Yan Lakeun testified before the US Senate Intelligence Committee in Washington on 19 September 2023, DC Kevin Dietsch/Getty Images
Yan Lakeun, the main AI scientist of Meta (Meta) and one of the “Godfather” of deep learning in its field, is widely afraid that powerful AI models are dangerous, which is largely imaginary, because current AI technology is not human-level intelligence information-even the cat-level intelligence is not available. And when it is certain that AI will eventually reach the stage of the so -called Artificial General Intelligence (AGI), this timeline can be much higher than that of more researchers.
“We are really far from human level intelligence. There were stories about the fact that you can use an LLM (large language model) so that you can be instructed to make chemical weapons or bioopons. It is wrong.”
“Those systems are trained on public data. They can’t really invent anything, at least today,” they explained. “Some time in the future can be really smart to those systems that may be better to give you useful information, as much as you can get with a search engine. But this is not true today.”
The French scientist, who won the 2018 Turing Award with Geoffri Hinton and Yoshua Bengio for his contribution to Artificial Neural Network Research, said that even today the most advanced AI system has less general meaning than house cats.
“A house cat’s brain has about 800 million neurons. You have to multiply that 2,000 to get the number of synaps, or to get a connection between neurons, which is equal to the number of parameters in the LLM,” Lakeun said, seeing that the largest LLM is similar to the number of a cat’s brain. For example, Openai’s GPT-3.5 model, which gives power to the free version of Chatgpt, has 175 billion parameters. More advanced GPT-4, run on eight language models, each with 220 billion parameters.
“So perhaps we are in a cat shape. But why are those systems not smart as cats?” Lakeun asked. “A cat can remember, understand the physical world, plan complex tasks, some level arguments – in fact it is much better than the largest LLM. Which tells you that we are missing anything that they are at all ideologically large that are getting machines to be wise in animals and humans.”
For reference, a dog brain has about 2 billion neurons and a mature human brain has about 100 billion.
Before we reach the human-level AI (HLA), we have to reach the cat-level and dog-level AI. We are not anywhere. We are still not missing anything big.
– Yan Lakeun (@ILECUN) February 5, 2023
Asked that AI will actually cross human intelligence, Lakeun said, “Perhaps over 10 years, perhaps within 20.” Then he quickly said, “When I am saying this, I am taking a big risk, because every AI researcter in the history of AI for the last 65 years has been highly optimistic about those types of prediction … and they are wrong, obviously.”
Lakun joined Meta in 2013 as the director of the company as director of AI Research, named Vice President and Chief AI scientist. He is part of the Meta’s Fundamental AI Research (Fair) team, led by Canadian computer scientist Jole Pineau. Lakeun is also a computer scientist professor at the University of New York, who teaches part -time at the NYU Center for Data Science and Cortecent Institute of Mathematical Sciences.