Yoshua Bengio

Stories

AI systems cross the creativity threshold

Founder and Scientific Advisor of Mila; Founder of LawZero

In January 2026, a Université de Montréal team compared 100,000 people with nine AI systems. GPT-4 beat the average human on a standard creativity test; the top 10% of humans still beat every model.

Updated Oct 26

The recursive loop begins

New Capabilities

As of September 2026: Led International AI Safety Report 2026; warns capabilities are outpacing safeguards while remaining optimistic about technical solutions at LawZero

In May 2025, DeepMind's AlphaEvolve became the first commercial AI to optimize its own training, shaving 23% off a critical computation kernel. A year later it was a product: Google runs it across data centers, chips, and Spanner, and six outside companies had published results.

Updated Sep 27

Notable Quotes

Frontier AI models already show signs of self-preservation in experimental settings today, and eventually giving them rights would mean we're not allowed to shut them down.

from The recursive loop begins · Updated Sep 27

These concerning behaviors increase with the reasoning capabilities of these systems.

from The recursive loop begins · Updated Sep 27

I'm more and more confident that it can be done in a reasonable number of years, so that there might be an impact before AI systems become so powerful that misalignment causes terrible problems.

from The recursive loop begins · Updated Sep 27

Since the release of the inaugural International AI Safety Report a year ago, we have seen significant leaps in model capabilities, but also in their potential risks, and the gap between the pace of technological advancement and our ability to implement effective safeguards remains a critical challenge.

from The recursive loop begins · Updated Sep 27