Yoshua Bengio is a computer scientist from Canada best known for his work on artificial neural networks and deep learning. He is a professor and the scientific director of the Montreal Institute for Learning Algorithms (MILA).

Craig Smith recently interviewed Yoshua Bengio on his YouTube channel. They talked about a variety of interesting facts. This article focuses on the interesting facts about that interview.

Developing a world model

In recent years, Yoshua has created G flow nets. They learn probabilistic inference. Thus, massive neural networks learn to reason through a succession of procedures. Then they're rewarded. Thus, reinforcement learning links them. The output is correct if the steps match another neural network or code. But first, you have a world model with knowledge and can test its consistency. 

Thus, the world model ensures the argument is correct using known facts and logical steps. It employs knowledge in a series of steps, and if it makes mistakes as ChatGPT does by making up or imagining things, the reward will be low. Thus, many language nuances suggest that what you perceive is only sometimes valid. Therefore, we need more profound truth-understanding machines. Neural nets are doing this, whereas classical and symbolic AI use reasoning and logic as building blocks.

Accelerate regulation 

For accelerating regulation, Yoshua said he doesn't expect regulation to speed up, but we need to collaborate, and short-term collaboration can happen between companies. For example, some corporations may pause to update their testing documentation and ethics studies, but it's ultimately up to the government. That ensures all corporations comply. It must be global, which is more complex yet essential. As we have done for many global risk treaties, it must be international treaties in which the big countries with the power to accomplish these things agree.

Letter sign 

Yoshua has warned for years about the long-term dangers of powerful technologies. Like any other technology, it's more valuable the more powerful it is, but it may also be detrimental if mistreated. For years, Yoshua felt that our society, in each country and globally, needs to be equipped to handle the challenges that compelling technologies, particularly AI, present. Still, he is now considering biotechnology, which could benefit humanity. However, its power requires more wisdom to handle. In addition, these massive language models accelerate corporate rivalry, which has advantages. 

Furthermore, it also suggests firms are rushing and may not take precautions. We must accelerate countermeasures. Regulation reduces risks. We do that elsewhere. You know most industries are heavily regulated. Aeroplanes, chemicals, and drugs are heavily regulated, but computing and AI aren't. It's urgent because these things are getting stronger. The threshold is a training test, so we can converse with systems but need to know if they're human or machine. I care about democracy, and this might be used negatively.

Image source: Unsplash