TRANSCRIPT 🔗
WEBVTT
00:00:10.240 --> 00:00:14.640
Friends, welcome and welcome back, I hope, to the Best2Go FinTech Podcast.
00:00:14.720 --> 00:00:15.919
I'm Fraser Anderson.
00:00:16.320 --> 00:00:24.079
Today I'm excited to be joined by Daniela Roos, Director of MIT's Computer Science and Artificial Intelligence Laboratory, or CSAL.
00:00:24.320 --> 00:00:29.039
And as you can imagine, we talk quite a lot about AI and some of the breakthroughs she's seeing at her lab today.
00:00:29.199 --> 00:00:32.640
Before we get to our conversation, a brief bit on Daniela's background.
00:00:32.799 --> 00:00:43.280
Daniela Roos is the director of MIT's Computer Science and Artificial Intelligence Lab, CSAL, and the Andrew Elma Viterbi Professor of Electrical Engineering and Computer Science.
00:00:43.439 --> 00:00:53.600
A trailblazer in robotics, artificial intelligence, and machine learning, Daniela's groundbreaking research spans soft robotics, underwater exploration, and autonomous systems.
00:00:53.759 --> 00:01:02.240
Her innovations include printable robots that unfold into functional machines, underwater robots for coral and fish studies, and algorithms for self-driving cars.
00:01:02.399 --> 00:01:12.719
As an IEEE fellow and member of the National Academy of Engineering, Daniela has earned numerous accolades, including the Engelberger Award for Robotics and the IEEE Edison Medal.
00:01:12.879 --> 00:01:16.239
She holds a PhD in computer science from Cornell University.
00:01:16.719 --> 00:01:18.400
Daniela, welcome to the show.
00:01:18.719 --> 00:01:20.159
Thank you so much, Fraser.
00:01:20.239 --> 00:01:22.000
I'm really excited to do this with you.
00:01:22.319 --> 00:01:34.640
I know we've been talking about it for some time, and you're so generous with your time with me, just given the demands, especially since your field seems to have made AI lab CEOs or AI lab directors rock stars.
00:01:34.799 --> 00:01:40.319
Could you maybe just tell the listeners a little bit about what CSAL is specifically and your role there?
00:01:40.799 --> 00:01:50.319
Maybe even the history of CSEL a little bit, because obviously those of us that are in the community know CSale extremely well by reputation, but I think our audience is a little bit broader.
00:01:50.799 --> 00:01:56.480
CSale stands for the Computer Science and Artificial Intelligence Laboratory at MIT.
00:01:57.200 --> 00:02:04.319
And it is the largest research unit at MIT with 1800 people, more than 1800 people.
00:02:04.799 --> 00:02:07.840
Now MIT has an interesting organization.
00:02:08.000 --> 00:02:09.680
We are organized as a matrix.
00:02:09.840 --> 00:02:18.719
We have departments where we conduct our teaching activity and we have research labs where we conduct uh the research.
00:02:18.960 --> 00:02:23.840
And CCL is a very, very special place with an extraordinary history.
00:02:24.240 --> 00:02:28.879
CSEL, you can think of it as having two parts, computer science and AI.
00:02:29.360 --> 00:02:33.759
And the AI side of computer science goes back to 1956.
00:02:34.159 --> 00:02:35.680
So that's the AI side.
00:02:36.000 --> 00:02:42.560
There is also an extraordinary history on the computer science side that goes back to 1963.
00:02:43.039 --> 00:02:47.520
When the big dream was for two people to use the same computer at the same time.
00:02:48.400 --> 00:02:51.840
And in those days, computers uh were big and bulky.
00:02:52.000 --> 00:02:55.439
In fact, in 1963, computers occupied a whole room.
00:02:55.520 --> 00:02:55.680
Yeah.
00:02:56.000 --> 00:02:59.280
And they were really curiosities for experts.
00:02:59.840 --> 00:03:19.680
And so that also is is really quite extraordinary because uh if you think about it, in 60 short years, we moved from dreaming about people using the same room-sized machine um at the same time to everyone computing on their smartphone devices.
00:03:19.759 --> 00:03:24.479
And computing is now so important in the world, it it affects everything.
00:03:24.719 --> 00:03:33.759
It uh it impacts how we live, it impacts how we work, it impacts how we play, it impacts scientific discovery, computing has become so important.
00:03:34.319 --> 00:03:42.159
And so along the way, MIT has made a lot of um important contributions that we now take for granted.
00:03:42.479 --> 00:03:49.520
Um, so for instance, Ethannet was born while Bob Metcalfe was a student at MIT.
00:03:50.000 --> 00:04:00.319
Along that trajectory, Tim Berners-Lee uh is now a member of CSEL at MIT, and MIT hosted for a long time the World Wide Web Consortium.
00:04:00.479 --> 00:04:07.599
Um, the first object-oriented programming language called Clue was invented at MIT by Barbara Liskov.
00:04:07.919 --> 00:04:11.520
And this is now the foundation of the whole software industry.
00:04:11.759 --> 00:04:18.639
Public key encryption that allows all of us to purchase things with our credit cards was invented at MIT.
00:04:19.120 --> 00:04:34.639
And the concept of mobile robots and computer vision systems and language, uh, natural language interaction between machine and people, all of these can be traced to important work that took place at CSEL.
00:04:35.040 --> 00:04:44.800
I would heartily recommend to anybody who's listening to this to go and look at a picture of CSEL because given what you've heard from Daniela, it looks just about exactly as you would expect.
00:04:44.879 --> 00:04:50.879
I mean, it's this incredibly futuristic, cavernous building with a kind of exoskeleton.
00:04:50.959 --> 00:04:52.399
I mean, it's just remarkable.
00:04:52.560 --> 00:04:58.959
And and uh I've had the pleasure of meeting with you in your office, and your office looks uh just about exactly as I would imagine it.
00:04:59.040 --> 00:05:01.759
There are sort of robots everywhere.
00:05:02.000 --> 00:05:20.480
And I think if we zoom back to kind of the founding of CSale, it's it's sort of easy to imagine back then how there's so much transformational change
Load full transcript
today, but even in the last 20 years, I mean, culturally, like, can you believe we're living in a world where like the CEOs of research labs are kind of treated like rock stars, they're speaking at Davos?
00:05:20.720 --> 00:05:25.360
Like, I mean, tell me kind of what your experience has been from when you got started to today.
00:05:25.839 --> 00:05:38.079
So a few decades ago, AI research was quite specialized, and it was primarily an academic pursuit that was associated with research labs and universities.
00:05:38.399 --> 00:05:49.519
Researchers like Marvin Minsky and John McCarthy were very influential within the field, but uh it's not clear how well known they were outside of the field.
00:05:49.759 --> 00:06:00.399
And so at that point in time, AI had an impact in science and in engineering, and this impact was uh theoretical and experimental.
00:06:00.560 --> 00:06:04.800
Um, but there was not such a direct influence on daily life.
00:06:05.279 --> 00:06:26.240
But today we see AI as having the potential to power everything, and we already see important applications from recommendation algorithms uh to self-driving cars, to support in medical diagnosis, and to large-scale language models that have become so uh prevalent.
00:06:26.720 --> 00:06:35.759
And so the influence of AI, um and in particular the influence of the AI CEOs, is no longer confined to technology.
00:06:35.920 --> 00:06:40.160
Yeah, because the AI leaders shape the economy.
00:06:40.399 --> 00:06:49.199
They are part of all the conversations around ethics, around education, uh around labor, and even geopolitics.
00:06:49.360 --> 00:06:49.600
Yeah.
00:06:50.000 --> 00:07:06.639
So AI leaders like Demis Hassabis and Daria Amode are recognized for their technical contributions, but they are also seen as thought leaders who are helping to navigate society's relationship uh with AI.
00:07:06.879 --> 00:07:14.879
Um it's it's really extraordinary because we are at the point in time where AI has become uh democratized.
00:07:15.040 --> 00:07:15.120
Yeah.
00:07:15.360 --> 00:07:20.000
But this doesn't mean that AI is going to solve all of our problems.
00:07:20.319 --> 00:07:23.279
It doesn't mean that AI is going to take down the world.
00:07:23.519 --> 00:07:31.600
But AI has the power uh to do so much for humanity and for the planet uh in economics.
00:07:31.759 --> 00:07:35.680
Uh the valuation of AI companies is skyrocketing.
00:07:35.759 --> 00:07:35.839
Yeah.
00:07:36.079 --> 00:07:43.040
And and the leaders of these companies are really seen as the architects of the future industries.
00:07:43.279 --> 00:07:52.480
And so this this uh capacity, this potential for economic uh impact surpasses everything we have seen in the past.
00:07:53.120 --> 00:08:00.639
You know, AI leaders are not just shaping their companies and uh and economies, they're also influencing global policies.
00:08:01.120 --> 00:08:06.000
Um they are influencing government uh regulation and control.
00:08:06.480 --> 00:08:09.120
And so there is a lot of there was a lot of impact.
00:08:09.199 --> 00:08:10.800
There was a lot of broad impact.
00:08:11.199 --> 00:08:29.680
Yeah, I mean I I recall very distinctly last year when you guys uh you and I were at um the World Economic Forum, you were on a panel with Jan Lakun and then a couple of other folks who maybe took the other side, and I won't mention their names, but you know, you two or you four were discussing AI safety, and I mean this was the World Economic Forum.
00:08:29.759 --> 00:08:34.799
I mean, the stage couldn't have been bigger, like it could not have been a less esoteric topic.
00:08:34.960 --> 00:08:48.639
You know, given the kind of breakthroughs we've seen maybe in the last two or three years, how are you thinking about AI safety today, you know, in terms of risks and and maybe to the extent to which some of the the sort of imminent fears are overblown?
00:08:48.720 --> 00:08:51.200
I'd I'd just be curious to get your latest thinking.
00:08:52.080 --> 00:09:08.000
Well, so my latest thinking has not changed much, and I remember very well that panel because it was a really uh special moment where the conversation felt quite balanced between ambition and responsibility.
00:09:08.240 --> 00:09:08.399
Yeah.
00:09:08.639 --> 00:09:14.879
And so my thinking about AI safety has evolved, but the core remains the same.
00:09:15.039 --> 00:09:20.480
Uh, it's all about building systems we can trust and we can certify.
00:09:20.639 --> 00:09:28.720
It's about building systems that align with our values and that enhance society without unintended harm.
00:09:29.200 --> 00:09:31.600
This last part is actually very important.
00:09:32.000 --> 00:09:39.039
The the computer science community is working hard on ensuring all of these aspects of AI.
00:09:39.200 --> 00:09:46.559
Uh, some of these problems already have solutions, some of them are very much uh core research topics.
00:09:46.879 --> 00:10:01.279
But I I personally am less concerned with dystopian ideas about how AI is going to take down the world and more focused on how do we put in place practical safeguards now?
00:10:01.440 --> 00:10:01.600
Yeah.
00:10:01.759 --> 00:10:06.559
Um so how do we ensure that AI is not biased?
00:10:06.799 --> 00:10:11.200
How do we ensure that the output of an AI solution is safe?
00:10:11.759 --> 00:10:22.399
How do we ensure that the um that we have good understanding of the error and uncertainty regimes uh of AI systems?
00:10:22.559 --> 00:10:26.720
How do we get robust mechanisms to prevent misuse?
00:10:27.120 --> 00:10:34.960
Yeah, and so these are all complex challenges, but there are challenges that are well recognized and we're beginning to make progress.
00:10:35.120 --> 00:10:51.919
So, for instance, on the topic of bias, we already have technical solutions uh that are possible to deploy immediately because the output of a machine learning model is related to the input that is used to
train that model.
00:10:52.240 --> 00:10:57.759
If you have bias in the input, you are going to have bias in the um in the output.
00:10:58.480 --> 00:11:13.120
And so we can take our data sets, we can analyze them according to the properties that we care about, and we can resample them so that what's overrepresented and underrepresented gets rebalanced in the data set.
00:11:13.440 --> 00:11:28.559
And so you see some problems have technical solutions, other problems will require a combination between technical solutions and uh governance in order for us to really be confident in what we have.
00:11:29.200 --> 00:11:35.200
And when I think about kind of outputs that are unsafe, it seems to fall into kind of two broad categories for me.
00:11:35.360 --> 00:11:51.200
One is sort of like knowledge that shouldn't be widely available, like how do I make a chemical weapon or how do I build a nuclear device and LLM being able to kind of walk me through how to do that, or just advice that is unreliable that may cause harm if I act on it.
00:11:51.440 --> 00:11:54.000
You know, are there others that you would call out?
00:11:54.320 --> 00:12:01.519
And uh of those two, where are you seeing the most promise in terms of kind of technical solutions to mitigate harm?
00:12:02.320 --> 00:12:08.399
So in my research group at MIT, we are working on a technique we call barrier net.
00:12:09.360 --> 00:12:17.679
And uh this technique ensures that the output of a machine learning prediction engine falls within a safe region.
00:12:18.320 --> 00:12:25.360
So that requires that humans who use the machine learning engine specify what is a safe region.
00:12:25.679 --> 00:12:36.559
So here's how roughly how it works: you take a machine learning model and you wrap it into another layer that introduces something called a control barrier function.
00:12:36.720 --> 00:12:40.720
This is a technique that's used in uh the controls literature.
00:12:41.360 --> 00:12:47.519
With a barrier function, we force the output of the machine learning model into a safe region.
00:12:47.919 --> 00:13:10.960
And so, if, for instance, if we want to deploy this machine learning model in a physical application, maybe to drive a car or to fly a plane, with Barrier Net, we can write a mathematical formula that will ensure that whatever forces and torques the model proposes are actually executable, are safe, are not going to prevent collisions.
00:13:11.360 --> 00:13:16.799
But you can take the same idea and apply it on other in other domains.
00:13:16.879 --> 00:13:22.960
So for instance, you can specify mathematically what might constitute toxic um speech.
00:13:23.120 --> 00:13:23.200
Yeah.
00:13:23.440 --> 00:13:32.320
And uh through BarrierNet, you will be able to eliminate toxic speech from the output uh of the uh machine learning engine.
00:13:32.399 --> 00:13:36.240
So that's one way of addressing um uh safety.
00:13:36.399 --> 00:13:38.639
Um but then there are other issues.
00:13:38.799 --> 00:13:49.279
There are issues that have to do with how these open source tools uh might fall in the in the hands of villains, of supervillains.
00:13:49.679 --> 00:14:01.919
And so I'm concerned about cybersecurity impacts, for instance, because today you you or maybe yesterday, you really needed to be a computer hacker.
00:14:02.000 --> 00:14:09.200
You needed to have a lot of uh experience and knowledge about computers in order to launch a cyber attack.
00:14:09.360 --> 00:14:09.600
Yeah.
00:14:10.159 --> 00:14:17.759
But uh the support from large language models and uh co-pilots can lower the barrier to entry.
00:14:18.080 --> 00:14:22.879
So the way I see it, uh AI can give us extraordinary powers.
00:14:22.960 --> 00:14:42.639
Uh it can help us with our speed and productivity, it can help us with our knowledge, with our insights, with our creativity, with our foresight, with our mastery, even with respect to empathy, AI can really improve how we connect uh with each other and with the rest of the planet.
00:14:42.960 --> 00:14:50.480
But the same tools that help the good people get better in in these areas also help the supervillains.
00:14:50.799 --> 00:14:57.519
And so here it's important to think about how do we build the guardrails that prevent that sort of thing.
00:14:57.679 --> 00:15:01.440
And I would say one more thing, I know you're I I can see that you're about to jump.
00:15:01.759 --> 00:15:10.480
But I would say that it's I mean, in some sense, we have this extraordinary tool that can do good and that can also do evil.
00:15:10.720 --> 00:15:17.519
Much like you can buy a knife and use it to prepare a healthy vegetarian dinner.
00:15:18.080 --> 00:15:22.399
Which I do tools, but that same tool can uh can also cause harm.
00:15:22.639 --> 00:15:38.399
And so in this case, it's important to keep in mind that we we have to combine our technical guardrails along with societal uh guardrails to ensure that we direct these tools for the greater good.
00:15:39.120 --> 00:15:47.679
Well, I mean, the mention of cooking a delicious vegetarian meal and some of the power of AI brings me to something which is very near to my heart, which is my stomach.
00:15:47.759 --> 00:15:54.720
And I love using LLMs to cook with, as do you, I believe, uh, which you mentioned in your book Minds Mirror, which everybody should read.
00:15:55.200 --> 00:16:04.799
And and so I think that gets us to the way that I think a lot of people have been experiencing AI, and that's really kind of through ChatGPT or Anthropic or any of these LLMs.
00:16:05.120 --> 00:16:16.000
And while this might be getting a little esoteric, you know, the the reason we're able to have this new functionality is is largely a consequence of scaling data and scaling compute.
00:16:16.240 --> 00:16:38.720
As someone who's not building foundational models, it it seems to me anyway that scale is this kind of like new vector that we can use to build capabilities we didn't
before and was probably somewhat non-obvious in a research context because it just takes so much energy, capital expenditure, um, you know, uh engineering to even test it.
00:16:39.039 --> 00:16:50.960
What I'm curious to get from you is um, you know, outside of the domain of LMs within your lab, are there exciting areas of research where it's like, oh, now we could use scaling for this.
00:16:51.120 --> 00:16:57.440
We could use scaling for this, whether that could be video or or robotics use cases.
00:16:57.600 --> 00:17:05.839
I don't know if anything kind of comes to mind there that you're thinking, wow, if we could tie together 64,000 GPUs, we could do something really interesting.
00:17:07.200 --> 00:17:13.599
So uh Frasier, let me let me start by clarifying what the scaling hypothesis is.
00:17:13.759 --> 00:17:14.079
Okay?
00:17:14.480 --> 00:17:25.920
So the scaling hypothesis says that AI models grow larger through increased parameters, increased training data, and increased compute power.
00:17:26.240 --> 00:17:32.799
And so the performance improves in ways that uh lead to surprising um capabilities.
00:17:33.359 --> 00:17:51.759
This idea is particularly significant because it suggests that breakthroughs in AI can be achieved by simply uh increasing scale, uh making the models increasingly bigger without considering new ideas or architectures or methods.
00:17:52.160 --> 00:17:59.279
And the current ideas that are being scaled up are ideas that have been around for a long time, for decades, in fact.
00:17:59.599 --> 00:18:04.160
Reinforcement learning, uh multilayer neural networks, deep networks.
00:18:04.400 --> 00:18:09.039
These are all ideas that have invented have been invented many decades ago.
00:18:09.200 --> 00:18:14.480
They did not work well back then because we didn't really have much computation.
00:18:14.640 --> 00:18:19.920
The models had like, you know, five neurons, yeah, or ten neurons, and there's only so much you can do with that.
00:18:20.000 --> 00:18:28.240
But now all of a sudden, through this expansive computation and data for training, we see extraordinary capabilities.
00:18:28.960 --> 00:18:34.640
But these extraordinary capabilities come as a at a cost, and there are also limitations.
00:18:34.880 --> 00:18:49.839
So in order to train and to uh do inference on these huge models, we need a huge amount of uh energy, and this actually has uh a negative impact uh on our climate.
00:18:50.000 --> 00:19:07.599
You see the computational energy and data requirements for larger and larger models grow exponentially, and that also makes them uh increasingly more expensive and more um restricted to the companies uh who have the means to do it.
00:19:07.839 --> 00:19:13.440
And also there is an element of sustainability uh that that goes with all this.
00:19:13.680 --> 00:19:22.079
So having uh having said that, let me say that there are extraordinary things we can do through this kind of uh scaling.
00:19:22.319 -->
00:19:22.480
Yeah.
00:19:22.880 --> 00:19:28.480
And in fact, the first moment in scaling was when OpenAI released uh GPT-4.
00:19:29.039 --> 00:19:35.119
And I remember that moment very well because it was really at CSEL that this announcement was made.
00:19:35.279 --> 00:19:45.359
Sebastian Bubeck came to CSEL in March of 2023, and he gave a talk that kind of energized and awed everyone.
00:19:45.599 --> 00:19:50.640
Um his talk was uh sparks of artificial general intelligence.
00:19:50.799 --> 00:20:06.160
And so this was the first time when we could see the extraordinary capabilities of uh GPT for to uh in in language generation, in generating code, in generating games, in in doing so many powerful things.
00:20:06.319 --> 00:20:12.400
The the tool was not so good at math back then, nor at common sense reasoning, but we were all blown away.
00:20:12.559 --> 00:20:12.720
Yeah.
00:20:12.960 --> 00:20:18.640
And since then, there has been so much work and so many extraordinary contributions.
00:20:18.960 --> 00:20:36.000
So, for instance, in computer vision, we have introduced vision transformers and stable diffusion algorithms that have leveraged scaling to create very high-quality images from textual prompts and even videos.
00:20:36.160 --> 00:20:36.319
Yeah.
00:20:36.559 --> 00:20:42.559
And this really demonstrates the power of combining scaling up with sort of new ideas.
00:20:42.640 --> 00:20:46.640
And in this case, the idea was to combine vision and text.
00:20:46.799 --> 00:20:59.359
And and these days we're using the same ideas to connect perception, action, and language into a kind of a more intelligent uh physical creature.
00:20:59.519 --> 00:21:19.039
And in fact, uh many people in my group and outside of my group at MIT and in the research labs of my friends from across the world are really looking at the ability of connecting language, connecting large language models into physical problems.
00:21:19.279 --> 00:21:30.559
And because you see, like if you think about a machine that needs to make sense of the world through vision alone, that machine has to reason about pixels and about geometry.
00:21:30.640 --> 00:21:32.000
It's very low-level reasoning.
00:21:32.079 --> 00:21:34.880
It's not akin to the human-level reasoning.
00:21:35.119 --> 00:21:57.359
But when we introduce large language models and when we have the ability to connect uh the concepts we can express in language uh to specific spatial patterns that we see in pixels, we kind of elevate the level at which a machine learning system reasons about the world and we make that much more level with humans.
00:21:57.599 --> 00:21:59.839
We really begin to aspire to machine learning.
00:22:00.160 --> 00:22:03.359
Machines that adapt to humans rather than the other way around.
00:22:03.759 --> 00:22:04.400
I like that.
00:22:04.640 --> 00:22:18.720
You mentioned briefly sort of like new architectures, given that a lot of, you know, a lot of the architectures we use today were invented a long time ago, and we've tweaked them to accommodate for the you know abundance of compute we have available to us today.
00:22:18.880 --> 00:22:21.599
You know, are are there new arch architecture that you're working on?
00:22:21.680 --> 00:22:24.079
You know, does it turn out that retention isn't all we need?
00:22:25.519 --> 00:22:26.799
That's a great question.
00:22:27.039 --> 00:22:35.519
So it turns out that attention might be great for headlines, but for true intelligence, we'll need a bit uh more.
00:22:35.759 --> 00:22:42.720
And so we will need feedback, we will need memory, and um even uh a sprinkle of common sense.
00:22:42.880 --> 00:22:48.319
So attention is still at the party, but attention cannot host the entire show.
00:22:48.720 --> 00:22:50.400
So uh what do I mean by this?
00:22:50.559 --> 00:22:56.160
So transformers operate in a kind of a very, very simple way.
00:22:56.319 --> 00:22:58.559
I mean, it's simple at a conceptual level.
00:22:59.039 --> 00:23:01.039
It's really complex behind this.
00:23:01.279 --> 00:23:21.920
But the idea is given a certain context, for instance, given a sequence of words, what is the next word that's most likely to uh to fit uh with the current context based on training the transformer on all the textual information that is out there, if the transformer is about uh language.
00:23:22.160 --> 00:23:32.160
And so transformers uh really operate um uh through statistical correlations and do this kind of next uh token prediction.
00:23:32.319 --> 00:23:49.440
Uh in fact, if you um read uh the attention is all you read paper, uh the second, third, or fourth sentence in the paper says we eliminate convolutions and recurrences in order to ensure efficiency in parallelism.
00:23:49.839 --> 00:23:58.000
And it's extraordinary what can be achieved with this with a statistically based, paralyzed, efficient approach.
00:23:58.240 --> 00:24:02.400
And uh, you know, we've all used the tools, we're in awe of them.
00:24:02.720 --> 00:24:10.319
But the the thing is that there is uh an important role for feedback, which transformers do not have.
00:24:10.559 --> 00:24:16.960
So feedback is so important for our own intelligence, just think about how we learn about anything and everything.
00:24:17.119 --> 00:24:17.359
Yeah.
00:24:17.839 --> 00:24:18.160
Right?
00:24:18.319 --> 00:24:24.240
So a straight-up feed-forward solution will not benefit from feedback.
00:24:24.640 --> 00:24:41.680
And so I'm interested in uh in mechanisms uh and architectures and algorithms that incorporate feedback uh that are efficient, uh, that are uh expressive and also that are causal, that are provably causal.
00:24:42.319 --> 00:24:49.599
Because I'm very interested in how we uh use machine learning to interact with the physical world.
00:24:50.079 --> 00:24:52.799
And this means we have to understand physics.
00:24:52.960 --> 00:24:56.079
Well, statistics is what defines transformers.
00:24:56.240 -->
00:25:05.200
And so uh transformers do not necessarily understand physics, uh, they do not have common sense, they do not have uh the rules of the world.
00:25:05.279 --> 00:25:10.559
In fact, they cannot capture spatiotemporal uh relationships.
00:25:10.799 --> 00:25:32.799
And this is a the these are areas that we need to uh improve on because transformers excel at processing sequential data, but they are not inherently designed to handle the kind of complex temporal relationship or spatial temporal relationships uh that are crucial for interacting with the world.
00:25:32.960 --> 00:25:52.240
And you can imagine, you know, learning how an object moves through space over time, weather prediction, understanding the forces that are required in order to get a machine to do laundry folding or to or to uh unload your dishwasher, all of these require something more.
00:25:53.039 --> 00:26:00.400
And so so we have been thinking about these ideas, and we came up with a new approach for machine learning.
00:26:00.480 --> 00:26:03.200
We call this approach liquid networks.
00:26:03.519 --> 00:26:12.400
And uh liquid networks are physics-based technologies, different from the more traditional transformer-based technologies.
00:26:12.640 --> 00:26:26.640
And liquid networks are also inspired by nature, they are inspired by the biology of what uh biologists have discovered happens in the neurons of small species.
00:26:27.200 --> 00:26:36.000
And surprisingly, uh the neurons of small species compute differential equations, uh, not uh not discrete computations.
00:26:36.160 --> 00:26:45.119
And so, based on this idea, we have a line of research in liquid networks and also oscillatory space networks.
00:26:45.279 --> 00:26:51.200
And with these ideas, we're actually showing that we can increase the expressive capacity of the models.
00:26:51.519 --> 00:27:00.000
We can get provably causal solutions, we can get adaptive solutions that continue to learn after training based on the inputs that they see.
00:27:00.160 --> 00:27:01.599
And we get efficiency.
00:27:01.759 --> 00:27:09.359
We can get hundreds of times more efficient training and uh and imprints for our models.
00:27:09.519 --> 00:27:36.480
And and we think that this actually has enormous capacity for impact because with these ideas you can imagine um machine learning solutions that do not require server farms, that you can train on your enterprise computers that can run on sensors and uh power grids and other edge uh devices and that do not have such a huge uh carbon footprint.
00:27:36.960 --> 00:27:41.279
And this is just a kind of like investor hat on for a second.
00:27:41.519 --> 00:27:49.279
Are there things that are exciting to you about the idea of taking this efficient architecture and then applying like a server farm to it?
00:27:49.440 --> 00:28:08.079
Like, is it able to, you know, if you took the same amount of compute with this architecture on a language data set, is that somehow deeper
understanding, or is it more interesting in things like video where you know sort of next pixel prediction is not really particularly feasible right now?
00:28:08.160 --> 00:28:16.240
We we're doing different things on the back end, but I mean, uh are there areas that like uh liquid neural networks plus scale are extremely exciting to you?
00:28:16.960 --> 00:28:19.519
Scale uh improves everything.
00:28:19.680 --> 00:28:21.839
So scale improves even liquid networks.
00:28:22.000 --> 00:28:22.160
Yeah.
00:28:22.400 --> 00:28:28.799
And so liquid networks are especially suited for applications that have sequence data.
00:28:29.039 --> 00:28:29.200
Okay.
00:28:29.359 --> 00:28:35.599
So this means language, this means videos, this means um sensors, financial transactions.
00:28:35.759 --> 00:28:42.079
There are a lot of applications that are are defined by sequential uh data streams.
00:28:42.319 --> 00:28:58.160
What we see with liquid networks is that scaling helps, but at the same scale, liquid networks are much more efficient and much more uh performant than similar models that are based on different technologies.
00:28:58.319 --> 00:28:59.680
So that's very exciting.
00:28:59.920 --> 00:29:12.640
We have begun to see some reports, some experimental studies that are questioning the uh scalability of our current uh existing models.
00:29:12.799 --> 00:29:16.079
And we are beginning to see reports of diminishing returns.
00:29:16.480 --> 00:29:33.119
So that means that continuing to increase the data and the computation and the size of the model used by uh other types of architectures, it's not clear that the investment is yielding a huge slump in performance.
00:29:33.599 --> 00:29:40.160
And so with liquid networks, we have the uh ability to kind of bridge that gap.
00:29:40.559 --> 00:29:52.319
And because we have feedback mechanism, because we have uh memory in the liquid network solutions, we really have increased capacity for performance.
00:29:52.400 --> 00:30:03.599
And in fact, the the numbers actually show that the models are highly performant at the fraction of uh energy consumption for both inference and training.
00:30:03.920 --> 00:30:09.200
There's gonna be a lot to come, I know, from from you out of the future on this specific model architecture.
00:30:09.279 --> 00:30:10.880
So we'll probably leave that there.
00:30:11.039 --> 00:30:24.640
Uh I did want to talk about Minds Mirror, your book, which I have read twice, uh, and I recommend that everyone reads it, because you're a wonderful writer, and just given the moment that we're at in technology, it's it's probably good for everybody to get grounded in this information.
00:30:24.960 --> 00:30:49.599
One thing that I detected throughout the book, which I think is particularly characteristic to you, I think given that you're a scholar and a teacher, the sort of relish with which you describe how LLM systems are kind of like knowledge acquisition machines, sort of hyper-teaching assistant or even a hyperteacher, from your perspective, is that a component of the capability we have?
00:30:49.759 --> 00:30:52.640
Does it make this field particularly special to you right now?
00:30:52.799 --> 00:30:58.400
Like was I reading too much into that given that we've uh I you know I know your background.
00:30:59.039 --> 00:31:02.400
Well, Fraser, I love this question and thank you for asking it.
00:31:02.559 --> 00:31:16.160
And so absolutely, the idea that large language models are knowledge acquisition machines really resonates uh with me, especially as I think about their parallel with how scholars work.
00:31:16.480 --> 00:31:23.920
They're much faster than us, uh, but they share a lot of our they share a lot of the scholarship uh activities.
00:31:24.000 --> 00:31:30.079
So for instance, large language models process and synthesize vast amounts of information.
00:31:30.240 --> 00:31:35.119
They distill insights, they identify patterns across domains.
00:31:35.359 --> 00:31:40.240
They operate much like a meticulous researcher would.
00:31:40.720 --> 00:31:47.359
So they're not just tools for retrieval, but they're also tools for exploration and expansion.
00:31:47.680 --> 00:31:59.359
And they they empower scholars to focus on the higher level aspects of their roles, like synthesis, interpretation, even creativity.
00:31:59.599 --> 00:31:59.839
Yeah.
00:32:00.160 --> 00:32:08.960
The other thing is that large language models don't just read one book, they read an entire library much faster than we can.
00:32:09.039 --> 00:32:22.480
And so that means they span disciplines, they span perspectives, they span styles, and and they really provide us with an extraordinary tool for connecting ideas from different uh fields.
00:32:22.720 --> 00:32:34.640
You you you can see that there is a parallel between this process and the process through which scholars generate breakthroughs by finding uh new insights and relationships.
00:32:35.119 --> 00:32:43.359
I I guess the difference is that scholars spend years reading uh and LLMs can read the same books in in in just moments.
00:32:43.519 --> 00:32:52.960
They provide summaries, they can provide analogies, uh, they even provide nuanced expectations tailored to how you create the prop.
00:32:53.759 --> 00:32:54.799
Tremendous.
00:32:55.359 --> 00:32:56.720
Why'd you write the book?
00:32:57.200 --> 00:33:05.039
You know, I wrote the book because I I wanted to uh share with the public what AI is and what AI isn't.
00:33:05.440 --> 00:33:12.559
And I wanted to explain to the public what AI can do for them, how they can dream of a better life uh with machines.
00:33:12.720 --> 00:33:12.960
Yeah.
00:33:13.200 --> 00:33:17.359
And so I structured the book in three parts for that reason.
00:33:17.519 --> 00:33:28.880
Now the first part is called powers, and this part uh talks about the the seven powers speed, knowledge, insight, creativity, foresight, mastery, and empathy.
00:33:29.039 -->
00:33:40.319
And each of these chapters is backed with existing work either in companies or in research labs and fantasizes a bit about what might happen in the future.
00:33:40.480 --> 00:33:40.720
Yeah.
00:33:40.960 --> 00:33:45.440
And so these are these are the good ways in which AI can help us.
00:33:45.680 --> 00:33:45.920
Yeah.
00:33:46.160 --> 00:33:50.000
Then I wanted for people to really understand how AI works.
00:33:50.160 --> 00:33:54.079
I wanted people to understand that there is more than generative AI.
00:33:54.240 --> 00:33:56.079
There are many different types of AI.
00:33:56.240 --> 00:34:03.920
So with AI, um, you can generate, you get support for decision making, you get support for optimization and for prediction.
00:34:04.000 --> 00:34:09.760
And there are different algorithms and techniques for each of these different aspects of AI.
00:34:10.000 --> 00:34:13.679
And the best business solutions are really a combination.
00:34:13.840 --> 00:34:20.079
They're not just one or the other, but uh they're a combination of these multiple approaches.
00:34:20.320 --> 00:34:20.559
Yeah.
00:34:20.800 --> 00:34:31.760
So I wanted to also present business leaders uh with a framework to decide whether AI is good for their business or not.
00:34:32.000 --> 00:34:45.519
And so there is something I call the AI playbook that um lists a suite of questions that um business leaders can uh uh pose before they get uh going with AI.
00:34:45.840 --> 00:35:06.559
And so uh the points come down to a few things uh defining why, in other words, the strategic objective for using AI, assessing readiness, starting small and then experimenting, scaling up and ensuring that there are feedback and training programs to continue to uh grow the solution.
00:35:06.719 --> 00:35:08.159
And so that's the second part.
00:35:08.320 --> 00:35:15.039
And then the third part is about what are the technical, societal, and economic challenges around AI.
00:35:15.280 --> 00:35:27.039
Because just because you have uh an AI solution for a problem, it doesn't mean that it makes uh economic sense to deploy AI because there's also a cost associated with AI.
00:35:27.199 --> 00:35:31.840
So some of these ideas are discussed uh in the third part.
00:35:32.159 --> 00:35:40.320
And then, of course, the book is called the Mind's Mirror because uh AI is kind of like our reflection in a pond.
00:35:40.480 --> 00:35:40.880
Yeah.
00:35:41.119 --> 00:35:44.960
So you can see it sort of looks like us, but it's not really us.
00:35:45.039 --> 00:35:46.480
And that's an important point.
00:35:46.880 --> 00:35:55.760
Uh AI is not us, AI does not reason according to the processes that um uh give intelligence to humans.
00:35:55.920 --> 00:35:57.760
Or at least not as of now.
00:35:58.000 --> 00:35:59.679
Who knows what will happen in the future.
00:36:00.239 --> 00:36:09.199
When I asked you to sign uh my copy of Mind's Mirror, which you were kind enough to do without making me feel too embarrassed, you wrote better future with AI.
00:36:09.360 --> 00:36:18.639
And uh when you were answering that that last question, you started by saying something like, I wanted people to imagine a better future with machines.
00:36:19.119 --> 00:36:29.840
And that makes me think of uh a young woman, presumably very precocious and very good at math, growing up in Romania and becoming fascinated with robots.
00:36:30.079 --> 00:36:40.880
You know, if it's not too personal, would you know, would you tell me a little bit about when you started to imagine like a brighter future with AI or a brighter future living alongside machines?
00:36:41.679 --> 00:36:42.559
Well, let's see.
00:36:42.719 --> 00:36:49.440
So I grew up in uh Romania uh during a very oppressive dictatorship.
00:36:49.679 --> 00:36:54.320
And at the time there was not much entertainment.
00:36:54.559 --> 00:37:02.239
Times were very different, there were food shortages, but we had really good uh education in schools and we had great books.
00:37:02.400 --> 00:37:15.039
So growing up reading was really the point where I I started fantasizing of superpowers, of of better uh of of better circumstances.
00:37:15.360 --> 00:37:38.000
And I find it really quite extraordinary that the kinds of things that were once confined to storybooks, the kind of powers that were once confined to storybooks, are now becoming reality uh through advancements in computation, in AI, in materials, in in robots, and you know, broadly in science.
00:37:38.159 --> 00:37:43.920
And you know, each one of us can now become our own superheroes because of these advancements.
00:37:44.079 --> 00:37:49.119
And I think this is really an extraordinary point in time for all of us.
00:37:49.599 --> 00:37:53.119
What and you know, robotics, it it sounds like was your first love.
00:37:53.280 --> 00:38:03.199
Do you think it like it always will be, or like uh is the is computation becoming so consuming based on the pace of innovation that's pulling you away from robots?
00:38:03.360 --> 00:38:06.000
Or do you think they're they're completely inseparable?
00:38:06.559 --> 00:38:14.000
I mean the reason I became interested in robots is because I was interested in computation that interacted with the physical world.
00:38:14.239 --> 00:38:14.400
Okay.
00:38:14.880 --> 00:38:18.960
When I started graduate work, computation was really digital.
00:38:19.199 --> 00:38:25.840
We were developing discrete algorithms that operated in the confines of of the digital domain.
00:38:26.480 --> 00:38:29.920
But the physical world uh is so much more complex.
00:38:30.079 --> 00:38:38.719
We have continuous, not discrete processes, we have uh uncertainty, we have errors, we have a lot of messiness in the physical world.
00:38:39.119 --> 00:38:45.360
And this is true whether you're talking about robots or any other physical interaction.
00:38:45.519 --> 00:38:56.800
And so my my PhD work um started with this aspiration to advance how we think
about computation that interacts with the physical world, not just the digital world.
00:38:56.960 --> 00:38:57.199
Yeah.
00:38:57.360 --> 00:38:59.679
And robots are one embodiment of that.
00:39:00.000 --> 00:39:04.800
Now, when we talk about robots, uh robots have a body and a brain.
00:39:05.199 --> 00:39:08.800
Um without the body, the robots do not do physical tasks.
00:39:08.960 --> 00:39:16.880
And in fact, the body is what the the configuration of the body is what defines the kinds of tasks that robots can do.
00:39:17.119 --> 00:39:22.079
But without the brain, uh the robots cannot really cannot really do anything.
00:39:22.159 --> 00:39:24.000
They they become inert mechanisms.
00:39:24.159 --> 00:39:24.480
Yeah.
00:39:24.639 --> 00:39:29.360
And so the the AI side, the brain side is extremely important.
00:39:29.920 --> 00:39:33.119
And as I think about so I I I cannot separate the two.
00:39:33.280 --> 00:39:33.440
Okay.
00:39:33.599 --> 00:39:37.760
Um as I think about AI, I will observe the following.
00:39:38.000 --> 00:39:48.559
I will observe that right now the solutions that are are giving us so much awe are primarily confined to the digital world, remain confined to the digital world.
00:39:48.800 --> 00:39:50.320
Those solutions are huge.
00:39:50.480 --> 00:39:53.039
They're black boxes that we cannot explain.
00:39:53.760 --> 00:40:01.679
They are not very robust, they hallucinate, uh, they consume huge amounts of power and they have a large carbon footprint.
00:40:01.920 --> 00:40:14.000
So I'm interested in new ideas that that solve some of these challenges, but also new models that uh bring the AI solutions closer to nature.
00:40:14.400 --> 00:40:29.679
Um, the idea that the AI solution can capture continuous process, uh, not necessarily a digital process, can be deployed in the physical world where we do not have server farms to put on every device that's running AI.
00:40:29.760 --> 00:40:34.400
Um, in the physical world, we also cannot tolerate hallucinations and mistakes.
00:40:34.639 --> 00:40:54.000
And so the questions I ask are how can I build better AI with better properties that is more energy efficient, uh, that is explainable, that is not a black box, and that is that can take us along the path of uh assurance and certifiability for AI.
00:40:54.239 --> 00:41:07.920
And uh these properties are important for physical applications, for safety-critical applications, but they're definitely applicable and important for all aspects of AI, for all types of deployments with AI.
00:41:08.320 --> 00:41:08.480
Yeah.
00:41:08.800 --> 00:41:26.000
I mean, you mentioned the natural world, and and that reminded me of, you know, you and I uh we met for a glass of wine last summer, and before that, before we'd met, I was I was reading that the philosopher, the Greek philosopher Democritus, was so wise that he could talk to animals, and it was sort of it was being described in an ironic way.
00:41:26.079 --> 00:41:28.719
I suspect he he couldn't, though apparently he was very wise.
00:41:28.960 --> 00:41:41.440
But then we met for a glass of wine, and I'd asked what you'd been up to, and he said, Well, yesterday, you know, I was talking to whales, because you know, you've developed a research to uh to start to interpolate what like whale sounds actually make.
00:41:41.599 --> 00:41:45.840
I don't know, like I don't know if you could talk a little bit about what we can expect to see there.
00:41:46.960 --> 00:41:53.519
Well, I I must have overspoken because I haven't had the pleasure of a conversation with a whale yet.
00:41:54.000 --> 00:42:02.880
But I have been very interested in their communication uh structures and their social uh structures.
00:42:03.039 --> 00:42:10.960
Um so I'm part of a project that is uh uh looking at recordings of sounds of sperm whales.
00:42:11.280 --> 00:42:15.280
Sperm whales vocalize by clicking, so they just go right.
00:42:15.360 --> 00:42:21.440
So they they cook they have a kind of a discrete uh way of uh exchanging information.
00:42:21.599 --> 00:42:27.039
They have huge brains and live in such a demanding and challenging um world.
00:42:27.280 --> 00:42:32.559
So you just wonder where is that like how how do they navigate?
00:42:32.719 --> 00:42:36.719
Um do they how do they stay connected with each other in society?
00:42:36.880 --> 00:42:47.280
And so in our project, we are uh analyzing the uh these recordings uh using machine learning, and we have discovered a lot of interesting things.
00:42:47.440 --> 00:42:55.119
We have discovered that the sperm whales uh make many more types of sounds that was previously known.
00:42:55.199 --> 00:42:56.960
You can think of this as the phonemes.
00:42:57.760 --> 00:43:15.599
We've also discovered that their vocalization meets the the axioms or the the the rules necessary for uh communication through language, and and there is a a certain pattern for uh for engagement, so the the communications are not random.
00:43:15.760 --> 00:43:22.079
We have no idea what they need, but we see a lot of structure, and it's a very fascinating area to think about.
00:43:22.480 --> 00:43:23.280
Tremendous.
00:43:23.519 --> 00:43:31.599
Well, we've gone from server farms to sperm whales, but before I let you go, I am going to have to put you through the rapid fire question round.
00:43:31.760 --> 00:43:32.880
Are you prepared?
00:43:33.280 --> 00:43:36.880
Well, I'm prepared, but I don't know if it's gonna be so rapid, so let's try.
00:43:37.199 --> 00:43:38.880
Alright, who impresses you?
00:43:39.199 --> 00:43:42.000
Okay, so this cannot be uh a short answer.
00:43:42.239 --> 00:43:59.760
Because I will tell you that I will tell you that I am impressed by people who stand out because they they have a level of excellence uh in their expertise and they use it their talents and convictions to create.
00:44:00.239 --> 00:44:02.400
Meaningful and lasting change.
00:44:02.800 --> 00:44:13.599
And so these
are the people who inspire us to aim higher, to think creatively, to act courageously, and really to contribute positively to society.
00:44:14.159 --> 00:44:26.000
And so to me, this comes down to at advancing the frontiers of knowledge, but also to upholding the truth and to creating knowledge and beauty.
00:44:26.320 --> 00:44:31.039
And I I really see these as very high callings for humans.
00:44:31.440 --> 00:44:36.559
And because you know to push the boundaries is to really illuminate the unknown.
00:44:37.119 --> 00:44:44.559
And to uphold the truth demands courage and integrity and the pursuit of clarity.
00:44:58.079 --> 00:45:02.320
So these are the kinds of people that I'm I'm very uh inspired by.
00:45:02.719 --> 00:45:04.639
So let me give you some examples.
00:45:04.880 --> 00:45:16.559
Steve Jobs, the visionary who brought us the iPhone, when the rest of the world was uh on flip phones, and this has changed how we live, how we work, how we play.
00:45:16.719 --> 00:45:19.039
It has changed the music industry.
00:45:19.199 --> 00:45:22.960
He made made contributions to animation, to personal computers.
00:45:23.119 --> 00:45:25.199
He's done a lot in our field.
00:45:25.920 --> 00:45:33.360
And uh in the same uh vein, Alan Turing, the father of modern-day uh computing and AI.
00:45:33.599 --> 00:45:36.800
Uh but then there are also some interesting historical people.
00:45:36.960 --> 00:45:38.559
So for instance, Copernicus.
00:45:38.800 --> 00:45:42.639
Copernicus altered humanity's understanding of the cosmos.
00:45:42.719 --> 00:45:49.119
And Galileo then brought the telescope and the experimental method uh in how we study the world.
00:45:49.280 --> 00:45:56.800
And Marie Curie discovered that the atom is not the smallest uh unit of matter.
00:45:57.039 --> 00:46:09.199
And I have to I have to refer to uh RBG, uh Luz Beda-Ginsberg, who really uh extended the Equal Protection Clause of the Constitution to women.
00:46:09.440 --> 00:46:11.360
These are really extraordinary people.
00:46:11.760 --> 00:46:18.960
And I I I will be remiss if I did not mention one MIT person, and this would be Claude Shannon.
00:46:19.199 --> 00:46:26.320
Okay, who revolutionized how information is quantified, how it's transmitted, and how it's processed.
00:46:26.639 --> 00:46:30.000
Um so I can carry on because people like this in every field.
00:46:30.400 --> 00:46:34.400
And I'm sure my name is on the tip of your tongue, but uh thank you for not embarrassing me.
00:46:34.639 --> 00:46:42.079
What uh what what piece of uh media, book, podcast, movie, article, essay have you gifted or shared the most?
00:46:42.639 --> 00:46:48.000
Well, I have to say that my two books have been broadly uh distributed recently.
00:46:48.239 --> 00:46:56.320
One book I have given uh regularly uh is uh Rainer Maria Rilke's book, Letters to a Young Poet.
00:46:56.480 --> 00:46:57.280
Do you know this book?
00:46:57.679 --> 00:46:58.239
Not yet.
00:46:58.559 --> 00:46:59.920
Well, I have to gift you one.
00:47:00.000 --> 00:47:10.320
Uh whose letters are very rich with profound insights about life, about scholarship, about creativity, and about the human experience.
00:47:10.639 --> 00:47:11.599
I love that.
00:47:11.920 --> 00:47:21.039
Well, we're about at time, and um, given the completeness of your answers, I don't think we have time for any more questions.
00:47:21.199 --> 00:47:22.880
So you mean I'm too worthy?
00:47:23.039 --> 00:47:23.760
No, not at all.
00:47:23.920 --> 00:47:24.480
Not at all.
00:47:24.639 --> 00:47:27.920
Your words are like pearls falling from heaven, full of wisdom.
00:47:28.159 --> 00:47:29.440
Oh, you embarrassed me.
00:47:29.760 --> 00:47:31.840
So thanks so much for making the time.
00:47:32.000 --> 00:47:35.199
I will see you in glamour Switzerland next week.
00:47:35.360 --> 00:47:37.440
So until then, uh Vita Zen.
00:47:40.559 --> 00:47:43.920
Hello again, friends, and thank you very much for listening till the end.
00:47:44.079 --> 00:47:45.679
One final thing before you go.
00:47:45.920 --> 00:47:56.320
If you enjoyed this episode, do you think you might also enjoy a monthly email from us where we share more stories, news, and perspectives from our early stage fintech ecosystem?
00:47:56.559 --> 00:48:04.159
If you think the answer might be yes, head over to our website, vestigoventures.com, and sign up for our Envisions newsletter.
00:48:04.320 --> 00:48:12.559
In addition to blog posts with our latest thinking and updates from our portfolio companies, we always include a short video interview with incredible people from our network.
00:48:12.719 --> 00:48:15.280
If you do sign up, I hope you enjoy it.