#47 Guru Jeff Dean: Artificial intelligence still has things to learn
February 22, 2022
Play Video

Artificial intelligence (AI) is supposed to (one day) be as smart and responsive as a human being. But if that’s the goal, why does it not work like a human brain? If we want the technology to emulate our intelligence, then why are we not building it in our image?

Jeff Dean knows AI. He’s the head of AI Research and Health at Google and he’s been with the tech giant for over 20 years. He has seen huge computing transformations in his time and one of the biggest ones is AI. While it’s come a long way, Jeff thinks we’re doing AI wrong in three specific ways.  


Artificial Intelligence has a one-track mind. For now. 

Most neural networks are trained to do one thing and only one thing. Many can accomplish extraordinary tasks, but none can work together and leverage knowledge from other networks. For something that is supposed to be the next generation of computational intelligence, that’s a pretty illogical problem to have. 

The way AI works today would be like forgetting everything you know about everything before you try a new task. Did you try to perfect your sourdough bread-making skills during the pandemic lockdown like everyone else? Did you have to relearn how to read a recipe, hold a spoon, or turn on an oven? Hopefully not, but this is exactly how today’s artificial intelligence works. 

Instead, Jeff says we should be training multitask models that can do thousands or even millions of different tasks. That way, when a new task comes along, it can be added to the model and it can begin to leverage the knowledge acquired through the other tasks. 


AI can only use one sense at a time. 

Most of the models today deal with a single modality of data only, either images, text, speech, or video, but not all of these all at once. But that’s not how you work, is it? You’re constantly using all of your senses to learn from, react to, and gather information to make decisions. 

AI models can, and should, be built in the same way. They should take in all different types of modalities of input data but then fuze them together so that regardless of whether the model sees the word ‘leopard’, sees a video of a leopard, or hears someone say the word ‘leopard’, it triggers the same response in the model. 


Artificial intelligence models are pretty dense.

Today’s AI models are really dense. Models are fully activated for every single task that we want to accomplish, whether that’s a really simple or a really complicated thing. This is also not how the human brain works. We have different parts of our brains that are good at different tasks and we call on those parts when they are needed. 

Jeff thinks AI models can work in this way if we want them to. Instead of a dense model, we can have one that is sparsely activated so that for particular tasks we can call upon different parts of the model. While it’s being trained, the model can also learn which parts of itself are good at which things so that it can continuously identify what parts it wants to call upon. 


Rome wasn’t built in a day, and AI won’t be built overnight. But making strategic changes now to how we approach AI development could very likely produce an even more powerful tool.