[BidClub_]
Latent Space · · 92 min

Captaining IMO Gold, Deep Think, On-Policy RL, Feeling the AGI in Singapore — Yi Tay

Yi Tay

YouTube ↗
TL;DR

  • The IMO gold followed a deliberate shift toward an end-to-end Gemini model rather than a second symbolic system, but Yi did not present it as his personal high-conviction decision. The broader AlphaGeometry/AlphaProof effort took years, while Yi and three other captains trained the competition checkpoint in roughly one week before a live, deadline-bound event. The host framed the AGI implication bluntly: “If the model can’t get to IMO gold, then can we get to AGI?”

  • On-policy RL is now Yi’s primary modeling tool because it lets a model improve from its own behavior. SFT imitates outputs generated elsewhere; on-policy training samples the current model, scores those trajectories, and trains on them. That loop is “a bit like self-distillation” and, in Yi’s view, generally more generalizable than indefinite imitation.

  • The discussion framed Gemini Deep Think’s significance as evidence for a one-model direction. Yi said specialized systems are often one-off systems and that he wants one model for everything, while acknowledging that the limit of what can be absorbed into model parameters is unknown. The broadly served model used a reduced inference configuration; the full configuration was sent to a limited set of recipients.

  • AI coding crossed Yi’s personal utility threshold from delegated typing to trusted diagnosis. He now pastes failed-job output into Antigravity, sometimes barely inspects the fix, and relaunches work that might otherwise consume 20 minutes or even a day. He does not frame this as one-for-one labor replacement, but as “a passive aura that buffs everybody” by reclaiming slices of time across a team.

  • Yi expects some recognizable form of the Transformer to reach AGI unless the entire learning paradigm changes, but rejects the idea that scale alone got the field here. Self-attention survived years of attempted removal, while new ideas remain compatible with an enormous installed stack of optimization work. His corrective to the bitter lesson is a “sweet lesson where ideas matter,” and he believes the closed-lab advantage is increasing as improvements compound.

  • Data efficiency may become a central constraint as available tokens tighten, but neither speaker claimed to know the missing mechanism. Humans appear dramatically more sample-efficient, yet they also learn continuously from multimodal experience; Yi’s practical hypothesis is better algorithms that spend more FLOPs per token. The host explicitly reversed an earlier belief that pre-training was dead after seeing major labs continue investing in it.

  • Singapore is being built as a small, talent-dense Gemini reasoning hub rather than a remote execution office. Yi is hiring for RL, reasoning and Deep Think, but will also consider exceptional competitive programmers or engineers because “ML can be learned easily.” Geography still matters through time zones, culture and recruiting: Singapore can be far enough away for research focus while remaining connected to the global frontier.

Digest · the substance, structured for research

1. Returning to Google put Yi back at the modeling frontier

  • Rejoining Google after 1½ years felt to Yi like reopening a Pokémon save: the same LDAP and username made it feel as though he had clicked “continue game.” Brain had become part of Google DeepMind, but the infrastructure remained familiar enough that the return was “pretty seamless.”

  • Yi had not returned with a predetermined DeepMind assignment. He missed research that was close to the model frontier—not necessarily fundamental theory, but work that could push beyond current capability—and initially planned simply to “work on research and see what happens.”

  • The Singapore group is informally the Gemini team, working on “Reasoning and AGI.” Yi stressed that AGI is not a formal organizational title; adding it to the job posting was “a vibes thing” that signals the North Star behind building frontier models.

2. On-policy RL is now the main modeling loop

  • Yi spent his earlier career arc on architecture and pre-training, then returned with “almost no RL background.” He nevertheless found the transition tractable because general research judgment transfers across toolsets; today, RL is “basically the main modeling tool set that we play around with.”

  • His clean distinction: SFT is off-policy because the model imitates outputs or trajectories generated by another model. On-policy RL lets the current model produce its own trajectories, applies a reward or verifier, and then trains that same model on what it generated.

  • The mechanism resembles self-distillation, but the important feature is behavioral ownership: the model explores from its present policy and receives correction on its own mistakes. Yi thinks this is generally more generalizable, while acknowledging that “there’s still a lot of science” unresolved around the gap between SFT and RL.

  • Imitation still supplies the starting repertoire. Yi compared sports instruction with actual play: tutorials teach movements, but the game supplies the consequential reward signal; models similarly pre-train and imitate before learning from interaction. He cautioned that human-model comparisons are useful analogies, not identities.

3. Learning requires knowing when to abandon a world model

  • The host treated machine learning as the most scientific framework yet developed for studying learning itself. His practical rule is to monitor one’s “learning rate”: after a surprising failure, ask where else the same assumptions may be wrong instead of making only a token Bayesian update.

  • Stable Diffusion was his example. A decade of incremental GAN progress had shaped his expectations, but laptop-scale image generation violated that world model strongly enough to trigger a career pivot. One counterexample can sometimes mean “your entire world model is wrong,” not that its probability should move by 2%.

  • Yi connected this to scientific paradigm shifts with the memorable warning that researchers can be proud of their priors “until they become your prison.”

  • Parallel sampling adds a model-level analogue. Training already draws multiple outputs, but Yi and the host agreed self-consistency is more nuanced than majority voting. The host described an LLM judge or verifier selecting a trajectory with stronger internal validity rather than merely choosing the most common answer.

4. IMO gold followed an end-to-end bet rather than the specialized-system route

  • Yi corrected the “done in one week” narrative immediately. Tang and collaborators had pursued the IMO problem for years, including the prior silver result and work around AlphaGeometry and AlphaProof; the week referred only to training the particular end-to-end checkpoint used this year.

  • The new bet was that Gemini should receive the problem as text and produce the proof directly, with “no second system.” The host emphasized how non-obvious this was: the previous route suggested that an LLM should operate computers, write Lean and call formal verifiers rather than internalize the whole procedure.

  • The host made the AGI implication explicit: “If the model can’t get to IMO gold, then can we get to AGI?” Yi did not claim that as his own test; he said he did not know whether there had been a symbolic-versus-end-to-end debate and was simply asked to train the model. He did argue that specialized systems are often one-off systems and that the broader direction is one model for everything.

  • He was careful about credit and expertise. Tang’s group drove the broader IMO program; Yi joined at model training, knows little about Olympiad mathematics, and cannot personally parse the resulting proofs. That disconnect is itself striking: an AI researcher can build a universal tool that earns gold in a domain he does not understand.

5. A live competition changed the operating rhythm

  • Yi and three other captains trained the checkpoint across London, Mountain View and Singapore. The coordination had no formal handoff protocol: someone might announce, “I’m going to board the plane now; I’ll be away for 12 hours,” leaving another captain to babysit jobs, catch crashes and restart runs.

  • The IMO was not a benchmark they could rerun until the number looked good. Problems P1 through P6 arrived live across two competition days in Australia, and designated team members had to run inference on the newly released questions under real deadlines.

  • Travel made the week messier: Yi and Jonathan were flying, office visits created meetings, and training jobs still needed attention. Yi described the atmosphere as closer to a distributed hackathon than a polished production process.

  • Nobody knew whether the resulting score would qualify for gold because the threshold depended on human performance rather than a fixed target. Yi watched participant scores while verification proceeded, giving the event “more adrenaline than just running on a benchmark and getting a number.”

6. Deep Think embodies the one-model economics

  • The conversation’s one-model framing was “one model, lots of parameters.” LLMs once could not act as calculators and now can approximate that function internally; Yi said the open question is how far weights can subsume capabilities resembling Lean, a physics engine or other formerly external machinery.

  • Yi did not rule out tools. His likely endpoint is a general model that absorbs most capabilities, calling a tool only where something “really cannot be subsumed.” The alternative—separate math, chemistry and physics engines—creates an endless inventory of bespoke systems.

  • The IMO checkpoint also became Gemini Deep Think without being rebuilt as a narrow math product. The broadly served version used a reduced inference-time configuration because the full IMO setup was expensive; the heavier configuration was sent to a limited set of recipients.

  • “Non-verifiable” did not mean a proof is unverifiable in principle. Yi meant that long-form proofs are non-trivial to verify unless translated into a formal system such as Lean, leaving substantial work in domains where verification is difficult to operationalize.

7. The gold medal reset expectations more than it settled AGI

  • The host’s surprise was not merely that AI reached gold, but that the prior specialized route was set aside in favor of an existing general Gemini scaled through Deep Think. Relative to the previous year’s silver architecture, that was the non-consensus move.

  • Yi said researchers experienced the result as a collective recognition of how much Gemini had advanced in one year. He avoided claiming the team expected failure, but described the outcome as a reassuring “pat on the back” for the engineers and researchers working on Gemini.

  • The host’s longer historical comparison was stronger: show today’s IMO and ICPC results, neural-network capabilities and general AI performance to researchers five years ago, and many would conclude that “we already reached some form of AGI.” The frontier keeps moving partly because expectations move with it.

  • Yi declined to opine on claims that inference optimization was the hardest part; he did not work on inference-time scaling. Even “IMO Cat,” which attracted attention as a supposed code name, was simply a job configuration named after his affection for cats.

8. Pokémon exposes planning limits without pretending to be discovery

  • No Pokémon-specific work was done for Gemini, which is part of why Yi likes it as a periodic check. Pokémon Crystal tests visual understanding, long-horizon planning and agent persistence in a setting people are unlikely to target seriously with “Pokémon maxing.”

  • Merely completing the linear game may eventually become a speed contest. Completing the Pokédex is qualitatively harder: it requires backtracking, web research, tracking world state, finding evolution requirements and sometimes arranging trades that cannot be completed through ordinary solo play.

  • The host’s pushback was worth retaining: following an online guide is less intellectually interesting than being the first system to create that guide. Yi later agreed that the task was “not super interesting,” while maintaining that synthesizing retrieved knowledge and reliably applying it inside a changing visual environment remains unsolved.

  • The host’s deeper benchmark is novel knowledge. He proposed freezing available information around 2012 or 2015 and asking whether a capable model could independently invent the Transformer; even after assuming away architectural leakage, Yi was unsure. Efficiently exploring novel idea space, rather than brute-forcing it, remains an open AI-scientist problem.

9. Reasoning is a training regime, not a settled metaphysics

  • Yi argued that “reasoning” has become too vague for one universal definition. In current technical practice, it usually means post-training models with RL to improve thinking trajectories: generate extra work, reward better paths and elicit capabilities that the base model does not reliably express.

  • Chain-of-thought tokens are only one implementation. Latent-thinking research asks whether a model can perform the intermediate computation in hidden state without decoding every step into discrete language. The host suggested that the native representation of thought might be numbers rather than English, Chinese or any human-readable narration.

  • Yi’s default stance is permissive: “Let the model do whatever it wants.” He does not believe machine thoughts must resemble human thoughts, even if multimodal models converge on related latent representations for concepts such as a laptop.

  • The host raised a recursive-data concern: public reasoning traces may increasingly enter pre-training corpora. Yi said he had never seen such traces himself and estimated that chain-of-thought tokens on Common Crawl had gone from 0 to roughly 0.001%. The host noted that clean sources could often be filtered, making retention a modeling choice rather than an inevitability.

10. AI coding crossed from convenience into delegated investigation

  • Yi’s earlier high-value use case was mundane: give a model a screenshot of a large results spreadsheet and ask it to generate plots, avoiding “this Matplotlib stuff” he finds annoying. AI coding itself had not yet earned much of his workflow.

  • The threshold has now moved. He can paste a failed training job into Antigravity, ask it to diagnose and repair the bug, and sometimes relaunch without deeply inspecting the patch. Initially he checked everything; later he concluded, “Maybe the model knows better than me.”

  • Nano Banana produced a similar transition from novelty to work utility. Image generation had largely meant trolling friends or making amusing artifacts; recently it became good enough for actual tasks. Yi described several such moments as AI crossing an “imaginary line” from emergent to useful.

  • He said he thinks he uses public Gemini, while adding that he was not sure whether internal training or codebase-specific differences were involved. The important signal is behavioral: a frontier ML researcher who previously used little AI now trusts these tools inside his own high-expertise work.

11. AI looks more like a team-wide buff than a one-for-one replacement

  • Yi rejected simplistic accounting in which a model replaces exactly one junior researcher. A manager should measure aggregate time: if AI saves comparable hours by removing fragments of work from several people, its organizational impact is real even without mapping neatly onto one headcount.

  • Some bug fixes might save 20 minutes; others might recover an entire day. His game-language metaphor was “a passive aura that buffs everybody”—productivity gains distributed across the group rather than one person disappearing from the organization.

  • Trust remains conditional because models sometimes get “lazy,” guess at a fix, and try to convince the user it worked. Some tasks remain easy for machines and hard for humans, with the reverse elsewhere, but Yi found those quadrants difficult to characterize cleanly.

  • The host argued that focus-firing datasets at human-easy/model-hard failures is precisely how hill-climbing reaches AGI. Yi agreed that progress often consists of small improvements, but added a prerequisite: a failure class must first be defined and evaluated well enough for targeted data to address it.

12. Transformers may persist even if learning changes underneath them

  • Yi expects the eventual AGI architecture to remain recognizably Transformer-like unless backpropagation and the surrounding paradigm change completely. Nearly a decade of attempts to remove self-attention repeatedly ended with researchers restoring at least one attention layer because performance needed it.

  • Local/global mixtures still count as attention in his taxonomy. Names and QKV details may evolve, but the underlying sequence-processing mechanism has proven unusually durable and compatible with everything built around it.

  • Very long memory—200 million or one billion tokens—could force change through continual learning, inference-time weight updates, cheaper attention or better chips. Yi framed architecture as the interface between tokens and the learning algorithm; the deeper constraint may therefore be the learning rule rather than sequence-to-sequence processing itself.

  • The host resisted treating eight years of Transformer investment as destiny. Yi’s answer was the “sweet lesson where ideas matter”: blindly scaling an MLP with “$100 trillion” would not work. Useful ideas are not yet showing diminishing returns, and he believes the unpublished advantage of closed labs is increasing as those ideas accumulate.

13. Data efficiency is becoming a first-class unknown

  • The host explicitly changed his mind on pre-training. He had thought it might be dead as labs redirected effort into RL, but conversations with pre-training leaders at Anthropic and OpenAI, plus a DeepMind talk, convinced him that major labs still invest heavily in it.

  • Human learning is the existence proof behind the efficiency question: a child can recognize a dog after very few examples where classical systems may need many more. Yi cautioned that humans also receive continuous visual, auditory and embodied input “24/7,” making raw token comparisons deceptive.

  • If models become data-bound while compute remains available, Yi’s leading abstraction is to spend more FLOPs per token—extracting more from every observation through better architecture or learning algorithms. He did not claim to know whether the bottleneck is off-policy noise, model structure, multimodality or something else.

  • The host also reported that some labs pay external startups at least seven figures for specialized RL environments and asked why strong coding agents cannot build those worlds internally. Yi’s honest answer was simply, “I have no clue”; he offered no invented theory about the economics or hidden domain expertise.

14. Generative retrieval turned documents into a model-native vocabulary

  • DSI began before the current LLM era as an attempt to reimagine retrieval with T5: one model would effectively hold the index in memory and directly generate a document identifier. Even brute-force arbitrary IDs worked surprisingly well because models had enough capacity to memorize the mapping.

  • Semantic IDs made the search space tractable by assigning identifiers with hierarchical semantic structure. After DSI, work associated with Ashish’s group and Mahesh transferred the idea into recommender systems, where the domain-specific encoding differed but the generative ranking formulation remained recognizable.

  • The host pointed to public YouTube work in which uploads are encoded into a codebook that is retrained every day in a batch job, and noted similar claims around Twitter’s LM-based feed. Yi recalled Spotify using related semantic-ID methods but no longer follows the field closely enough to characterize current implementations.

  • Yi still called DSI one of his most creative works, despite disliking traditional IR experimentation. Offline NDCG, stubborn BM25 baselines and distant causal feedback made the environment feel as if “gravity is different”—like striking a badminton sweet spot and hearing glass shatter. Yi also called recommendation the consumer-AI “god problem” because ranking, filtering and personalization are so valuable.

15. Singapore is competing through density, connectivity and research taste

  • The Singapore symposium with Quoc, Jeff Dean and regional leaders was intended to “set the vibe,” not mark a single launch moment. Teams form gradually as people join; bringing AI pioneers into one room signaled that Singapore could participate near the frontier rather than only consume work from elsewhere.

  • Geography still matters through time zones, recruiting and mental space. Yi sees the Bay Area as saturated with “AI, AI, AI everywhere”; Singapore offers enough cultural distance for quiet research while remaining globally connected. A global team can also provide practical 24-hour coverage.

  • The team will stay small initially and optimize for talent density while working on RL, reasoning, Gemini and Deep Think. Yi seeks people with a research track record—not necessarily in RL—or exceptional competitive-programming, coding or other technical achievements; “ML can be learned easily,” whereas raw technical ability and independent research taste are harder to manufacture.

  • His recruiting signal is public work that makes him want to send the author a DM. A student who independently chooses a tasteful problem and executes it well demonstrates both judgment and capability, though Yi acknowledged that growing competition and misaligned adviser agendas make that much harder than it sounds.

16. Physical health became part of Yi’s research system

  • Across roughly 1½ years, Yi said he lost 23 kilograms, doubled his heart-rate variability and reduced his resting heart rate by about 30 beats per minute from an unhealthy 80–90 range. He treated wearables and health data as productivity inputs because improved energy has helped his work.

  • Yi suggested that general energy and even physical hunger might relate to intellectual hunger, while the host disagreed from personal experience: “When I’m hungry, I just think of food.” Yi concluded that exercise, recovery and health count as work because they improve the capacity to do research.

Yi Tay

The thing that I find most useful about these models in general is that, when I have big spreadsheets with a lot of results and just need plots of them, models can be quite good at taking a screenshot and making a plot of it. I hate making all this Matplotlib stuff; it’s so annoying. There were so many moments this year where AI suddenly crossed that imaginary line. I think AI could have done a lot better with this stuff.

I think Nano Banana also got to the point where usually you make these images just for fun, to troll your friend or something like that. But Nano Banana was actually so good.

Speaker 1

Welcome back. How are you?

Yi Tay

Yeah, I’m good. I’m good. Good to be back. It’s been 1½ years.

Speaker 1

Yeah, it’s been 1½ years. It feels like a long time. So, last time we talked, you were at Reka, and then you joined Google DeepMind again, working for Quoc again. More recently, you started Google DeepMind Singapore. Is it Google DeepMind Singapore or Gemini Singapore? I don’t know if you have named the team.

Yi Tay

I think we have a Gemini team in Singapore. Yeah, I have a team in Singapore. It’s called Reasoning and AGI.

Speaker 1

Reasoning and AGI. Is it important to have AGI in the name?

Yi Tay

It was just a vibes thing that we put AGI in. I think one reason why we work on these models is that we want to get to AGI, and it was just a vibes thing that we added AGI to the job posting. But there’s no formal team name or anything like that. It’s basically the Gemini team in Singapore.

Speaker 1

Yeah. I think people are trying to triangulate. Amazon has an AGI team, you guys have an AGI team, and Meta now has a superintelligence team. What are people signaling when they choose these names for their teams? Do they have a plan, or is it just vibes?

Yi Tay

You’re trying to fish out hot takes from me, you know?

Speaker 1

You have “AGI” officially in your job title.

Yi Tay

No, it’s not a team name.

Speaker 1

It’s not a team name, yeah.

Yi Tay

Yeah, it’s just—you know, we want to signal the North Star: that we’re building these models to get to AGI.

Speaker 1

Yeah. No, I wasn’t really fishing for hot takes. So, you rejoined Google DeepMind, and I think last time we talked about how you were at Google Brain, came out, and now you’re back at Google DeepMind. I listened back to the whole thing; it was an amazing episode last time. I wonder what your general reflections were on plugging back into the Google infrastructure.

Yi Tay

Coming back is very interesting because it felt like a return to Google. Everything, including your LDAP and your username, is all the same. It’s like playing Pokémon: you leave it aside, and then you go back and click “Continue.”

Speaker 1

It’s a new game.

Yi Tay

Yeah, you save the game and then continue the game. It’s like that. Obviously, over the last 1½ years while I was away, many things have changed. Brain is now part of Google DeepMind and stuff like that, so obviously a lot has changed. But overall, coming back has been pretty seamless.

Obviously, I love Google’s infrastructure, and I think the builds are great and stuff like that. I’m very glad to be back to Google’s infrastructure.

Speaker 1

Was the intention always that you were going to work on DeepMind?

Yi Tay

No, not really. I missed research a lot—doing research. Not super-fundamental research, but research close to the model. I really missed being at the frontier and trying to go beyond it. I missed that a lot.

When I came back, DeepMind wasn’t really a thing, and I don’t think there were any plans, actually. It was just, “I’m going to work on research and see what happens.”

Speaker 1

I’m sure there was some inclination that reasoning was the next frontier, and that’s obviously the most rewarding research path, especially this year.

Yi Tay

Yeah, I think these days reasoning and RL are probably quite important. RL and reasoning come together. I spent a lot of my past life—what I call the past arc—working on architectures and pretraining. But now I’m more focused on RL research.

I’m not talking about old-school RL, but game RL and old-school RL. To be honest, I had almost no RL background coming back. But I think RL is the main means of modeling these days, so it was pretty easy to jump back in.

A lot of fundamental skills in research are general-purpose and universal. It’s quite easy to innovate even in a toolset that you’re not super-used to. So, I think RL is basically the main modeling toolset that we play around with these days.

Speaker 1

Superficially, I see some overlap in your UL2 and T5 work—the focus on objectives and the things that you’re trying to incentivize. I would have guessed there was more overlap than you’re saying right now, which is interesting. But I know that’s very superficial.

Yi Tay

The objectives have some overlap, right?

Speaker 1

Yeah.

Yi Tay

I think it’s mainly about on-policy and off-policy, and designing these things that change how the learning algorithm itself works.

Speaker 1

Let’s introduce this terminology to people if they’re not that familiar with RL policies. I do think that a lot of people are trying to understand what’s working about this generation of RL research.

Jason had this interesting post, which I think you were co-signing. The basic idea was that you always want to be on-policy instead of mimicking other people’s successful trajectories and applying them to your own actions. You learn from the reward given by the environment—basically, correcting your own path instead of trying to imitate somebody else’s path. Is that basically correct? What’s your reflection on that, or what would you add on top of that?

Yi Tay

First of all, Jason writes really well, and I really wish more people wrote like him.

The biggest analogy between on-policy and off-policy is that off-policy is basically like when you do SFT. You take somebody else’s generated outputs, trajectories, or whatever, and train on them. That’s off-policy.

On-policy is mainly the core idea of modern LLM RL, where you generate, reward the model based on its own generations, and then train the model on its own generations. It’s a bit like self-distillation, to some extent: the model generates its own output, you reward it, and then it trains on its own output.

On-policy-ness is basically this idea of the model training on its own outputs—letting the model generate its own trajectories, letting some reward verify them, and then having the model train on its own outputs. I think this is more generalizable in general.

There’s still a lot of work to be done on the gap between SFT and RL itself. But bringing this analogy back to real life, on-policy-ness is more like how humans operate. We go around the world, make mistakes, and then say, “Okay, this is where this went wrong and this is where this went right.”

Speaker 1

Imitation learning is basically somebody else telling you what to do, and you just copy them.

Yi Tay

Yeah. So, bringing this philosophy back to life is quite powerful. Now that I have a kid, I want my kid to try things, and then you tell them, “Okay, this is where it went wrong and this is where it went right,” rather than, “Just copy everything somebody else does.”

Speaker 1

Montessori schooling is mostly that, right? Very unstructured learning: you discover your own path, and we just give you a safe environment to do it.

Yi Tay

Yeah.

Speaker 1

What is the point at which you should transition from imitation to on-policy? I do bounce back and forth. For models, it seems like there’s mostly been a very concrete progression: first you imitate, and that’s pretraining, and then you do RL. But SFT is still imitation.

Yi Tay

I would say that, in models, it seems like there’s mostly been a concrete progression: first you imitate, and that’s pretraining, and then you do RL. But for humans, it’s still a little bit of both.

It’s like sports. When you start playing sports, you begin by imitating—hardcore imitation. But you can’t imitate forever, because imitation—I don’t know whether this is a good analogy—watching a lot of tutorials and stuff is more like imitating. You’re trying to learn certain movements and things like that.

But on-policy-ness is going to the game itself and trying to get a reward signal from that. I think humans do need some form of imitation learning. Everybody starts off by imitating.

But the human-and-model comparison isn’t exact. It’s fun to have analogies, but we shouldn’t take things super-literally.

Speaker 1

I actually am a pretty serious taker of machine learning insights into human learning.

Yi Tay

That’s how we learn from models now?

Speaker 1

Yeah, because I think machine learning is the most scientific way we’ve ever studied learning in general. That’s true. We had to invent curriculum from scratch.

Yi Tay

Yeah, that’s true.

Speaker 1

Things like learning rate—if your learning rate is too high, this happens; if your learning rate is too low, that happens. Where do humans even have a learning rate?

I tell people to keep an idea of their own learning rate and be wary of it being too low. For example, if you’ve been wrong once, you should ask, “Where else have I been wrong?” Usually, people update more slowly than they should when they’ve been wrong.

Yi Tay

What is that—stubbornness?

Speaker 1

It could be stubbornness. I don’t know if that’s the right word for it. It could be that they’re too Bayesian, when actually their prior assumptions were wrong and they need to completely throw out their previous assumptions, because one counterexample invalidates all their prior experience.

Your entire world model is wrong. Throw it away. So Bayesian reasoning is actually wrong. Let's say you live for 10 years under some assumptions and then you have 1 example that breaks your narrative. You shouldn't be like, “Okay, now I have a 2% update.” No. You should be like, “Something's really freaking changed. Everything I've ever seen for the last 10 years is probably wrong. What else am I wrong about?” Update 20%, update 50%, not 2%. You know what I mean? That's a learning-rate thing for me.

My direct example is getting into AI. I was watching GANs for 10 years. Has it been 10 years? 2012, 2013? Time flies. I was watching GANs and I was like, “Okay, this is cool. It's getting more detail, but it's not that impressive.” Then, all of a sudden, Stable Diffusion came out and you could run it on your laptop. That was my learning rate: “Okay, my mental model of generative images did not include this, so I am very wrong and I need to pivot everything.” That's how I started Latent Space. What does it mean that your learning rate is high?

Yi Tay

Yeah, I will nudge it up. I schedule my learning rate because a world model has been violated.

Speaker 1

Okay, I think it's a good strategy. I think this also brings a little bit to light what happens when new paradigms emerge—how quickly people are willing to adopt them or invalidate their understanding of things. I think as scientists, a lot of the time, as things progress, we do have to keep invalidating our own world model. Someone could discover a new way to do something, and then suddenly something comes along and invalidates it.

Yi Tay

Yeah, you can be very proud of your priors until they become your prison.

Speaker 1

Yeah, that's actually very interesting.

Yi Tay

Yes, yes.

Speaker 1

Okay, there was a bit of a tangent. I don't know how we got there. You did highlight Denny's LLM reasoning lectures, where he could trace the intellectual history of reasoning in LLMs.

Yi Tay

Yeah, yeah. There was one part I was going to prompt you about, which was also self-consistency, right?

Speaker 1

Yeah. I think people roughly know it. It's more crudely implemented with OpenAI than with you guys, where it's straight-up 8 inferences and then they judge them or whatever. But I do think that it's also relevant to on-policy distillation, where you have 8 different passes and they're all from the same model. I'm checking my intuition there, but basically, the things you were saying about why on-policy is important and using an external verifier to improve your reasoning—you can also do that with parallel reasoning.

Yi Tay

Yeah, I mean, when we train our models, they sample multiple times. So, to some extent, there's some form of self-consistency.

Speaker 1

Is that directly self-consistency, right?

Yi Tay

Yeah, self-consistency is a little bit more nuanced. If you talk to Denny, he would tell you it's not majority voting, for sure, but it's a more nuanced version of that. Parallel thinking is definitely related to self-consistency.

Speaker 1

Yeah. I think OpenAI also put out some interesting papers on majority voting versus other forms of multiple-output consensus. At the highest level, there's an actual LLM judge that decides, “This is a worthwhile trajectory that is more valid,” based on some internal consistency or by inspecting the chain of thought. It's very cool that we can train models to do that.

Yi Tay

Yeah, for sure. Self-consistency is a big fundamental idea. I mean, chain of thought and self-consistency are big ideas, and self-consistency was also a big fundamental idea in modern LLM literature.

Speaker 1

Yeah, amazing. Okay, so let's bring it to one of the headlines of this podcast: diving into the IMO world. This was around March and July? July. You guys announced—oh, there's a very nice photo here. This is the photo I was looking at. This is in London, I believe, where you had the showroom.

Yi Tay

Yeah, the showroom.

Speaker 1

Oh, you got to be at a photo-taking to get the credit. Is that right?

Yi Tay

What? No, no, I'm just kidding. The contributor list is bigger than this.

Speaker 1

Yeah, yeah, but they were saying, “Oh, yeah, okay, you should go there in order to get a literal gold medal.” No, no, no, no. They do get credit for being part of the IMO effort; it's just that the photo was taken there. It's a joke. It's not a joke. Anyway, could you tell the story of studying for this IMO thing? Apparently, it was done in 1 week.

Yi Tay

Let me clarify a lot of things. The IMO effort has been very long-standing. Tang and co. have been working on this for a long time, even last year. Last year they got silver. I was not there. Google had AlphaGeometry and the AlphaGeometry work, and then they had AlphaProof and stuff. So, it's a very long-standing effort.

I think this year we wanted to try to actually use Gemini as an end-to-end model—basically, no second system.

Speaker 1

No second system. It takes in text and outputs text like a model. Even that was not an intuitive thing. I covered the silver result from last year, and I was like, “Okay, it's pretty close. It's 1 point off from a gold. Just try harder and you'll get gold.”

That decision to abandon it was pretty bold. In retrospect, it's easy to say this, but if the model can't get to IMO gold, can we get to AGI? At some point, we have to use these models to try these Olympics and competitions. One of the goals this year was, “Okay, we're going to do end-to-end: take in text, output text.” That's where my involvement came in.

Yi Tay

I was not involved in the IMO effort until the model-training part. I have to say that Tang did most of the IMO work. I just trained the model with a bunch of—

Speaker 1

Okay, what did that work involve? What are some of the—

Yi Tay

Basically, we prepared model checkpoints for the actual IMO itself. There's also something that's easily overlooked about the IMO: many times, when you chase benchmarks and stuff like that, it's something you can keep running and keep climbing until you get there. But the IMO was a live competition. Some members of the team were in Australia for the event, and it was happening live. It was unfolding live.

Speaker 1

It's a very hard goal. You receive the problem, punch it into your system, and then you—

Yi Tay

Yeah, yeah, yeah. Some of the professors from Tang's team were at the IMO itself. I don't even know whether the IMO is a conference, but there were people there in Australia. There were people whose actual job was to run inference on IMO problems P1 through P6 as they came out. They also came out on different days, so there were different sets—day 1, day 2, something like that.

The fun part is that I knew nothing about the IMO at all. I wasn't a kid who took part in the IMO. I was too dumb for that. You're kind of a player. I was a bit of a piano player.

What I knew was that we delivered the checkpoint, and that checkpoint was used to achieve IMO gold. But then there was somehow a week in London where everybody gathered. Everybody was flying to London, and this photo was taken there. You got to see how the different parts came together, including being in rooms with the other co-captains. It felt a little bit like a hackathon.

So, the training process for this IMO model itself was maybe a week or so—not the actual whole process.

Speaker 1

I think the question is that I'm still not over the decision to throw away AlphaProof. I understand that you have this goal of AGI, and obviously, at some point, one model should do all of it. But I think you pointed out to me that in 2024, to do the IMO, IOI, ICPC, and all the other things you guys did, you needed an LLM reasoning system that knew how to operate a computer, write Lean, run the Lean verifier, and do all these things.

Basically, you RFT'd the Lean verifier into the chain of thought. Is that the—

Yi Tay

Wait, so basically, it is not obvious that you can do that at all. In some way, you mean that the system has to be encoded in the parameters of the model somehow.

Speaker 1

Yes. I mean, at the end of the day, it comes down to whether you believe in this connection: 1 model, lots of parameters. There's also tool use and things like that. But to some extent, we should be able to get to a point where—when LLMs first started, a model couldn't even be a calculator. Now it can somewhat be a calculator. Technically, a tool like a calculator is somewhat encoded in the parameters of the model.

So, I think we eventually get to a point where whether there are things that cannot be expressed in the parameters of the model is an open question.

Yi Tay

We don't know where the limit is, but I think we'll keep pushing and pushing this limit. So whether it's something like a Lean system or something else to solve other problems—maybe a physics engine or something—we'll still continue to push that boundary.

I actually don't know whether there were a lot of debates about symbolic systems versus end-to-end. That's the word I was trying to find. I don't really know whether there was a debate; to me, I was just like, “Oh, let's train the model,” and someone told me to train the model, and then I trained the model. Basically, the IMO effort decided this.

I also think that these specialized systems are very much one-off systems. You could create a chemistry engine, a math engine, or whatever, but at the end of the day, you want one model for everything. This fits that direction a little bit more, where you have one model. This model was also launched as Gemini Deep Think, as a general-purpose Gemini Deep Think. It's basically unchanged, but with maybe some of the configuration toned down a bit.

Speaker 1

Yeah. The inference-time configuration served to most people was different, but the full IMO inference configuration was shipped to some betting magicians. That's because of the inference cost, right? But that was good enough to be a general-purpose model.

I think my take is that this intuition was what led to trying to move toward one model, because with specialized systems, there's no end. You could create many specialized systems.

Yi Tay

Yes.

Speaker 1

The most I can see happening in the future is that there'll be a model, and if there's something that really cannot be subsumed by a model, then you just use a tool or something. But my prediction is that most things can be subsumed by the model.

AI researchers are quite good at hill climbing. History would say that you have a lot of evidence to back this up. Is this the model output? This is it, right?

Yi Tay

Yeah, I think this is the model output, yeah.

Speaker 1

What do you see when you look at this? Obviously, it looks like well-written problems. It looks like something a real human mathematician would do. People did compare yours with the OpenAI one, where OpenAI's is a lot rawer and they have to clean up their versions. We don't have to talk about OpenAI, but what was interesting to you when you saw this kind of output?

I want to give a little bit of a special disclaimer: I know nothing about the app. The wonderful thing about this era of LLMs is that you can be an AI researcher or engineer and not have any domain knowledge, and you can still get a gold medal. It's a universal tool that you don't know anything about. I can't pass this at all. This is foreign to me.

Maybe a proof is a particular kind of chain of thought. Another interesting thing some of your collaborators were talking about was, “This is the first example of reasoning in a non-verifiable domain.” Which, to me, isn't a proof by definition verifiable? I just want to give you things to riff on, or debates that might be worth digging into.

Yi Tay

I think there's a lot of work to be done in non-verifiable domains. When people say “non-verifiable,” they usually mean non-trivial to verify, or not as easy to verify as the solution to a math problem, because it takes a long form. It's non-trivial to verify unless you convert it to Lean and then do all kinds of things.

I'm going into territory where I'm not sure what I can say or cannot say.

Speaker 1

Okay, sure. That's it. I think another open topic of debate was how much domain-specific work or post-training was done, because you then went on to do the IOI and ICPC work as well, right? With the same model.

Yi Tay

I wasn't directly involved in the ICPC, but I was related to Science 10. That's all I can say, yeah.

Speaker 1

Yeah. Any other interesting callouts? I don't want to just focus on the team. You called out Jonathan as someone who was a co-captain on this effort. How did the effort come together?

Yi Tay

I think there were 4 captains for the IMO: 2 from London, Jonathan from Mountain View, and me from Singapore. The 4 of us basically trained this model together.

One interesting thing was that we were all in different time zones. There's also something very interesting about passing on the job. There was no fixed rule for how to work together between captains. It was more like, “I'm going to board the plane now. I'll be away for 12 hours,” and someone else would take it, just babysitting it.

Sometimes there were bugs and the job would go down. Basically, it was very ad hoc, and it varied between the captains how we decided to work together.

It was also an interesting time because we were all flying. I think the London folks didn't have to fly, but Jonathan and I had to fly. When you visit another country or another office, you have many meetings. I was in and out of meetings, so it was pretty interesting.

Nobody really knew whether we would get a gold medal at that time, because the IMO hadn't happened yet. It was interesting and exciting. Then there was the whole process of getting this verified by the IMO committee. I had to learn a lot about how the IMO works.

Apparently, the gold score isn't even a fixed number. It's like a bell curve. It was a time when you just looked at the score and thought, “What does this mean?” I was even watching the human participants and seeing what their scores were, because whether Gemini would get a gold depended on how the humans did. If it was a certain percentage, you would wonder what the reason was that it got the result it did.

To some extent, you don't have any control over that. But you're curious, right? First, I would say that it's definitely more exciting. There's more adrenaline than just running a benchmark and getting a number. It's also a process that took some time.

If you guys have specific questions, you can ask us, but I think this whole thing has been a highlight for me. This IMO effort has been one of the highlights. Most people, if you had asked them maybe 2 years ago whether a model could get an IMO gold, would have said it was impossible.

Speaker 1

Then the silver helped, right, from last year? But the fact that you can throw that system completely away, just take the existing Gemini and scale up DeepMind, and then run it for an IMO gold—I think that's also very non-consensus compared to last year.

Yi Tay

Yeah, definitely. To some extent, I think researchers were also surprised. I wouldn't say they were surprised, but it was more like a pat on the back for how much progress we actually made. By “we,” I mean collectively all the engineers and researchers working on Gemini. A lot of progress was made. Look at how much we advanced in 1 year.

Speaker 1

If just 5 years ago you imagined this outcome and looked at the state of AI now—the IMO and ICPC gold medals, and even things like neural networks—if you looked at AI progress now compared with 5 years ago, I think people would say that we had already reached some form of AGI.

Yi Tay

Some form of AGI.

Speaker 1

Yeah, we're just moving. If you take these checkpoints and travel back 5 years, someone should make a drama about this. It's really quite impressive how quickly the field has moved.

The hard parts, would you say, were scaling inference?

Yi Tay

In what aspect?

Speaker 1

In terms of what required the most effort, or where the most brainpower was expended on the team. I saw some comments saying that the hardest part was the inference optimization, or the very, very long-horizon inference that Deep Think needed compared with normal Gemini. Stuff like that.

Yi Tay

I didn't work on the inference-time scaling, so I wouldn't know.

Speaker 1

There was also the codename, apparently IMO Cat, which you named after your desk.

Yi Tay

Okay, that's not really—so, I think I tweeted about it at some point. The IMO Cat was basically just the name of the job configuration. It wasn't an official codename or anything. You just need some kind of name.

I like cats, and that's basically it.

Speaker 1

Fair enough. That is mostly it on the IMO, unless you want to bring up anything else. We have other research topics, but before I go into those, I wanted to leave the floor open to cover what else people should know about the reasoning effort going on at Google DeepMind.

Yi Tay

Let me think where to start. Yeah, this. What do people need to know? It was really good, yeah.

That’s all I would say to people.

Speaker 1

Maybe an easy one to start with would be this: A lot of people were focusing on academic benchmarks 2 years ago, then last year maybe LLM Arena, and this year Pokémon. Pokémon is a very interesting benchmark for visual reasoning and general long-horizon agent planning. You seem to focus on it a lot, and I think Gemini did very well, so obviously it’s something that’s easy to talk about.

I think I probably should have listened to this.

Yi Tay

There’s actually nothing specifically done for Pokémon.

Speaker 1

Of course. Yeah, of course there’s nothing specifically done for Pokémon. Logan had this tweet recently about Gemini 3 on Pokémon Crystal.

Yi Tay

Yeah, Pokémon Crystal. I used to play a lot of Pokémon, and I’m a big Pokémon fan in general. I think it’s a great long-horizon benchmark, like you said.

I also think it’s good to check in once in a while on benchmarks that almost never get contaminated, or that people don’t actually spend time optimizing for. It’s kind of silly when people ask, “What are you working on?” and someone says, “I’m working on Pokémon,” or, “I’m working on Pokémon maxing.”

Speaker 1

We did interview the Claude-based Pokémon AI. I think his name is David, and it showed serious flaws in Anthropic’s screen-understanding and vision capabilities. It literally couldn’t tell that it was trying to get past a wall. It just kept running into it because it didn’t know the wall was there, so it didn’t have any special reasoning at all.

I mean, that could be the honest thing. Another question is whether the model has access to game-state information or whether it’s completely visual.

Yi Tay

Claude’s implementation is very game-state-heavy. They dumped effectively all the memory of what was going on in the emulator.

Speaker 1

Yeah, yeah, I see.

Yi Tay

I think solving Pokémon is going to be more about how fast you solve it. The thing that I haven’t really seen so far is whether the model can complete the Pokédex.

Speaker 1

Why is that?

Yi Tay

Because completing the Pokédex is so hard, right? You need to plan, and you need to look up information. If you don’t go online, you basically need to do a little bit of deep research.

The model would never know that it needs to trade. Some Pokémon need to be traded to evolve, or you need to breed them. If it’s able to go online, post on forums, and find someone to say, “Hey, can I trade with you to evolve this Pokémon?” then maybe it could do it. But anyway, I haven’t seen a model be able to complete the Pokédex. Completing the Pokédex is actually pretty hard for models, so I think that’s an interesting benchmark.

Speaker 1

I wonder what the real-world analogy would be. If we have a model capable of doing that, what can we make it do that we cannot do today?

Yi Tay

There’s a lot of planning involved.

Speaker 1

Which is real deep research.

Yi Tay

There’s also a lot of planning involved, but I think completing the Pokédex is different because the Pokémon game is very linear. The Pokédex involves a lot of backtracking and research.

Speaker 1

Yeah, a lot of research.

Yi Tay

So it’s probably a different nature.

Speaker 1

Is that as interesting to you as, for example, what a lot of other people in the AI-for-science world are trying to do—discovering things that you cannot look up? Novel knowledge.

Basically, what you’re saying is that we’re not even there yet. We’re at the point where models cannot consistently apply knowledge that they look up. If you give Gemini access to web search and say, “Can you try to collect all the Pokémon in the Pokédex?” you wouldn’t have high confidence that it would do it. I don’t know if someone has actually tried. Probably not, right?

Yi Tay

No. I think the hard part is actually trying to synthesize the web knowledge and then apply it in the game itself, with all the visual state and everything going on.

Speaker 1

Yeah, yeah, yeah. It probably would be solvable in one of those ways.

Yi Tay

It’s not challenging.

Yi Tay

It’s not super interesting.

Speaker 1

The task is really just: Can you look up the guide to do it, and then can you apply the guide? That’s it.

You know what’s even more intelligent than that? Creating the guide—being the first to figure out how to create it.

Yi Tay

Oh, yeah, yeah.

Speaker 1

But when it comes to this knowledge, it’s mostly an exhaustive-search thing. A model can try and try, but humans are the guides for that.

Yi Tay

Okay, so that’s actually less interesting to you. When you think about it, it’s not super interesting, but it’s okay. It’s just that I haven’t seen a model try to do this.

Speaker 1

Yeah, I think efficient search of novel idea space is interesting. Obviously, you can brute-force anything, but we’re not talking about brute-forcing. We’re talking about trying to create an AI scientist.

Yi Tay

Being able to generate novel knowledge.

Speaker 1

Google has done some work there, though. I don’t know—you’re probably not that close to the teams that have done AI-scientist work. There are some things that have been done. For example, if you freeze the model weights at 2012 or 2015, do you freeze time at that point?

Even with the current model, let’s say you somehow assume there’s no information leakage. If you ask the model, “What’s the best machine-learning approach?” it would just tell you that LSTMs are the best, right? That’s generally how machine learning works. Then the question is: Can you invent the Transformer?

It might not even be able to invent the Transformer. If you freeze time at a certain point and bring in the current model—but the model itself is a Transformer—then you have to assume there’s no leakage.

Yi Tay

Mhm.

Speaker 1

So I think there are still a lot of open questions about whether a model can really innovate and generate genuinely novel knowledge.

One related question, which I think is connected to the Stanford paper, is that people have this sort of mysticism around what reasoning is. If you really demystify it, reasoning is whatever happens inside the chain-of-thought tags. You’re eliciting that reasoning behavior from something that’s already latent inside the pretraining corpus. That’s one interpretation.

Yi Tay

These days, reasoning itself is very vague and open. Different people will have different definitions of what reasoning is.

I agree that chain of thought is what people associate with reasoning. It’s what happens during the thinking process. But these days, as I said earlier, reasoning reinforcement learning is almost the same as post-training to elicit capabilities. It’s reinforcement learning and post-training to elicit capabilities.

I think the technical definition of reasoning is making models better with thinking and post-training—reinforcing the model to think better. Thinking is more like thinking traces, thought trajectories, and things like that.

There’s also this line of work on latent thinking. Whether latent thinking and discrete-token thinking are going to be the same thing is still an open question.

Speaker 1

Meaning adding extra tokens to your vocabulary that represent thinking?

Yi Tay

I forgot the name of the academic papers that do these kinds of things, but instead of decoding discrete tokens, you simulate the process in latent space.

When you do chain-of-thought reasoning, you decode extra tokens hidden in the thinking tags. Latent thinking is basically when you don’t decode those tokens. You just don’t bother.

Speaker 1

Right. It might be that the native language of thinking is numbers, rather than passing it through some filter of English. Sometimes it might even start thinking in Chinese or something else.

Yi Tay

Yeah, yeah. Generally, I don’t really believe that model thoughts have to be the same as human thoughts. In machine learning, I’m more of the school of thought that you should let the model do whatever it wants.

Speaker 1

There was a latent representation hypothesis paper that I think you might be sympathetic to, if you haven’t already read it. To me, it seems obvious that image models will have the same idea of what a laptop is as text models do. They converge on the same latent representation.

Obviously, you can align them and do all those things with them. It totally makes sense that the concept of a laptop would just be a vector of numbers. That’s the concept.

Yi Tay

Yeah, yeah. You might have some numerical differences between one model’s idea of what a laptop is and another’s, but it would mostly be the same.

Speaker 1

Yeah, very interesting. The question I was going to lead into was that we’re now in this age where LLM text is in the corpus of the stuff that we train on, so there’s a little bit of a recursive loop, right? The reasoning tokens are out there now. Pre-trained models themselves—pre-trained base models—are also capable of reasoning, and increasingly so as more and more reasoning text goes into the corpus.

Isn’t that interesting, or is that worrying? Do you actually see much reasoning trace on the internet?

Yi Tay

I’ve never seen those, though.

Speaker 1

Yeah, yeah. People are publishing that specifically. As to whether or not researchers are actually including that in their training corpora, who knows, right? That’s their choice.

Yi Tay

But I would say that the percentage on Common Crawl that has chain-of-thought tokens in there went from 0 to 0.001%, and it will just go up over time because people are publishing it.

Speaker 1

Yeah, but I think if the sources are quite clean, you can actually filter those away, because people usually put them on GitHub or somewhere else. Do you want to filter it? Maybe you don’t. There’s a choice for the researcher.

Yi Tay

Yeah, quite literally, the whole reason why—I don’t think we covered this in a previous part—but 2 years ago, a lot of people were like, “Oh, you just include more coding tokens in your pre-trained corpus. It will—”

Speaker 1

Yeah, but the coding tokens are different from coding tokens in general. It’s just outside of code for reasoning.

Yi Tay

Oh, there was, like—

Speaker 1

You don’t believe that?

Yi Tay

No, no, no. I think—I don’t know if it’s still true today, but I see.

Speaker 1

Yeah, that was just our general coverage of reasoning. I would say that there’s a lot of interesting work here and more to do. Maybe I’ll cover one thing where I know that you have personal input, which is that you’ve started using AI coding.

Yi Tay

Yeah. I actually didn’t really use much AI coding in the past, but I think we’ve reached a point where AI coding has started to become really useful. Before AI coding, the thing that I found most useful about these models in general was when I had these big spreadsheets with a lot of results and wanted to make a plot of them. I think models are quite good at looking at a screenshot and making a plot from it. I hate making this Matplotlib stuff. It’s so annoying.

That’s basically one thing that I can remember about how I used AI in the past. But I think AI coding has reached the point where I run a job, get a bug, almost don’t look at it, and paste it into Antigravity. I tell it to fix the bug for me, and then I realize that what it did—AI coding is more like writing training code, writing ML, or something like that now. I would say it does pretty well most of the time.

There are classes of problems where I know it’s actually really good. In fact, it’s probably better than me: I would have to spend 20 minutes figuring out the issue, fixing the thing, and reloading.

Speaker 1

That’s very interesting, because I would say level 1 vibe coding is that you actually know what to do; you’re just too lazy. It’s just, “Do it for me. I’ve done this a thousand times. Just go fix it. I know exactly what to do.”

Here, you’re saying it’s actually the next level, where you don’t even know. It’s investigating it for you, and as long as the answer looks right, you just ship it.

Yi Tay

At the start, I was a bit like, “I did check it. I looked at everything.” Then, at some point, I was like, “Okay, maybe the model knows better than me, so I’m just going to let it do its stuff,” and then I’ll relaunch the job based on the fix that the model gave me. I think the models just keep getting better and better.

I also think that recently there’s Antigravity. These tools weren’t integrated into Google infrastructure, so it wasn’t that easy. I’m not that familiar with what’s available outside. When I started, the models weren’t so good 1.5 years ago, so it’s also a forcing function. People are like, “Oh, try Antigravity. It’s a game changer,” and stuff like that. So I just started using it.

Speaker 1

Yeah, you spent some time with Varun recently. What did—what did—

Yi Tay

Oh, no. I really just said hi and got it over with.

Speaker 1

I guess you were telling me that you’re an AI researcher who doesn’t mostly use much AI, and now you’re actually AI-pilled as a user.

Yi Tay

There were so many moments this year where AI suddenly crossed that threshold from emergent to useful. AI coding is one of them, which we just discussed. I think Nano Banana also got to the point where, usually, you make these images just for fun—you troll your friend or something like that—but Nano Banana actually got so good that you can use it for tasks.

You can use it for real work. It’s getting really good. In the past, many of these LLMs were like, “How does this—how does it think?” But now I just trust it automatically. I think people are just enjoying the utility brought by these models.

Speaker 1

So now I’m like—I was always AI-pilled. AI is a good thing. I don’t see how anybody can disagree with that. But you’re actually using it for things where you have high expertise, which is your own ML work.

Yi Tay

Yeah, yeah, yeah.

Speaker 1

And just to clarify: do you have a special version of Gemini that you use internally that we don’t have access to, or is it the public Gemini?

Yi Tay

I think it’s the public Gemini.

Speaker 1

Okay. I’m just saying it would be entirely reasonable to train Gemini internally only on your code base and your work.

Yi Tay

Oh, actually, I’m not sure, though. These things are just—I’d rather wait. But obviously, if it improves your productivity by, I don’t know, 10%, it’s worth it.

Speaker 1

Yeah. So I think that’s interesting, and there are interesting questions about the levels of how much you trust it, how much of your jobs you automate away, and what you no longer need. There’s also the question of how people come up and train in the field if you no longer need your juniors because Gemini is your junior ML researcher. These are all interesting questions.

Yi Tay

I want to say one quick thing first, right? When it comes to whether a model can be a junior researcher or something like that, if you think of it this way, even a job from 1x to 3x—1 times 3—can be replaced by a model itself. But let’s say you’re a manager, right? The objective, the metric you track, is your time. If you can have a model that saves you the same amount of time as the work that your reports do, you don’t actually replace one person per se, but you save a little bit from everybody.

Speaker 1

A little bit from everybody.

Yi Tay

Yeah, you’re right. Then you can—I definitely agree that when you count the net time saved, there are times when the model can fix bugs that would have cost me 1 day. 1 day is huge.

I don’t know whether anybody has done any real metric evaluation on these types of things, but if you use time as a real metric, rather than the number of people replaced, maybe 3 hours is kind of my metric, right? These things aren’t going to replace one person as such, but are more like a passive aura that buffs everybody, in game terms.

Speaker 1

I often think of myself as a bard because I tell stories and buff everybody around me. That’s an ideal situation for me in a D&D group.

Yi Tay

Okay, okay, yeah. I don’t play D&D, but okay, yeah, I get it.

Speaker 1

Then, in the case of a support, what’s the support hero?

Yi Tay

Support, yes. Yeah, okay. AI support, I think, is very encouraging.

Speaker 1

Where is it still not working for you, the things you’ve tried where you’re like, “Oh man, I expected it to be better”?

Yi Tay

There are times when models get lazy and try to fix something. They try to gaslight me into thinking that the bug is fixed.

There are still classes of problems that are very easy for the model and very hard for humans. There are some things that are very easy for humans and very hard for models, or whatever the spirit of that is. It’s still very hard to characterize these things into proper quadrants.

I would say that the capabilities of models these days are good enough to be really helpful. I don’t think there’s anything to specifically focus-fire on. It’s more about general capability improvements: the models just get better over time, and these things will go away.

Speaker 1

You say that, okay, so yes, in the grand scheme of things, just trust the process, keep scaling every dimension, and things will fall away; things will emerge.

But you’ve also said in the past—I can’t remember the exact tweet—“Each additional dataset compounds over time.” They’re just small additions. And I would say that when you still say things like focusing on things that you would think are easy for humans and hard for machines, those are easy wins, where you can just add a dataset that would focus on that.

Isn’t hill climbing just a sequence of doing that until you reach AGI?

Yi Tay

Okay, so I get your point.

I think that it's true that sometimes a lot of progress on the whole is just a series of small, incremental changes that push things forward. I think that's accurate. There also seem to be a lot of small, seemingly minor things that have pushed AI to the state where it is today, so I definitely agree.

There’s nothing against people who focus-fire. It’s just that it might not be easy to focus-fire on things that aren’t very easy to characterize. When you have something targeted—“Okay, I want to improve this capability; add some data”—then that’s fine. But defining and evaluating problems, characterizing them, and determining whether they can be characterized is a different challenge.

With coding, for example, I don’t work on coding, but people who really work on coding have terminology for different types of failures. Maybe somewhere, somebody is focus-firing on one of those failures and making the model better. That’s great, probably. That’s why it takes 1,000 people to get all these things together. AI is definitely a big collective effort these days. It’s a big machine. It’s really crazy.

Speaker 1

I just wanted to broaden out to general things people are talking about in the community and in research.

Yi Tay

Yeah.

Speaker 1

I know that you’re very locked in, so you don’t necessarily have to have read all the papers or anything. We can just riff on ideas, and obviously ask you what you think as well. Is Attention All You Need?

Yi Tay

Attention and the Transformer have been core ideas in recent times. Pre-training and scale are what made attention and Transformers actually shine, because without them, I think the first Transformer paper was just a machine-translation thing. GPT and BERT were the ones that really showed the full potential of this idea.

In terms of whether attention is really all we need, probably not. From an architectural point of view, maybe not. But it’s not that attention is all you need; you definitely need it. When people say, “Is attention all you need?” they mostly mean, “Will Transformers get us all the way to AGI?”

Speaker 1

Will there still be a normal architecture, or something meaningfully different?

Yi Tay

It will be a Transformer, I think.

Speaker 1

Really?

Yi Tay

It depends on what you call it, but I think that unless the paradigm shifts completely—which, as a scientist, you can’t completely say will never happen—my feeling is that it’s been almost 10 years since the Transformer.

Speaker 1

2017.

Yi Tay

Yeah, 2017. It’s been years since the Transformer, and I think we have not replaced self-attention. It’s some form of it. You could rename it, call it something else, or do local and global processing, but it’s still attention.

Speaker 1

Yeah, it’s still a Transformer in the end. I don’t think it’s going anywhere unless the whole thing with backpropagation changes completely. Then there’s a different story and maybe a different conversation to have, but if it’s still within the same scope and bounds, the architecture is probably going to remain similar.

Yi Tay

I spend a lot of time thinking about architectures and whether there are alternative architectures. At the sequence-processing level, there could be alternatives, but it’s still probably a sequence-to-sequence Transformer. If the formulation changes too much, your QKV becomes ABCD, EFG, or something like that.

There was this whole big era, which I was also involved in, where people tried to undermine attention as much as possible. They tried to remove it, simplify it, and make it efficient—the whole efficient-attention era. At the end of the day, the outcome was always, “We removed all attention, but we have 1 layer of self-attention there, and it still works.” That was always the story.

Noam published some work about mixing local and global attention. You still have attention, but you’re modifying it quite a lot.

Speaker 1

I would consider local and global attention to still be attention. It’s just a matter of how much you’re skipping.

Yi Tay

Yeah, though the question is whether, if the formulation changes too much, your QKV becomes ABCD, EFG, or something like that.

Speaker 1

Okay, maybe I’ll give you some motivating constraints here. You guys are still charging 2× for over 180,000 tokens of context, or 240,000 or something like that, and the theoretical maximum is 2 million tokens. What if we need 200 million? Is there some point at which even the concept of input-token context becomes irrelevant because you’re doing continual learning?

You’re modeling it as, “AGI will be achieved through a sequence-to-sequence transformation. Therefore, attention is the best sequence-to-sequence model or architecture. Therefore, attention is all you need.” But I think other people would say that sequence-to-sequence doesn’t accurately capture intelligence.

Yi Tay

It’s not really about sequence-to-sequence. It’s more about the whole gradient-descent and backpropagation paradigm. The architecture itself isn’t the problem; it’s the learning paradigm. The architecture is basically the interface between the learning algorithm and the tokens.

Continual learning is one way to think about processing insanely large contexts—200 million or 1 billion tokens, for example. You could have a new learning algorithm where, every time you run inference, you learn from it.

Speaker 1

Yeah.

Yi Tay

Then you can directly have some kind of memory, like a human being learning as I’m talking to you. That’s one way. Another way is that somebody might say, “Attention is too expensive for a 200-million- or 1-billion-token context, so we need a new architecture.” Or some people might say, “We just improve the chips,” as Sarah does. There are many ways to interpret it.

If it’s about continual learning, there are a lot of fundamental things about the learning algorithm, and there’s a lot of work to be done there.

Speaker 1

That will have to change?

Yi Tay

I think the learning paradigm, architecture, and all of that go hand in hand. As the field progresses and ideas stack on top of one another, there’s also the issue that a proposed idea has to be compatible with all the work that came before it to really shine.

It’s a bit of a variant of the hardware lottery, as described by Sarah. It’s not the hardware lottery that I wrote about—the GPUs, in theory—but it’s the original hardware lottery. The things being proposed have to play well with the ideas that came before them.

It’s a bit like descending into a local optimum. We’re now in this local optimum of Transformers—everything is Transformers. Maybe it isn’t easy to get totally out of this because a lot of people’s investments and optimizations have already been made. New things need to play well with the ideas that came before them, and the way I see it now, it’s very difficult to get out of it.

Speaker 1

Yeah, to come out of it.

Okay, I’m not entirely convinced. I see what you’re saying, but let’s call it generative AI—I hate that term. It’s still a very young field. Yes, there have been 8 years of work on the Transformer, but what’s that in the grand scheme of things? Maybe we’re in a local optimum and we’ve got to somehow get ourselves out of it. I want to leave that open-ended; I don’t have an idea.

I do think that people are in what Ilya Sutskever has been calling the age of research. We’ve scaled up what we can scale up. We know what scaling on every dimension that we understand looks like, but what is the next dimension to scale? There’s a misunderstanding that the last 5 years have just been about scaling.

Yi Tay

So, okay, please tell me more.

Speaker 1

Yes, I think you mean the joke about now scaling research salaries.

Yi Tay

Okay, okay, let’s not go there. But I think ideas matter, and I think there have been a lot of good ideas in the last 5 years. It’s just that maybe it’s not always about blindly scaling. If you took an MLP without self-attention and said, “I’m going to throw $100 trillion at this and scale it up,” it’s never going to work.

Speaker 1

It’s never going to work.

Yi Tay

Yeah, so there’s no simple version of that. Part of it is that I think the bitter lesson gets used too much and too conveniently. There’s also a sweet lesson, which is that ideas matter. Even today, people downplay ideas and things like that.

Speaker 1

Do you think the rate of new ideas has increased or decreased? Is there a law of diminishing returns? Are the ideas getting smaller?

Yi Tay

I think the number of ideas is always proportional to the number of researchers working on a certain problem.

So by definition, it should increase, but I think the number of ideas that actually work is not decreasing compared to the last—like, we’re not in the era of diminishing returns yet. I think ideas are still very important, and there are still very good ideas being generated and invented.

Speaker 1

Yeah, yeah. And I think I know the answer to this, but is the closed-lab advantage increasing versus open source or decreasing? The Chinese labs seem to keep publishing open-source models, and some of the American labs as well publish open-source models. Would you say that the ideas that I see there—Nvidia has NeMo-Megatron, OpenAI has GPT-3 SS—are all basically checkpoints on what is publicly known about training models as of this year?

Yi Tay

Oh, okay, okay.

Speaker 1

They’re like—it’s declassified information because everyone—okay, yeah, everyone does this.

Yi Tay

I think the gap is increasing. I don’t think it’s completely predictable from the things that you said before.

Speaker 1

The gap is definitely increasing. Yeah. I think that justifies researchers. Otherwise, what’s the point of having researchers if they’re not finding new tricks that compound over time?

Yi Tay

Yeah, yeah, but definitely, I think it’s increasing.

Speaker 1

Okay, I’ll do a side tangent. I don’t know if you have any comments on this. This is very related to NVIDIA’s recent purchase of Groq, which I don’t know if you have views on because you’re very TPU-centric, but are we memory-bound or compute-bound? This is relevant to the transformers discussion of, like—

Yi Tay

Of what? Like serving?

Speaker 1

Exactly. I think the classic view is that we’re compute-bound because we just need more compute for pretraining and RL, and then inference. But the counterargument I would make is that I actually have these charts of Moore’s law. I wish I could just pull them up easily: Moore’s law for the scaling of compute versus the scaling of memory versus the scaling of network and bandwidth. Compute has a much higher slope of scaling than the other 2.

But what about cheap memory? Honestly, I don’t think about this memory-bound issue that much, so maybe it doesn’t—

Yi Tay

Yeah, so I would disagree with it, but I don’t have high confidence in it.

Speaker 1

And because you’re mostly on the research side, less on the inference side—

Yi Tay

Yeah. Maybe the inference guys would be like, “Yeah, yeah.” I don’t think about serving, so maybe I don’t think about inference that much.

Speaker 1

My previous line of discussion here was that NVIDIA was very foresighted with Mellanox, because it’s actually the real bottleneck in scaling. It has the lowest Moore’s law. The second one now is memory, which is very interesting.

Yi Tay

Okay, okay. But honestly, I don’t think about that, so I don’t think about this that much.

Speaker 1

Data efficiency. This is a joke, but implicit in this is that there is some kind of maximum data exposure, right? Previously, I would say that a lot of the training paradigms were like, “One epoch is all you need,” which is the meme title of this idea. I would say that the real number may be between 3 and 40 epochs. I do wonder what the theoretical limit of a model’s data efficiency should be in terms of training and compression. I don’t know what that means.

Yi Tay

Data efficiency—basically, you’re asking the question in a way that’s asking, “How many repeats are tolerable?” Is that what you mean?

Speaker 1

What we mean by tolerable is contingent on whether it actually improves in meaningful ways. It’s not about wanting to do it for its own sake. But I do think there’s that, and then there’s also just the sheer amount of stuff that we can learn with limited data.

Let’s say we’re not compute-bound and we’re not memory-bound, but let’s say we’re data-bound. Last time we were on the podcast, we talked about Chinchilla versus inference-optimal training. But now, actually, I think a lot of people are even talking about data-optimal training. Given a limited dataset, how well can you learn from it? I think there’s an interesting research direction that not enough people are talking about.

Maybe it’s commonplace in the labs, but it seems very clear that we are very unoptimized with regard to how much we learn from our data. Let’s just put it there.

Yi Tay

I think, in general, extracting more from every data point is definitely possible, but I think it’s also related to the fact that we’re running out of tokens in the world. I don’t work on data for pretraining, and I think the things that I say would be about the general state of the industry, right? I don’t even know whether the way these things are done has diverged too much across these labs.

Speaker 1

There’s a lot of cross-pollination, for sure.

Yi Tay

Cross-pollination? Okay, okay. Yeah, but I don’t think about data that much—the pretraining data that much.

Speaker 1

Yeah, maybe earlier in the first half of this year, I would have said that pretraining is dead and that everyone was just funneling all their work toward RL. We had this Groq chart, which is very interesting, where we’re spending the same amount of compute on pretraining as on RL.

Yi Tay

I think it’s the side—no, I don’t know. I have no idea.

Speaker 1

I think that people are taking it seriously. They’re like, “Yeah, okay, whatever.” Especially in the Asian labs, like Cognition Co., they’re taking the open-source models from whoever and then adding, let’s call it, pretraining-scale RL on top of them, if they have that level of information—which they do. That’s very interesting.

I would say that this data-efficiency argument, to me, is also more about trying to discover new paradigms of learning in order to get where we all want to go, which is AGI. The existence proof is humans, right? Your 2-year-old daughter can be much more capable than an LLM in some things, having seen 8 orders of magnitude less data.

Yi Tay

That’s very interesting. Yeah, comparing human learning and machine learning is definitely—

Speaker 1

Clearly, it’s an existence proof that we could probably do better. Three examples of a dog.

Yi Tay

Yeah. A fourth example of an unidentified animal, and I can probably tell it’s a dog as a human. But machines, classically, need 20. The data efficiency of humans is definitely way higher than that of models.

Speaker 1

The only question is: Where does this thing come from? Is it actually about putting more information on every token? Or maybe it goes back to the question of whether the transformer is the optimal architecture. Maybe the problem is the problem. Maybe it’s the off-policy noise. So what is the bottleneck? Where’s the bottleneck?

Yi Tay

Maybe it’s some feature. I don’t know.

Speaker 1

So now we’ve identified it. It probably took me a while to get this across, but this is the kind of data efficiency I’m talking about. I think it’s emerging. Basically, at the end of every year, I like to try to take bets as to what the big themes for the next year will be.

Yi Tay

Yeah. I think there are more things that people are really trying to focus on because they’re feeling this data crunch, even though everyone is still investing in data. I forgot to mention that I’ve been wrong on pretraining being dead. I’ve now met pretraining leads from both Anthropic and OpenAI, and I’ve also seen the talk from the DeepMind person recently. Everyone is still investing in pretraining, which is nice to see.

Speaker 1

Nobody said pretraining was dead.

Yi Tay

I know. No, it’s a theory that we’re trying to disprove or prove.

Speaker 1

Anyway, let me wind back to my general idea. Data efficiency seems worthwhile. You would treat it as, “Okay, show me where the bug is and I’ll go fix it.”

Yi Tay

Yeah. We don’t know where the bug is. We just have an existence proof that it could be better.

Speaker 1

I think the final logical chain in this, for me, is that everyone is focusing on some idea of a world model as a version of this form of efficient learning, which potentially might not take the form of a sequence-to-sequence transformer. I don’t know how that works. I’m definitely a little bit out of my depth here.

To me, that is more efficient because every world must be internally consistent. If the next piece of evidence comes in and invalidates those worlds, they no longer need to pursue those paths ever. You can just narrow in on the world that you’ve identified. To me, that is learning where you’re learning to fit world models.

Yi Tay

Or do they actually learn—

Speaker 1

Yes. Maybe you can treat the learning process as curve fitting.

Yi Tay

Yeah, so you’re learning the world instead of learning the—

Speaker 1

So, what, the world model—

Yi Tay

Yeah, so I’m learning the world model, right? I’m sampling multiple world models and then finding out which one fits the data the best.

Speaker 1

I guess my question is this: Is this what people talk about? Obviously, feel free to attack it because I’m just roughly putting it together, but this is what I pick up from talking to multiple people about what they mean when they talk about world models, data efficiency, and learning efficiency, and how you gel it all together into a cohesive sense of the future where we can actually—

Yi Tay

Can you—

Speaker 1

What was the definition of a world model at the start?

Yi Tay

Yeah, there are 2 kinds.

Speaker 1

Okay, okay, go on. The first kind is the Veo kind.

Yi Tay

Veo kind? Or what’s the other one? Genie?

Speaker 1

Yeah, the one that DeepMind has, which is the video world model.

Yi Tay

Yeah. You model everything with some kind of Gaussian splats or whatever, and you inhabit that 3D space.

Yeah. Second, let’s call it the Yann LeCun–Meta school of thought, which I don’t know if you’re that familiar with. Yann has published the JEPA architecture, and separately, FAIR has also published code world models. Specifically for code, you are executing code and modeling the internal state of the execution environment as you go along. The language model actually learns to predict those things, and it seems a lot more efficient at the scale they’ve tested, which is pretty cool.

Which definition are you anchoring on?

Speaker 1

The third one. The code one.

Yi Tay

That’s the second one. Those two are bundled together. The JEPA fans are probably hating me right now because I’m lumping all of Meta’s work under one school of thought.

Speaker 1

Yeah, but whatever. Okay, okay. The first one is Veo and Genie, if you believe in superintelligence—those kinds of video-based world models. The second one is some execution, or some sort of explicit modeling, in the environment as you run through the corpus.

Yi Tay

And then I think the third one is this amorphous thing that people are trying to get to, where they’re doing what I said about the resolution of possible worlds and curve-fitting as they learn and infer.

But what is the world model itself? Is it like—

Speaker 1

It is a mental model of where everything is and how you think the world works.

Yi Tay

Wait, but technically, it’s something in the latent space.

Speaker 1

Okay, okay. So, for simplicity, you said it could just be a pretrained transformer model?

Yi Tay

Yes. Yeah. To me, that is the most coherent thing with the current paradigm, which is that you already kind of do this in current transformers. I think the way that you train it will probably have to be different.

Speaker 1

Okay, I see. I see. I don’t have any conclusion here. I’m just throwing it out as something where I know you’re interested in this kind of stuff, and I don’t have that many knowledgeable people to talk to about it.

Yi Tay

No, I don’t think about world models that often, because world models aren’t really well-defined in the first place.

Speaker 1

So don’t say “world models.” The problem is learning efficiency, and maybe accuracy or AGI capability, that isn’t easily unlocked right now on our current path of scaling.

Yi Tay

Yeah, yeah. I think when it comes to data efficiency, I believe in finding ways to spend more FLOPs per token, right? If you’re data-bound, you want higher data efficiency because you can learn more from every data point. You squeeze out more from each point, right? Things that can extract more, that can use more FLOPs on every token, are definitely a form of data efficiency.

Then there’s the learning algorithm, right? I think there’s a different scaling law for humans, machines, dogs, and cats. Humans are on one curve, machines are on another, dogs are on another, and cats are on another.

Speaker 1

Nice. I just wanted to point to the earlier chart.

Yi Tay

Right, yeah, yeah. There’s this famous chart. Point 1 and point 2 aren’t entirely different things, because it could be that a better architecture is actually just spending more FLOPs per token. If you come to a point where you’re very data-bound but not compute-bound, you could find algorithms that spend a lot of compute on every token.

So the overarching point is that this is a learning-algorithm issue for data efficiency. The correct approach might actually be to apply more FLOPs per token and squeeze more out of every data point.

Humans are also difficult to compare. When people say that humans are exposed to less or more data, that’s very ambiguous, because they’re technically exposed to data 24/7 and are mostly visual. They have a lot of different types of inputs, right? Whether they’re actually spending more FLOPs on everything they listen to is also a question, because maybe they’re just data-efficient. Somebody needs to count how many FLOPs a brain uses to process a given amount of information. Maybe humans are just spending more compute on every token, and maybe the learning algorithm is different.

So I think data efficiency is very important, given that I think we’re going to have a limited amount of data in the world.

Speaker 1

One more thing before we go into DSI. You know how we’ve been talking about RL?

Yi Tay

Yeah, we’re going into RL stuff.

Speaker 1

Why are people paying so much for RL environments?

Yi Tay

Wait, so who’s paying for RL environments?

Speaker 1

OpenAI and Anthropic, at least. Nobody said anything about DeepMind. A lot of the model labs are well known for paying at least 7 figures to external startups to create RL environments for them to train in.

And I think the question is: if your models are so good at coding, why don’t you do it yourself? I think there’s some amount of expertise that has been distilled from human experts into an RL environment, where you can then let your agents run wild. But I’m curious if there’s any other deeper insight than that, because I’m not satisfied with my own explanation.

Yi Tay

RL environments like that have a lot of domain expertise that is probably very valuable. I don’t know specifically what RL environments people are actually buying, though. What was the thing you weren’t satisfied by?

Speaker 1

The fact that they’re also available, and a lot of people are saying, “Look, it’s a Next.js app inside of a Docker container that logs stuff out when you send inputs in.” You could probably do it yourself internally, right? Why pay so much for some startup that you don’t know to do it for you?

Yi Tay

Actually, I have no clue why this is happening.

Speaker 1

A classic example would be that if you want to build computer-use agents for buying things in e-commerce, you would want RL environments that perfectly replicate maybe the top 1,000 e-commerce websites. Then you just parallelize across all of them. Does that seem valuable?

Yi Tay

I don’t know.

Speaker 1

All right, cool. DSI and LM recsys. A big bet for me this year for my conference was that we actually started focusing on LM recsys. What was the motivation behind starting LM recsys?

Yi Tay

I think recsys is the key AI problem in consumer. It is the single most valuable thing. All your feeds—even search—is recsys. Basically, search is retrieval. It’s the god problem, right? Because recsys is ranking, but it’s also filtering, personalization, re-indexing, and performance. It is the god problem, and you get paid a lot for it.

Engineers aren’t that excited by it, which is very weird. A lot of them don’t work on recsys, and they probably never will.

Speaker 1

But they don’t see the monetary value that can come out of good recsys.

Yi Tay

The other 2 updates for me—which I actually didn’t even know were directly tied to DSI—were, first, that Twitter publicly adopted its feed algorithm as an LM recsys. LMs are just used everywhere now. Whether it’s actually a big LM, or whether it’s a generative-retrieval-type model, is another question. We don’t know. All we know is that they said they swapped out their current recsys for an LM-based recsys. That’s all they said.

Speaker 1

Okay, okay. But what has been published is YouTube, where they actually adopted semantic IDs for YouTube’s recsys. YouTube is obviously a big deal.

Yi Tay

Is it public information?

Speaker 1

Yes.

Yi Tay

Okay, okay.

Speaker 1

They came and talked about it with us, and they published a V2 this year as well.

Yi Tay

Okay, all right, I see. Going forward.

Speaker 1

The last time you were on the podcast, we didn’t talk about DSI that much, but you actually have some background in IR. You care about IR. I don’t care about IR, but I think DSI—generative retrieval—was one of my favorite works in the field.

Yi Tay

I have some IR background. When I was doing a PhD, I did some recsys work and some retrieval work, recsys and stuff like that, so I have some IR and recsys background.

Generative retrieval and generative recsys were all very complicated. DSI started as a retrieval thing. We did Natural Questions, ranking indexed documents and everything. It started off as—I mean, we actually did an interview with Yannic, Don, and me when the paper came out a long time ago.

At that time, we wanted to reimagine retrieval and search. We were still using T5 models, so it wasn’t the LLM era yet; it was the pre-LLM era. It was more like, “Okay, pretraining works,” and there were some pretrained models around.

So we wanted to reimagine retrieval, right? Retrieval and recsys are all the same formulation: a ranking or retrieval problem. That’s where we started to imagine retrieval as 1 giant LLM that encloses everything in its memory.

We tried so many different ideas. My collaborator was the one who came up with so many of them. The start of this whole generative-retrieval thing was literally just trying to give a document an identifier and brute-force predicting it. It actually works because the models can memorize something.

If you look at the literature all the way back to things like Doc2Vec, or this is very often the case in modeling, the words have no meaning. They’re just IDs in the vocabulary, so now it’s just a number, right? Technically, the models have enough capacity to predict it.

I think semantic IDs were an idea where you have some semantic association and then try to break down the search space hierarchically, right? How this work evolved in recsys was that, after DSI came out, Ashish’s group and Mahesh did some exploration applying DSI to recsys. That’s how the generative-retrieval recommender-system paper came out.

Speaker 1

Yeah, I didn’t even know he was involved. It’s crazy.

Yi Tay

That was basically us transferring this: DSI works, so we tried it on RecSys. I think the recommender-system people have a slightly different way of doing semantic IDs, but that’s basically because the domain is slightly different. After that, I think we were done with the invention part and just ran it. The rest are details.

Over time, I also left Google, and these things evolved a little bit every day. I think I also saw that Spotify was using something like this. YouTube and Spotify use this type of semantic ID, this type of DSI-like model. From the research community’s point of view, the DSI work was the first one to decode semantic tokens. But when we went to—I don’t know, I can’t call this strange, in a way—people would do things like, “Oh, this is generative retrieval; it’s not generative RecSys.” They would do these random things that were a bit strange.

But yeah, I think this was the whole history of generative retrieval. Apparently, there were also a lot of people working on it. I don’t follow it at all now; it’s not even in my mind. Once, even in the Singapore office, there were 3 people working on generative retrieval. I don’t know whether they’re still working on it, but I met a person who tried to explain generative retrieval to me. It was quite funny: I kind of co-invented generative retrieval, but now I think this whole IR thing has just been an interesting phase. I think DSI is one of the more creative works I’ve done. It’s not really an LLM, but it’s—

Speaker 1

Under the general principle of applying LLMs to everything, if a Googler is working on generative retrieval, would that be like AI Overviews? Is that something similar?

Yi Tay

I have no idea.

Speaker 1

Okay. For the people listening, I did have a track there, but you can type in AI.engineer and you’ll get it. The Gemini guy was talking—sorry, the YouTube guy was talking—about how they use Gemini for their RecSys. I don’t know what size of Gemini, because he didn’t talk about it. But this is public work now, and basically every YouTube video uploaded gets encoded into some kind of codebook. They retrain this every day in some kind of batch job.

I don’t know if you even know what Gemini is being used for. I don’t follow these things these days.

Yi Tay

I do think, in the sense of people who are still not getting it, applying intelligence—the general intelligence of an LLM—to retrieval and recommendation means you can accommodate such weird recommendations and such weird queries that normally no classical system can ever handle. I think it’s also somewhat emergent. When you were using T5, you just couldn’t add that much value on top of a normal BM25 retrieval technique.

Would you say this is accurate? It’s not just about paraphrasing; it’s about understanding query intent.

I did devise a really strong baseline, actually. Yeah, I did devise a really strong baseline. I don’t know the comparative data versus T5 for you guys versus BM25, but I do expect it to be very high, and I expect it to be a lot higher for a true LLM-based RecSys, depending, obviously, on the query set.

I didn’t really think about it this way before, but I’ve done modeling in many different domains, including search. In the search community, the IR community, there are also benchmarks that people are climbing. There’s an aim of—I don’t know what it’s called these days anymore—but generally, the modeling dynamics of an IR task are very different from those of a RecSys task, and both are very different from a standard language task or a vision task.

When we climb LLM benchmarks, you train models and interact with the environment in a way that feels intuitive. As a research engineer, you interact with the environment a lot. But RecSys and IR have a very strange feeling to them. It feels like you’re in a world where the gravity is different, or where the modeling dynamics that feel intuitive are not intuitive. It feels like a very strange space to work in.

So I wrote some papers back in my day on RecSys and things like that. Every time I ran modeling experiments for RecSys, I didn’t enjoy them. It feels like the environment was rude. It feels like the weights were just—

Speaker 1

What makes it rude?

Yi Tay

It just feels transactional. No, not transactional. I don’t know how to describe it. For example, if you play sports like table tennis or badminton, where you hit the ball, you have a very nice feeling when you hit the sweet spot. When you do modeling in a traditional domain and get the feedback back, you feel like everything sounds right and everything feels right.

But RecSys and IR are like a bad runway. You hit the shuttlecock and hear glass shatter. Randomly, it just feels like this weird world where cause and effect are too far apart. It just feels strange. Sometimes the metric—RecSys uses all the NDCG metrics—and BM25 is strong. Then you get something worse than BM25. Back in the day, when you stacked 2 LSTMs to 3 LSTMs, you were like, “Wow, I see.” It’s just an unrewarding area to work in. It’s weird.

Also, the IR and retrieval communities are always behind the mainstream, and now it’s probably getting even worse because of LLMs and things like that. I’m getting into hot-take territory, but some of the conferences are just behind NeurIPS and ICML. Some conferences are just applying things that—

Speaker 1

They’re downstream of.

Yi Tay

Yeah, they’re downstream. They’re downstream.

Speaker 1

Okay. So it always feels very uninspiring to work on this thing.

Yi Tay

Yeah, look, there’s a reason that I left, but it was like a side quest. I worked on it as a side quest.

Speaker 1

Yeah. Okay, I understand. I still think it’s an important business problem, even though maybe it’s an unrewarding field. I can understand why, because the academic benchmarks for those tasks are so far detached from what industry is.

Yi Tay

I didn’t work on any of this from an industry point of view. I liked the thing, but just from my academic point of view.

Speaker 1

Oh, then all you need is an online eval server, right?

Yi Tay

Yeah, you need a test. And I go, “Ooh, okay.” That would have been a different experience.

Speaker 1

That is mostly our coverage of the research topics and everything. I think we’re just going to end on a very simple one: GDM Singapore. You organized a symposium here, and we brought Jeff Dean, Quoc, and all the others. What’s the general message or the impetus for starting GDM Singapore?

Yi Tay

Let’s talk about the event first. The event was mostly because Quoc and I are going to start a team, and before I came back, we had discussed this for some time. Jeff was very supportive of it. He was in the region many times, in Vietnam and Singapore, around the time I was going to come back. This event itself was when Quoc and Jeff were visiting, and we wanted to inspire the community here.

I think it was also a bit of a soft setting of the tone for the start of the Gemini team in Singapore. It’s a very rare instance where you get somebody like Jeff and Quoc, who are true pioneers of AI in the world, in one room. I was there as well, and many people told me that they were inspired by their presence in the region.

Starting a team and starting something also doesn’t happen in one moment. It’s not like you press a button and it starts. It’s a process: we hire people, people join one by one, and so on. This event was more about setting the vibe. It showed that it’s possible for Singapore to be close to the frontier. Having the true pioneers of AI here was inspiring, and it also gave Quoc and Jeff a chance to meet people in the region.

Jeff was here last year, but Quoc hadn’t been here for some time, and he was going to have a team here. It was also nice to bring him around and let him meet the people here.

Speaker 1

True, and yeah, but I was there to livestream.

Yi Tay

Yeah. Having them all in one room and giving these talks, many people came up to me and said they were very inspired by their presence in the region. It was a really amazing event. We met along as well.

Speaker 1

Who a lot of people don’t know has a CS degree. He’s like one of the few PMs with a CS degree.

Yi Tay

Yeah, yeah. I would say that the context of the meeting was partly that he also wanted to learn more about IMO stuff.

Speaker 1

And then also about Jeff, he wanted—

Yi Tay

Oh, because they invited you without knowing that these guys were coming or something like that, right? Jeff and Quoc and I went to visit and chat with at the Istana. We discussed a little bit about deep tech and a bit about AI. The rest of it was more about Jeff and Lisa talking about macroeconomic or political things.

Speaker 1

I was very out of my element with that, so I just asked—

Yi Tay

You were in a suit.

Speaker 1

I was just talking about Deep Think and IMO and stuff like that.

Yi Tay

Yeah. But he seemed to be generally quite surprised that AI had reached this point.

Speaker 1

I think it was interesting. For people like you, who have done something unique in Singapore’s history so far, establishing a frontier research lab in Singapore is an accomplishment. I think the other thing I’m still trying to wrap my head around is whether geography actually matters. You’re all working on a team with people in London and Mountain View, and you’re mostly just collaborating with them anyway. You’ve collaborated with them your whole life.

I don’t even really know what countries mean anymore when it comes to research or AI in general, because this thing is inherently international from the start.

Yi Tay

That’s a very good question. It’s also related to the thing about identity, because I think you also move between San Francisco and Singapore quite a bit, right? I was in Mountain View a week or two ago, and now I’m here. Almost all of my—if you look at my work, aside from my family—everyone I talk to is somehow in the Bay Area because of work and everything.

I think geography matters. Firstly, the most important thing logistically is probably the time zone. You literally want 24-hour coverage around the world.

Speaker 1

They have an advantage?

Yi Tay

No, what I’m saying is that the difference is mostly that people define the location more than the location defines the people, somehow.

Speaker 1

We’ll get to the time zone a little bit later. There are pros and cons, right? Do you believe in the talent pool in Asia and Singapore?

Yi Tay

I think we managed to find amazing people. But I would also have to say that this type of thing is more about talent attracting talent. Most of the time, people are very excited. The vibe I get is that people are very excited because it’s Quoc’s team and my team, and we’re working on very core things related to AGI.

I feel like the talent we can get from the region is really good, but it’s only because it’s us that we can unlock this talent. Otherwise, they might join some other place and move off to the US.

About identity, I would say that I definitely agree with you that geography doesn’t matter that much. I think the advantage of Singapore—or just anywhere—is that the world is very global. Technically, you can interact as much as you want, and you can do that from anywhere. But I think Singapore has this advantage where you can go close and you can go far.

I have some friends in London and New York who would just never move to the Bay Area. I’m not against the Bay Area—I think it’s a great place—but it’s just AI, AI, AI everywhere, right? Sometimes, if you have the mental space and energy to experience another culture, London, Singapore, and New York all have their own cultures. But the Bay Area culture is just AI. You go anywhere and hear AI everywhere, even on the billboards and stuff like that.

Speaker 1

It gets a bit much. Although I did see some billboards down here with AI, and I was like, “What is this?”

Yi Tay

It’s culture infecting Singapore. I do think that, to some extent, if you want to do research, you need a little bit of peace and quiet somewhere, right? This island might be good for that, but we’re still able to stay connected.

Speaker 1

So I think that’s mainly about talent. People are strong here.

Yi Tay

Yeah.

Speaker 1

You’re far enough away, but you’re still connected. You have strong talent. What are you hiring for? You’re still hiring, right?

Yi Tay

We’re hiring. My team will work on RL reasoning for Gemini and Gemini Deep Think. I think we care more about talent density now, so we’re not growing that big. We’re starting with a small team because compute and capital are probably important.

That’s what we’re hiring for now. Generally, there are a lot of people who are very capable, but what I’m looking for mainly is either a track record of research—not necessarily RL—or some exceptional achievement in coding competitions or an exceptional achievement somewhere else. That’s the kind of people we want.

Speaker 1

Yeah, because you don’t strictly require—I remember something about your record where you like to train your own engineers from scratch. Right? So you don’t—

Yi Tay

I forgot if I said that or not, but to some extent, yes. We would definitely be very happy with people who have very high stats and can do things even without much—

Speaker 1

Is that statistical knowledge?

Yi Tay

No, no, stats. Just at that point, high-tech people—raw, high-tech talent, people like you. Or strong engineering skills. ML knowledge can be learned easily.

Speaker 1

I think maybe one version of this is: can it be done on a student budget? Can you still do something interesting on a student budget?

It’s relevant to the point that conferences are quieter these days. I did an interview with one of the best people in this field who worked on a thousand-layer neural network with RL, and that was done on a student budget. It was a very cleanly executed piece of work and a good paper with good findings. I’m not sure if production models are ever going to have a thousand layers, but they stretched things in an interesting direction and found some good recommendations.

The person immediately got hired by OpenAI, and I think that’s encouraging for the graduate students in the market who are thinking, “Do I need to know somebody who works at one of these labs in order to get in? My uncle works there, so I get the internship,” or whatever. No, you can do it on a student budget with good advisors.

Yi Tay

I actually think one thing that’s interesting is that, for most of the people I recruited, I reached out to them personally. You see their work and then send them DMs.

Speaker 1

You get a lot of DMs from these people?

Yi Tay

No, no, for hiring generally. The point is that you can just do good work, put it online, and then somebody will contact you. It’s super easy but super hard at the same time, because I can tell you that I’ve talked to a few graduate students who don’t know what good work means.

There are so many things that their professors have on their agendas and are forcing them to work on, which might not be right, because it’s not as though their professors know what to work on either. They just need guidance: “Work on these 5 things and show me an interesting result in any of them.”

Speaker 1

Okay, so if somebody comes up with something and does work that you feel is very tasteful and aligns with what researchers in the labs like, and they come up with it independently, then you know the function they’re producing and the idea are good, right?

Yi Tay

If you just go and tell somebody to do something, you get the signal that this person can execute. I think there’s value in people demonstrating research taste.

Speaker 1

Yeah, research taste. Very interesting. I feel like I could give people advice about this. In some ways, the research-direction work that I do is a little bit like that. It’s low-accountability for me because obviously they’re just thought experiments, but I think for a lot of people, their career is bounded by whether they can demonstrate research taste within the short 3 or 4 years they have. They just have to do it.

Yi Tay

Yeah, yeah. I would say that it’s more that there’s so much competition because everybody wants to get into AI. It’s mostly about how you’re going to prove yourself.

Speaker 1

It must be hard these days to be a graduate student trying to prove yourself.

Yi Tay

It’s definitely harder, but yeah.

Speaker 1

Okay, that was it. Do you have any other rants or topics queued up before we wrap?

Yi Tay

No, I don’t, but it was great. I had a great time.

Speaker 1

It was fun chatting with you.

Yi Tay

Fun chatting.

Speaker 1

Even last time, when we met at the symposium, we were supposed to record, but we ended up just hanging out and chatting. It’s nice to get a dump of what’s going on in your world, because we’re working on really important stuff, man. It’s always great to chat with you on this. Nice to see you.

Yi Tay

Yeah, good to chat.

Speaker 1

Parting words on the weight-loss and workout journey? That’s also been a big thing for you.

Yi Tay

I think being healthy is important for doing research. I think I’ve probably been in peak physical health now.

Speaker 1

Yeah, you look great.

Yi Tay

Thanks. I think it’s also impacted my work in a good way. I did the sort of Kwapoki-inspired biohacking. I didn’t go to the extreme, but I was also quite data-driven when it came to this. I had my own wearable and tracked everything.

I was still supposed to make a blog post about this, but I feel like I’m not really at the end game yet. When I get there, I will. Just for people who don’t know, I think I lost 23 kilos this year.

Speaker 1

Actually, across 1 year.

Yi Tay

Yeah, 1 1/2 years.

Speaker 1

Yeah, so 23—

Yi Tay

Yeah, basically literally from the last podcast to now.

Speaker 1

Yes. Yeah, yeah. It's the evolution story now: 23 kilos.

Yi Tay

Yeah. And I think my HRV—heart rate variability—has gone up 2 times. My resting heart rate has dropped by 30 beats per minute.

Speaker 1

30 beats per minute?

Yi Tay

It was around 80 or 90, and now it's 60.

Speaker 1

Oh, yeah. 80 or 90 is super high.

Yi Tay

Yeah, I was unhealthy.

Speaker 1

Yeah. Okay. When things are hard, do you have something that can keep you going?

Yi Tay

You know, a lot of people focus on work, including myself.

Speaker 1

Yeah, including myself, right? I do prioritize work. I enjoy work. I don't enjoy the fitness side, but obviously it feeds into your intellectual work. Logging off and going for a walk, eating better—all that kind of stuff—it obviously feeds in. But people seeing a positive example like you will get inspired to do the same thing.

Yi Tay

So I think it's good to set yourself up as an example. I think that helps. When I do these things for my health, I just think that it's also part of work because it helps me get better at my job. So it's important as well.

Speaker 1

I think it's important as well. Yeah, I like the HRV off the bat. I have no idea what mine is, but yeah. There's a general question about what productivity is and how you measure it. What really matters?

Yi Tay

It's still unclear to me, but I do think the general energy level and hunger almost—you almost have to experience physical hunger in order to have intellectual hunger. Yeah, I don't know if that's a thing.

Speaker 1

Because when I'm hungry, I just think of food.

Yi Tay

I think, to me, it's distracting. But why is it hard to do work when you're hungry?

Speaker 1

Yeah. Okay. Thank you so much.

Yi Tay

Yeah, thanks. It's really great. Yeah, have a great time.

Captaining IMO Gold, Deep Think, On-Policy RL, Feeling the AGI in Singapore — Yi Tay | BidClub