The Canada Brief

Tech · 19 d ago

Illustration of a neural network building another

Tech firms want AI to design better versions of itself

The term

recursive self-improvement

Many firms are aiming for recursive self-improvement, a process in which AI models design and train new, improved versions of themselves.

Google

75 per cent

Anthropic

Over 80 per cent

AI agents already write a growing share of new code: 75 per cent at Google and over 80 per cent at Anthropic, as of this spring.

What we know

  • AI writes code and runs experiments
  • Google used AI to refine a Gemini model's instructions

Not yet

  • Whether AI can choose which problems to solve next

No one has yet achieved true self-improvement, according to Hugo Larochelle of the Mila AI institute in Montreal. AI can code and run experiments, but cannot yet decide which problems to tackle next.

Yoshua Bengio

Maryse Boyce · CC BY 4.0 · Wikimedia Commons

And us?

Yoshua Bengio, who co-signed a warning of extreme risks from automated AI research, runs LawZero. The non-profit received $300 million from the Canadian and German governments to build an AI that could check other models.