes the way your brain’s learning or at least the rate.

AHMAD: Yeah, there’s a flood of neurotransmitters and neuromodulators, as well, that invoke change, sometimes very rapidly. Another example, you know, if you touch a hot stove—that’s the canonical example—you will learn that very, very quickly. So there’s a lot of chemical changes that happen. But it’s also really interesting that we can update things and update our world knowledge without impacting everything else that we know. This is something that’s very, very different, again, from today’s AI models. We’re able to make these changes in a very contextual and very, sort of, fine-grained way.

BURGER: So, Nicolò, I want to go and talk a little bit now to transformers. So I think, you know, you and I and Subutai were all working in the AI field, you know, many years before 2017, when the transformer hit. You know, I was building, you know, with my team hardware to accelerate RNNs [recurrent neural networks], LSTMs [long short-term memory], you know, which had this awful loop-carried dependence, you know, the bottlenecked computation, and then the transformer was just much more parallelizable.

So what do you think’s really going on in these things? And maybe we could start—I know you and I have talked a lot about this—maybe just start with the major blocks. You know, you’ve got the attention layer. You’ve got the feedforward layer. You’ve got, you know, the encoder stack and the decoder stack and the latent space in between. Can you just, kind of, walk us through those pieces at a high level and tell us what you think is going on?

FUSI: Yeah. Yeah, I mean, I have a very opinionated view of why transformers are so great.

BURGER: That’s why you’re here. [LAUGHS]

FUSI: Maybe, like, yeah, maybe I’ll inject it. I don’t know. I don’t think it’s a super novel creative opinion, but it is an opinion. So I guess the two principal … the two main components you already described: the, you know, the transformer [read: attention] layers and the feedforward layers. One way to think about them is, how does information in your context relate to each other and what is every token referring to, for instance, in the case of transformers in language models?

So by context, we mean, like, the information you feed through the model, that the model keeps continuously generating and appending to.

BURGER: So like your chat history.

FUSI: Your prompt. Your what? Your chat history or your particular prompt in a chat session.

BURGER: OK.

FUSI: That prompt, which is a sequence of words, gets discretized in a series of tokens. Tokens can be individual words, can be multiple words, kind of, connected together. The way we go from words to tokens typically is through an algorithm that tries to basically collapse as much as possible. Multiple words, like “the dog,” may be just one token as a first, kind of, level of compression to feed into the model. So it just tries to bring things together as efficiently as possible.

Then there is, you know, within these models, there is a transformer layer. This transformer layer or this attention layer, sorry, tries to basically figure out what the “the” refers to—the term “the” in “the dog,” or “the dog jumps on the table,” “jumps” refers to the dog. So there is this kind of, like, mapping that happens.

And then there is, like, feedforward layers, which in modern large language models, they store a lot of information. Like, that’s kind of, like, where the knowledge typically kind of sits in, the things that the model just knows. You know, that, I don’t know, if you slam your arm against [the] cup of water on your table, that cup of water falls off the table. That’s something that the model, kind of, has baked in through reading a lot about cups falling off of tables when they’re hit.

So that’s, kind of, those are, for me, the two fundamental components, and the reason why I have an opinionated view is that, you know, honestly, I do believe that RNNs and, you know, even state-space—modern incarnations of state-space models—are good enough to learn over these, you know, language data or whatever or vision data or audio data.

The good thing about transformers is that they do two things very well. One is they get out of the way. They don’t have this notion of “everything has to be encoded through a state” like recurrent networks. And two, they do that very computationally efficiently as you were saying. There isn’t a computational bottleneck. And so they created this nice overhang where they happen to be the right architecture at the right time to unlock enough flow of information through the model …

BURGER: Yeah.

FUSI: … that we could get through these amazing things.

BURGER: Let me press you on one thing. Like, you know, in the attention blocks, you can figure out which words or which tokens relate to which tokens. So I put in the prompt and it’s finding all the relations and then feeding those relations up to, you know, the feedforward layer—well, the feedforward unit within a layer. And you said that knowledge is encoded there, but then what does it really mean for those maps to then access knowledge, but then you project it back into, you know, the output and then feed it up to the attention block in the next layer?

FUSI: Again, yeah.

BURGER: So it seems kind of weird that I’d be, like, accessing knowledge and then taking that knowledge, merging it, and going back to another attention map.

FUSI: Well, you can see it as a mixing operation that happens in the feedforward part of the layer. You know, like, you’re attending, then you’re mixing, and, kind of, like, reprojecting to some space with higher-information content or, like, a different level of information extraction. And then you’re putting it back into, “OK, so let me do another round of processing” and, kind of, attending and then a mix again. And then I do it again and then I do it again.

So I think that the information that is present in the prompt and in the, you know, that has been baked into the weights gather further and further refined. Whether that refinement is extraction of structure or aggregation into higher-level concepts, I’m not sure. I think it’s just structure gets extracted and things that are irrelevant get kind of pushed away. But that doesn’t necessarily mean that it gets aggregated through the architecture.

BURGER: So now I’m going to try to, like, restate what I think I hear you saying. So, you know, we’re adding information and we’re kind of adding information at a higher level but not necessarily throwing away the low-level information, at least that’s not relevant, right?

FUSI: Yeah.

BURGER: Because, you know, if the higher-level stuff depends on the low-level stuff, I have to have that first. And so then you get to the top of the encoder block and you’re in the latent space with all of that information kind of maximized. Is that a way to think about it? And if you agree, can you talk about what the encoder block really is and what the latent space is?

FUSI: I tend to agree, yes. I mean, there is … you’re describing … I think you’re describing what I think is happening, which is there is given the context in your prompt and given the task that the model perceives or, like, figures out that you’re doing, it has to highlight and pull out the relevant information. And it does that not by summarizing layer by layer, but it does it by, you know, increasing the prominence of that information and suppressing other things. So I think that’s ultimately what happens up to the point where you reach this beautiful point in concept space, which identifies both your intent and the things in the prompt and in the knowledge of the model that are necessary to solve it.

BURGER: And so one last question, and then I want to go to Subutai for a second.

So now when we go through the decoder stack, are we just going the other way and stripping out the high-level concepts early and then getting down to the granular tokens? Or, you know … because you go up through the encoder stack, those attention blocks and feedforward layers, to get to that magical latent space. And now we’re going to go the other direction. How do you think about that other direction through the decoder stack, which is the same primitives as the encoder stack?

FUSI: Same primitives. You can think of it as kind of the reverse operation. Like you, you never lost information throughout. You just kind of suppress or privileged different kinds of information. And now you’re basically just projecting it back out to a space that is, you know, intelligible. And it’s, kind of, where the model gets it’s … I hesitate to use the term reward because it has a particular implication, but that’s, kind of, where the loss gets computed and then gets pushed back through the model.

BURGER: Right, as you’re trying to evolve and train all those parameters—the relationship between words, the information in the feedforward layers, the design of that latent space, and the extraction of the knowledge from it.

FUSI: That’s right. And so in encoder-decoder model, you push through the whole thing, you decode back to a particular token, which for people who don’t know, it’s, like, literally a number out of a vocabulary, like word No. 487. And if it was word No. 1,500, you get, you know, like, …

BURGER: Something else.

FUSI: … a bad reward. Yeah. Yeah. And then … and if you got it right, you get a positive signal that then just flows back through the model.

BURGER: I’d like to go over to Subutai now. So after hearing this, you’ve studied, you know, neuroscience and the neocortex and cortical columns and all of this for a long time, and you and I have had lots of debates. Is the human brain doing something different than that? You know, are we just building latent spaces, then extracting? The architecture is very different, but what’s going on under the hood?

AHMAD: Yeah, the architecture is very different. You know, as Nicolò was describing what happens throughout a transformer stack, I was trying to relay and relate, you know, what we know in the brain, as well.

In a typical, you know, transformer model, t