A Copycat or A Competitor? Gato Faces Off GPT-3


GPT-3The focus of GPT-3 and Gato will never attain AGI thanks to DeepMind’s amazing new ‘Gato’ AI 

DeepMind introduced a new multi-modal AI system today that can handle over 600 distinct tasks. It’s called Gato, and it’s possibly the world’s most astounding all-in-one machine learning package. Gato, as the agent is known, is a multi-modal, multi-task, multi-embodiment generalist policy. The same network can play Atari, caption images, talk, stack blocks with a real robot arm, and much more, deciding whether to output text, joint torques, button clicks, or other tokens dependent on its context.

While it’s unclear how well Gato will perform until researchers and people outside of the DeepMind laboratories get their hands on it, it appears to be everything GPT-3 wishes for and more. OpenAI, the world’s most well-funded artificial general intelligence (AGI) company, created GPT-3, a large-language model (LLM). However, before we can compare GPT-3 and Gato, we must first comprehend OpenAI and DeepMind’s respective business models.

Elon Musk created OpenAI, Microsoft has invested billions in it, and the US government could care less about legislation and control. Given that OpenAI’s entire objective is to construct and control an AGI (an AI capable of doing and learning whatever a human could if given the same access), the fact that all the business has managed to produce is a fancy LLM is a little concerning.

GPT-3 is impressive, don’t get me wrong. In reality, it’s on par with DeepMind Gato in terms of performance, but that judgment requires some nuance. Because no one understands how to make AGI work, OpenAI has taken the LLM road on its path to AGI.

It will take time to figure out how to go from deep learning to AGI, just as it took time for the fire to be discovered and the internal combustion engine to be invented. GPT-3 is an example of an AI that can at least mimic human behavior by generating text. What DeepMind has done with Gato is essentially the same thing. It’s taken something that operates similarly to an LLM and made it into a prestidigitator capable of over 600 different tricks.

It sounds wonderful that AI can perform all of these seemingly disparate activities because creating words is very different from directing a robot to us. However, this isn’t all that unlike GPT-3 being able to distinguish between the plain English text and Python code. This isn’t to suggest it’s simple, but it may appear to an outside observer that the AI can also brew a cup of tea or easily learn another ten or fifty activities, which it can’t. In essence, Gato and GPT-3 are both capable AI systems, but neither is capable of general intelligence. 

Here’s the issue: Unless you’re betting on AGI appearing as a consequence of a freak accident think of the movie Short Circuit it’s time for everyone to rethink their AGI timetables. “Never” is one of science’s only cursed words, so I wouldn’t say that. However, this makes it appear that AGI will not be developed in our lifetimes. 

For almost a decade, DeepMind has been working on AGI, and OpenAI 2015. And neither has been able to solve the first challenge on the road to AGI: creating an AI that can learn new things without needing to be trained. Gato, I believe, might be the most advanced multi-modal AI system in the world. But I believe DeepMind has simply produced a more marketable version of OpenAI’s dead-end-for-AGI notion. 

Final thoughts: DeepMind has accomplished something extraordinary that will almost certainly result in the firm making a lot of money. If I were the CEO of Alphabet (DeepMind’s parent company), I’d either spin Gato off as a standalone product or shift DeepMind’s focus from research to development. Gato has the potential to be more profitable than Alexa, Siri, or Google Assistant in the consumer market (with the right marketing and applicable use cases). 

Gato and GPT-3, on the other hand, are no more plausible AGI entry points than the aforementioned virtual assistants. Gato’s multitasking abilities are more akin to a video game console that can store 600 different games than a game that can be played in 600 various ways. It’s not a general AI; instead, it’s a collection of pre-trained, narrow models neatly packaged. 

If that’s what you’re after, that’s not a bad thing. But there’s nothing in Gato’s accompanying research report that suggests this is even a first step toward AGI, let alone a stepping stone. Companies like DeepMind and OpenAI have built up a lot of goodwill and cash by insisting that AGI was just around the horizon. At some point, that goodwill and capital will have to pay off in some way.