Gato Might Not Be the Sharpest in AGI Race: Looking Back at AlphaZero

It's time to ask if Gato has a better chance is being an AGI than AlphaZero.

Gato, as the agent is known, is the generalist AI of DeepMind that can execute a wide range of jobs that humans can, without specializing in a single skill. Gato can do over 600 various things, including play video games, caption photos, and move real-world robotic arms. It is a generalist policy that is multimodal, multi-task, and multi-embodiment.

How Does Gato Work?

Gato operates by normalizing and modulating all the inputs and data streams from various jobs into flat token sequences. It can interact with languages, and images, play games and interact with mechanical objects when treated as weights. It is accomplished by sampling the tokenized weights from the first step into autoregressive action vectors one token at a time. The action is decoded and delivered to the environment, producing a new observation, when all tokens composing the action vector have been tested (as stated by the environment's action specification). After that, the method is repeated. Within its 1024-token context window, the model is always aware of all previous observations and actions. Gato's main design principle is to train on as many different types of data as possible, including photos, text, proprioception, joint torques, button presses, and other discrete and continuous observations and activities. It serializes all data into a flat series of tokens to facilitate the analysis of this multimodal input. Gato can be trained and sampled from this representation in the same way that a normal large-scale language model can. Sampled tokens are constructed into dialogue responses, captions, button presses, and other actions based on the context during deployment. The tokenization, network design, loss function, and deployment of Gato are described in the subsections below.

Is It Being Overhyped?

While Gato is undeniably fascinating, some researchers have gotten a bit carried away in the week since its release. One of DeepMind's top researchers and a coauthor of the Gato paper, Nando de Freitas, couldn't contain his excitement. "The game is over!" he tweeted, suggesting that there is now a clear path from Gato to artificial general intelligence, or AGI, a vague concept of human- or superhuman-level AI. The way to build AGI, he claimed, is mostly a question of scale: making models such as Gato bigger and better. Unsurprisingly, de Freitas's announcement triggered breathless press coverage that DeepMind is "on the verge" of human-level artificial intelligence. This is not the first-time hype has outstripped reality. Other exciting new AI models, such as OpenAI's text generator GPT-3 and image generator DALL-E, have generated similarly grand claims. For many in the field, this kind of feverish discourse overshadows other important research areas in AI.

Is Gato better than Alphazero?

The strength of "Gato" lies in one key thing: it never forgets what it has been taught.

In recent years, many AI models have begun to combine different skills. Examples include "DALL-E" or "Imagen," capable of generating images from a simple text description. Recently, the French artificial intelligence model "NooK" managed to beat several world bridge champions.

"AlphaZero," another model already built by Deepmind, has learned to play Go, chess, and shogi. But there's one difference: "AlphaZero" could only learn one task at a time. After learning to play a strategy game, it had to forget what it had learned to move on to the next game.

What is AlphaZero?

AlphaZero was developed by the artificial intelligence and research company DeepMind, which was acquired by Google. It is a computer program that reached a virtually unthinkable level of play using only reinforcement learning and self-play in order to train its neural networks. In other words, it was only given the rules of the game and then played against itself many millions of times (44 million games in the first nine hours, according to DeepMind). AlphaZero uses its neural networks to make highly advanced evaluations of positions, which negates the need to look at over 70 million positions per second (like Stockfish does). According to DeepMind, AlphaZero reached the benchmarks necessary to defeat Stockfish in a mere four hours.

Can AlphaZero be Called AGI?

In simple words, the answer will be a 'no'. In the race to achieve AGI, AlphaZero can only be considered a narrow superintelligence in that it exceeds all human performance in a single problem. However, AGI does not only consist of superintelligence as it is not only about surpassing human task abilities but rather also requires the concept of general intelligence or what can be called common sense.