Working towards the safe development of AI for the benefit of all at Université de Montréal, LawZero and Mila.
A.M. Turing Award Recipient and most-cited AI researcher.
https://lawzero.org/en
https://yoshuabengio.org/profile/ ..
more
Working towards the safe development of AI for the benefit of all at Université de Montréal, LawZero and Mila.
A.M. Turing Award Recipient and most-cited AI researcher.
https://lawzero.org/en
https://yoshuabengio.org/profile/
Yoshua Bengio is a Canadian computer scientist, and a pioneer of artificial neural networks and deep learning. He is a professor at the Université de Montréal and scientific director of the AI institute MILA. .. more
Reposted by Yoshua Bengio
(1/4)
Thank you to all contributors for their dedication.
(19/19)
internationalaisafetyreport.org/publication/...
(18/19)
(17/19)
Attackers can still often find ways to evade them fairly easily. One initiative crowdsourced over 60,000 successful attacks against state-of-the-art models. When given 10 attempts, testers can still generate harmful responses about half the time.
(16/19)
(15/19)
(14/19)
For example, in 2025 multiple companies added safeguards after pre-deployment testing could not rule out the possibility that new models could assist novices seeking to develop biological weapons.
(13/19)
(12/19)
(11/19)
(10/19)
Misuse:
→ AI-generated content & criminal activity
→ Influence & manipulation
→ Cyberattacks
→ Bio & chemical risks
Malfunctions:
→ Reliability issues
→ Loss of control
Systemic risks:
→ Labor market impacts
→ Risks to human autonomy
(9/19)
At least 700 million people now use leading AI systems weekly. In the US, use of AI has spread faster than that of computers and the internet.
(8/19)
(7/19)
Leading models now achieve gold-medal performance on the International Mathematical Olympiad.
AI coding agents can complete 30-minute programming tasks with 80% reliability—up from 10-minute tasks a year ago.
(6/19)
3️⃣ Many safety measures improved, but remain fallible. Developers increasingly implement multiple layers of safeguards to compensate.
(5/19)
In 2025:
1️⃣ Capabilities continued advancing rapidly, especially in coding, science, and autonomous operation.
(4/19)
Acting too early risks entrenching ineffective policies, but waiting for strong evidence may leave society vulnerable to risks.
(3/19)
internationalaisafetyreport.org/publication/...
(2/19)
Reposted by David Autor, Brigitte Nerlich, Jack Stilgoe
(1/19)
www.weforum.org/videos/davos...