Matt Turk: We have maxed out the features. We have maxed out the hardware. That's what we get.

Matt Turk: By almost any definition anyone could have written down, let's say 5 years ago or 10 years ago, we basically have the vision of AGI that that we had back then.

Matt Turk: Everything that grows exponential will level off because if you need resources, the resources will be exhausted.

Matt Turk: You can see up to two orders of magnitude more compute available, 100x more compute.

Matt Turk: If you don't know how to use agents well, you will be left behind.

Matt Turk: How far can you push it? I think it's never been a more exciting time to work in AI. Hi, I'm Matt Turk. Welcome to the Matt podcast. Today we have a special reality check on AGI with two guests who are very close to the computational reality of AI. Tim Dettmers of AI2 and Dan Fu of Together AI. In this episode, the team argues that we are hitting diminishing returns and running into hard physical constraints while Dan argues that we are still leaving huge performance on the table and that today's models are lagging indicators of hardware progress. Then we shift into a fun practical discussion on how to use agents and what to expect from AI in 2026. Please enjoy this fun conversation with Tim and Dan.

Matt Turk: Tim and Dan, welcome.

Tim Dettmers: Thanks for having us.

Matt Turk: So Tim, a few weeks ago you wrote a great provocative blog post entitled Why AGI will not happen. And then Dan, a few days later you replied with your own blog post equally fascinating entitled Yes, AGI will happen. I'd love to go into your backgrounds. You both have the very interesting characteristic of being having a foot in industry and a foot in academia. So Timmy if you want to start with yours.

Tim Dettmers: I'm assistant professor at Carnegie Mellon University machine learning and computer science department and also research scientists at the Allen Institute for AI. My past research has been mostly on efficient deep learning quantization. That means model compression. Take large models, compress them down from like 16 bit to something like 4bit. Key research has been there. For example, it's a very efficient fine tuning. We compress to 4bit, use adapters on the model and then use up to 16 times less memory than if you have dense both handling and now I'm working on coding agents that we have a very exciting release in about two weeks. State-of-the-art agents. You can quickly specialize to private data, get strong performance on any code base that you like. Um and yeah, that's very exciting.

Matt Turk: Dan,

Dan Fu: Hey. So I'm an assistant professor at UC San Diego. And also my total VP of kernels at Together AI. So in industry, I focus a lot on basically making models go fast. So GPU kernels are the things that actually translate the models to how they run on on the GPU. You can think of them as as basically specialized GPU programs. A lot of my research in my PhD in my lab focused on that. So I I developed things like Flash Attention which was a efficient kernel for the one of the core operations of of a lot of the language models that that we use today. I also did research on sort of alternative architectures to transformers, things like state space models and and things like that. And Together I'm really focused on how do you make the best language models that we have today. How do we make them go faster? I think as of this morning's recording, we actually just released a blog post with Cursor about how we accelerated a bunch of their models and and helped them launch Composer 2.0 on Nvidia's Blackwell GPU. So that that's a a bit of a flavor of what I do. So let's get into this AGI discussion and then in the second part of this conversation we'll talk about agents and coding agents and your thoughts there because I want to make sure we cover that AGI obviously it's a term that everybody uses and I think we can all agree that nobody really knows what that means but for for purposes of this discussion what is a useful definition of AGI from your perspective?

Dan Fu: Sure. Yeah. I think so. One one of the things that that we kind of discussed back and forth in in this set of blog posts is sort of what AGI means. For me, I think one of the things that I've been thinking about recently is that if you took where we are today with the models that we have today with the language models and I think you know we'll probably talk about this a bit more with the later with the agents by almost any definition anyone could have written down let's say 5 years ago or 10 years ago. Certainly when you know Tim you and I started our PhDs we basically have the vision of AGI that that we had back then. We have things that can write code. They can write you know human text even though you know maybe the they they use too many m dashes or or something like that but they can uh do do these really really amazing things. I think one of the things that I think about is at what level does this kind of become a new industrial revolution where you can where this technology is really going to change a lot of what the way that we do things today the the and have a huge you know really really great economic impact in terms of software engineering I feel like we're already there or almost there like uh there are there are things that may be super specialized like I don't know if they're going to be able to write like the best forran and cobalt code in the world um but for web development a lot of even a lot of low-level systems engineering, they're they're already really great. One of the reasons that that that I wrote my blog post was um if you kind of think about where we are today, you know, we maybe already have AGI or like some some form of AGI. And if not, then certainly the next generation of models, the models that today are training already. Um if they're at all better than what we have today, then we're we're we we've already hit something that's that's really amazing and and and and pretty wild.

Tim Dettmers: When I wrote my blog post actually I um I noticed like oh I I forgot to put the definition of AGI my blog post even though my blog post is very much about AGI and and I think that sometimes sort of reflects how we think about AGI we don't think carefully about the definition I mean there are sort of um and I I thought about it before and I think there sort of uh different kind of definitions that have sort of advantages disadvantages I wouldn't say I mean as you said before there's not one definition that people agree on just to mention sort of a couple um and I think one that's sort of quite widespread is to see AGI as sort of cognitive abilities, cognitive tasks. What can you do cognitive? I mean and software engineering very cognitive writing very cognitive moving a robot in in space um that's more genetic um sort of you can you could also say like hey you also need to think about how you move that's also part of cognition but I think most people would separate that and um say everything digital is kind of cognitive and if you have physical that goes beyond that what I think makes sense is sort of this economic angle can we at another ind industrial revolution. What it means is is AI useful and it's so broad and useful that you want to use it everywhere kind of um it accelerates all kinds of things um similar to where computers were introduced productivity increase not initially a productivity actually went down and you need your diffusion in the economy to pick it up again. we might see something like that with AGI more broadly in task for like software engineering scale it's going up uh pretty significantly but yeah I think that is useful.

Matt Turk: Let's jump into the the heart of the argument Tim I was amused by what you said about um where where all those ideas of a super intelligence uh come from if you want to uh talk to that.

Tim Dettmers: Yeah and to sort of lay out sort of the entire narrative there are certain thoughts about AGI that is sort of very much rooted in a certain kind of thinking. Uh it comes from effective altruism communities of the rationality communities. I was part of this communities like long time ago that is now 15 years ago. If you look on Twitter there's always like oh we get AGI in 2 years and then one year later oh we get AGI in two years and one year later we get AGI in two years. I feel like it's a little bit of lazy thinking, a little bit of being in a bubble and not being exposed to different ideas. And that was one of the main motivations for me writing the blog post because I feel like there are some things some ideas that if you think about them um they might provide a counterpoint to a lot of the thinking that's out there.

Matt Turk: Yeah. And you and your c your core thinking is that um there is a tension between those ideas and the computational reality. Is that is that a fair way to put it?

Tim Dettmers: Yeah, there's some there's a physical component and then there's an idea component, but there similar structure and this structure is basically diminishing returns. Everything that grows exponential will level off because if you need resources, the resources will be exhausted. Resources can mean different things and if you look at the physical aspect um it gets just more and more difficult to advance technology. That is the case almost with any field of uh research or development. Um things get more more easily. You need more resources to make further progress and the pro progress sort of and goes uh lower and lower and so if you look at the physical reality of computational devices and then also computation itself has particular structure and so basically computation useful computation is two things. The first is you need to gather data from one location and aggregate it in a certain location where you then put this new information together to compute a transformation of that information. You basically want to combine um sort of known things and compute some new sort of um things that you didn't know before. Useful information. useful information can and needs to be transformed from information that you already know. If you move a lot of information around but you don't transform it, you can't make new information. If you do a lot of computation on the information that you already have, you miss out on the long distance insight, the indirect insights. I think a lot of this actually maps to the neural network architectures that we have. Um the beginnings we had convolutional networks they are very effective and what they doing they don't move much memory they do a lot of computation and that means your device needs a lot of flops and memory bandwidth is not that important once you go to very dense computation very large matrices then it goes in the direction um of the current neural networks but there you have still sort of this component of your recurrent um basically paying attention to previous states. And but because it's recurrence, the memory reuse of that computation is minimal. And with um transformers, you basically then had these large matrices that compute basically that transform the incoming information from the previous layer. And then you had a tension that now computes the information across time or space. And um I would argue these are the most two fundamental ways of computing information. You want relate information to itself or transformation of that uh information. But then you also want to basically relate information to um distantly related of information. So you want long-term relationships and you want a transformation based on what you already know.

Matt Turk: And you say this is slowing down, right? like in your um blog post you have a pretty striking sentence where you say uh GPUs will no longer improve meaning fully. We have essentially seen the last generation of significant GPU improvements.

Tim Dettmers: Yeah. So this has two components and so one is also sort of a very fundamental thing and it's physical in the sense that um I mentioned these two um components uh memory movement and computation and so computation can only be useful if you move memory to this sort of local neighborhood where you do this computation. Now this is a geometric problem. You need to have a large store of information and then use this large store to move information closer to where you do want to do the computation. And we have figured out how to physically do is optimal. We have like a large slow memory that's DRAM. Then we move it to a cache. If you look at the geometry, that's how you do it fast. If you have a certain size of computation, this is optimal. If you have a different size of computation matrix multiplication, then you want to use not a CPU but more like a GPU which has higher latency but more throughput. You can move more data but more slowly. And um yeah, if you look at all of that, you can push around a little bit how you structure everything like the caches and uh how large they are and how much cores are they shared. But in the end uh the fundamental problem uh remains the same. You have a geometric problem. You can only fill the space in a certain way and that means you always have certain access pattern with certain um latencies and the biggest latency is a big block of DRAM that is the major bottleneck. This is also called the uh Von Neumann bottleneck bottleneck based on computers almost all computers that we have and this is a bottleneck of um moving a program to where you execute the program and for neuronet networks that's basically move the weights in the inputs to uh the execution where you execute the program that would be the tensor cores. there are not many ways how you can go around this bottleneck. Um the only way is to store the memory locally and do the local computation there. And there are some processes that do something like that. For example, this reburse processor. So they don't have this Von Neumann bottleneck in a major way on the ship, but then they need to also pipe data into that ship. And so the Von Neumann bottleneck moves basically away from the ship to your storage or 2J network. And so you just move it away but still the same bottleneck. You need to load the program which might side on the disk or memory through the network to the ship and same physical problem. You just move a couple of variables around. That is sort of one part of the problem. We don't have architectures that can solve this problem. That's sort of the second part where um my argument kicks in and that is you need new technology to overcome bottlenecks. But once you have leveraged that technology, you need new technology to get over that. If you look at what we can do is we move from DAM to uh HPM. So that's DM that's stacked. That's much faster, but you can only stack it that high because it's very difficult to manufacture and test in correctness. um the yields is very low and we run actually in 2026 there's not enough HDM you can't build the nice processes and more because you run out and it's just too difficult to manufacture them and with that um we have all these innovations one of those tensor core big step up then have 8 bit precision another step up then we have four bit precision uh with particular blockwide quantization particular data types from my research another research we know that's those two information theoretically opt able in a practical sense. If you train on enough data, four bit precision is not up enough. You need actually 8 bit precision. So you can't go further. The hardware is maxed out. We have no new technology. We can make it easier to manufacture and a little bit cheaper but not faster. And we have maxed out on the additional features. Sparity could be something people tried it for 50 years. I tried it present well. And so that might be the last thing but 4bit precision is the end of quantization. So, um, that's the end of it. Like, we have maxed out the features, we have maxed out the hardware, that's what we get.

Matt Turk: Okay. Fascinating. All right, Dan, what is your perspective on all of this?

Dan Fu: I really appreciated Tim's post because I think one one thing that that I really appreciated is that there's some some AJI talk that if you just kind of like trace the exponential, at some point you get, you know, the the thing that will eat up the the universe or or whatever. Um, which I I always found it a little bit odd to to think that way. I I appreciate the the thing in terms of the actual physical constraints because uh you know like Tim said these are physical systems with physical inputs um and uh and actually doing physical computation. I think my perspective was that if you look at where the systems are today and you look at the the models that we've trained uh we are just so far from being from even using the last generation of hardware as as efficiently as as possible. So not to mention all the new hardware that that's being built out. So I think on the technical side I'd say there there are two major points I I wanted to make in my post which was one if you look at the models that are kind of the the really great ones um that uh that that we know today and I in my blog post I mostly talked about open source models because they talk a little bit more about how they train the resources behind it. um we don't have public figures behind you know how much OpenAI and Anthropic are using um but if you look at the DeepSeek model for instance this is one of the best open source models we have out there today it was trained at the end of 2024 on last generation kind of nerfed GPUs H800's instead of H100's the 800 is nerfed I by all sorts of ways from from Nvidia um to to get around the the export restrictions at the time um and they were trained with let's call it like about 2,000 H800s uh according to the report for I I think about a month and when you compute how much how long that took when you see how much compute was actually available on the chip you get something like a 20% um uh effective chip utilization or something like that the the number that the the term of art is called MFU model flop utilization um but basically that's a a 20% um utilization number um meanwhile you know in I think in the early earlier in the 2020s we were seeing lots of training runs on on older hardware different model architectures that were easily achieving 50 60% MFU. So if you just take that number and then say hey maybe there's a way to to to to get it out there since then um you know my good friend Tri's released a whole new set of kernels on how to train the these models better and you say okay there's a 3x there just just from from from that one piece. Um then the other thing to realize is that so that is a model that is being used today in early 2026 as the base for some of the best um open- source or or open source adjacent models out there. It would have started training the base model at least a year and a half ago. So let's call it mid 2024. Since then we started building out completely new clusters with the current generation of hardware. So in in on Nvidia these are the Blackwells. Um, we've started, you know, there there companies like Poolside that are building out tens of thousands of B200, GB200 chips. You know, there's there's uh other folks like Reflection who are b who are building out tens of thousands of B200 chips. So this is comparing we have a new set a new generation of hardware where even if you take the exact same precision as you had before exact same everything uh 2 to 3x faster compute 10x larger clusters um plus you know maybe 3x lurking in terms of just pure optimization that's 3 * 3 times maybe another 10 that's like what another 90x of compute available um and that's not even looking at future buildouts that is literally clusters that you can point to today um that that people have started training on that that you might hope that at the end of that you'll you'll get much better models. The point I really wanted to make was if you just look at it from those basic inputs, you can look around, you can squint a little bit, you can see up to two orders of magnitude more compute available um compared to the models that we are indexing on today. Now we can argue about you know is there are there going to be diminishing diminishing returns in ter in terms of scaling up um are there going to be do we expect the scaling curves to hold and and all that um but you you can just look around and see it and that's that's you know 100x more compute. So I think from the physical just a pure compute perspective there's a lot more available a lot more um that that we're not doing. Uh this is not even to mention a bunch of the the great points that you mentioned Tim. So, these are all 8bit training runs. We've just started writing the papers about how to do a 4-bit training run properly. Um, there's new things like the uh on the GB200, you have 72 of these really connected really really quickly. I don't think we've even seen the first pre-trained model come out of that yet. GPT 5.2.2 I think was the first time that that you saw in one of OpenAI's reports, hey, this was trained on H100, H200, and GP200, which to me suggests that was actually pre-trained on one of the really old clusters, maybe some fine-tuning was done um on on the new GP200s.

Matt Turk: You make the point that not not only is the hardware underutilized, but you also say that uh models themselves are a lagging indicator.

Dan Fu: Yeah. So the models that we see today that we can play with today have been pre-trained on clusters that were built out uh a year or two ago. Um because you know you you need enough time to get the cluster running. You need enough time to do the large pre-training run and then you need enough time to really post-train it, fine-tune it, do all the LHF and and all that stuff. Um so the the models that we have a snapshot today that at the beginning of a conversation it's like h maybe it is AGI maybe it isn't are already trained on uh clusters that are a year and a half old we've built out much larger clusters since then we can expect you know you you can expect that they're going to use them for for pre-training the models that we see today that we index on quality today are actually trained on pretty old hardware um and we've got new generations of hardware more software uh choices we can make not to mention architectural choices like so so Tim you were mentioning that this thing about you need to move data and then compute on data um we've actually seen the transformer change in architecture um for over the last a little bit slowly for like you know for for researcher a little bit slowly for my taste but you've seen the fundamental way we do the computation change even if you find another 1.5x or 2x there now you're talking you know 100 150x more compute like um so there there there's a lot more compute out there um to to train better higher quality models.

Matt Turk: If I understand this whole discussion correctly, all of this is about pre-training, right? And whether we can train bigger model with uh you know more data and more compute. But in conversations on this pod, a lot of the conversation have been about the importance of post-raining and the you know building AI systems with pre-training plus RL. Where does that fit?

Dan Fu: That's a great question and and I think another piece that that we we didn't you know I don't think either of our blogs particularly hit on. One way I like to think about it is that pre-training is like the general strength training that you do in the gym. You go lift heavy weights, you improve your strength, improve your your your general ability and then post-training is like the specific drills that you run to um to to get a good at at a particular task. So historically the vast amount of compute has gone to pre-training. Suggest building models that are more generally capable of doing many things have a lot of knowledge get to a point where maybe they they have more knowledge than than your average person you know I I certainly don't know as much as as ChatGBT for for instance and then the post training is both how do you make it helpful so uh you know ChatGBT you ask it to do something and then it actually listens to you um and and tries to tries its best to do it but I think the other thing that that we've started to see increasingly in post training is that you can start to postrain specific speific skills. So, uh the model that's really good at helping you code uses a lot of the knowledge that you got from pre-training but is actually adapted to be particularly good for coding or um or the or the the model that's really good for legal work, for instance, has a lot of the the pre-training backbone, but then the post- training is really what gets it to that to that place where it's really useful. From a pure computational perspective, pre-training is usually much more compute expensive than post-training. um post-training the work that you have to do I think um I I'm not a post- trainining expert but the work ends up looking a lot more like how do you build a useful product how do you get user feedback how do you do things like that even then there's there there's a world where uh you know maybe the the next generation of models is the the next generation of pre-trained models is is a strong enough base that if you go tackle each vertical of the economy that you care about you could actually post- train it to um to something quite useful Um, so I think that that that's a whole other computational aspect of it. So maybe we don't even need that 100x more compute that that may that that may be out there. Maybe it's more a a more traditional work of you know let's understand this problem deeply to understand how to train in almost the human sense like how how would you take an intern and and train this intern to to do this specific task? How do you get this very powerful pre-trained model to do something really useful? Um in in this post- training sense.

Matt Turk: Is this concept of usefulness that you both mention where both of your point of views maybe converge some ways AGI is something but what ultimately matters is is where you land in terms like usefulness in the industry and therefore even though you one may not be able to reach that kind of ethereal definition of AGI that nobody really understands uh due to diminishing returns in some ways doesn't matter because we still have like so much juice to squeeze that uh we have enough to go until we get to a place where this is truly useful not just for coding but for the rest of the economy.

Tim Dettmers: Yeah. The main conclusion of sort of my blog post was exactly that like we shouldn't pay too much attention to AGI but more about thinking about how can we make it most useful that might go beyond of how useful is a model. I mean Dan mention mentioned like post training as a product. An important part we saw with computers is diffusion in the economy that requires a very different mindset. US mindset is build the best model and then everybody will use it but it can help to really figure out how can you benefit the most people in the most pragmatic way and I think that's sort of more the Chinese mindset and so that that sort of mindset so if I think of usefulness one is model the other is sort of the mindset but I would agree I think I think that both Dan and I I think most people would agree that if you have an AI that does very impressive things like math olympiate things and that sort of thing but it can't do anything useful is at AGI um mean um and so models are already useful so that that scenario will not happen but I think what we really want is very very useful models and I think we have that and I think we can prove that but um I don't think we get to AGI by SER definitions but we will see significant impact.

Dan Fu: Yeah I think I I would just add to that that um Tim you had this point about you know how much of the economy is is physical how much of it is um um knowledge work and I think that the US China you know contrast is is really interesting there and you know there's been these analyses this book by Dan Wong going around about um the manufacturing economy the engineering economy versus the more lawyerly economy um and I think there's certainly a lot of great knowledge work to to be done in the US um I think there there's also if you look at what the actual sectors of the economy are. So a large portion of it is healthcare, a large portion of it is education. Um tech is certainly a also a large portion that's kind of leading the stock market and and driving the stock market. We have a there there's a lot of great people who are trying to use the new models to try to do things like develop new drugs or or understand how to how to make a real impact in in healthcare or if we can get robotics off the ground and do things like start taking um helping with some of the the the you know the the physical labor um maybe not necessarily building houses but the the day-to-day household labor um that that could be large untapped portions of the economy. Those pieces are are are really great. can you can almost start to see the first pieces towards it. Uh the self-driving analogy is really interesting to me because early on I'll say early on in my PhD I was quite skeptical about uh self-driving. So let's call this 2018 2019 it felt like self-driving was always a year away or two years away or if you ask the the experts they say oh five years away. Um and then last year I rode in a Whimo and today I can I can I I just actually got got access to Whimo on the highway. So now conceivably I could, you know, potentially sell my car and I live in in the Bay Area in California. Um I won't because I personally like driving. Um but there's there's a lot of the progress is funny in this way where it's kind of it's not there. It's not there. It's not there. And then one day something a switch flips and then you're suddenly like, oh, not only is this thing pretty good, it's actually a lot better than the service that I'd get in an Uber or a taxi or something like that. That's a really exciting thing if we see that happen. if that we see that flips that switch flip for like you know household cleaning or or putting away the dishes or things like that I think it would it would it would really be really exciting it would change a lot of folks perspectives I'm not a roboticist myself but I'm really watching that space with a lot of excitement.

Matt Turk: Dan as a quick tangent do you think we evolving towards a multi-h hardware multi-chip kind of world based on what you see you know obviously there's been Groq and Nvidia there's Cerebras there's like a bunch of sort of special AS6 companies coming up from from your kind of like lowlevel in the stack vantage point. What do you see?

Dan Fu: Yeah, that that's a great question. So, it's something that I spend quite a bit of time thinking about um more so I'd say on the lab side than than necessarily um in on on the industry side. Um although of course we're we're paying close attention kind of on both sides. I think it's it's at a really exciting time where um the Nvidia chips are really strong, really reliable. Uh there's a lot of software support around them that that that has built around we're starting to see the same things happen for example on AMD chips um with with some some of the research there. So on the lab side, we put out a recently a library called Hipkittens led by my great friend Simron Aurora. Um, and she was really looking at, okay, how can we take what are the right software abstractions to to program on these AMD GPUs? And turns out they're not exactly the same as the Nvidia GPUs. So even even two GPUs that have relatively similar specs certainly compared to Grock or Cerebras or Sabanova or one of these other chips um even though they're relatively similar they they actually have pretty different software abstractions you need to use. Um and I think more people are getting excited by that and and investing time and and energy into that. Um we saw the Grock acquisition from Nvidia. A lot of people are excited about TPUs today. I think uh Cerebras and OpenAI just just announced um their partnership. Um so I think certainly it's going to be a wave of things coming forward that that you're going to see um a lot more. I'm sure Nvidia will still do great and still grow beyond their $5 trillion company or whatever it is at the time of recording. But I think you're you're going to see a lot more diversity um especially around I think inference of the model. So training and inference are actually quite different computations and as a result you might actually want quite different chips to do it. Um on the inference side you might want for example your models to live locally on your phone on your laptop. You know my phone my my iPhone which is a few years old at this point is already more powerful than some of the GPUs that that I had when when I was starting my PhD. That growth of of that hardware power is is really exciting to see.

Matt Turk: And Dan, you mentioned a second ago in reference to uh self-driving cars, that moment where where things flipped, switches is turned on. Has that happened with agents already? You talked about software singularity. Are we are we at that moment for agents?

Dan Fu: Yeah, I think that so personally in in my life, I'd say that moment was last Juneish. So June 2025 was the moment that that it really flipped for me. to give some context here. So what I do in my day job at Together AI is we write a lot of these GPU kernels. I I don't know how popular, but the in the in the general ML zeitgeist GPU kernels are thought of as kind of like the final boss of the thing that you learn how to program. They're very hard. They're very highly parallel. Uh you don't write them like you have to write in C++ which is this old language that the old systems people used, you know, decades ago or whatever. They're not in Python, etc. When you're trying to hire for people who can write kernels, it's very hard. It's a very challenging skill set. It's certainly, you know, the the the the the the tip of the spear in terms of um the their programming strength. And last June, we had this really interesting realization where we realized that cloud cloud code, Cursor agent, these agentic coding assistants were actually very good at writing these kernels. So, um, there was one week where I think I wrote like three or four different features that usually would have taken me a week each and I wrote all of those in in a single day. Um, and I was like, "Oh my god, this thing is making me five times more productive as a kernel expert. Um, I got my team on it. Now my team has all these really complex systems that they've built where they can write a whole feature that I think would have taken um months of a of a whole team's time before. Um, and this is kind of the the that that final boss of of programming challenge that that was really that was really challenging. So from from our perspective for coding for this really technically challenging GPU um kernel programming, it kind of crossed that that kind of crossed the Rubicon for us already. And I think you know uh I gave this talk a few months ago um at Slush about what we're calling the software singularity where we realized hey in terms of software engineering even for these really niche skills um it's it's certainly better than the average programmer. It's at a place where it can accelerate the really expert programmer. So um it's right now as of you know today's recording is at a place where if I just let it on its own it might not generate the right thing for you. But if you give an expert programmer this set of tools, they can go 10 10 times faster than than they were able to go before. Um, and I think that that that's a really exciting uh place to be.

Matt Turk: And on that topic of agents, Tim, you just wrote another great blog post called uh use agents or be left behind. And part of what you talk about is coding agents versus agents for the rest of tasks. Where are we in that arc where um agents are are transitioning from being excellent at code to useful for the rest of our lives?

Tim Dettmers: That blog post was also as as a reaction to what I see is there are a lot of productivity gains if you use coding agents for all kinds of tasks and as a professor you don't code that much. You can actually code more easily which probably other professors previously would not do. so easy now. But yeah, also for non-coding tasks, it's super useful. And when I look at um the productivity gains that I have, some of it is smaller like two or three. Sometimes it's like 10 times faster. I do task 10 times faster. The quality is not degraded. Sometimes the quality is um higher. Um an agent might not be as good as I am, but the agent doesn't get tired. Agent doesn't uh make sort of bad mistakes or Yeah. and you cognitively struggle like with like complicated information that you put together similar to Buddha kernels what what um Dan mentioned all of that is working and um I mean Matt as you put it is coding agents and agents for other stuff but how I would see it is it's just coding agents coding agents are general agents uh coding agents can write programs that solve other problems and code is so general if there's a digital problem you could solve it with code and coding agents make the things so easy that um now you can solve a variety of problems in a way that you couldn't solve before and this angle makes you productive. I would say this is the main the main thing that has been changing. Coding agents allow you to check a problem in a way that you could think before and is at a pace that you couldn't think before. You can paralyze a lot of different tasks. The agent doesn't get tired. Um you just keep going. the works much easier.

Matt Turk: One bit I love in your post is that you're careful to separate hype from reality at the beginning but then quickly you land from your experience of experimenting uh with agents uh for the live stream in particular you land at the conclusion that more than 90% of code and text should be written by agents you need to do so or you will be left behind.

Tim Dettmers: And I think for for many people that are engineers that's already true and there is sort of this thinking oh if you produce text or code and everything sent by Asians must be low quality much must be bad but um the key thing is you inspect the code you inspect the text you might might make some slide edits the 10% that you do might make a big big difference through this um basically um sort of editing just reviewing of output you kind of make it your own um AI generated things um are not less personal than things you've written on know I see that if if I write a grant proposal with the help of an agent it's sort of alive it I can feel like it's exciting like if person reading it said like yeah this is good research I want to fund this I think that's just the reality like if you just generate things and don't look at it and just say like yep that's good that will not help you but um you can quickly review content you can skin it you can look at like ah this doesn't look right or I want it different and edit it and you're good to go that will be the reality and the skills that you need to um work in that way. They're not fully developed for most people. They're also not fully developed for me. It's still sort of a face of experimentation. Models move, frameworks move. And so you need to adapt, you need to learn, there's lots to learn, but if you do it, payoffs are huge. And I think there was a thinking that software engineers will no longer exist. But I think people no longer believe that. It's like subconscious are so productive. That is exactly what you need to learn. If you use agents well, you can do so many things and I think that's the core thing. If you don't know how to use agents well, you will be left behind. That will become a critical skill.

Matt Turk: Practically, how do I do that? If I'm uh not a coder and I think about automating some parts of of my job, what are some of your recommendations on how to approach that problem?

Tim Dettmers: Yeah, I mean the best thing is just sort of being sort of very pragmatic. Just think about things and trying to code them. particular if you're not a coder that's very difficult and there's sort of this barrier where say like I haven't coded before like I don't know this and but if you interact with agents um they can just build stuff and with minimal learning I mean they can also explain stuff with minimal learning you can get there execute programs build websites of this visual you get quick feedback it's not that difficult anymore I mean often I mentioned you need to inspect things but if you build simple tools for yourself to make your life easier. Often you don't need to do that. The agents write good code. If you work in a company, you need to integrate it in a good code base. You probably should review it. But if you build a small program on your own to make your work more productive, um that's easy. Um just to give you an example that might be random here I bet for example a tool that if I have a video where I talk so I record videos of how I interact with agents then there are certain phases where I just look at the outputs and try to understand things and there are phases where I talk so I just build a tool that recognizes the speech that when I'm saying gets a timestamps then it slices the video so basically I have an entire video where I talk rather than sort of moments where uh nothing happens and that's very easy to do. I built this like in 20 minutes. I think everybody can do it because I didn't look at the code. The agent that did it and then I look at the video I'm like, "Oh yeah, let's do it." Right. If you get started with a feedback loop, um you don't need to code. You just need to inspect the output that you can understand or learn how to execute a Python program or a bash shell and you're there.

Matt Turk: How do you pick what you want to automate? Like how do you how do I think about automation in my life?

Tim Dettmers: Yeah. So I also um talked about that in a blog post and it can be yeah sort of um sort of uh more intuitive thing and then a more nuance thing. I think the more intuitive thing is like you just think about what could be useful and then can be like something more complex. You say I want an Android app or an iPhone app that does this thing and you initially might think that's complex but then you throw a coation it works immediately. the world's your uh voice. It's like there's so many things that you can just do and you can be very creative and say like what I always wanted to have and it wasn't there. Nobody built this product. Can I build it now? And I think that mindset gives you useful things that makes you more productive. But it also flexes your muscle and sometimes it doesn't work and then you understand like okay add agent cycle this or this is what I still need to learn to make both these kind of things and I think that is sort of the more intrusive perspective that's very useful and that quickly get you started on the path where you say like um I mean first there's excitement then it's sort of sober the sober reality but then you pick up again and it's like you realize okay I do like this I get more and more productive day by day that's sort of the more intuitive part the more nuance part is the sort of part that I I learned in automation industry I worked like three years in automation industry in Germany automating factories and it's sort of a very calculated sort of approach where you you look at you how you work you time each of these steps and then you say if I automate this exact step in such a way what could be the payoff how much time would I save and then you calculate what is the productivity gain and then you calculate how much time do I need to develop this automation and if you do that you can quickly realize that certain automating certain things will not make a difference the blog post I mentioned emails doesn't really work um and there might be other things like a big thing is like always calendar invites nobody likes to create like invites for a meeting but then if you think about it is you're also very particular about meetings like some days you want more meetings you meeting day or you say like ah I can put this in before lunch an agent doesn't know that and if you specify that to an agent you could also just create the calendar invite this meeting by flight and it it doesn't increase productivity by much and so there are a lot of problems if you think about this nuance way and can say like okay I don't do that but ah this this will help me.

Matt Turk: And then on your end what have you learned or observed in terms of agents what works what currently doesn't work but will work soon how to manage them.

Dan Fu: I think there there there's two broad um things I I I've noticed for agents. So the first is um making the agents effective uh ends up being a lot like managing um junior folks on your team or or at a company. So for example, the new intern who who shows up on your team um you're not going to go to the intern and say, "Hey, go fix our revenue for the year, double our revenue for the year," or something like that. like maybe you'll try that once, but you're you're unlikely to to see um the the payoff from that. Um instead, what you often do with junior folks is you say, "Hey, here's a first little task that you can do that in to get to know this complicated codebase and here are the things that you might run into." Um because you you've kind of done it before. And when you give the agents that context, give them that ability to look at those things, then then they can usually figure things out. The other bit is that when you have a a new person on your team, you maybe won't give them access to all the production credentials and and all the production database and all those things, but you're going to give them enough tools to to be productive. Um, so sometimes there there's this tension between, oh, I don't want my agent to go delete my everything in production, so I'm just going to have it be hamstrung and and watch every little thing it does. Whereas, if you did that with a person, you would never expect that person to be productive. That's another big you kind of want to you know think about the agents as at least today as you know maybe interns or or more junior folks. The other really interesting thing that I've noticed and I think when I think more the educational role of a of a professor and how do you prepare people for this future where agents are going to be such an important part of of workflows is that how do you train for that? And one of the things that I've noticed is that the more expertise somebody has, so whether that's in for example Tim with process automation um or uh expertise in in what I do in kernels and writing these very highly specialized programs, the more expertise you have, the more powerful the agent makes you. And that's because you can work at such a higher level of abstraction. You know what the important things are. You know how to set the direction. um you know what the common pitfalls are like what is easy what's kind of hard what you need to break up into multiple steps. One piece of conversation that um was coming up for for a while was like are agents going to replace all software engineers or things like that or are they going to replace all junior people or something like that. Um I think where we are today that's probably you know clearly not the case anymore. Um where you know if I have a tool to make my team 10x stronger I'm not going to fire nine people on my team. I'm going to say, "Okay, go do this. Become a 100x more productive than you were before." That's one bit. But then also kind of the script for how you become an expert at something. Uh is is probably pretty similar to to the way that it was before. You're going to study things deeply. You're going to try to understand things a lot. Um you're going to want to do things yourself with with your um you know, hands-on and and really get things done. Uh in this world, you know, the Chat GPT can teach you a lot of things. I was personally um uh personally I was trying to get Catch APG to teach me all the little ins and outs of how a car works. I don't know how effective it was so far, but um you know there it's it's a lot easier to to learn things now than it was um certainly even five two three years ago. Um so th those are kind of some of the the the things I'd say. So you know you want to treat the agents as if you're you are in that manager role. You want to help them get unstuck. You don't want to just say like throw the agent at the problem, walk away and and never look at it again. Um, but you also want to kind of figure out how to level yourself up so that you can be a better manager, have more domain expertise and and really understand things in a deeper way.

Matt Turk: So the fact that you need to learn and be an expert uh and that doesn't change uh I think that's very interesting and that makes a lot of sense. The question is like if you show up on the job as a you know young kernel engineer and that's your first day typically they would be okay well you do this simple task and that other simple task and then by year two you graduate to a more complex task. How does this sort of hands-on job training look like?

Dan Fu: Right. Yeah. So we we think about this a lot together because you know we're we're still hiring aggressively even in this world where the the the models are very good, the agents are very good. the way we think about it. So, first I actually went, you know, the professor and me went and actually recorded a bunch of lectures on how GPUs work. Um, so I I make everybody watch those and then I still give them um a a task from scratch, which is, okay, go take this Flash Attention kernel and modify it to do some other thing. You can pick the the extra feature. The nice thing about the agents is that you can dive into that higher impact role in a way that you weren't able to before. So it's really impactful I think when a junior IC goes to try to manage someone for the first time um because you're suddenly starting to think in much more precise terms. So um you know the the classic software engineer thing is hey the PM asked me to do this and wrote this super long doc with all these requirements but then the minute that you try to go ask someone else to do something you realize the specificity that you need when you want to address a feature or something like that. The nice thing about agents is you can almost start to shortcut that process where you can have the junior IC, still be an IC, still do IC style contributions, but they can now act as that manager role and act as their own PM. Um, because when you're communicating with the agents, you need to be as precise about what what you're trying to get done. So, in some sense, I've seen um with the with the really uh you know, the junior folks who are joining my team. So, these are folks fresh out of college or or fresh out of a masters. Um, when they are really gung-ho about understanding and being able to use the AI agents, they're they're able to communicate so much better than in in the olden days. They're they're able to level up their um their level of understanding a lot faster. Um, and then of course they they can do things and build tools at a speed that that um would have been really really hard to do, you know, 5 10 5 10 years ago.

Tim Dettmers: And maybe I'll add sort of the educational perspective there because I think it's sort of quite interesting. It's a little bit sort of contrasting. So what is also quite interesting is the educational perspective sort of from agents. I talk quite a bit that um basically use agents or be left behind and that is also true for students but just like as Dan said um you need the domain expertise you need to have some knowledge to use agents. Well, what we're seeing is if we allow students to use agents, they are very productive, but sometimes they build solutions that look correct that are actually very bad or just wrong and they don't realize. We at this point where it's almost very difficult to learn both domain expertise and uh a agent use. That's a very difficult balance to achieve uh because we don't want to have students that don't understand things, but we also want to have students that um and basically can use agents and so if they can't do that, they will not be effective in the workforce. What Dan said is you already have a pretty good person with strong background knowledge and then that person can level up. They're equipped with agents. But what do you do if you have someone that just learns computer science? How much AC should they learn? how much just um work should you do without agents and that's a very tricky balance and we don't know how to solve it. If we let people use agents, they perform very poorly on basic knowledge. And if we let people just do the balancing knowledge, they don't know how to use agents and they can't compete. So they can't do useful work in the in the workforce nowadays. Maybe the their solution is do all sort of the basic knowledge first and then agents. But that's not what students do. Students have access to these AI tools. They will use them because it's easy. And so maybe the solution is just you need a way of thinking and of working with information and knowledge that you don't understand and develop. I mean there's critical thinking. I think this goes beyond critical thinking. You basically need to uh basically know the unknown unknowns things that you didn't consider and don't understand and that you didn't even think about. you need to have that ability to think more about that to really keep up with agents because I think in the future it's realistic that we work on problems that we don't understand that agents understand but we need to keep up in some way um that would be difficult.

Matt Turk: All right. So to switch ts as we uh get uh closer to the end of this conversation what are you guys currently working on? what's top of mind for you at um Allen Institute on the one hand, together on the other hand, whoever wants to take it first.

Tim Dettmers: Well, we have actually sort of a very exciting project that will be released uh very soon in the next weeks and um so I worked quite a bit of efficiency. Um, I've been switching my work basically to coding agents and so we will have a major release of an open source coding agent that has a couple of key features. For one, training is 100 times cheaper. Um, you need to generate synthetic data, then you need to train on it. And so we have a method that's 100 uh roughly 100 times cheaper, but still get state-of-the-art performance. And then uh we have another sort of major result uh which um you can almost see as a holy grail of open source um models and that is we can take a private codebase like you have a company you have this codebase uh cloud doesn't know your codebase but um you have the data you can find a model in it so uh what we have is you can just point our method to that repository you don't need to have any sort of tests or um need to understand how to generate the data. It's just automatic. You quickly generate the data and then you have um an agent that is as a frontier model, but you can have like a 32 billion agent. You can deploy locally. You can have an army of specialized models for particular tasks for particular code bases and so forth. And uh yeah, I think that that is a very powerful results. All that is also packaged with um a science of coding agents. There are a lot of confounding factors, a lot of hidden things in papers that are not mentioned. We sort of unearth them, uh, build scaling laws and show what does matter and what doesn't matter. And if you put all of that together, I think the sheet agents, very few GPUs, everybody can use it and very easily. We unblock people by revealing all the sort of secrets. I think there will be a vast change in terms of how quickly we can progress in coding agents.

Matt Turk: Very cool. What about you, Dan?

Dan Fu: Yeah. So Together I think the the the major question that we're trying to answer today is we have all these powerful AI models that that can do all these amazing things. They're very expensive to run today. So you know the the question in all the public markets is like is OpenAI going to be able to turn a profit um ever. Um and what's really exciting about Together is that we are kind of on the forefront of getting the these models to to use the hardware as as good as it can. Um, we talked a little bit about training, you know, ear early on in the podcast. At inference time, when the when when you have the model, when it's already been trained, already been post-trained, the hardware utilization is like less than 5%. Um, so it's it's at a place where there's so much more that we can do. There's so much more efficiency that that we can push out of these. And we're really excited about figuring out how to use the hardware the the best way that we can. Um, whether it's serving, you know, customers like Curser or or whatever the the next great foundation model company is. Um, but yeah, we're we're really excited about pushing that frontier and then really getting us to this future that the that, you know, if these models are going to be as impactful as you think, if they're going to be running everywhere, running your daily lives, uh, you know, running your toaster, uh, well, we better make it the the best possible toaster that we can.

Matt Turk: Do you want to talk about uh, Mega Kernels and Together Atlas for a minute?

Dan Fu: Together Mega Kernels, Together Atlas. These are both um, projects along these lines. So, let me dive into the mega kernels first. So to understand this uh the the first thing when we say kernels is we usually mean we are going to write a specialized GPU program for a single operation in a model um you can think of as model as one of these train models is like ABC different operations in a row and there will be hundreds of these and the way that we've been writing kernels for the whole history of let's say call it Nvidia hardware is that you really specialize a single kernel for a single operation with these mega kernels we're doing something quite interesting which is we can take the entire model, however many billions of parameters, and put it into a single GPU kernel. Um, and with that, you can start to do a lot more fine grain optimization than you were able to do before. Uh, it actually starts to make the Nvidia GPU look a little bit more like a Cerebras chip or look a little bit more like a Samanova chip in terms of the the optimization that you're able to do. And this is really critical at inference time. So, we're able to see 2x sometimes 3x speed ups um over even highly optimized inference engines. Um, so we're working on bringing that to um uh to to really work in production, bring bring it to fruition and and use it across our whole stack together. Atlas is another really great uh interesting research project that we've done recently where we can get the model. There's this uh there's this technique called speculative decoding that we use at inference where basically we have a little model that is trying to guess what the big model is going to do and because of the ways that we've designed language models. If the little model guesses correctly, you basically get those tokens for free. Um so if you do the speculative decoding right, you can get again 2x 3x speed ups um over over you know just just running a a vanilla model. Um, with Together Atlas, we do one extra thing, which is we say, "Okay, we can get that little model and we can actually adapt it to your traffic." The longer you use this model, the more it learns the patterns of what you're asking, of what you're saying, and it can actually get faster over time. Um, so all these things, you know, we we have efficiency in mind, inference efficiency in particular, and we're all kind of pushing towards um making making all these things faster.

Matt Turk: What are you excited about for 2026 at a reasonably granular level? Like what what do you think happens? What do you think doesn't happen?

Tim Dettmers: Yeah. So I think they're both sort of um I'm sort of split. I think a lot of things will be very boring and not much innovation, but then we also surprised by a couple of things that we maybe don't see. And I think actually sort of in the frontier we will be less surprised. I mean it's no secret that we ran out of pre-training data and as Dan said like these are sort of the muscles. Then you can sort of smooth over and you smooth over with synthetic data and um that's how you build coding agents um on lots and lots of different environments and combine the data. U we make some progress there but I think you already see the machine learns. I don't think coding agents will be that much better. The user experience will probably improve. uh but um you see it that um all these models get almost equally good like if I use um I had my u config like uh GLM 4.7 setup and I used it and I thought I was using Opus 4.5 and then later realized oh wait I use a different model because they keep quite similar and so I think we see less progress there where I think we see more progress is um actually the small models if you train smaller models on more specialized data, they can do quite well. And the smaller models you get, they're pretty powerful. A 100 billion parameter models, you can fit it pretty well. Even sort of low-grade data center GPU like an RTX 6,000 cost $6,000. I think for a lot of companies, it will be very interesting. They don't need to rely on the frontier models. The small models might even be better because of specialized. Big problem and like uh Anthropic CEO pointed it out you have these powerful open rate models but nobody uses them because the deployment is so complex and that is because once you go beyond 8 GPUs that you first need the users to make it efficient but then also very complicated in systems there's no open source system that can do that at the moment you need disagregated infrance separation across sequence lengths and so forth perhaps um we can build this we can build this also for an HGPU machine for smaller models and then the efficiency that you see with a 100 billion model will rival what frontier models have. So you will get the efficiency of small models. You get the flexibility of small models. Performance on the frontier will stagnate but on the smaller level we get more more powerful models still because you can distill from these large models into these small models and taken together. I think that will change things.

Dan Fu: I'm also you know really excited about small models. I think uh we're going to see a lot more capability out of them. Um I'll be watching the open source uh models pretty closely. I think uh you know with things like JM 4.7, you're starting to see it. You're starting to see the open source models rival some of um at least our current best Frontier models. Uh I think we're going to see another big jump in open source capabilities this year. Um I'm really excited to see new hardware. So we're starting to hear a little bit about Reuben, uh the next generation of Nvidia uh GPUs. I think we're starting to hear a bit about the AMD 400 series of GPUs. Um really excited to see kind of what that next jump in hardware capabilities is even as we haven't fully kind of used even the current generation of hardware. I'm excited to see what people do kind of with all the other modalities. I think last year video generation models had a little bit of a moment with Sora 2 with with Gemini um and and VO I think they called it. Really excited to see what what they can do. Um, and yeah, really excited to just just see, you know, what is that frontier of of intelligence that you can get on your laptop or on your phone and how fast can you push it? How how far can you push it? I think it's never been a more exciting time to work in AI.

Matt Turk: You both mentioned state space architectures earlier in the conversation. Do you think that's part of the near future that we sort of evolve to post transformer architectures whether state space ja world models whatever direction is that something that you see in the near-term horizon and that you think is uh desirable?

Dan Fu: I think uh in a lot of places they're they're already there. So some of the best audio models in the world are at least partially based on state space models. I think Nvidia released a bunch of really great hybrid models recently. I think uh Neotron is is what they called them. And so there there's a lot of really great um work kind of there already. Um I think we will see the architectures continue to evolve. So um in some sense the DeepSeek MLA compression takes some of those ideas. Uh one of the miniax models had sort of a linear attention idea. So I think you're going to see a lot more diversity in architectures. You're kind of already seeing it. Um but certainly certainly I think out of the Chinese labs where um there isn't really an OpenAI of China right so there there isn't like an OpenAI orthropic or or or a Google gem right that kind of brings all these you know these um centers of like product and model and revenue and all those together. So um I think you see a lot more risk takingaking out of the Chinese labs where you're trying to differentiate the next model your next open source model. One way to do that is architecture. Um another way to do that of course is just pure quality. So I think we're we're going to see a lot more explosion of of different architectures.

Matt Turk: All right. Well, this has been a fascinating conversation. Really appreciate the time and insight and thoughts. Thank you so much. Really appreciate it.

Tim Dettmers: Yeah, thank you so much, Matt.

Dan Fu: Thanks so much for having us. Great to see you, Tim. This was a lot of fun.

Matt Turk: Hi, it's Matt Turk again. Thanks for listening to this episode of the Mad Podcast. If you enjoyed it, we'd be very grateful if you would consider subscribing if you haven't already, or leaving a positive review or comment on whichever platform you're watching this or listening to this episode from. This really helps us build a podcast and get great guests. Thanks, and see you at the next episode.