What is missing from the “AI world”? Oxford University professor Michael Wooldridge: The real world

The infinite monkey theorem states that if a monkey were allowed to press keys randomly on a typewriter for an infinite amount of time, it would almost certainly be able to type any given text , such as the complete works of Shakespeare.

In this theorem, "almost certainly" is a mathematical term with a specific meaning, and "monkey" does not refer to a real monkey, but is used to refer to an abstract device that can generate an infinite sequence of random letters.

Figure | A chimpanzee typing randomly, given enough time, can almost certainly type out every book in the French National Library. (Source: Wikipedia)

This theory shows that it is wrong to infer that a large but finite number is infinite. Even if the observable universe is full of monkeys typing non-stop, the probability that they can type out a "Hamlet" is still less than 1/10^183800.

Moreover, even if countless monkeys were given infinite time, they would not learn how to appreciate the poetic words of a bard .

" The same is true for artificial intelligence (AI) ," said Michael Wooldridge, professor of computer science at the University of Oxford.

Photo: Michael Wooldridge

In Wooldridge's view, although AI models such as GPT-3 have demonstrated amazing capabilities with the help of tens or hundreds of billions of parameters, their problem lies not in the size of processing power, but in the lack of experience from the real world .

For example, a language model might learn “rain is wet” very well, and when asked whether rain is wet or dry, it will likely answer that rain is wet, but unlike humans, this language model has never actually experienced the feeling of “wet”. To them, “wet” is just a symbol that is often used in conjunction with words like “rain” .

However, Wooldridge also emphasized that the lack of knowledge of the real physical world does not make AI models useless , nor does it prevent an AI model from becoming an empirical expert in a certain field, but it is indeed doubtful to think that AI models have the same capabilities as humans in issues such as understanding.

The related research paper titled “What Is Missing from Contemporary AI? The World” has been published in the journal Intelligent Computing.

In the current wave of AI innovation, data and computing power have become the foundation for the success of AI systems: the capabilities of AI models are directly proportional to their size, the resources used to train them, and the size of the training data.

Regarding this phenomenon, DeepMind research scientist Richard S. Sutton previously stated that the "painful lesson" of AI is that its progress is mainly due to the use of larger and larger data sets and more and more computing resources.

Figure｜AI generated works (Source: Wired)

Wooldridge affirmed the overall development of the AI ​​industry: “Over the past 15 years, the pace of progress in the AI ​​industry, and in the field of machine learning (ML) in particular, has repeatedly surprised me: we have to constantly adjust our expectations to determine what is possible and when it might be realized.”

However, Wooldridge also pointed out the problems in the current AI industry. "While their achievements are commendable, I believe that most current large-scale ML models are limited by one key factor: AI models have not truly experienced the real world.

In Wooldridge's view, most ML models are built in virtual worlds such as video games . They can be trained on massive data sets, but once they are applied to the physical world, they lose important information and are just AI systems that are detached from the entity .

Take the artificial intelligence that powers self-driving cars. It’s not practical for self-driving cars to learn on their own on the road, and for this and other reasons, researchers often choose to build their models in virtual worlds.

“ But they simply don’t have the capacity to operate in the most important environment of all, which is our world ,” Wooldridge said.

(Source: Wikimedia Commons)

Language AI models, on the other hand, are subject to the same limitations. They have arguably evolved from ridiculously terrible predictive text to Google’s LAMDA, which made headlines earlier this year when a former Google engineer claimed that the AI ​​program LAMDA was sentient.

“Whatever the validity of this engineer’s conclusions, it’s clear that he was impressed by LAMDA’s conversational abilities — and for good reason,” Wooldridge said, but he does not believe LAMDA is sentient, nor is the AI ​​anywhere near such a milestone.

“These base models demonstrate unprecedented capabilities in natural language generation, can generate relatively natural-sounding snippets of text, and appear to have acquired some commonsense reasoning capabilities, which is one of the major events in AI research over the past 60 years.”

These AI models require massive input parameters and are trained to understand them. For example, GPT-3 is trained using hundreds of billions of English texts on the Internet. The combination of large amounts of training data and powerful computing power allows these AI models to behave similarly to the human brain, going beyond narrow tasks, beginning to recognize patterns, and establishing connections that seem unrelated to the main task.

(Source: OpenAI)

But Wooldridge said the underlying models are a bet that “they are trained on huge amounts of data to have useful capabilities across a range of domains, and can then be specialized for specific applications.”

“Symbolic AI is based on the assumption that ‘ intelligence is mainly a knowledge problem ’, while the underlying model is based on the assumption that ‘ intelligence is mainly a data problem ’. Inputting enough training data into a large model is believed to have the potential to improve the model’s capabilities.”

Wooldridge believes that this “might is right” approach continues to expand the size of AI models in order to produce smarter AI, but ignores the real physical world knowledge needed to truly advance AI.

“To be fair, there are some signs that this is changing,” Wooldridge said. In May, DeepMind announced Gato, a model based on large language and robotics data that can operate in simple physical environments.

“It’s great to see the foundational models taking their first steps into the physical world, but it’s only a small step: the challenges to making AI work in our world are at least as great as those to making it work in simulation, and likely greater.”

At the end of the paper, Wooldridge wrote: "We are not looking for the end of the road to AI, but we may have reached the end of the road at the beginning."

What do you think about this? Feel free to leave a message in the comment section.

References:

https://spj.sciencemag.org/journals/icomputing/2022/9847630/

https://www.eurekalert.org/news-releases/966063