IN BRUSSELS FOR A BRILLIANT TALK
Came to Brussels to talk about his book as provocative as it is fun and interesting to read, “AI does not exist! », Luc Julia, best known for having co-created the voice recognition of Siri, today on iPone, took his audience on a historical and scientific journey on the concept of artificial intelligence. Without addressing China directly, he paved the way for it by admitting that his quantum computers, the fastest in the world, would deploy a new future for an A.I. still in its infancy in the West.
First of all, who is Luc Julia? He is one of the two creators of Apple’s voice assistant, Siri. He was also Senior Vice-President and Chief Technical Officer of Samsung’s brand new artificial intelligence laboratory (SAIL) in Paris. Today, the brilliant computer scientist and mathematician works at Renault to develop AI in general.
In fact, the French inventor says that at the age of nine, he had already built a robot to make his bed! Then, he takes us from his village near Toulouse, in France, to the mythical Silicon Valley, in the footsteps of this famous “artificial intelligence” about which we hear so much nonsense, to understand what it’s all about. act exactly and anticipate what it may hold in store for us in the future.
“Because today I say it loud and clear: artificial intelligence does not exist! “, he launches in a fun and provocative way!
AI doesn’t exist?
Yes, it all started with a huge misunderstanding. In 1956, at the Dartmouth conference, mathematician John McCarthy convinced his colleagues to use the term “artificial intelligence” to describe a discipline that had nothing to do with intelligence. All the fantasies and misconceptions we are fed today stem from this unfortunate designation.
We understand directly that Mr. Julia’s objective is to debunk the received ideas and fantasies that surround AI today. “Afraid of evil robots taking over and eventually taking over the world?” Or “You are interested in AI and new technologies, but you would like to hear a different story”?
Beat the World Chess Champion
He insists. What we call generalized AI does not exist and with current techniques, it will never exist. What we call weak AI is today’s AI, it’s what we do with machine learning, deep learning and it has nothing to do with intelligence.
But it’s up to us to guess what he means by intelligence?
Take the example of Kasparov, he says, Russian genius who was beaten by Deep Blue in 1997. Is this proof of the intelligence of the machine? Not at all !
For what exactly have we done here? We modeled all the possibilities in chess (1053 in all). For a man, that’s a lot. But in 1997, it is possible to do it on Deep Blue. From there, Kasparov may be brilliant, he lost, inevitably.
Even the Game of Go?
Asians don’t believe it… Because with the game of Go, it’s a little different: you can’t model all the possibilities. However, some of them have been modeled and statistical models help to fill in the gaps.
Ah, after combinatorics, statistics.
It’s not as clean as chess but overall you have a lot more possibilities in memory than a human player, he says.
But for Julia, there is still no real intelligence! “It’s just a lot of data and a bit of statistics,” he remarks with delight.
Just look at the numbers: DeepMind is 1500 CPUs, around 300 GPUs, a few TPUs and 440 kWh. The human in front is 20 Wh. And he, the human, knows how to do a lot more than play go!
This proves that they are completely different approaches and that it is ridiculous to think that we are close to intelligence. CQFD?
Visual Recognition
But then … if we look at the side of vision? Not better he advances. Systems capable of recognizing cats have been built with a success rate of 95% by providing them with data from 100,000 cat images. But “a child only needs two pictures of a cat to recognize one for his entire life, with a 100% success rate. ! »
In effect ! And a driverless car if it “sees” a passerby in the distance with a T-shirt that looks like a STOP! sign, it will simply stop.
Make combinations from data provided by humans; if no data, make statistics; stupidly recognize an image or a sound from still data provided by man: is all this thought?
The Chinese have understood this (here I speak on my behalf): they call AI « ren gong zhi neng”, meaning the capacity for intelligence through human labor.
Medicine, Smart Home but…
But hey, there are areas in which AI has made progress? Yes, admits Luc Julia. Still the vision but here applied to medicine. Today, with a few tens of millions of images of breast cancer, he says, we can diagnose something very early with a success rate of 99.9%. A radiologist who will see a few thousand breast X-rays during his career cannot diagnose with such a success rate, at such an early stage. The other area in which there has been a lot of progress is voice recognition. Says the co-creator of Siri! And he also admits, “At home in California, my whole house is programmed by home automation. My daughter lives in a total smart home. When she goes to her friends and the house is classic, she is snobbish and contemptuous. I made it a monster! he said, laughing.
But despite these certain advances, there are still major problems. Let’s skip over the bugs of course… But more annoying because it brings technology back to the human, all too human: biased data sets. The scientists or programmers who select the famous big data do so according to their point of view… cultural, political, gender, etc. This can quickly turn a “chatbot” into a racist or misogynist!
The ecological cost
Luc Julia then asks the public of the Talk at Silversquare “Central” not to do AI for nothing! It’s stupid and it consumes a lot of energy! Remember, a human being thinks: 20 Wh. A machine: 440 kilo Wh!
“The ecological danger is real.” It is necessary to regulate as for the political questions of the Chatbot.
The future of AI?
But apart from this “smart object assistance”, which are ultimately not as “smart” as human beings but more practical, how can AI evolve? Current systems work thanks to two things: a mathematical model or a logical model. Things will get serious when we can effectively use the dreaded quantum computers…”This is no longer about mathematics, but about…physical processes” Without directly addressing China, he paved the way for it by admitting that its quantum computers, the fastest in the world, would deploy a new future to an A.I. still in its infancy in the West.
In 2016, Google’s AphaGo reportedly impressed Chinese military strategists by beating the Go world champion 7 times out of 11 games. A colossal achievement as the game combinations are almost “endless”. But will Asia have its revenge in the future?
Maybe Luc Julia would say, but he concluded with the challenge of challenges: “If we want to create a real AI, we have to copy the processes of the human brain. But the greatest scientists only understand 20% of how the brain works.”
What exciting challenges for the future!
