For years, artificial intelligence has made companies scramble to create their own versions of it in ways relevant to their niche. But the new Deepmind model: Gato gives other AI models a run for their money. Unlike the other preceding models, this smarter and better AI model can allegedly do multiple tasks.
Nando de Freitas, one of Gato’s top researchers, has announced the model’s close to superhuman-level AI, suggesting that its artificial general intelligence will reach groundbreaking heights.
DeepMind announced the new model earlier this month and dubbed it a “generalist.” Gato can do approximately 604 different tasks compared to the other AI models. Gato can chat, caption images, play video games, and stack blocks. However, some had suspected Gato might not live up to the hype amidst their fascination with the new model.
What can the Deepmind model Gato do?
The researchers of DeepMind had trained Gato by collecting various tasks and modalities. These data were then put into a sequence in the form of tokens through a neural network. The researchers claim that any model with a general sequence can work in the next token prediction, so they selected a simple and scalable transformer specifically for Gato.
They used a 1.2 billion parameter transformer that focuses on decoding and comes with 24 layers and an embedding size of 2048.
Image datasets and natural language were also leveraged during the training. This gained Gato enough knowledge about agent experience in simulated and real-world conditions.
It works by creating an initial sequence through a tokenized prompt during the deployment stage. Then the environment gathers the initial observation, which is attached to the sequence. Gato understands each token, and once all tokens have been processed, Gato makes the action and delivers it to the environment.
The final step is the environment produces a new observation, and the circular process is repeated.
How Gato differs from other AI models
Freitas had Tweeted “The game is over” to announce that Gato might be the machine with human-level intelligence. He claims that achieving AGI or artificial general intelligence means making models like Gato bigger and better. However, some would beg to differ.
Although other models have stolen the limelight in AI, Gato’s researchers consider it a more advanced model than DALL-E and GPT-3, an image generator and text generator, respectively.
The way these models worked before was thought to be linear, which means focusing on one task at a time. But the newer models are now starting to combine various skills. Another DeepMind innovation called AlphaZero can also play shogi, chess, and Go. However, AlphaZero can only perform one task at a time. Once it learns to play chess, it must forget playing Go and other tasks it previously learned.
On the other hand, Gato is somewhat unique as it can learn several tasks simultaneously without discarding any of the other tasks. The drawback is that because Gato is a “jack of all trades” in task learning, it might not give a stellar performance compared to other models that focus on one task.
Dismissive researchers
A few external researchers dismissed Freita’s claim and said Gato is far from being intelligent. He claims that the hype in the world of AI is often accounted to a “triumphalist culture.”
Freita’s colleagues might agree with a few external researchers dismissing Freita’s claim. Both Scott Reed and Jackie Kay prefer not to give a straightforward answer.
Reed said, “I think most machine-learning people will studiously avoid answering. Very hard to predict, but, you know, hopefully we get there someday.”
