Why Frontier AI Still Sees Like a Toddler, w/ Andrew Dai

0:00An elementary school kid can beat all the Frontier models. If you have an elementary school kid that can beat any model, that seems to indicate we're not quite there yet, at least in the visual world. So let's say there's a table and there's some place settings around it and I ask you how many people is the place settings set up for, right? And then I give you a certain amount of time to look at that image. And this is also at the level of which current Frontier models can operate at. So they use mostly pattern matching, very well known behavior. Same with kids. We talked to an engineering firm who's spending, that team is spending hundreds of hours in the design software.

0:45And this one component is taking them like 200 to 300 hours of mechanical engineering time. It's very odd to think about as a software engineer, because, you know, easy to automate, but these like more traditional things still take a long time.

1:02Corey:Welcome humans to the Neuron AI Explained. I'm Corey Knowles, and I'm joined today by the man, the myth, and the follow-up question, Grant Harvey. How are you today? I do ask a lot of follow-up questions, so it's an accurate description of me. Well, today we're going to talk about how AI can write code, pass exams, summarize the internet, But you show it a real-world image that requires some amount of spatial judgment, maybe physical reasoning or engineering context, and suddenly that magic starts to wobble. Well, Elorian is a new AI research and product lab focused on that problem, visual reasoning.

1:36Corey:And we're joined today by its co-founder and CEO, Andrew Dye. Andrew spent years at Google Brain and DeepMind, including work connected to Gemini and sparse mixture of expert systems. Put simply, he's one of the researchers whose career runs straight through the modern AI stack, from pre-training and instruction following to sparse scaling and Gemini-era multimodal systems. Basically, we're incredibly excited to talk to him. Quick note that today's episode is sponsored by Dell Technologies and NVIDIA. You're going to hear more about them here in just a little bit. Andrew, welcome to the Neuron.

2:07Corey:Glad to have you here. Thanks, Kari. It's an honor to be on here. Awesome. Well, I guess to get started, the thing that when I first went and looked, you guys up that really jumped out at me was your core claim that frontier models still can't reason images much better than a three-year-old. And I was wondering, what do you mean by reason about images? And what is it that models are failing to do today? Yeah, I would say that reasoning is kind of like a complex subject, as you can imagine. I think you see some of these video generation models and they say these models do reasoning and therefore they create great video you can also some people might think of oh it can identify exactly what species this flower is as a kind of reasoning and yeah reasoning is involving all these problems but really what we're talking about is more complex reasoning so one example can be if I ask you a question about a picture.

3:11For example, let's say there's a table and there's some place settings around it and ask you how many people is the place settings set up for? And then I give you a certain amount of time to look at that image. So I would guess that if it's only, say, a table for two or three people, it would take you less than a second to say, well, this is set up for two people or three people or four people. It's a signal that will tell you the answer. And this is also at the level of which current Frontier models can operate at. So they use mostly pattern matching so that you can instantly pattern match and get an accurate number if it's less than five or so.

3:48It's also called, yeah, it's a very well-known behavior. Same with kids. But then if I show you the image and it's laid out for like 20 people, say it's a really long table, like Harry Potter style environment or banquet, then I'm going to guess it's going to take you longer than one second to count those place settings. And it is exactly in these types of scenarios that these models fall flat. So anything we say that takes longer than a second, these models just can't handle.

4:17Corey:What's the difference between, say, like an image description when you give a chatbot, like Chatubichi or Claude, an image, and you ask it to describe it versus actual visual reasoning? Like what is the difference? Yeah, so visual reasoning is kind of more than just saying what is in an image, right? So we can take this to a more complex level. So in the place setting scenario for a banquet, you can say that it's set up for 20 people. Everyone has a fork, a knife, and a plate. But there are lots of scenarios where description just doesn't work. So, for example, say there's a metro map, right? 2D metro map.

4:56Say it's London. London has one of the most complex metro maps in the world. for you to fully describe that metro map and say I'm only allowing you to use language, no diagrams, no visuals whatsoever. I'm expecting it's going to be pretty hard for you to have every detail of that metro map laid out. So there are just things that are simply much easier to describe with visual imagery. The language mazes are another one. or like say you have a tangle of USB cords and you want to figure out which one to unplug to get your power adapter. It's going to be very hard to describe that situation in text.

5:40Corey:I don't even know how to do that without physically pulling them apart. So here's the thing about enterprise AI right now. Every leadership team has the pilot. Every company has the proof of concept, but actually scaling AI from the developer's desktop to the data center all the way out to the edge, that's where things start to break. In fact, 95 % of organizations say they can't even get their data ready for AI workloads. Not the model, the data. That's why we teamed up with Dell AI Factory with NVIDIA to build a resource hub we're actually excited about. It includes articles, videos, case studies that cover the full AI journey from strategy and data foundations to infrastructure decisions to real-world use cases and ROI.

6:22Corey:You'll find stories on what an AI-native factory looks like when physical AI hits the production floor to where hybrid AI workloads should actually run and what sovereign AI means when you're operating under real compliance constraints. And the whole thing is backed by the industry's first end-to-end enterprise AI portfolio. AI-ready workstations, servers, storage, networking, all jointly engineered with NVIDIA. So you can start small on a ProMax workstation, scale out to the data center, and get to production up to 86 % faster than going it alone. So if you're tired of hype and just want to understand what it actually takes to make Gen AI work, head over to the Enterprise Guide to Scalable AI Hub on techrepublic.com or click the link in the description of this video.

7:06Corey:I guess maybe a slightly more technical direction here. Where do current models, specifically let's say multimodal models, where do they break first? Is it around spatial relationships, object permanence, physical causality? Are there different areas like that where it's like they do well at this, but this is just awful? Yeah, spatial relationships they do very well at. One way I like to think about this is the difference between pattern matching, pattern recognition versus more complex reasoning. So a lot of these models, anything that is pattern matching, they can do very well. And that includes identifying what species this flower is or what breed of dog this is.

7:49All you need to do there is pattern matching. But pattern matching, of course, breaks down when you have things that are more combinatoric, when you have many possible combinations of some things, many possible ways things can be arranged. You can describe this as a spatial reasoning problem, or you can just describe it as just a reasoning problem. The spatial combinations, obviously, there's a huge number of spatial combinations as soon as you get more than five objects or so. So that's where it shows up quite obviously, but it also shows up in things that you don't need to strictly describe as spatial, but has some spatial element like counting, following chords, navigating, navigation, these kinds of problems.

8:29Corey:That makes sense. I feel like that's a really, you know, in a practical setting, that's really important for, let's say, like even just a simple use case of, hey, I'm trying to design a user interface. And, you know, can the agent that I'm working with actually look at the user interface and know, you know, what everything is, like keep track of all of it, reason about it? Because anecdotally, I've seen some of the top models lie to my face about an image, especially in visual design work, when you needed to actually see what it's doing. So what's going on there and how can this be improved based on what you're building?

9:03Yeah, design, I think, is one of the most interesting areas for us in design. It is definitely much more about pattern matching, right? To create a great UI, you need to understand space, 2D space, how to arrange things so things are intuitive and clear. And you don't end up with an interface from like the 90s or something like that.

9:27Corey:unless that's your goal some people do like that oh yeah they're creating like a retro game or retro experience yeah yeah so yeah design props up everywhere so we talk currently about like these models replacing white collar work right but white collar work some of it is coding definitely not the majority and then there's some math and then there's a lot of like document analysis summarization writing but the rest of white collar work is all visual like a part of that is design going from web design through to marketing design but also engineering mechanical engineering things like iPhones EV batteries Starship rockets and I'm pretty sure that no one is designing the next iPhone using code designers just they don't code they don't you can't design with code these things like they have a presence in the physical world and they're like every angle every centimeter right for an iphone every millimeter matters and code just isn't precise enough to do that in a nice way for the designer and then there's other things like understanding photographs understanding charts bar charts line graphs say like a stock market charts right these things are all inherently visual and the way a lot of these current models are working with these this that the chain of reasoning or chain of thought is entirely in the tech space, which leads to a lot of these issues that we see with hallucination and other things happening.

10:56Corey:So is what you're building essentially creating some sort of visual chain of thought? Is that part of it? Yeah, exactly. That's a core part of what we're doing. And we believe that the visual chain of thought is a necessary component. So say when you're buying a sofa, you're probably going to visualize how it looks like in a living And when an architect is designing a new floor plan for a new office building, they might need to figure out how a wheelchair can navigate through the space is everything. And I think it's very hard to do that just through text. Wow. Well, you know, we've scaled text models.

11:32Corey:We've scaled multimodal data. We've scaled compute. Why hasn't scaling solved this already? What makes the visual world more stubborn, I guess? Yeah, great question. So all frontier models model language fundamentally different to vision. So language is done with this pre-training next token prediction setup, which is the setup we proposed 11 years ago. We had the first paper on pre-training and fine tuning back in 2015. And that approach has turned out to be very scalable once it's combined with transformers from 2017. And obviously, we've just been able to scale and scale that approach. But vision is very different.

12:11So language naturally lends itself to next token prediction because we communicate in 1D, right? Which is this sequence of words, which is very, very good for communication, but very different to the visual world, right? Visual world is inherently, we're perceiving it through 2D or 3D through our eyes. So squashing the visual world into a 1D sequence to predict the next token in visual space doesn't really make sense. And so I don't think we've really hit upon the right setup there to have vision models that scale. So you do see that scaling is not as effective for vision as it is for language.

12:53But that also creates opportunity because for us, it means we don't need to build trillion parameter models to have a state-of-the-art visual model. That's great. There's a lot of opportunity to find something that does scale.

13:08Corey:Nice. Does this look something like, you know, instead of predicting the next token, you're predicting the next image, predicting the next frame? How do you actually do the visual reasoning part of it? Right. So the visual reasoning part of it, we're taking a more, like I would say, guided approach. so for kids say kids that are like preschool age if you ask them how many how many balls are in a box if there's three or two say there's three two one or four four balls usually they can get the answer right immediately instantly without any kind of reasoning or anything and they have a pretty decent accuracy and that's because that's from pattern matching because with an arrangement of three balls is very you just intuit what the number is yeah you don't even think about it it just exactly exactly but if you have more than five balls if you start having 10 or 20 what you see these kids doing is they they start pointing to the balls to track what they've counted already and what they have left to count for adults obviously we sometimes do this in our mind but for kids you can still see this happening and you see this happening for various other tasks like following a line or following a maze, they will use their finger to point or they will use a pen to draw.

14:30But if you think about it, this data isn't really available on the web. No one has like counting data on the web because it's just so obvious to adults how to do it. And also it's more work to edit an image, right? And so this kind of data isn't available. So what we're doing is we're both constructing the data through synthetic data And we are also working with vendors to collect this data where people manually label exactly how they are solving that problem.

15:01Corey:Wow. That just really made me think of, like, you used to see these contests in a store where there would be, like, a great big jar full of marbles. Yes. And you have to, you know, it'd be like, whoever gets the closest gets the jar or gets, you know,$50 or something. And this makes me think about the difference in counting what's laying flat in a box versus needing to be able to intuit the volume possible inside this jar, for example, and be able to intuit how many M &Ms or marbles or whatever are likely in the jar. Is that kind of the sort of difference we're talking about? Yeah, that's exactly right.

15:38That's a great example because I've also seen problems where there are varying sizes of marbles, balls in a jar. Some of them are like jawbreakers, you know, size of a small fist and some are smaller. And then you get a peek at, you know, how many there are, but then you have to do a little bit of basic statistics and volumes and everything. But yeah, this is exactly the kind of problem that needs this combination of visual reasoning and linguistic reasoning.

16:09Corey:So that's awesome. That's really what I was wondering. I remember those from like movies and seeing them as a kid back in the 80s and 90s in places. It was a thing. Are you hitting the limits of siloed AI? Just as humans once transformed society by sharing intent, knowledge and innovation, AI faces a similar inflection point. To achieve distributed superintelligence, we must move beyond simply scaling up. We need to scale out, too. OutShift by Cisco is building the Internet of Cognition, an open infrastructure enabling agents and humans to collaborate in real time. Visit outshift.com to learn more about the Internet of Cognition.

16:50Corey:So you worked on Gemini, which was designed as a multimodal model pretty much from the beginning, if I recall correctly. And what did that experience teach you about the limits of today's multimodal training recipes? Yes, I was a data area lead for Gemini. I've been leading Gemini since day one. And what we've seen is that language and coding capabilities have really advanced rapidly, both in Gemini and outside. But we do see that visual capabilities, multimodal capabilities on the reasoning side have been slower to advance. on the generation side obviously as being in leaps and bounds there's a lot of great generation models like VO and Genie and others that can generate very high quality video but this isn't necessarily reflected in the reasoning part and there was a benchmark that came out a few months ago that exemplified that that benchmark showed that these models say the capabilities are still at a level of a preschooler yeah that was a visual reasoning benchmark right from yeah Yeah, we have a question about that, actually.

17:59Corey:Yeah. Yeah, and so, yeah, even an elementary school kid can beat all the frontier models. So to me, when we're talking about AGI and already being an AGI in the visual world, if you have an elementary school kid that can beat any model, that seems to indicate we're not quite there yet, at least in the visual world. Yeah. Yeah, yeah. Yeah, the paper specifically was like, you know, there's a spatial ceiling for folding, reflection, rotation, and spatial matching. Why are those problems so difficult or why are they AI's weakness? Yeah, so I think it comes from a combination of things. One is, like I said before, about these models only having the text chain of thought and no visual chain of thought.

18:46But another reason is that if you look at the development of computer vision, it's been very focused on just a few tasks. So these are object recognition. That's why Google Lens is so great. And we can identify any plant or flower. And another one is like localization, seeing where things are in image, segmentation, detection. These are things that are very useful for specific use cases, like surveillance, monitoring, like when your doorbell goes off and when your doorbell app goes off and says, oh, there's a mouse running across your yard or a scroll or something. That's all the traditional setups and objectives.

19:33But these like more complex visual reasoning tasks, like spatial reasoning tasks, has, I'd say there's been less attention paid to those. I think partly because they are so hard And another part I'd say is probably they do require some degree of like text reasoning. And that text reasoning just didn't exist. One example is one and a half years ago, all one didn't exist. We didn't even have text reasoning. Sounds like a lifetime ago, but it was only 18 or 20 months ago. That's crazy. Yeah. And then we had that and now and then we have great math and coding models. and yeah that line of progression it just was clear to me while I was while I was in Gemini that the next step would be incorporating visual reasoning now that we have a very good foundation

20:23Corey:that makes a lot of sense when I think about how I think of you know if I take a piece of paper and I start to fold it there's some part of it that I intuit there's some part of it that I I don't know if I say in my head if I fold it this way it's gonna look like this but or if it's just memory like there are a lot of parts of your brain that you're putting to work there when you think about how something will fold or or rotate it makes me wonder if there needs to be like a 3d element of this or like you have to train models in 3d for them to understand the space i don't know the technical behind that but curious what your thoughts are there yeah that's uh great yeah that's very perceptive i'd say yeah i i agree that there is an element of like 3d modeling in our brains.

21:08There's a 2D for a bunch of problems like recognizing characters and a bunch of things. But yeah, I agree there's some kind of 3D modeling going on. So that needs to be represented in the chain of thought as well.

21:21Corey:Yeah, because like I see that light just above your head, that strip light. And I don't think that's a light that's crooked on a ceiling. I know that that light is at that angle because it's going away, that it's getting farther away from you. Wow. really interesting so it sounds like high thoughts but you have to think of it that way like you have to think about all the things you take for granted to try and think how how they think i always think could you could you watch a video while you folded that paper i think might be a good test like could you truly do something else while that's going on or or is it a thing that you are actually consciously considering as you go i would be curious to know and honestly i don't know it's I'm just spitballing here ideas.

22:04Corey:Speaking of, something I'm curious about is like, what's the goal for a Lorient? Are you seeing this as like, we're going to build a foundation model? Is this going to be a reasoning layer that would exist within other models? Or maybe some type of something I'm not even considering that maybe doesn't fit neatly into the buckets we know already? Yeah, so we are building a visual reasoning frontier model. And we believe that it's very important to integrate a language and visual reasoning into the same model. So our model can be called independently and it will do all these kinds of reasoning by itself.

22:40And yeah, as part of that, like I gave with the example with the floor plan or the sofa, it will need to generate images. It will need to be able to edit images. Eventually, it might need to like rotate things in 3D space, but that'll probably be later. very very but yeah these are all things that we believe needs to be done at the same time as a text-based reasoning and the text-based reasoning is important to connect with humans right because that's how we communicate right that's how you tell a model to do something or that's how you you get the model output if it's a question answering setup so that's why we're not just doing a visual train of thought that's why it needs to be quite integrated into the text part as well.

23:25Corey:I suppose it makes sense then why the ability to communicate with it via text is the place to start. Like as in terms of AI development as a whole, like to do the other things, you will need to be able to communicate with it. So you, I guess, start there. Makes sense. Yeah, exactly. Yeah. One saying, wow, was it Gemini that I thought nailed this quite well is that Gemini is spread into several areas as a data area there's a eval area and really data and eval is about communicating with the model right it's the only way we have to communicate with the model in a sense totally yeah yeah well it's kind of interesting that we're starting with language it makes sense from for where we're at now as a society but it seems almost like every other you know intelligent being animal human what have you started with maybe perhaps a visual reasoning first so it's kind of funny that we're doing it in reverse if you think of it that way yeah because they don't all have language you're right yeah like well i was just thinking about flies you know in my in my house today it's like well the fly you know it can't communicate with me but it can see me it can navigate this 3d space and and you know its brain is i don't know the order of magnitude but much much smaller than than ours and it still has that kind of encoded in it so it is interesting to build this like to build the visual reasoning second you would think maybe you know everything else built it first yeah yeah that's a great observation but yeah ultimately i think the reason this happened is as the compute costs get bigger there's definitely a need for this to be economically useful and it's going to be very hard to interact with a fly brain even if we could well related to this so some of your stated use cases include engineering design satellite analysis and robotic navigation.

25:14Corey:So could these models be small enough to operate on device, like on a satellite or on a robot? Or will it be, you know, still like cloud for the frontier and, you know, you stream it to the device? How are you envisioning that? I think initially we are building something for the cloud, for streaming. But yeah, I think it's definitely like edge devices, like there's a huge range of applications there. So we will be looking to build smaller distilled models as well that are compact. That seems to be the path for AI is that, you know, you're always pushing the frontier, which means it's going to be large.

25:52Corey:But then like as it gets large, you start working to make it smaller and more condensed. But you keep pushing bigger, too, at the same time to further the advancement itself. So I think what you're working on is really, really unique. And it's not a conversation we've had before. So we're both really excited to talk about this. We're starting basic and working our way up. Exactly. Related to this, so you've also worked around sparse mixture of experts or MOE style scaling. Does visual reasoning require a different kind of specialization than language reasoning? Almost like different experts for the different elements like geometry, physics?

26:29Yeah. Yeah, the mixture of experts architecture that's currently popular around all the labs is quite language-tuned and language-specific. So as soon as you have multimodal inputs, there are some odd things that happen to the experts. So the allocation becomes uneven. And this is because fundamentally the underlying representations, call them activations or embeddings or whatever, But fundamentally, they are different between visual inputs versus linguistic inputs. Because obviously for language inputs, they fit very well into this neatly organized embedding space. But for multimodal inputs, it's not the case.

27:12These days, people don't use these kind of discrete representations. They use continuous representations. So fundamentally, the behavior is quite different, especially with MOE setup. So there are more things to think about in terms of optimization, balancing the experts, all those kind of things. So we're doing some adjustments to the architecture there. And we can do it because we are specializing our model to these multimodal use cases. Whereas if you needed to trade off and balance multimodal with coding, with math, with language, then you can't make the kind of changes that we want to make.

27:47Corey:That makes sense. Yeah, because it's almost like a jack of all trades, master of none. kind of situation if you're trying to do general purpose it's like maybe you can get away with a lot more when you're changing it yeah in a niche setting yeah exactly something i've wondered is how do you evaluate progress with what you're working on like benchmarks are tricky they can be gamed models learn them a little bit over time and real world imagery is messy it could be anything so like how do you know when your model is improving yeah yeah that's right evaluation is a hot topic and also very hard with these models.