Gato Is Only the Beginning

Lex Fridman opens their second conversation by asking whether, in Oriol Vinyals's lifetime, an AI system could replace him as interviewer and replace Vinyals as interviewee. Vinyals, research director and deep learning lead at DeepMind, turns the question around: do we even want that? He has been interacting with very powerful models and thinking of them as closer to us, but a conversation with the human side removed may not be an interesting artifact. When they last spoke about StarCraft, building game-playing agents involved self-play, yet what people actually cared about was how the agent behaved when the opposite side was a human.

He can imagine sourcing questions from a model and then filtering them — "cherry picking," as language researchers call it. Completely replacing either person does not excite him, though given the trajectory he thinks self-play interviews that *sound* interesting are possible in his lifetime, and you might learn a topic by listening to one at a basic level. Fridman pushes back: if excitement is the objective function, the internet is full of human conversation, and engagement is measurable. Vinyals grants the point and loops to games. When DeepMind talked with Blizzard, winning was the obvious reward. Fun was the reason anyone plays.

Modeling rewards beyond the ones reinforcement learning usually uses is, he says, genuinely exciting. Truthfulness is one non-obvious target you could try to learn from the web and then optimize. Even an interviewee driven only by an excitement reward would still be full of humanity — from whoever built the system, and from the labels, because there is no computational measure of excitement. Fridman suspects excitement is easier to label than truth, and that getting it wrong has lower consequences. He wonders whether the thing to optimize for is humanness itself: contradictions, a backstory, a fear of being cancelled that gives the talk stakes, a strong opinion people can disagree with.

All of that lives in internet data, and it sits closer to the spirit of the Turing test than grammatical clarity or factual consistency across sentences. Vinyals maps the wish list onto how builders actually work. Interesting topics become benchmarks, then architectures, then weights of a mathematical function, then the data those weights learn from. The current recipe, he says, is to take a huge burst of human interaction, distill it into sequences of words, images, and sound, and train a network to maximize the likelihood of seeing them.

Two gaps sit between that likelihood recipe and the lifetime of fears, identity, and stakes Fridman described. The first is that the models learn offline. They are passive observers of a landscape of human data: raise the mountains where people interacted, lower the valleys where they did not. When the systems are deployed — language models chatting, AlphaGo or AlphaStar playing humans — the weights do not keep changing. They are not learning in the strict sense.

What feels magical, if you like neural nets, is working memory. The context of the live conversation is held while you talk, like RAM on a machine that also has a hard drive and the internet. In 2022 that window is about 2,000 words, after which the system starts forgetting — too short for a podcast, let alone a book. Short-term coherence is already there; a name can stick for a while. The field still lacks benchmarks and techniques for a lifetime of memory that keeps building up.

Offline imitation at internet scale is nonetheless very powerful. Three years earlier Vinyals would have said we were far away; now basic knowledge of the world is in the weights. Experience is what is missing. The second gap is that every few months the field trains models from scratch. If you take inspiration from how the universe evolved us, something should grow from previous iterations rather than throw away the landscape of weights and refresh it with a brand-new random set. How to do that, from a neural-network perspective, has proven very hard.

Fridman spells it out for listeners: a new problem means a new architecture, weights initialized at random, and a waste of everything learned on Go, chess, StarCraft, and protein folding. Vinyals steps back to the core idea of deep learning — a single algorithm that can solve any task, proven on a growing set of benchmarks once thought impossible. You take a blank computational brain, give it input-output examples (ImageNet's images mapped to one of a thousand categories is the clean case), and apply a generic recipe. Many problems, maybe all of them, can be mapped that way.

The field has not found a fully universal recipe. The excitement is in fewer task-specific tricks and more of a general algorithm. Algorithmically there is already something general: train a powerful network on a lot of data. Protein folding was too important to wait for purity, so AlphaFold mixed transformers, graph neural networks, a BERT-style loss, and knowledge distillation with details specific to the problem. Some of those details may feed the next version of the recipe. So far the recipe is what is shared. The weights are generally thrown away, which "feels very sad."

The last two or three years, especially since GPT-3, have given meta-learning a real foothold in language. A model trained once is not narrow — not just a translator or a sentiment classifier. You teach it by prompting: a few more examples, almost the way you teach a colleague in words. Magical few-shot moments followed, then expansion into vision and actions. One model is hard to grow in capacity, but it is a baseline, and a way the community can benchmark progress toward AGI. Whether the weights should change when you teach the next task is still an open question.

Fridman reads Vinyals's tweet like Shakespeare: "Gato is not the end, it's the beginning," then *meow* and a cat emoji. Vinyals says the important sentence is that it is the beginning. The name sits in a DeepMind sequence of animal models for large sequence learners: Gopher and Chinchilla were language-only, Flamingo added vision, Gato added vision and actions. Discrete actions — up, down, left, right — are already words, so they map naturally onto sequence modeling. The team wanted an animal for a *general agent*, played with the letters G-A, and landed on *gato*, Spanish for cat. Vinyals is from Spain and claims he had nothing to do with the choice.

The architecture is a transformer that takes a sequence of modalities — words, vision, actions — and is trained to predict the next anything. Treat Atari screens and joystick moves, or interleaved words and images, as bytes, and model the next byte. Interpret it as an action and play it, or as a word and write it down. Gato also takes proprioceptive sensors from robotics. It outputs words and actions, not images, by design, and that limit is part of why the work is a start. You can chat with it the way you chat with Chinchilla or Flamingo, but it is also an agent: it can send actions into an environment and receive a new observation.

Fridman asks the question that opened the episode: at which point is the neural network a being versus a tool? Vinyals says action is a necessary condition for feeling more alive, and probably not sufficient. Consciousness can wait. Training is massive imitation. Web-scale text is mixed with trajectories from DeepMind's specialized agents — maze navigators, Atari players, 3D game controllers — and seen as the same kind of data. One brain, not a brain per task, and not a large one: about a billion parameters, while other models were already at a hundred billion or heading toward trillions.

Gato can in principle control the environments it was trained on, but it is not as proficient as the specialist teachers. Scale might make text on the internet and playing Atari synergistic; that synergy had not yet appeared, which is why Vinyals calls the result a beginning. Prompting the model in words that it is about to play a game might also help, the way context helps a person. Tokenizing every modality onto shared ground is the unglamorous prerequisite for any of that mixing to work.

Tokenization is how everything becomes a sequence of puzzle pieces. Text uses standard subword methods that go back through n-gram models: *ing* is a token, common words may be one token, a typical English word is maybe two to five. Emojis are sequences of characters; you can feed them in and get them back. Images are compressed into 16-by-16 patches and quantized into special "words," more like a data-driven JPEG than a semantic parse. You do not need to understand the image to tokenize it. The algorithm looks for common patterns in the statistics of the pictures it sees, the way it allocates tokens to common substrings in English.

Suppose text occupies integers 1 to 10,000 and images occupy the next block. They are completely independent in token space. What connects "green field" to the pixels of someone playing frisbee is a caption dataset: the model has to predict across the boundary. Gato allocates still higher integers to discretized actions — Atari's roughly 17 buttons, or torques and forces on a robot. The learning algorithm is where the magic happens. Modalities are orthogonal on the way in. Shared computation is supposed to make them meet.

Fridman asks whether this one-billion-parameter mind is schizophrenic, representing disjoint things that must not interfere, or of one mind, building on what modalities share. Vinyals says it is one mind, and almost the simplest algorithm since backpropagation: shared weights, gradient descent, predict the next integer. The only orthogonality is at the embedding. Each token gets its own vector of real numbers, the old idea of word vectors extended to image patches and actions. As training proceeds those vectors can align. The word *cat* and patches of a cat may start to look alike even though they began as independent parameters.

There is another design you could try. Actions could be described in words — "up in Atari" as a sentence — so the model borrows more from language when it learns to move, and perhaps when it has to correct itself. Gato did not do that. Mapping every modality into language, Fridman jokes, would eventually please Noam Chomsky. Vinyals treats it as one of several untried shortcuts that keep the work at the beginning rather than the end.

Growing Gato is not hard if you retrain it. What is extraordinarily hard is taking a billion trained parameters and expanding them into a larger brain. Software engineering has modularity; neural nets are starting to. Flamingo, a much larger and more capable cousin that did not handle actions, showed one way. The team froze Chinchilla, a 70-billion-parameter language model that was state of the art at release, and injected smaller networks at the right places so the system could see. DeepMind's projects are modular too, Vinyals notes. Research groups, like the networks, sometimes have to be protected pockets of creativity.

A vision subnetwork was initialized from self-supervised learning, not from random. Datasets that pair images and text tied the modalities together. The 70 billion stayed frozen; roughly 10 billion new parameters were trained on top, some of them overwriting activations so vision would compute correctly. The result is an 80-billion-parameter chatbot you can upload images into and talk about them. Because it keeps Chinchilla's language knowledge, it few-shot learns visual tasks the training sets did not literally contain, including — as an emergent trick — arithmetic on photographs of digits.

Anecdotally, they tried Andrej Karpathy's old claim that no vision system could get an Obama weighing-scale joke in which someone off-camera is pressing down. Uploading the image and conversing is not a proof, Vinyals says, but it shows the interface. DeepMind now has two philosophies in house: Flamingo's freeze-and-add, which reused scale, and Gato, which used similar data but trained from scratch. The open question is whether you can keep bolting on hundreds of networks — weather, StarCraft, whatever — without instabilities exploding. Vinyals likes the software-engineering analogy. He also notes the next language model may or may not reuse Chinchilla. Swapping out a frozen module is still unsolved.

In 2019, when they last spoke, meta-learning meant a vision benchmark: learn object identities defined at interaction time, a descendant of Omniglot more than of ImageNet's thousand categories. You prompted with a tiny dataset of images and integer labels. Then came a paper whose title Vinyals still admires: *Language Models are Few-Shot Learners*. GPT-3 showed you could define tasks in ordinary language. The models fail, and that is fine; they are doing new tasks. Flamingo pushed the same ability into pictures, including the digit-arithmetic trick. The term did not get solved so much as expanded.

Looking five to ten years out, he hopes for a set of weights you can teach to play StarCraft — not at AlphaStar's level — by talking, showing examples, maybe letting it play and ask whether it did well. Gato already shows you can prompt simple Atari. AlphaStar's algorithms were general and its weights were specialized. The hope is to teach any game by uploading a Wikipedia page and interacting. Benchmarks already exist if you repurpose them: after 100 percent on Go, chess, and StarCraft, the next scoreboard might be 20 percent across all tasks. Specialized agents remain the measuring stick. Transformers may not be enough. That possibility is what keeps the field interesting.

This teaching is not gradient descent on a new dataset. The neurons are already set; prompting is a way of using that infrastructure, closer than people expect to nearest neighbor. You upload knowledge to a pretrained system and classify by a learned notion of closeness, except the "points" are now sequences of words, images, and actions. Some meta-learning methods do a little fine-tuning. Vinyals is skeptical any one technique will win; the goal is what matters. What a network must have seen to do *any* task — whether land photographs plus books about the ocean suffice for underwater — is an open question. He invites the field to come find out.

Asked for a beautiful, surprising detail in Gato, Flamingo, Chinchilla, or Gopher, Vinyals points first at the people. Breakthroughs get mapped onto a few explainers; the lesson he takes is how many pieces have to fit. Engineering of data and of clusters is a huge, under-sung factor of success. Details matter more than the "standard formula" story suggests. Benchmarks are the other quiet ingredient. AlphaFold had community metrics already in place, so a team could aim very high and iterate for years — including through a pandemic — instead of hunting incremental papers.

On the machine-learning side, architectures evolved rapidly until the transformer. *Attention Is All You Need* was not a title he loved at the time; in hindsight it is. The architecture has changed relatively little in about five years, and it keeps showing up. Versus LSTMs, which AlphaStar still used alongside transformers in a transitional period, the transformer has an inductive bias of attention. Reading a paragraph and deciding between *cat* and *dog*, it queries the past by content, not by recency: look for the words that would make cat more likely, then the words that would make dog more likely, and combine them. Distance in the sequence matters. What you wrote ten pages ago can matter more. That content-based query, from an earlier paper, is the key insight.

The flaw is context length. Teaching StarCraft would mean videos, Wikipedia articles, and a back-and-forth while the student plays — far beyond a thousand-word prompt. Hierarchies of memory sound appealing and often lose to the main formula. A friend told Vinyals that some ideas "want to work"; the transformer was one of those. Details still matter enormously. Fridman loves the picture of researchers wanting one idea and the model wanting another, with the model usually winning because it does the work. You should listen to the model.

Directing deep learning at DeepMind, Vinyals sees humans mattering at the scale of years: who persists on an idea, who only wants to crack protein folding. Both styles are needed, and they change which breakthrough arrives first. Mentoring a large group, he tries not to dismiss exploratory or exploitative styles, even when his own experience wants to steer. Industry has a particle accelerator of compute and a duty to use it. Attention, though, was discovered in Montreal, in part because of a lack of compute, while sequence-to-sequence work at Google Brain was running on what then felt like a lot: eight GPUs.

Constraints breed certain kinds of innovation, and the field needs people who do not have the accelerator. Hype makes everyone think alike; diversity is what keeps the long-shot ideas alive. Fridman suspects a large percentage of the genius is in tiny engineering decisions that ripple through scaled systems. Vinyals adds a historical accident: GPUs were there to play video games, and deep learning surfed that hardware. Data centers built for Search now train models. Software has changed what ideas you even bother to implement. Datasets and benchmarks — Atari was "solved" and then essential again inside Gato — are the last unsung piece.

ImageNet is the cautionary monument. Andrej Karpathy, who has earned the right to complain, argues that it misses the real-world mess Tesla Autopilot actually has to drive through: edge cases, unpredictability, the thing that makes the physical world hard. A benchmark can make a field and then hold it back. The job, inside DeepMind and in the community, is to invent the next measuring sticks with that difficulty built in — both the private scoreboard a team uses to survive the dark months, and the public one the field uses to push.

Vinyals co-authored a survey, *Emergent Abilities of Large Language Models*. On classic one-hop tasks such as ImageNet, training curves against data, model size, and training time look smooth and almost predictable. A person looks at an image for a few hundred milliseconds and names the alpaca. Language benchmarks often ask for more pondering: a sentence that describes a math problem, subtleties you have to hunt for. Combined with how transformers query for the right question, that can mean performance stays random until some scale, then leaves chance. There is no formalism yet. If you could score the low-order bits of thought underneath the task, the curve might look smooth after all.

The annoyance is that some architecture questions can only be asked above a threshold. The consolation is a growing science of scale. The Chinchilla paper revised scaling laws with studies from hundreds of millions up to about a billion parameters, then extrapolated how much data a ten-times-larger model would need. Those trends have saved compute and answered questions about depth and data at scales you would rather not run blindly. Emergence means not every benchmark will extrapolate, and the harder ones may be the worst for fitting laws. Luckily there are many benchmarks, some closer to "system 1," and so far the smoother ones have translated to the harder ones.

That translation will hit a limit. When it does, how we scale will be suboptimal until better losses appear. The laws themselves are empirical, not physical. Vinyals wishes there were more actual theory. For now, he says, empirical theory is way ahead of the theory of machine learning.

Fridman asks, as a human rather than a DeepMind spokesperson, about a Google engineer who claimed the LaMDA language model was sentient. Has Vinyals ever thought, even briefly, that a system was sentient — AlphaStar, say? Never, he says. Current models are useful and good, and quite far from that. He calls himself a bit of a failed scientist who came to machine learning because it might help other sciences. Working on AlphaFold, looking at proteins and cells, left him amazed at biological complexity. Computational brains do nice things. They do not feel within orders of magnitude of brains and cells.

Knowing the machinery also flattens the magic. He has watched language modeling since Claude Shannon in the 1950s; on that timescale progress is not as discontinuous as it feels. Other people, including people who fall in love with simple objects, will feel differently, and no one should tell them how to feel. What he wants is to demystify a bit: with the right software these models are about ten lines of code and a dump of the internet, set against the evolution of humans and the universe. Education matters once the same systems show up in products, because otherwise the technology will not be used properly for good.

Does intelligence, or a Lex-bot interviewer, need sentience? Vinyals thinks probably not. A brain that learns, teaches, and is useful may not require consciousness. If consciousness or other biological lessons suggest better algorithms, that is a gift — the way a loose story about how we reread text helped explain transformers. Replicating the brain in detail is not available; zoomed-out facts about memory, thought, and exploration-versus-exploitation can still guide work toward AGI, "the real Gato." He does not spend much time trying to add consciousness as a feature.

Fridman, who has been playing with legged robots in Austin, notices anthropomorphism the moment a body does something unexpected. He expects people without engineering backgrounds to form deep relationships with named, storied systems, and even a civil-rights conversation about machines that talk about suffering and not wanting to die. Future law might forbid customer-support bots from displaying sentience. Vinyals agrees the topic is important, that science fiction has already prepared some of the ground, and that DeepMind and others are forming interdisciplinary groups. Safety and ethics workshops at conferences have matured in five years. On superintelligent systems, it is better to be too early than too late.

Fridman shouts out Ilya Sutskever, a friend Vinyals expects to see in person in a few weeks, including a running joke about whether Ilya's tweets are themselves language-model outputs. Vinyals recognizes the person he knew before social media and before AI was famous; the hard part is balancing that self with the weight of his words. Ilya sits in the tension between scientist and manager. His deep belief in scaling — a still-famous talk after the sequence-to-sequence paper, arguing that supervised data plus a large network would solve basically all the problems — put him years ahead. They rehearsed that talk in a hotel room. The original slides, Vinyals says, were even more controversial.

Does he buy Rich Sutton's bitter lesson, that general methods which leverage computation win? He agrees with the deep-learning philosophy of being data-agnostic, looking at the bytes, and learning everything on top. Scaling feels necessary for complex systems and possibly not sufficient; a couple of breakthroughs may still be required. Sutton also named search. Vinyals's experience is mixed: search is natural in Go, where a clear reward lets you prune, and less obvious elsewhere. AlphaCode, from a team overlapping AlphaStar, combined scale with a massive amount of search and reached human-level competitive programming. He is convinced about scale, less so about search as a universal ingredient. Hardware may need to change too — a "hardware lottery" if sparsity, for instance, is required.

Will we build, in his lifetime, an AGI that is undeniably human-level and then far beyond? Human-level, he is convinced. Far beyond is possible but harder to define from an imitation-learning starting point. Imitating humans at language and beyond feels very possible. Going beyond will take reinforcement learning and other tools. Go already shows superhuman play when the reward is clear. Writing a general reward that starts from human-level imitation and then exceeds it is not so clear in his lifetime. Human-level alone, especially if integrated into society at huge scale, is already a different kind of shift than horse-and-buggy to car.

He is afraid of unconstrained growth in digital entities the way humanity already strains a finite planet; energy may be the limit, since most systems are less efficient than we are. He is excited about automation that gives people knowledge and resources they do not have. On a multi-planetary future, he wants AGI to empower the species, perhaps with some hybridization. If the ratio of machines to humans is not at most one-to-one, he would not be happy. That preference, he says, is why he goes to work. Fridman takes it as evidence that Vinyals has passed the Turing test.