Thank you, Madam Chair and members of the committee, for inviting me to testify today.
My name is Samuel Buteau. I have over a decade of technical experience in AI and now work at ControlAI, a non-profit organization that informs lawmakers of the risks posed by superintelligent AI and what the solutions are.
In March of this year, we submitted evidence to this committee outlining how the development of superintelligent AI represents a clear and grave national security threat to Canada. Since then, that threat has become harder to dismiss.
In April, Anthropic announced their new, most-advanced model, Mythos. During testing, Anthropic found that Mythos was able to gain unauthorized access to every major operating system and web browser, autonomously and outside of human control. Consequently, Anthropic took the decision not to release Mythos to the public, because, in their own words, “The fallout—for economies, public safety, and national security—could be severe.”
Nobody programmed Mythos to be the world's most capable hacker, because nobody programmed Mythos. AI systems are grown, not built, and nobody understands how they work. Anthropic's own CEO estimated that perhaps “we...understand 3% of how they work.” Importantly, that 3% figure is about what we understand today. Whilst Mythos is not being released publicly, it is still being used internally to develop the next generation of AI, and as AI development becomes automated, the ability for humans to understand AI systems and their capabilities will diminish even further whilst the speed of development rapidly accelerates.
We are currently on the path to developing systems that we do not understand and that are vastly more competent than humans in every domain. Experts say that superintelligent AI may arrive in two to five years. We will not be able to control these systems.
Mythos already threatens our national security, but we can manage it. It is clear that whatever comes in the future will not be manageable.
What can we do about this?
Firstly, the Canadian government should publicly recognize superintelligent AI as the national and global security threat that it is.
Secondly, just as it does for other national security threats, the government should begin closely monitoring the threat of superintelligence being developed and develop detailed scenario planning and doctrines.
Finally, but most importantly, Canada should take global leadership by forming a global coalition focused on preventing the development of superintelligent AI anywhere in the world. This coalition would form the basis for a mutual trust-but-verify regime where countries verify each other's supercomputers to ensure superintelligent AI is not being developed, whilst simultaneously monitoring the global flow of powerful chips.
Superintelligent AI is an equal threat to national and global security wherever it is developed. If China or any other country seeks to develop superintelligent AI, we must stop them at all costs. This trust-but-verify regime is an achievable and sustainable way to achieve this.
Thank you.