Beyond GPT-3, DeepMind launched its new favorite Gato, but was questioned as "replacing the soup without changing the medicine"

Inspired by large-scale language modeling, Deepmind applied a similar approach to build a single "generalist" agent Gato, which is multi-modal, multi-task, and multi-embodied. (embodiment) features, capable of performing more than 600 different tasks. This artificial intelligence is arguably the most impressive all-in-one machine learning suite in the world to date.

DeepMind explained in the official blog:

"Gato is a multi-mode, multi-task, multi-component general artificial intelligence. Under the same network conditions, you can play Ya Dali games, captioning images, chatting with people, and controlling a robotic arm to stack blocks, etc. It will decide whether to output text, rotate the robotic arm, press a button, or make other marks based on the current situation."

While it remains to be seen how exactly Gato will perform, it appears to be more than everything GPT-3 could hope to achieve.

Image source: DeepMind official website

Does Gato really surpass GPT-3?

GPT-3 is a large language model (LLM) produced by OpenAI, a well-funded artificial general intelligence (AGI) company. Not only does it have billions of dollars in backing from Microsoft, but U.S. government regulation allows it to basically do anything.

The artificial general intelligence (AGI) it focuses on is a type of artificial intelligence that has human intelligence and can perform any intellectual task that humans can do. Some researchers call general artificial intelligence strong AI (strong AI) or complete AI (full AI), or that machines have the ability to perform general intelligent actions. Compared with weak AI, strong AI has a full range of human cognitive abilities.

Originally, OpenAI’s mission was to develop and control an AGI, but the reality is that everything the company managed to create was very fancy LLM, which was somewhat contrary to its “original intention”.

While GPT-3 is equally impressive as DeepMind’s Gato, the public’s standards for evaluating them require some nuance.

Because OpenAI is taking the LLM route on the road to AGI, the reason is simple: no one knows how to make AGI work. Just like it took a lot of time between the "discovery of fire" and the "invention of the internal combustion engine," figuring out how to go from deep learning to AGI didn't happen overnight. Still, GPT-3 can do some things that look like humans, such as generating text.

Gato, which bills itself as a "general artificial intelligence", does almost the same thing as GPT-3. It only integrates something that works very much like LLM into a "magician" that can perform more than 600 tricks. The highlight is using a single sequence model to solve all tasks, but requires increasing the amount and diversity of training data.

The Gato's ability to perform multiple tasks is more like a console that can store 600 different games than like a game you can play 600 different ways. It is not a general artificial intelligence as the introduction says, but consists of a bunch of pre-trained, narrow models neatly bundled together.

Image source: DeepMind official website

As Mike Cook of the Knives and Paintbrushes research group recently told Kyle Wiggers of TechCrunch:

“It’s exciting that an AI like Gato can do all these very different-sounding tasks, because to us, writing text and controlling a robot sound very different.

But in reality, this is not that different from GPT-3 understanding ordinary English text and Python code.

Not that this is easy to implement, but to an outside observer, it may sound like It’s an AI that can make tea or easily learn ten or fifty other tasks, but in reality, it can’t do any of those things.”

In short, both Gato and GPT-3 are Powerful artificial intelligence systems, but none of them have the capability of general intelligence.

When will the AGI era arrive?

DeepMind has been developing in the direction of AGI for more than ten years, and OpenAI started in 2015. But neither solves the first problem on the road to AGI: building an AI that can learn new things without training.

Unless you bet that the emergence of AGI is the result of luck, it is time to re-evaluate the progress of these companies in the field of AGI.

Perhaps, Gato may be the most advanced multi-modal artificial intelligence system in the world. But DeepMind adopts the same concept of making AGI a dead end as OpenAI, and just makes it more marketable.

Gato may be able to gain more traction in the consumer market through marketing than Alexa, Siri or Google Assistant. However, Gato and GPT-3 are not more viable entry points to AGI than the virtual assistants mentioned above.

If this is the type of AI you're looking for, then that's not a bad thing. However, in Gato's accompanying research paper, there is no evidence at all that it is moving in the right direction of AGI, let alone a stepping stone to AGI.

The above is the detailed content of Beyond GPT-3, DeepMind launched its new favorite Gato, but was questioned as "replacing the soup without changing the medicine". For more information, please follow other related articles on the PHP Chinese website!