Yoshua Bengio has been thinking for some time about what would happen if the technology he helped develop became smarter than humans – and beyond our control.
“We could fundamentally create new types of living entities whose own preservation constitutes a greater value than ours,” he said.
Entities, he worries, which, with the help of robots, could one day “travel the planet”.
But Bengio, scientific director of Milathe Montreal-based artificial intelligence institute he founded in 1993 is increasingly considering policy solutions to ward off such a grim scenario.
Based in his spacious but unpretentious 1950s residence on the outskirts of Mount Royal Park, the AI guru clearly feels that time is running out, both when it comes to establishing his own bucket list. do for 2024 when it comes to containing increasingly powerful governments. artificial intelligence systems.
“It will be regulation first,” he says, “but eventually they will want to regain some control, perhaps initially by building their own infrastructure.”
This infrastructure includes building much more powerful computers, stacked with thousands of graphics processing units (GPUs), ideal components for training or testing large AI language models like ChatGPT.
He would like to see this class of machines built in Canada, funded by governments, so that public entities have the digital firepower to keep pace with the private tech giants they will be tasked with monitoring or regulating.
“I think the government will have to understand at some point, hopefully sooner rather than later, that it’s important for (them) to have that muscle,” Bengio said.
Bengio says such a supercomputing resource would cost about a billion dollars, and when he presented the idea to governments across Canada, the response so far has been: “we’re listening.”
“It’s a lot of money,” he admitted.
The UK is investing big
However, other governments around the world are already spending a lot of money to build more powerful public computers for AI, including the United Kingdom, which last fall announced a computer called Isambard-AI that would be built in the University of Bristol as part of building more powerful public computers for AI. of a £900 million plan to “transform the UK’s IT capacity”.
This computer would be 10 times faster than any other computer currently operating in the United Kingdom, and approximately 20 times faster than the most powerful publicly available supercomputer in Canada, the Narval, housed at the University of Montreal. Superior Technology School”, according to the director of the organization that manages it, Suzanne Talon of Calcul Québec.
The non-profit organization, funded by Ottawa, the province and academia, is one of five National Host Sitescalled clusters, for Canadian public supercomputers as well as other facilities at the universities of Victoria, Vancouver, Waterloo and Toronto.
Between its two supercomputers, Narval and Beluga, Calcul Québec currently has a total of 1,300 GPUs available to university researchers in Canada.
But that’s dwarfed by tech giants like Meta, which alone plans to have the equivalent of 600,000 of these GPUs by the end of the year in an attempt to develop artificial general intelligence.
Talon says that while bigger is not necessarily better, there is a lack of public computing resources available to Canadian researchers, with Calcul Québec’s servers already running full time except for maintenance.
“There’s no question that there hasn’t been a truly massive investment in AI in Canada and so the main question is: ‘What is the right amount?’ ” “, she says.
Although Talon says this could involve working on more “frugal” models, it is important that public research is not left behind.
“I mean we need to understand how AI works, so we need to have access to it,” she said.
Competition for GPUs
Siva Reddy, assistant professor of linguistics and computer science at McGill University and also a senior academic member at Mila, estimates that the total combined resources available for public AI research in Canada are about a tenth of what a The only major American technology company has done so – just for itself.
He says that even if researchers have access to GPUs through computing clusters like Narval, scale remains an issue.
While Reddy says it’s possible to run smaller AI models like Meta’s Llama with what’s currently available through public computing clusters in Canada, that’s not the case for larger models like ChatGPT .
“This model, forget it, we can’t run it in our clusters,” he said.
When it comes to, for example, research into systemic discrimination or bias, Reddy says that while analyzing AI models is possible with current public computing power, creating new models from scratch or adjusting existing models is not, unless it monopolizes resources shared by hundreds. of researchers across the country.
He claims that training a large language model like ChatGPT requires continuous access to 1,000 GPUs for 34 days.
“So imagine if you want to take this whole cluster, that means no one can do any work for a month,” says Reddy. “So we need a supercluster on its own for priority projects.”
Although he “absolutely” supports Bengio’s idea of building one or more public supercomputers to work on large language models, Reddy stresses that it is important to recognize the environmental impact of the energy required to operate them .
“Operating these systems also requires a lot of carbon emissions.”
Governments are working on a solution
James Peltier, who oversees the Research Computing Group and IT services at Simon Fraser University, site of the CEDAR supercomputer, says the cluster’s five sites do a “decent job” of meeting the needs of Canadian researchers, but that demand far exceeds supply.
“Currently, we can only meet about 20% of GPU requirements and 40% CPU requirements,” he says. CPUs are central processing units, which are not as well suited to training AI models.
But Peltier says that when it comes to spending on AI infrastructure, profit-driven private companies are not subject to the same budgetary constraints as governments, which have other public policy priorities to fund, such as fighting COVID-19 or coping with the crisis. opioid crisis.
“They don’t face the same competing challenges,” he says.
The office of Industry Minister François-Philippe Champagne says the government is working with partners to support Canadian researchers with the “secure and affordable computing capacity” needed for AI.
Spokeswoman Audrey Champoux says Champagne speaks regularly with Bengio, who is co-chair of the AI Ministerial Council, and that Bengio has raised the issue of AI computers and security.
In Quebec, a spokesperson for Economy Minister Pierre Fitzgibbon said that although Bengio had not raised the question of building an AI supercomputer on the scale of what is being built in United Kingdom, the minister was available to discuss such a project with him.
Regulate for the present and the future
In addition to advocating for greater public AI infrastructure, Bengio is also doubling down on calls for more democratic control and greater government regulation of the artificial intelligence sector.
Last summer, he appeared before a U.S. Senate subcommittee on AI regulation, and in the fall he was tasked with chairing a “State of the Science” review. Advanced AI Capabilities and Risks Reportfrom a global security summit hosted by the United Kingdom
In Canada, Bengio criticized the federal government for moving too slowly on Bill C-27, which includes provisions to partially regulate AI and is currently under study before the standing committee on industry and House technology.
That’s because some take issue with putting so much emphasis on the existential risks of AI by saying so. distracted more real problems that technology already causes, or that excessive rules could stifle innovation.
Asked about his warnings that catastrophic consequences would be dismissed as “AI hype,” Bengio says it is possible to mitigate the current harms caused by the technology, while preparing for more hypothetical risks.
“For me, there is no separation,” he says. “We need to manage risks and minimize damage. Now, tomorrow, five years, ten years and twenty years (from now).”