AI pioneer Yoshua Bengio implores Europe to build frontier models
BERLIN — Europe must build its own proprietary frontier models of artificial intelligence to avoid risks in relying on U.S. or Chinese models, world-leading AI developer Yoshua Bengio told POLITICO.
Working with open-weight AI models, built in China or elsewhere, would be a dangerous strategy for European companies and governments, locking in long-term economic and military dependence and multiplying cyberattacks, Bengio said.
The European Union is currently locked in a debate about whether to invest the billions necessary for homegrown frontier models, or to build applications on top of proprietary American or open-weight Chinese AI.
“Adapting models you don’t control and you may not have doesn’t work,” said the French-born Canadian computer scientist, who has been vocal in warning about the risks of the breakthrough technology that is largely based on the neural network systems he helped develop.
Asked whether that meant Europe needed to develop its own advanced models, Bengio’s answer was unequivocal: “Of course, because they are the key to everything.”
He said the current framework where developers and other AI users can freely use open models isn’t bound to last. “There’s a good chance we will not have the ability to take existing free models and just download them and use them in our industry, in even just as little as one year from now,” he said.
Bengio, a Turing Award winner along with the other two “godfathers of AI” Geoffrey Hinton and Yann LeCun, also had a stark warning about the risks of open-weight models in multiplying cyberattacks.
“It’s going to get worse if we continue on the current trajectory,” he said, amid news reports last week of German politicians’ iPhones being hacked, and supermarkets across Europe targeted by cyberattacks. “We should have seen this coming.”
Bengio has an easy explanation for the uptick. “It’s very simple: Because there are open-weight models that any criminal, any terrorist can download, and they don’t need to be an expert.”
In the face of both the long-term strategic and more immediate security risks — after comparatively “minuscule” investment in AI — the EU’s private and public sectors must now commit to major new spending to develop frontier models. He doesn’t think Europe should stop working with American providers, but it should not become entirely dependent on them.
“We need alternative options,” he said.
