Distillation as a Modern Scaling Engine

From rewriting Google’s search stack in the early 2000s to reviving sparse trillion-parameter models and co-designing TPUs with frontier ML research, Jeff Dean has quietly shaped nearly every layer of the modern AI stack. As Chief AI Scientist at Google and a driving force behind Gemini, Jeff has lived through multiple scaling revolutions from CPUs and sharded indices to multimodal models that reason across text, video, and code.

Jeff joins us to unpack what it really means to “own the Pareto frontier,” why distillation is the engine behind every Flash model breakthrough, how energy (in picojoules) not FLOPs is becoming the true bottleneck, what it was like leading the charge to unify all of Google’s AI teams, and why the next leap won’t come from bigger context windows alone, but from systems that give the illusion of attending to trillions of tokens.

We discuss:

* Jeff’s early neural net thesis in 1990: parallel training before it was cool, why he believed scaling would win decades early, and the “bigger model, more data, better results” mantra that held for 15 years

* The evolution of Google Search: sharding, moving the entire index into memory in 2001, softening query semantics pre-LLMs, and why retrieval pipelines already resemble modern LLM systems

* Pareto frontier strategy: why you need both frontier “Pro” models and low-latency “Flash” models, and how distillation lets smaller models surpass prior generations

* Distillation deep dive: ensembles → compression → logits as soft supervision, and why you need the biggest model to make the smallest one good

* Latency as a first-class objective: why 10–50x lower latency changes UX entirely, and how future reasoning workloads will demand 10,000 tokens/sec

* Energy-based thinking: picojoules per bit, why moving data costs 1000x more than a multiply, batching through the lens of energy, and speculative decoding as amortization

* TPU co-design: predicting ML workloads 2–6 years out, speculative hardware features, precision reduction, sparsity, and the constant feedback loop between model architecture and silicon

* Sparse models and “outrageously large” networks: trillions of parameters with 1–5% activation, and why sparsity was always the right abstraction

* Unified vs. specialized models: abandoning symbolic systems, why general multimodal models tend to dominate vertical silos, and when vertical fine-tuning still makes sense

* Long context and the illusion of scale: beyond needle-in-a-haystack benchmarks toward systems that narrow trillions of tokens to 117 relevant documents

* Personalized AI: attending to your emails, photos, and documents (with permission), and why retrieval + reasoning will unlock deeply personal assistants

* Coding agents: 50 AI interns, crisp specifications as a new core skill, and how ultra-low latency will reshape human–agent collaboration

* Why ideas still matter: transformers, sparsity, RL, hardware, systems — scaling wasn’t blind; the pieces had to multiply together

Show Notes:

* Gemma 3 Paper

* Gemma 3

* Gemini 2.5 Report

* Jeff Dean’s “Software Engineering Advice from

Building Large-Scale Distributed Systems” Presentation (with Back of the Envelope Calculations)

* Latency Numbers Every Programmer Should Know by Jeff Dean

* The Jeff Dean Facts

* Jeff Dean Google Bio

* Jeff Dean on “Important AI Trends” @Stanford AI Club

* Jeff Dean & Noam Shazeer — 25 years at Google (Dwarkesh)

—

Jeff Dean

* LinkedIn: https://www.linkedin.com/in/jeff-dean-8b212555

* X: https://x.com/jeffdean

Google

* https://google.com

* https://deepmind.google

Full Video Episode

Timestamps

00:00:04 — Introduction: Alessio & Swyx welcome Jeff Dean, chief AI scientist at Google, to the Latent Space podcast00:00:30 — Owning the Pareto Frontier & balancing frontier vs low-latency models00:01:31 — Frontier models vs Flash models + role of distillation00:03:52 — History of distillation and its original motivation00:05:09 — Distillation’s role in modern model scaling00:07:02 — Model hierarchy (Flash, Pro, Ultra) and distillation sources00:07:46 — Flash model economics & wide deployment00:08:10 — Latency importance for complex tasks00:09:19 — Saturation of some tasks and future frontier tasks00:11:26 — On benchmarks, public vs internal00:12:53 — Example long-context benchmarks & limitations00:15:01 — Long-context goals: attending to trillions of tokens00:16:26 — Realistic use cases beyond pure language00:18:04 — Multimodal reasoning and non-text modalities00:19:05 — Importance of vision & motion modalities00:20:11 — Video understanding example (extracting structured info)00:20:47 — Search ranking analogy for LLM retrieval00:23:08 — LLM representations vs keyword search00:24:06 — Early Google search evolution & in-memory index00:26:47 — Design principles for scalable systems00:28:55 — Real-time index updates & recrawl strategies00:30:06 — Classic “Latency numbers every programmer should know”00:32:09 — Cost of memory vs compute and energy emphasis00:34:33 — TPUs & hardware trade-offs for serving models00:35:57 — TPU design decisions & co-design with ML00:38:06 — Adapting model architecture to hardware00:39:50 — Alternatives: energy-based models, speculative decoding00:42:21 — Open research directions: complex workflows, RL00:44:56 — Non-verifiable RL domains & model evaluation00:46:13 — Transition away from symbolic systems toward unified LLMs00:47:59 — Unified models vs specialized ones00:50:38 — Knowledge vs reasoning & retrieval + reasoning00:52:24 — Vertical model specialization & modules00:55:21 — Token count considerations for vertical domains00:56:09 — Low resource languages & contextual learning00:59:22 — Origins: Dean’s early neural network work01:10:07 — AI for coding & human–model interaction styles01:15:52 — Importance of crisp specification for coding agents01:19:23 — Prediction: personalized models & state retrieval01:22:36 — Token-per-second targets (10k+) and reasoning throughput01:23:20 — Episode conclusion and thanks

Transcript

Alessio Fanelli [00:00:04]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, founder of Kernel Labs, and I’m joined by Swyx, editor of Latent Space.

Shawn Wang [00:00:11]: Hello, hello. We’re here in the studio with Jeff Dean, chief AI scientist at Google. Welcome. Thanks for having me. It’s a bit surreal to have you in the studio. I’ve watched so many of your talks, and obviously your career has been super legendary. So, I mean, congrats. I think the first thing must be said, congrats on owning the Pareto Frontier.

Jeff Dean [00:00:30]: Thank you, thank you. Pareto Frontiers are good. It’s good to be out there.

Shawn Wang [00:00:34]: Yeah, I mean, I think it’s a combination of both. You have to own the Pareto Frontier. You have to have like frontier capability, but also efficiency, and then offer that range of models that people like to use. And, you know, some part of this was started because of your hardware work. Some part of that is your model work, and I’m sure there’s lots of secret sauce that you guys have worked on cumulatively. But, like, it’s really impressive to see it all come together in, like, this slittily advanced.

Jeff Dean [00:01:04]: Yeah, yeah. I mean, I think, as you say, it’s not just one thing. It’s like a whole bunch of things up and down the stack. And, you know, all of those really combine to help make UNOS able to make highly capable large models, as well as, you know, software techniques to get those large model capabilities into much smaller, lighter weight models that are, you know, much more cost effective and lower latency, but still, you know, quite capable for their size. Yeah.

Alessio Fanelli [00:01:31]: How much pressure do you have on, like, having the lower bound of the Pareto Frontier, too? I think, like, the new labs are always trying to push the top performance frontier because they need to raise more money and all of that. And you guys have billions of users. And I think initially when you worked on the CPU, you were thinking about, you know, if everybody that used Google, we use the voice model for, like, three minutes a day, they were like, you need to double your CPU number. Like, what’s that discussion today at Google? Like, how do you prioritize frontier versus, like, we have to do this? How do we actually need to deploy it if we build it?

Jeff Dean [00:02:03]: Yeah, I mean, I think we always want to have models that are at the frontier or pushing the frontier because I think that’s where you see what capabilities now exist that didn’t exist at the sort of slightly less capable last year’s version or last six months ago version. At the same time, you know, we know those are going to be really useful for a bunch of use cases, but they’re going to be a bit slower and a bit more expensive than people might like for a bunch of other broader models. So I think what we want to do is always have kind of a highly capable sort of affordable model that enables a whole bunch of, you know, lower latency use cases. People can use them for agentic coding much more readily and then have the high-end, you know, frontier model that is really useful for, you know, deep reasoning, you know, solving really complicated math problems, those kinds of things. And it’s not that. One or the other is useful. They’re both useful. So I think we’d like to do both. And also, you know, through distillation, which is a key technique for making the smaller models more capable, you know, you have to have the frontier model in order to then distill it into your smaller model. So it’s not like an either or choice. You sort of need that in order to actually get a highly capable, more modest size model. Yeah.

Alessio Fanelli [00:03:24]: I mean, you and Jeffrey came up with the solution in 2014.

Jeff Dean [00:03:28]: Don’t forget, L’Oreal Vinyls as well. Yeah, yeah.

Alessio Fanelli [00:03:30]: A long time ago. But like, I’m curious how you think about the cycle of these ideas, even like, you know, sparse models and, you know, how do you reevaluate them? How do you think about in the next generation of model, what is worth revisiting? Like, yeah, they’re just kind of like, you know, you worked on so many ideas that end up being influential, but like in the moment, they might not feel that way necessarily. Yeah.

Jeff Dean [00:03:52]: I mean, I think distillation was originally motivated because we were seeing that we had a very large image data set at the time, you know, 300 million images that we could train on. And we were seeing that if you create specialists for different subsets of those image categories, you know, this one’s going to be really good at sort of mammals, and this one’s going to be really good at sort of indoor room scenes or whatever, and you can cluster those categories and train on an enriched stream of data after you do pre-training on a much broader set of images. You get much better performance. If you then treat that whole set of maybe 50 models you’ve trained as a large ensemble, but that’s not a very practical thing to serve, right? So distillation really came about from the idea of, okay, what if we want to actually serve that and train all these independent sort of expert models and then squish it into something that actually fits in a form factor that you can actually serve? And that’s, you know, not that different from what we’re doing today. You know, often today we’re instead of having an ensemble of 50 models. We’re having a much larger scale model that we then distill into a much smaller scale model.

Shawn Wang [00:05:09]: Yeah. A part of me also wonders if distillation also has a story with the RL revolution. So let me maybe try to articulate what I mean by that, which is you can, RL basically spikes models in a certain part of the distribution. And then you have to sort of, well, you can spike models, but usually sometimes... It might be lossy in other areas and it’s kind of like an uneven technique, but you can probably distill it back and you can, I think that the sort of general dream is to be able to advance capabilities without regressing on anything else. And I think like that, that whole capability merging without loss, I feel like it’s like, you know, some part of that should be a distillation process, but I can’t quite articulate it. I haven’t seen much papers about it.

Jeff Dean [00:06:01]: Yeah, I mean, I tend to think of one of the key advantages of distillation is that you can have a much smaller model and you can have a very large, you know, training data set and you can get utility out of making many passes over that data set because you’re now getting the logits from the much larger model in order to sort of coax the right behavior out of the smaller model that you wouldn’t otherwise get with just the hard labels. And so, you know, I think that’s what we’ve observed. Is you can get, you know, very close to your largest model performance with distillation approaches. And that seems to be, you know, a nice sweet spot for a lot of people because it enables us to kind of, for multiple Gemini generations now, we’ve been able to make the sort of flash version of the next generation as good or even substantially better than the previous generations pro. And I think we’re going to keep trying to do that because that seems like a good trend to follow.

Shawn Wang [00:07:02]: So, Dara asked, so it was the original map was Flash Pro and Ultra. Are you just sitting on Ultra and distilling from that? Is that like the mother load?

Jeff Dean [00:07:12]: I mean, we have a lot of different kinds of models. Some are internal ones that are not necessarily meant to be released or served. Some are, you know, our pro scale model and we can distill from that as well into our Flash scale model. So I think, you know, it’s an important set of capabilities to have and also inference time scaling. It can also be a useful thing to improve the capabilities of the model.

Shawn Wang [00:07:35]: And yeah, yeah, cool. Yeah. And obviously, I think the economy of Flash is what led to the total dominance. I think the latest number is like 50 trillion tokens. I don’t know. I mean, obviously, it’s changing every day.

Jeff Dean [00:07:46]: Yeah, yeah. But, you know, by market share, hopefully up.

Shawn Wang [00:07:50]: No, I mean, there’s no I mean, there’s just the economics wise, like because Flash is so economical, like you can use it for everything. Like it’s in Gmail now. It’s in YouTube. Like it’s yeah. It’s in everything.

Jeff Dean [00:08:02]: We’re using it more in our search products of various AI mode reviews.

Shawn Wang [00:08:05]: Oh, my God. Flash past the AI mode. Oh, my God. Yeah, that’s yeah, I didn’t even think about that.

Jeff Dean [00:08:10]: I mean, I think one of the things that is quite nice about the Flash model is not only is it more affordable, it’s also a lower latency. And I think latency is actually a pretty important characteristic for these models because we’re going to want models to do much more complicated things that are going to involve, you know, generating many more tokens from when you ask the model to do so. So, you know, if you’re going to ask the model to do something until it actually finishes what you ask it to do, because you’re going to ask now, not just write me a for loop, but like write me a whole software package to do X or Y or Z. And so having low latency systems that can do that seems really important. And Flash is one direction, one way of doing that. You know, obviously our hardware platforms enable a bunch of interesting aspects of our, you know, serving stack as well, like TPUs, the interconnect between. Chips on the TPUs is actually quite, quite high performance and quite amenable to, for example, long context kind of attention operations, you know, having sparse models with lots of experts. These kinds of things really, really matter a lot in terms of how do you make them servable at scale.

Alessio Fanelli [00:09:19]: Yeah. Does it feel like there’s some breaking point for like the proto Flash distillation, kind of like one generation delayed? I almost think about almost like the capability as a. In certain tasks, like the pro model today is a saturated, some sort of task. So next generation, that same task will be saturated at the Flash price point. And I think for most of the things that people use models for at some point, the Flash model in two generation will be able to do basically everything. And how do you make it economical to like keep pushing the pro frontier when a lot of the population will be okay with the Flash model? I’m curious how you think about that.

Jeff Dean [00:09:59]: I mean, I think that’s true. If your distribution of what people are asking people, the models to do is stationary, right? But I think what often happens is as the models become more capable, people ask them to do more, right? So, I mean, I think this happens in my own usage. Like I used to try our models a year ago for some sort of coding task, and it was okay at some simpler things, but wouldn’t do work very well for more complicated things. And since then, we’ve improved dramatically on the more complicated coding tasks. And now I’ll ask it to do much more complicated things. And I think that’s true, not just of coding, but of, you know, now, you know, can you analyze all the, you know, renewable energy deployments in the world and give me a report on solar panel deployment or whatever. That’s a very complicated, you know, more complicated task than people would have asked a year ago. And so you are going to want more capable models to push the frontier in the absence of what people ask the models to do. And that also then gives us. Insight into, okay, where does the, where do things break down? How can we improve the model in these, these particular areas, uh, in order to sort of, um, make the next generation even better.

Alessio Fanelli [00:11:11]: Yeah. Are there any benchmarks or like test sets they use internally? Because it’s almost like the same benchmarks get reported every time. And it’s like, all right, it’s like 99 instead of 97. Like, how do you have to keep pushing the team internally to it? Or like, this is what we’re building towards. Yeah.

Jeff Dean [00:11:26]: I mean, I think. Benchmarks, particularly external ones that are publicly available. Have their utility, but they often kind of have a lifespan of utility where they’re introduced and maybe they’re quite hard for current models. You know, I, I like to think of the best kinds of benchmarks are ones where the initial scores are like 10 to 20 or 30%, maybe, but not higher. And then you can sort of work on improving that capability for, uh, whatever it is, the benchmark is trying to assess and get it up to like 80, 90%, whatever. I, I think once it hits kind of 95% or something, you get very diminishing returns from really focusing on that benchmark, cuz it’s sort of, it’s either the case that you’ve now achieved that capability, or there’s also the issue of leakage in public data or very related kind of data being, being in your training data. Um, so we have a bunch of held out internal benchmarks that we really look at where we know that wasn’t represented in the training data at all. There are capabilities that we want the model to have. Um, yeah. Yeah. Um, that it doesn’t have now, and then we can work on, you know, assessing, you know, how do we make the model better at these kinds of things? Is it, we need different kind of data to train on that’s more specialized for this particular kind of task. Do we need, um, you know, a bunch of, uh, you know, architectural improvements or some sort of, uh, model capability improvements, you know, what would help make that better?

Shawn Wang [00:12:53]: Is there, is there such an example that you, uh, a benchmark inspired in architectural improvement? Like, uh, I’m just kind of. Jumping on that because you just.

Jeff Dean [00:13:02]: Uh, I mean, I think some of the long context capability of the, of the Gemini models that came, I guess, first in 1.5 really were about looking at, okay, we want to have, um, you know,

Shawn Wang [00:13:15]: immediately everyone jumped to like completely green charts of like, everyone had, I was like, how did everyone crack this at the same time? Right. Yeah. Yeah.

Jeff Dean [00:13:23]: I mean, I think, um, and once you’re set, I mean, as you say that needed single needle and a half. Hey, stack benchmark is really saturated for at least context links up to 1, 2 and K or something. Don’t actually have, you know, much larger than 1, 2 and 8 K these days or two or something. We’re trying to push the frontier of 1 million or 2 million context, which is good because I think there are a lot of use cases where. Yeah. You know, putting a thousand pages of text or putting, you know, multiple hour long videos and the context and then actually being able to make use of that as useful. Try to, to explore the über graduation are fairly large. But the single needle in a haystack benchmark is sort of saturated. So you really want more complicated, sort of multi-needle or more realistic, take all this content and produce this kind of answer from a long context that sort of better assesses what it is people really want to do with long context. Which is not just, you know, can you tell me the product number for this particular thing?

Shawn Wang [00:14:31]: Yeah, it’s retrieval. It’s retrieval within machine learning. It’s interesting because I think the more meta level I’m trying to operate at here is you have a benchmark. You’re like, okay, I see the architectural thing I need to do in order to go fix that. But should you do it? Because sometimes that’s an inductive bias, basically. It’s what Jason Wei, who used to work at Google, would say. Exactly the kind of thing. Yeah, you’re going to win. Short term. Longer term, I don’t know if that’s going to scale. You might have to undo that.

Jeff Dean [00:15:01]: I mean, I like to sort of not focus on exactly what solution we’re going to derive, but what capability would you want? And I think we’re very convinced that, you know, long context is useful, but it’s way too short today. Right? Like, I think what you would really want is, can I attend to the internet while I answer my question? Right? But that’s not going to happen. I think that’s going to be solved by purely scaling the existing solutions, which are quadratic. So a million tokens kind of pushes what you can do. You’re not going to do that to a trillion tokens, let alone, you know, a billion tokens, let alone a trillion. But I think if you could give the illusion that you can attend to trillions of tokens, that would be amazing. You’d find all kinds of uses for that. You would have attend to the internet. You could attend to the pixels of YouTube and the sort of deeper representations that we can find. You could attend to the form for a single video, but across many videos, you know, on a personal Gemini level, you could attend to all of your personal state with your permission. So like your emails, your photos, your docs, your plane tickets you have. I think that would be really, really useful. And the question is, how do you get algorithmic improvements and system level improvements that get you to something where you actually can attend to trillions of tokens? Right. In a meaningful way. Yeah.

Shawn Wang [00:16:26]: But by the way, I think I did some math and it’s like, if you spoke all day, every day for eight hours a day, you only generate a maximum of like a hundred K tokens, which like very comfortably fits.

Jeff Dean [00:16:38]: Right. But if you then say, okay, I want to be able to understand everything people are putting on videos.

Shawn Wang [00:16:46]: Well, also, I think that the classic example is you start going beyond language into like proteins and whatever else is extremely information dense. Yeah. Yeah.

Jeff Dean [00:16:55]: I mean, I think one of the things about Gemini’s multimodal aspects is we’ve always wanted it to be multimodal from the start. And so, you know, that sometimes to people means text and images and video sort of human-like and audio, audio, human-like modalities. But I think it’s also really useful to have Gemini know about non-human modalities. Yeah. Like LIDAR sensor data from. Yes. Say, Waymo vehicles or. Like robots or, you know, various kinds of health modalities, x-rays and MRIs and imaging and genomics information. And I think there’s probably hundreds of modalities of data where you’d like the model to be able to at least be exposed to the fact that this is an interesting modality and has certain meaning in the world. Where even if you haven’t trained on all the LIDAR data or MRI data, you could have, because maybe that’s not, you know, it doesn’t make sense in terms of trade-offs of. You know, what you include in your main pre-training data mix, at least including a little bit of it is actually quite useful. Yeah. Because it sort of tempts the model that this is a thing.

Shawn Wang [00:18:04]: Yeah. Do you believe, I mean, since we’re on this topic and something I just get to ask you all the questions I always wanted to ask, which is fantastic. Like, are there some king modalities, like modalities that supersede all the other modalities? So a simple example was Vision can, on a pixel level, encode text. And DeepSeq had this DeepSeq CR paper that did that. Vision. And Vision has also been shown to maybe incorporate audio because you can do audio spectrograms and that’s, that’s also like a Vision capable thing. Like, so, so maybe Vision is just the king modality and like. Yeah.

Jeff Dean [00:18:36]: I mean, Vision and Motion are quite important things, right? Motion. Well, like video as opposed to static images, because I mean, there’s a reason evolution has evolved eyes like 23 independent ways, because it’s such a useful capability for sensing the world around you, which is really what we want these models to be. So I think the only thing that we can be able to do is interpret the things we’re seeing or the things we’re paying attention to and then help us in using that information to do things. Yeah.

Shawn Wang [00:19:05]: I think motion, you know, I still want to shout out, I think Gemini, still the only native video understanding model that’s out there. So I use it for YouTube all the time. Nice.

Jeff Dean [00:19:15]: Yeah. Yeah. I mean, it’s actually, I think people kind of are not necessarily aware of what the Gemini models can actually do. Yeah. Like I have an example I’ve used in one of my talks. It had like, it was like a YouTube highlight video of 18 memorable sports moments across the last 20 years or something. So it has like Michael Jordan hitting some jump shot at the end of the finals and, you know, some soccer goals and things like that. And you can literally just give it the video and say, can you please make me a table of what all these different events are? What when the date is when they happened? And a short description. And so you get like now an 18 row table of that information extracted from the video, which is, you know, not something most people think of as like a turn video into sequel like table.

Alessio Fanelli [00:20:11]: Has there been any discussion inside of Google of like, you mentioned tending to the whole internet, right? Google, it’s almost built because a human cannot tend to the whole internet and you need some sort of ranking to find what you need. Yep. That ranking is like much different for an LLM because you can expect a person to look at maybe the first five, six links in a Google search versus for an LLM. Should you expect to have 20 links that are highly relevant? Like how do you internally figure out, you know, how do we build the AI mode that is like maybe like much broader search and span versus like the more human one? Yeah.

Jeff Dean [00:20:47]: I mean, I think even pre-language model based work, you know, our ranking systems would be built to start. I mean, I think even pre-language model based work, you know, our ranking systems would be built to start. With a giant number of web pages in our index, many of them are not relevant. So you identify a subset of them that are relevant with very lightweight kinds of methods. You know, you’re down to like 30,000 documents or something. And then you gradually refine that to apply more and more sophisticated algorithms and more and more sophisticated sort of signals of various kinds in order to get down to ultimately what you show, which is, you know, the final 10 results or, you know, 10 results plus. Other kinds of information. And I think an LLM based system is not going to be that dissimilar, right? You’re going to attend to trillions of tokens, but you’re going to want to identify, you know, what are the 30,000 ish documents that are with the, you know, maybe 30 million interesting tokens. And then how do you go from that into what are the 117 documents I really should be paying attention to in order to carry out the tasks that the user has asked? And I think, you know, you can imagine systems where you have, you know, a lot of highly parallel processing to identify those initial 30,000 candidates, maybe with very lightweight kinds of models. Then you have some system that sort of helps you narrow down from 30,000 to the 117 with maybe a little bit more sophisticated model or set of models. And then maybe the final model is the thing that looks. So the 117 things that might be your most capable model. So I think it has to, it’s going to be some system like that, that is really enables you to give the illusion of attending to trillions of tokens. Sort of the way Google search gives you, you know, not the illusion, but you are searching the internet, but you’re finding, you know, a very small subset of things that are, that are relevant.

Shawn Wang [00:22:47]: Yeah. I often tell a lot of people that are not steeped in like Google search history that, well, you know, like Bert was. Like he was like basically immediately inside of Google search and that improves results a lot, right? Like I don’t, I don’t have any numbers off the top of my head, but like, I’m sure you guys, that’s obviously the most important numbers to Google. Yeah.

Jeff Dean [00:23:08]: I mean, I think going to an LLM based representation of text and words and so on enables you to get out of the explicit hard notion of, of particular words having to be on the page, but really getting at the notion of this topic of this page or this page. Paragraph is highly relevant to this query. Yeah.

Shawn Wang [00:23:28]: I don’t think people understand how much LLMs have taken over all these very high traffic system, very high traffic. Yeah. Like it’s Google, it’s YouTube. YouTube has this like semantics ID thing where it’s just like every token or every item in the vocab is a YouTube video or something that predicts the video using a code book, which is absurd to me for YouTube size.

Jeff Dean [00:23:50]: And then most recently GROK also for, for XAI, which is like, yeah. I mean, I’ll call out even before LLMs were used extensively in search, we put a lot of emphasis on softening the notion of what the user actually entered into the query.

Shawn Wang [00:24:06]: So do you have like a history of like, what’s the progression? Oh yeah.

Jeff Dean [00:24:09]: I mean, I actually gave a talk in, uh, I guess, uh, web search and data mining conference in 2009, uh, where we never actually published any papers about the origins of Google search, uh, sort of, but we went through sort of four or five or six. generations, four or five or six generations of, uh, redesigning of the search and retrieval system, uh, from about 1999 through 2004 or five. And that talk is really about that evolution. And one of the things that really happened in 2001 was we were sort of working to scale the system in multiple dimensions. So one is we wanted to make our index bigger, so we could retrieve from a larger index, which always helps your quality in general. Uh, because if you don’t have the page in your index, you’re going to not do well. Um, and then we also needed to scale our capacity because we were, our traffic was growing quite extensively. Um, and so we had, you know, a sharded system where you have more and more shards as the index grows, you have like 30 shards. And then if you want to double the index size, you make 60 shards so that you can bound the latency by which you respond for any particular user query. Um, and then as traffic grows, you add, you add more and more replicas of each of those. And so we eventually did the math that realized that in a data center where we had say 60 shards and, um, you know, 20 copies of each shard, we now had 1200 machines, uh, with disks. And we did the math and we’re like, Hey, one copy of that index would actually fit in memory across 1200 machines. So in 2001, we introduced, uh, we put our entire index in memory and what that enabled from a quality perspective was amazing. Um, and so we had more and more replicas of each of those. Before you had to be really careful about, you know, how many different terms you looked at for a query, because every one of them would involve a disk seek on every one of the 60 shards. And so you, as you make your index bigger, that becomes even more inefficient. But once you have the whole index in memory, it’s totally fine to have 50 terms you throw into the query from the user’s original three or four word query, because now you can add synonyms like restaurant and restaurants and cafe and, uh, you know, things like that. Uh, bistro and all these things. And you can suddenly start, uh, sort of really, uh, getting at the meaning of the word as opposed to the exact semantic form the user typed in. And that was, you know, 2001, very much pre LLM, but really it was about softening the, the strict definition of what the user typed in order to get at the meaning.

Alessio Fanelli [00:26:47]: What are like principles that you use to like design the systems, especially when you have, I mean, in 2001, the internet is like. Doubling, tripling every year in size is not like, uh, you know, and I think today you kind of see that with LLMs too, where like every year the jumps in size and like capabilities are just so big. Are there just any, you know, principles that you use to like, think about this? Yeah.

Jeff Dean [00:27:08]: I mean, I think, uh, you know, first, whenever you’re designing a system, you want to understand what are the sort of design parameters that are going to be most important in designing that, you know? So, you know, how many queries per second do you need to handle? How big is the internet? How big is the index you need to handle? How much data do you need to keep for every document in the index? How are you going to look at it when you retrieve things? Um, what happens if traffic were to double or triple, you know, will that system work well? And I think a good design principle is you’re going to want to design a system so that the most important characteristics could scale by like factors of five or 10, but probably not beyond that because often what happens is if you design a system for X. And something suddenly becomes a hundred X, that would enable a very different point in the design space that would not make sense at X. But all of a sudden at a hundred X makes total sense. So like going from a disk space index to a in memory index makes a lot of sense once you have enough traffic, because now you have enough replicas of the sort of state on disk that those machines now actually can hold, uh, you know, a full copy of the, uh, index and memory. Yeah. And that all of a sudden enabled. A completely different design that wouldn’t have been practical before. Yeah. Um, so I’m, I’m a big fan of thinking through designs in your head, just kind of playing with the design space a little before you actually do a lot of writing of code. But, you know, as you said, in the early days of Google, we were growing the index, uh, quite extensively. We were growing the update rate of the index. So the update rate actually is the parameter that changed the most. Surprising. So it used to be once a month.

Shawn Wang [00:28:55]: Yeah.

Jeff Dean [00:28:56]: And then we went to a system that could update any particular page in like sub one minute. Okay.

Shawn Wang [00:29:02]: Yeah. Because this is a competitive advantage, right?

Jeff Dean [00:29:04]: Because all of a sudden news related queries, you know, if you’re, if you’ve got last month’s news index, it’s not actually that useful for.

Shawn Wang [00:29:11]: News is a special beast. Was there any, like you could have split it onto a separate system.

Jeff Dean [00:29:15]: Well, we did. We launched a Google news product, but you also want news related queries that people type into the main index to also be sort of updated.

Shawn Wang [00:29:23]: So, yeah, it’s interesting. And then you have to like classify whether the page is, you have to decide which pages should be updated and what frequency. Oh yeah.

Jeff Dean [00:29:30]: There’s a whole like, uh, system behind the scenes that’s trying to decide update rates and importance of the pages. So even if the update rate seems low, you might still want to recrawl important pages quite often because, uh, the likelihood they change might be low, but the value of having updated is high.

Shawn Wang [00:29:50]: Yeah, yeah, yeah, yeah. Uh, well, you know, yeah. This, uh, you know, mention of latency and, and saving things to this reminds me of one of your classics, which I have to bring up, which is latency numbers. Every programmer should know, uh, was there a, was it just a, just a general story behind that? Did you like just write it down?

Jeff Dean [00:30:06]: I mean, this has like sort of eight or 10 different kinds of metrics that are like, how long does a cache mistake? How long does branch mispredict take? How long does a reference domain memory take? How long does it take to send, you know, a packet from the U S to the Netherlands or something? Um,

Shawn Wang [00:30:21]: why Netherlands, by the way, or is it, is that because of Chrome?

Jeff Dean [00:30:25]: Uh, we had a data center in the Netherlands, um, so, I mean, I think this gets to the point of being able to do the back of the envelope calculations. So these are sort of the raw ingredients of those, and you can use them to say, okay, well, if I need to design a system to do image search and thumb nailing or something of the result page, you know, how, what I do that I could pre-compute the image thumbnails. I could like. Try to thumbnail them on the fly from the larger images. What would that do? How much dis bandwidth than I need? How many des seeks would I do? Um, and you can sort of actually do thought experiments in, you know, 30 seconds or a minute with the sort of, uh, basic, uh, basic numbers at your fingertips. Uh, and then as you sort of build software using higher level libraries, you kind of want to develop the same intuitions for how long does it take to, you know, look up something in this particular kind of.

Shawn Wang [00:31:21]: I’ll see you next time.

Shawn Wang [00:31:51]: Which is a simple byte conversion. That’s nothing interesting. I wonder if you have any, if you were to update your...

Jeff Dean [00:31:58]: I mean, I think it’s really good to think about calculations you’re doing in a model, either for training or inference.

Jeff Dean [00:32:09]: Often a good way to view that is how much state will you need to bring in from memory, either like on-chip SRAM or HBM from the accelerator. Attached memory or DRAM or over the network. And then how expensive is that data motion relative to the cost of, say, an actual multiply in the matrix multiply unit? And that cost is actually really, really low, right? Because it’s order, depending on your precision, I think it’s like sub one picodule.

Shawn Wang [00:32:50]: Oh, okay. You measure it by energy. Yeah. Yeah.

Jeff Dean [00:32:52]: Yeah. I mean, it’s all going to be about energy and how do you make the most energy efficient system. And then moving data from the SRAM on the other side of the chip, not even off the off chip, but on the other side of the same chip can be, you know, a thousand picodules. Oh, yeah. And so all of a sudden, this is why your accelerators require batching. Because if you move, like, say, the parameter of a model from SRAM on the, on the chip into the multiplier unit, that’s going to cost you a thousand picodules. So you better make use of that, that thing that you moved many, many times with. So that’s where the batch dimension comes in. Because all of a sudden, you know, if you have a batch of 256 or something, that’s not so bad. But if you have a batch of one, that’s really not good.

Shawn Wang [00:33:40]: Yeah. Yeah. Right.

Jeff Dean [00:33:41]: Because then you paid a thousand picodules in order to do your one picodule multiply.

Shawn Wang [00:33:46]: I have never heard an energy-based analysis of batching.

Jeff Dean [00:33:50]: Yeah. I mean, that’s why people batch. Yeah. Ideally, you’d like to use batch size one because the latency would be great.

Shawn Wang [00:33:56]: The best latency.

Jeff Dean [00:33:56]: But the energy cost and the compute cost inefficiency that you get is quite large. So, yeah.

Shawn Wang [00:34:04]: Is there a similar trick like, like, like you did with, you know, putting everything in memory? Like, you know, I think obviously NVIDIA has caused a lot of waves with betting very hard on SRAM with Grok. I wonder if, like, that’s something that you already saw with, with the TPUs, right? Like that, that you had to. Uh, to serve at your scale, uh, you probably sort of saw that coming. Like what, what, what hardware, uh, innovations or insights were formed because of what you’re seeing there?

Jeff Dean [00:34:33]: Yeah. I mean, I think, you know, TPUs have this nice, uh, sort of regular structure of 2D or 3D meshes with a bunch of chips connected. Yeah. And each one of those has HBM attached. Um, I think for serving some kinds of models, uh, you know, you, you pay a lot higher cost. Uh, and time latency, um, bringing things in from HBM than you do bringing them in from, uh, SRAM on the chip. So if you have a small enough model, you can actually do model parallelism, spread it out over lots of chips and you actually get quite good throughput improvements and latency improvements from doing that. And so you’re now sort of striping your smallish scale model over say 16 or 64 chips. Uh, but as if you do that and it all fits in. In SRAM, uh, that can be a big win. So yeah, that’s not a surprise, but it is a good technique.

Alessio Fanelli [00:35:27]: Yeah. What about the TPU design? Like how much do you decide where the improvements have to go? So like, this is like a good example of like, is there a way to bring the thousand picojoules down to 50? Like, is it worth designing a new chip to do that? The extreme is like when people say, oh, you should burn the model on the ASIC and that’s kind of like the most extreme thing. How much of it? Is it worth doing an hardware when things change so quickly? Like what was the internal discussion? Yeah.

Jeff Dean [00:35:57]: I mean, we, we have a lot of interaction between say the TPU chip design architecture team and the sort of higher level modeling, uh, experts, because you really want to take advantage of being able to co-design what should future TPUs look like based on where we think the sort of ML research puck is going, uh, in some sense, because, uh, you know, as a hardware designer for ML and in particular, you’re trying to design a chip starting today and that design might take two years before it even lands in a data center. And then it has to sort of be a reasonable lifetime of the chip to take you three, four or five years. So you’re trying to predict two to six years out where, what ML computations will people want to run two to six years out in a very fast changing field. And so having people with interest. Interesting ML research ideas of things we think will start to work in that timeframe or will be more important in that timeframe, uh, really enables us to then get, you know, interesting hardware features put into, you know, TPU N plus two, where TPU N is what we have today.

Shawn Wang [00:37:10]: Oh, the cycle time is plus two.

Jeff Dean [00:37:12]: Roughly. Wow. Because, uh, I mean, sometimes you can squeeze some changes into N plus one, but, you know, bigger changes are going to require the chip. Yeah. Design be earlier in its lifetime design process. Um, so whenever we can do that, it’s generally good. And sometimes you can put in speculative features that maybe won’t cost you much chip area, but if it works out, it would make something, you know, 10 times as fast. And if it doesn’t work out, well, you burned a little bit of tiny amount of your chip area on that thing, but it’s not that big a deal. Uh, sometimes it’s a very big change and we want to be pretty sure this is going to work out. So we’ll do like lots of carefulness. Uh, ML experimentation to show us, uh, this is actually the, the way we want to go. Yeah.

Alessio Fanelli [00:37:58]: Is there a reverse of like, we already committed to this chip design so we can not take the model architecture that way because it doesn’t quite fit?

Jeff Dean [00:38:06]: Yeah. I mean, you, you definitely have things where you’re going to adapt what the model architecture looks like so that they’re efficient on the chips that you’re going to have for both training and inference of that, of that, uh, generation of model. So I think it kind of goes both ways. Um, you know, sometimes you can take advantage of, you know, lower precision things that are coming in a future generation. So you can, might train it at that lower precision, even if the current generation doesn’t quite do that. Mm.

Shawn Wang [00:38:40]: Yeah. How low can we go in precision?

Jeff Dean [00:38:43]: Because people are saying like ternary is like, uh, yeah, I mean, I’m a big fan of very low precision because I think that gets, that saves you a tremendous amount of time. Right. Because it’s picojoules per bit that you’re transferring and reducing the number of bits is a really good way to, to reduce that. Um, you know, I think people have gotten a lot of luck, uh, mileage out of having very low bit precision things, but then having scaling factors that apply to a whole bunch of, uh, those, those weights. Scaling. How does it, how does it, okay.

Shawn Wang [00:39:15]: Interesting. You, so low, low precision, but scaled up weights. Yeah. Huh. Yeah. Never considered that. Yeah. Interesting. Uh, w w while we’re on this topic, you know, I think there’s a lot of, um, uh, this, the concept of precision at all is weird when we’re sampling, you know, uh, we just, at the end of this, we’re going to have all these like chips that I’ll do like very good math. And then we’re just going to throw a random number generator at the start. So, I mean, there’s a movement towards, uh, energy based, uh, models and processors. I’m just curious if you’ve, obviously you’ve thought about it, but like, what’s your commentary?

Jeff Dean [00:39:50]: Yeah. I mean, I think. There’s a bunch of interesting trends though. Energy based models is one, you know, diffusion based models, which don’t sort of sequentially decode tokens is another, um, you know, speculative decoding is a way that you can get sort of an equivalent, very small.

Shawn Wang [00:40:06]: Draft.

Jeff Dean [00:40:07]: Batch factor, uh, for like you predict eight tokens out and that enables you to sort of increase the effective batch size of what you’re doing by a factor of eight, even, and then you maybe accept five or six of those tokens. So you get. A five, a five X improvement in the amortization of moving weights, uh, into the multipliers to do the prediction for the, the tokens. So these are all really good techniques and I think it’s really good to look at them from the lens of, uh, energy, real energy, not energy based models, um, and, and also latency and throughput, right? If you look at things from that lens, that sort of guides you to. Two solutions that are gonna be, uh, you know, better from, uh, you know, being able to serve larger models or, you know, equivalent size models more cheaply and with lower latency.

Shawn Wang [00:41:03]: Yeah. Well, I think, I think I, um, it’s appealing intellectually, uh, haven’t seen it like really hit the mainstream, but, um, I do think that, uh, there’s some poetry in the sense that, uh, you know, we don’t have to do, uh, a lot of shenanigans if like we fundamentally. Design it into the hardware. Yeah, yeah.

Jeff Dean [00:41:23]: I mean, I think there’s still a, there’s also sort of the more exotic things like analog based, uh, uh, computing substrates as opposed to digital ones. Uh, I’m, you know, I think those are super interesting cause they can be potentially low power. Uh, but I think you often end up wanting to interface that with digital systems and you end up losing a lot of the power advantages in the digital to analog and analog to digital conversions. You end up doing, uh, at the sort of boundaries. And periphery of that system. Um, I still think there’s a tremendous distance we can go from where we are today in terms of energy efficiency with sort of, uh, much better and specialized hardware for the models we care about.

Shawn Wang [00:42:05]: Yeah.

Alessio Fanelli [00:42:06]: Um, any other interesting research ideas that you’ve seen, or like maybe things that you cannot pursue a Google that you would be interested in seeing researchers take a step at, I guess you have a lot of researchers. Yeah, I guess you have enough, but our, our research.

Jeff Dean [00:42:21]: Our research portfolio is pretty broad. I would say, um, I mean, I think, uh, in terms of research directions, there’s a whole bunch of, uh, you know, open problems and how do you make these models reliable and able to do much longer, kind of, uh, more complex tasks that have lots of subtasks. How do you orchestrate, you know, maybe one model that’s using other models as tools in order to sort of build, uh, things that can accomplish, uh, you know, much more. Yeah. Significant pieces of work, uh, collectively, then you would ask a single model to do. Um, so that’s super interesting. How do you get more verifiable, uh, you know, how do you get RL to work for non-verifiable domains? I think it’s a pretty interesting open problem because I think that would broaden out the capabilities of the models, the improvements that you’re seeing in both math and coding. Uh, if we could apply those to other less verifiable domains, because we’ve come up with RL techniques that actually enable us to do that. Uh, effectively, that would, that would really make the models improve quite a lot. I think.

Alessio Fanelli [00:43:26]: I’m curious, like when we had Noam Brown on the podcast, he said, um, they already proved you can do it with deep research. Um, you kind of have it with AI mode in a way it’s not verifiable. I’m curious if there’s any thread that you think is interesting there. Like what is it? Both are like information retrieval of JSON. So I wonder if it’s like the retrieval is like the verifiable part. That you can score or what are like, yeah, yeah. How, how would you model that, that problem?

Jeff Dean [00:43:55]: Yeah. I mean, I think there are ways of having other models that can evaluate the results of what a first model did, maybe even retrieving. Can you have another model that says, is this things, are these things you retrieved relevant? Or can you rate these 2000 things you retrieved to assess which ones are the 50 most relevant or something? Um, I think those kinds of techniques are actually quite effective. Sometimes I can even be the same model, just prompted differently to be a, you know, a critic as opposed to a, uh, actual retrieval system. Yeah.

Shawn Wang [00:44:28]: Um, I do think like there, there is that, that weird cliff where like, it feels like we’ve done the easy stuff and then now it’s, but it always feels like that every year. It’s like, oh, like we know, we know, and the next part is super hard and nobody’s figured it out. And, uh, exactly with this RLVR thing where like everyone’s talking about, well, okay, how do we. the next stage of the non-verifiable stuff. And everyone’s like, I don’t know, you know, Ellen judge.

Jeff Dean [00:44:56]: I mean, I feel like the nice thing about this field is there’s lots and lots of smart people thinking about creative solutions to some of the problems that we all see. Uh, because I think everyone sort of sees that the models, you know, are great at some things and they fall down around the edges of those things and, and are not as capable as we’d like in those areas. And then coming up with good techniques and trying those. And seeing which ones actually make a difference is sort of what the whole research aspect of this field is, is pushing forward. And I think that’s why it’s super interesting. You know, if you think about two years ago, we were struggling with GSM, eight K problems, right? Like, you know, Fred has two rabbits. He gets three more rabbits. How many rabbits does he have? That’s a pretty far cry from the kinds of mathematics that the models can, and now you’re doing IMO and Erdos problems in pure language. Yeah. Yeah. Pure language. So that is a really, really amazing jump in capabilities in, you know, in a year and a half or something. And I think, um, for other areas, it’d be great if we could make that kind of leap. Uh, and you know, we don’t exactly see how to do it for some, some areas, but we do see it for some other areas and we’re going to work hard on making that better. Yeah.

Shawn Wang [00:46:13]: Yeah.

Alessio Fanelli [00:46:14]: Like YouTube thumbnail generation. That would be very helpful. We need that. That would be AGI. We need that.

Shawn Wang [00:46:20]: That would be. As far as content creators go.

Jeff Dean [00:46:22]: I guess I’m not a YouTube creator, so I don’t care that much about that problem, but I guess, uh, many people do.

Shawn Wang [00:46:27]: It does. Yeah. It doesn’t, it doesn’t matter. People do judge books by their covers as it turns out. Um, uh, just to draw a bit on the IMO goal. Um, I’m still not over the fact that a year ago we had alpha proof and alpha geometry and all those things. And then this year we were like, screw that we’ll just chuck it into Gemini. Yeah. What’s your reflection? Like, I think this, this question about. Like the merger of like symbolic systems and like, and, and LMS, uh, was a very much core belief. And then somewhere along the line, people would just said, Nope, we’ll just all do it in the LLM.

Jeff Dean [00:47:02]: Yeah. I mean, I think it makes a lot of sense to me because, you know, humans manipulate symbols, but we probably don’t have like a symbolic representation in our heads. Right. We have some distributed representation that is neural net, like in some way of lots of different neurons. And activation patterns firing when we see certain things and that enables us to reason and plan and, you know, do chains of thought and, you know, roll them back now that, that approach for solving the problem doesn’t seem like it’s going to work. I’m going to try this one. And, you know, in a lot of ways we’re emulating what we intuitively think, uh, is happening inside real brains in neural net based models. So it never made sense to me to have like completely separate. Uh, discrete, uh, symbolic things, and then a completely different way of, of, uh, you know, thinking about those things.

Shawn Wang [00:47:59]: Interesting. Yeah. Uh, I mean, it’s maybe seems obvious to you, but it wasn’t obvious to me a year ago. Yeah.

Jeff Dean [00:48:06]: I mean, I do think like that IMO with, you know, translating to lean and using lean and then the next year and also a specialized geometry model. And then this year switching to a single unified model. That is roughly the production model with a little bit more inference budget, uh, is actually, you know, quite good because it shows you that the capabilities of that general model have improved dramatically and, and now you don’t need the specialized model. This is actually sort of very similar to the 2013 to 16 era of machine learning, right? Like it used to be, people would train separate models for lots of different, each different problem, right? I have, I want to recognize street signs and something. So I train a street sign. Recognition recognition model, or I want to, you know, decode speech recognition. I have a speech model, right? I think now the era of unified models that do everything is really upon us. And the question is how well do those models generalize to new things they’ve never been asked to do and they’re getting better and better.

Shawn Wang [00:49:10]: And you don’t need domain experts. Like one of my, uh, so I interviewed ETA who was on, who was on that team. Uh, and he was like, yeah, I, I don’t know how they work. I don’t know where the IMO competition was held. I don’t know the rules of it. I just trained the models, the training models. Yeah. Yeah. And it’s kind of interesting that like people with these, this like universal skill set of just like machine learning, you just give them data and give them enough compute and they can kind of tackle any task, which is the bitter lesson, I guess. I don’t know. Yeah.

Jeff Dean [00:49:39]: I mean, I think, uh, general models, uh, will win out over specialized ones in most cases.

Shawn Wang [00:49:45]: Uh, so I want to push there a bit. I think there’s one hole here, which is like, uh. There’s this concept of like, uh, maybe capacity of a model, like abstractly a model can only contain the number of bits that it has. And, uh, and so it, you know, God knows like Gemini pro is like one to 10 trillion parameters. We don’t know, but, uh, the Gemma models, for example, right? Like a lot of people want like the open source local models that are like that, that, that, and, and, uh, they have some knowledge, which is not necessary, right? Like they can’t know everything like, like you have the. The luxury of you have the big model and big model should be able to capable of everything. But like when, when you’re distilling and you’re going down to the small models, you know, you’re actually memorizing things that are not useful. Yeah. And so like, how do we, I guess, do we want to extract that? Can we, can we divorce knowledge from reasoning, you know?

Jeff Dean [00:50:38]: Yeah. I mean, I think you do want the model to be most effective at reasoning if it can retrieve things, right? Because having the model devote precious parameter space. To remembering obscure facts that could be looked up is actually not the best use of that parameter space, right? Like you might prefer something that is more generally useful in more settings than this obscure fact that it has. Um, so I think that’s always attention at the same time. You also don’t want your model to be kind of completely detached from, you know, knowing stuff about the world, right? Like it’s probably useful to know how long the golden gate be. Bridges just as a general sense of like how long are bridges, right? And, uh, it should have that kind of knowledge. It maybe doesn’t need to know how long some teeny little bridge in some other more obscure part of the world is, but, uh, it does help it to have a fair bit of world knowledge and the bigger your model is, the more you can have. Uh, but I do think combining retrieval with sort of reasoning and making the model really good at doing multiple stages of retrieval. Yeah.

Shawn Wang [00:51:49]: And reasoning through the intermediate retrieval results is going to be a, a pretty effective way of making the model seem much more capable, because if you think about, say, a personal Gemini, yeah, right?

Jeff Dean [00:52:01]: Like we’re not going to train Gemini on my email. Probably we’d rather have a single model that, uh, we can then use and use being able to retrieve from my email as a tool and have the model reason about it and retrieve from my photos or whatever, uh, and then make use of that and have multiple. Um, you know, uh, stages of interaction. that makes sense.

Alessio Fanelli [00:52:24]: Do you think the vertical models are like, uh, interesting pursuit? Like when people are like, oh, we’re building the best healthcare LLM, we’re building the best law LLM, are those kind of like short-term stopgaps or?

Jeff Dean [00:52:37]: No, I mean, I think, I think vertical models are interesting. Like you want them to start from a pretty good base model, but then you can sort of, uh, sort of viewing them, view them as enriching the data. Data distribution for that particular vertical domain for healthcare, say, um, we’re probably not going to train or for say robotics. We’re probably not going to train Gemini on all possible robotics data. We, you could train it on because we want it to have a balanced set of capabilities. Um, so we’ll expose it to some robotics data, but if you’re trying to build a really, really good robotics model, you’re going to want to start with that and then train it on more robotics data. And then maybe that would. It’s multilingual translation capability, but improve its robotics capabilities. And we’re always making these kind of, uh, you know, trade-offs in the data mix that we train the base Gemini models on. You know, we’d love to include data from 200 more languages and as much data as we have for those languages, but that’s going to displace some other capabilities of the model. It won’t be as good at, um, you know, Pearl programming, you know, it’ll still be good at Python programming. Cause we’ll include it. Enough. Of that, but there’s other long tail computer languages or coding capabilities that it may suffer on or multi, uh, multimodal reasoning capabilities may suffer. Cause we didn’t get to expose it to as much data there, but it’s really good at multilingual things. So I, I think some combination of specialized models, maybe more modular models. So it’d be nice to have the capability to have those 200 languages, plus this awesome robotics model, plus this awesome healthcare, uh, module that all can be knitted together to work in concert and called upon in different circumstances. Right? Like if I have a health related thing, then it should enable using this health module in conjunction with the main base model to be even better at those kinds of things. Yeah.

Shawn Wang [00:54:36]: Installable knowledge. Yeah.

Jeff Dean [00:54:37]: Right.

Shawn Wang [00:54:38]: Just download as a, as a package.

Jeff Dean [00:54:39]: And some of that installable stuff can come from retrieval, but some of it probably should come from preloaded training on, you know, uh, a hundred billion tokens or a trillion tokens of health data. Yeah.

Shawn Wang [00:54:51]: And for listeners, I think, uh, I will highlight the Gemma three end paper where they, there was a little bit of that, I think. Yeah.

Alessio Fanelli [00:54:56]: Yeah. I guess the question is like, how many billions of tokens do you need to outpace the frontier model improvements? You know, it’s like, if I have to make this model better healthcare and the main. Gemini model is still improving. Do I need 50 billion tokens? Can I do it with a hundred, if I need a trillion healthcare tokens, it’s like, they’re probably not out there that you don’t have, you know, I think that’s really like the.

Jeff Dean [00:55:21]: Well, I mean, I think healthcare is a particularly challenging domain, so there’s a lot of healthcare data that, you know, we don’t have access to appropriately, but there’s a lot of, you know, uh, healthcare organizations that want to train models on their own data. That is not public healthcare data, uh, not public health. But public healthcare data. Um, so I think there are opportunities there to say, partner with a large healthcare organization and train models for their use that are going to be, you know, more bespoke, but probably, uh, might be better than a general model trained on say, public data. Yeah.

Shawn Wang [00:55:58]: Yeah. I, I believe, uh, by the way, also this is like somewhat related to the language conversation. Uh, I think one of your, your favorite examples was you can put a low resource language in the context and it just learns. Yeah.

Jeff Dean [00:56:09]: Oh, yeah, I think the example we used was Calamon, which is truly low resource because it’s only spoken by, I think 120 people in the world and there’s no written text.

Shawn Wang [00:56:20]: So, yeah. So you can just do it that way. Just put it in the context. Yeah. Yeah. But I think your whole data set in the context, right.

Jeff Dean [00:56:27]: If you, if you take a language like, uh, you know, Somali or something, there is a fair bit of Somali text in the world that, uh, or Ethiopian Amharic or something, um, you know, we probably. Yeah. Are not putting all the data from those languages into the Gemini based training. We put some of it, but if you put more of it, you’ll improve the capabilities of those models.

Shawn Wang [00:56:49]: Yeah.

Jeff Dean [00:56:49]: So, or of those languages.

Shawn Wang [00:56:52]: Uh, yeah, cool. Uh, it’s, uh, I have a side interest in linguistics. I, I, I did, uh, uh, a few classes back in college and like, uh, part of me, like if I was a linguist and I could have access to all these models, I would just be asking really fundamental questions about language itself. Yeah. Like, uh, one is th there’s one very obvious one, which is Sapir-Whorf, like how much does like the language that you speak affect your thinking, but then also there’s some languages where there’s just concepts that are not represented in other languages, but some others, many others that are just duplicates, right. Where, uh, there’s also another paper that people love called the platonic representation where, you know, like the, the, an image of a cup is, uh, if you say learn a model on that and you, you, you have a lot of texts with the word cup eventually maps to it, like roughly the same place. And so like that should apply to languages except where it doesn’t. And that’s actually like very interesting differences in what humanity has discovered as concepts that maybe English doesn’t have.

Shawn Wang [00:57:54]: I don’t know. It’s just like my, my rant on languages. Yeah.

Jeff Dean [00:57:58]: I mean, I, I did some work on a early model that fused together a language based model with you have, you know, nice word based representations and then an image model where you have. Trained it on image net like things. Yes. And then you fuse together the top layers of, uh, no, this is devise, uh, uh, the, you do a little bit more training to fuse together those representations. And what you found was that if you give a novel image that is not in any of the categories in the image model, it was trained on the model can often assigns kind of the right cat, the right label to that image. Um, so for example, um, I think, uh, telescope and, uh, binoculars were both in the training, uh, categories for the image model, but, um, microscope was not. Hmm. And so if you’re given an image of a microscope, it actually can come up with something that’s, uh, got the word microscope as the label that it assigns, even though it’s never actually seen an image labeled that.

Shawn Wang [00:59:01]: Oh, that’s nice. That’s kind of cool. Yeah.

Jeff Dean [00:59:04]: Um, so yeah.

Shawn Wang [00:59:07]: Useful. Uh, cool. Uh, I think. There, there’s more general, like broad questions, but like, I guess what, what do you, uh, wish you were asked more in, in, in general, like, you know, like you, you have such a broad scope. We’ve covered the hardware, we’ve covered the, the, the models research. Yeah.

Jeff Dean [00:59:22]: I mean, I think, uh, one thing that’s kind of interesting is, you know, I, I did a undergrad thesis on neural network, uh, training, uh, uh, parallel neural network training, uh, back in 1990 when I got exposed to neural nets and I always felt kind of, they were the right abstraction. Uh, but we just needed way more compute than we had then. Mm-hmm. So like the 32 processors in the department parallel computer, you know, could get you a, a little bit more interesting, uh, model, but not, not enough to solve real problems. And so starting in 2008 or nine, you know, the world started to have enough computing power through Moore’s law and, you know, larger, interesting data sets to train on to actually, you know, start training neural nets that could tackle real problems that people cared about. Yeah. Speech recognition. Vision, and eventually, uh, language. Um, and so, um, when I started working on neural nets at Google in, in late 2011, um, you know, I really just felt like we should scale up the size of neural networks we can train using, you know, large amounts of parallel computation. And so, uh, I actually, uh, revived some ideas from my undergrad thesis where I’d done both model parallel and data parallel, uh, training and I compared them. Uh, I, I called them. I’ve been doing this since I was eight. It was something different. There was like pattern partitioned and, you know, model partitioned or something.

Shawn Wang [01:00:43]: Well, I have to, is it, is it public? And we can go dig it up?

Jeff Dean [01:00:45]: Yeah, it’s on, it’s on the web. Okay, nice. Um, but, uh, you know, I think combining a lot of those techniques and really just trying to push on scaling things up over the last, you know, 15 years has been, you know, really important. And that means, you know, improvements in the hardware. So, you know, pushing on building specialized hardware like TPUs. Uh, it also means, you know, pushing on software, abstraction layers to let people express their ideas to the computer. Thank you for having me.

Jeff Dean [01:01:40]: Thank you for having me.

Shawn Wang [01:07:10]: If that’s something you would agree with at the time, or is there a different post-mortem?

Jeff Dean [01:07:15]: The brain marketplace for compute quotas.

Shawn Wang [01:07:18]: Compute quotas, where basically he was like, okay, David worked at OpenAI as VP Engine and then he worked at Google. He was like, fundamentally, OpenAI was willing to go all in, like, bet the farm on one thing, whereas Google was more democratic. Everyone had a quota. And I was like, okay, if you believe in scaling as an important thing, that’s an important organizational-wide decision to do.

Jeff Dean [01:07:41]: Yeah. Yeah, I mean, I think I would somewhat agree with that. I mean, I think I actually wrote a one-page memo saying we were being stupid by fragmenting our resources. So in particular, at the time, we had efforts within Google Research. And in the brain team in particular, on large language models. We also had efforts on multimodal models in other parts of brain and Google Research. And then Legacy DeepMind had efforts like Chinchilla models and Flamingo models. And so really, we were fragmenting not only our compute across those separate efforts, but also our best people and our best. And so I said, this is just stupid. Why don’t we combine things and have one effort to train an awesome single unified model that is multimodal from the start, that’s good at everything. And that was the origin of the Gemini effort.

Shawn Wang [01:08:52]: And my one-page memo worked, which is good. Did you have the name? Because also for those who don’t know, you named Gemini.

Jeff Dean [01:08:58]: I did. There was another name proposed. And I said, you know what? You know, it’s sort of like these two organizations really are like twins in some sense coming together. So I kind of like that. And then there’s also the NASA interpretation of the early Gemini project being an important thing on your way to the Apollo project. So it seemed like a good name. Twins coming together. Right.

Alessio Fanelli [01:09:27]: Yeah. Nice. I know we’re already running out of time, but I’m curious how you use AI. Today to code. So, I mean, you’re probably one of the most prolific engineers in the history of computer science. Um, I was reading on through the article about you and Sanjay’s friendship and how you work together. And you have one quote about, you need to find someone that you’re going to pair program with who’s compatible with your way of thinking so that the two of you together are a complimentary force. And I was thinking about how you think about coding agents and this, like, how do you shape a coding agents to be compatible with your way of thinking? Like. How would you rate the tools today? Like, where should things go? Yeah.

Jeff Dean [01:10:07]: I mean, first, I think the coding tools are, you know, getting vastly better compared to where they were a year or two, two years ago. So now you can actually rely on them to do more complex things that you as a, as a software engineer want to accomplish. And you can sort of delegate, you know, pretty complex things to these tools. And I think one really nice aspect about the, uh, interaction between, uh, uh, human, uh, software engineer and, uh, uh, coding model that they’re working with is your way of talking to that, uh, coding model actually sort of, uh, dictates how it interacts with you, right? Like you could ask it, please write a bunch of good tests for this. You could ask it, please help me brainstorm. Performance ideas and your way of doing that is going to shape how the model responds, what kinds of problems it tackles, you know, how much do you want the model to go off and do things that are larger and more independent versus interact with it, uh, more to make sure that you’re shaping the right kinds of, of things. And I think it’s not the case that any one style is the right thing for everything, right? Like some kinds of problems you actually want, uh, maybe a more frequent interaction style with a model. And other ones, you’re just like, yeah, please just go write this because I, I know I need this thing. I can specify it well enough, um, and go off and do it and come back when you’re done. And so I do think there’s going to be more of a style of having lots of independent, uh, software agents off doing things on your behalf and figuring out the right sort of human computer interaction model and UI and so on for when should it interrupt you and say, Hey, I need a little more guidance here, or I’ve done this thing. Now what, now what should I do? Um, I think we, we’re not at the end all answer to that question. And as the models get better, that, uh, set of decisions you put into how the interaction should happen may, may change, right? Like if you, if you have a team of 50 interns, how would you manage that if they were people? And I think it’s not, do you want 50 interns? You might, if they’re really good, right?

Shawn Wang [01:12:23]: It’s a lot of management. But it’s a lot of, uh.

Jeff Dean [01:12:25]: Uh, yeah. I mean, I think that is probably within the realm of possibilities that lots of people could have 50 interns. Yeah. And so how would you actually deal with that as a person, right? Like you would probably want them to form small sub teams, so you don’t have to interact with 50 of them. You can interact with five, five of those teams and they’re off doing things on your behalf, but I don’t know exactly what the, how this is going to unfold.

Alessio Fanelli [01:12:52]: Hmm. Yeah. How do you think about bringing people? Like the pair programming is always helpful to like get net new ideas in the distribution, so to speak. It feels as we have more of these coding agents, write the code, it’s hard to bring other people into the problem. So you go to like, you know, you have 50 interns, right? And then you want to go to Noam Shazier be like, Hey, no, I’m, I want to like pair on this thing. But now there’s like this huge amount of work that has been done in parallel that you need to catch him up on. Right. And I’m curious, like if people are going to be in a way more isolated in their teams, where it’s. It’s like, okay, there’s so much context in these 50 interns that it’s just hard for me to like relay everything back to you.

Jeff Dean [01:13:33]: Maybe. I mean, on the other hand, like imagine a classical software organization without any AI assisted tools, right. You would have, you know, 50 people doing stuff and their interaction style is going to be naturally very hierarchical because, um, you know, these 50 people are going to be working on this part of the system and not. Not interact that much with these other people over here. But if you have, you know, five people each managing 50 virtual agents, you know, they might be able to actually have much higher bandwidth communication among the five people, uh, then you would have among five people who are also trying to coordinate, you know, a 50 person software team. Each.

Alessio Fanelli [01:14:15]: So how, how do you, I’m curious how you change your just working rhythm, you know, like you spend more time ahead with people going through SPACs and design. Goals. Like,

Jeff Dean [01:14:26]: um, I mean, I do think it’s interesting that, you know, whenever people were taught how to write software, they were taught that it’s really important to write specifications super clearly, but no one really believed that. Like it was like, yeah, whatever. I don’t need to do that. I’m going to really, I don’t know. I mean, writing the English language specification was never kind of an artifact that was really paid a lot of attention to. I mean, it was important, but it wasn’t sort of the thing. That drove the actual creative process quite as much as if you specify what software you want the agent to write for you, you’d better be pretty darn careful of and how you specify that because that’s going to dictate the quality of the output, right? Like if you, if you don’t cover that it needs to handle this kind of thing, or that this is a super important corner case, or that, you know, you really care about the performance of this part of it, you know, it may, uh, not do what you want. Yeah. And the better you get at interacting with these models. And I think one of the ways people will get better is they will get really good at crisply specifying things rather than leaving things to ambiguity. And that is actually probably not a bad thing. It’s not a bad skill to have, regardless of whether you’re a software engineer or a, you know, trying to do some other kind of, uh, task, you know, being able to crisply specify what it is you want. It’s going to be really important. Yeah.

Shawn Wang [01:15:52]: My, my joke is, um, you know, good. Yeah. I think one thing is in, uh, indistinguishable from sufficiently advanced executive communication, like it’s like writing an internal memo, like weigh your words very carefully and also I think very important to be multimodal, right? I think, uh, one thing that, uh, anti-gravity from, from Google also did was like, just come out the gate to very, very strong multimodal, including videos, and that’s the highest bandwidth communication prompt that you can give to the model, which is fantastic. Yeah.

Alessio Fanelli [01:16:20]: How do you collect things that you often you will have in your mind? So you have this amazing, like performance sense thing that you’ve heard about how to look for performance improvements. And is there a lot more value in like people writing these like generic things down so that they can then put them back as like potential retrieval artifacts for the model? Like, or do I have like the edge cases is like a good example, right? It’s like, if you’re building systems, you already have in your mind, specific edge cases, depending on it. But now you have to like, every time repeat it. Like, are you having people spend a lot more time writing? Are you finding out more generic things to bring back?

Jeff Dean [01:16:56]: Or, um, I mean, I do think well-written guides of, of how to do good software engineering are going to be useful because they can be used as input to models or, you know, read by other developers so that their prompts are, you know, more clear about what the, the underlying software system should, should be doing. Um, you know, I think it may not be that you need to create a custom one. For every situation, if you have general guides and put those into, you know, the context of a coding agent, that, that can be helpful. Like in, you can imagine one for distributed systems, you could say, okay, think about failures of these kinds of things. And these are some techniques you can deal with failures. You know, you can have, uh, you know, Paxos like replication, or, you know, you can, uh, send the request to two places and tolerate failure because you only need one of them to come back. You know, a little. Description of 20 techniques like that in building distributed systems, probably would go a long way to having a coding agent be able to sort of cobble up more reliable and robust distributed systems.

Shawn Wang [01:18:07]: Yeah. Yeah. I wonder when Gemini will be able to build Spanner, right?

Alessio Fanelli [01:18:12]: Probably already has the code inside, you know?

Alessio Fanelli [01:18:16]: Yeah. That, I mean, that’s a good example, right? When you have like, you know, the cap theorem and it’s like, well, this is like truth and you cannot break that. And then you build something that broke it.

Shawn Wang [01:18:26]: Like, I’m curious, like models in a way are like, would he say he broke it? Did you, would you say you broke cap theorem? Really? Yeah. Okay. All right. I mean, under local assumptions. Yeah. Under some, some, yeah. And they’re like, you know, good clocks. Yeah. Yeah.

Alessio Fanelli [01:18:41]: It’s like some, sometimes you don’t have to like always follow what is known to be true. Right. And I, I think models in a way, like if you tell them something, they’re like really buy into that, you know? Um, yeah. So yeah, just more. Thinking than any answer on how to fix it.

Jeff Dean [01:18:57]: Yeah, my, my, uh, you know, it’s just on this, like, like big prompting and, and, uh, iteration, you know, I think that coming back to your latency point, um, I always, I always try to one, one AB test or experiment or benchmark or research I would like is what is the performance difference between, let’s say three dumb fast model calls with human alignment because the human will correct human alignment, being human looks at the first one and produces a new prompt.

Shawn Wang [01:19:23]: For the second one. Correct. Okay. As opposed to like, you spec it out, you know, it’s been a long time writing as a pro a big, big fat prompt, and then you have a very smart model. Do it right. Right. You know, cause, uh, really is, is, uh, our lacks in performance, uh, an issue of like, well, you just haven’t specified well enough. There’s no universe in which I can produce what you want because you just haven’t told me. Right.

Jeff Dean [01:19:44]: It’s underspecified. So I could produce 10 different things and only one of them is the thing you wanted. Yeah.

Shawn Wang [01:19:49]: And the multi-turn taking with a flash model is enough. Yeah.

Jeff Dean [01:19:54]: Yeah, I’m, I’m a big believer in pushing on latency because I think being able to have really low latency interactions with a system you’re using is just much more delightful than something that is, you know, 10 times as slow or 20 times as slow. And I think, you know, in the future we’ll see models that are, and, and underlying software and hardware systems that are 20X lower latency than what we have today, 50X lower latency. And that’s going to be really, really important for systems. That need to do a lot of stuff, uh, between your interactions.

Shawn Wang [01:20:27]: Yeah. Yeah. There, there’s two extremes, right? And then meanwhile, you also have DeepThink, which is all the way on the other side. Right.

Jeff Dean [01:20:33]: But you would use DeepThink all the time if it weren’t for cost and latency, right? If, if you could have that capability in a model because the latency improvement was 20X, uh, in the underlying hardware and system and costs, you know, there’s no reason you wouldn’t want that.

Shawn Wang [01:20:50]: Yeah.

Jeff Dean [01:20:52]: But at the same time, then you’d probably have a model. That is even better. That would take you 20X longer, even on that new hardware. Yeah.

Shawn Wang [01:21:00]: Uh, you know, there, there’s, uh, the Pareto curve keeps climbing. Um, yeah, onward and outward, onward and outward. Yeah. Should we ask him for predictions to, to go? I don’t know if you have any predictions that you, that you like to keep, you know, like, uh, one, one way to do this is you have your tests whenever a new model comes out that you run, uh, what’s something that you’re, you’re not quite happy with yet. That you think we’ll get done soon.

Jeff Dean [01:21:29]: Um, let me make two predictions that are not quite in that vein. Yeah. So I think a personalized model that knows you and knows all your state and is able to retrieve over all state you have access to, that you opt into is going to be incredibly useful compared to a more generic model that doesn’t have access to that. So like, can something attend to everything I’ve ever seen? Yeah. Every email, every photo, every. Yeah. Video I’ve watched, that’s going to be really useful. Uh, I think, uh, more and more specialized hardware is going to enable much lower latency models and much more capable models for affordable prices, uh, than say the current, current status quo. Uh, that’s going to be also quite important. Yeah.

Shawn Wang [01:22:16]: When you say much lower latency, uh, people usually talk in tokens per second. Is that a term that is okay? Okay. Uh, you know, we’re at, let’s say a hundred. Now we can go to a thousand. Is it meaningful to go 10,000? Yes. Really? Okay. Absolutely. Right. Yeah. Because of chain of thought and chain of thought reasoning.

Jeff Dean [01:22:36]: I mean, you could think, you know, uh, many more tokens, you could do many more parallel rollouts. You could generate way more code, uh, and check that the code is cracked with a chain of thought reasoning. So I think, you know, being able to do that at 10,000 tokens per second would be awesome. Yeah.

Shawn Wang [01:22:52]: At 10,000 tokens per second, you are no longer reading code. Yeah. Like you will just generate it. You’ll, I’m not reading it.

Jeff Dean [01:22:58]: Well, remember, it may not, it may not end up with 10,000 tokens of code. Yeah. It may be a thousand tokens of code that with 9,000 tokens of reasoning behind it, which would actually be probably much better code to read. Yeah.

Alessio Fanelli [01:23:11]: Yeah. If I had more time, I would have written a shorter letter. Yeah. Yeah. Yeah. Um, awesome. Jeff, this was amazing. Thanks for taking the time. Thank you.

Jeff Dean [01:23:20]: It’s been fun. Thanks for having me.

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe