Terence Tao on AI risks: "The dual use capability of these most recent tools is indeed extremely concerning. There is a significant correlation between the capability of AI tools to solve pure mathematical problems, and their ability to potentially perform malicious acts such as exploiting security weaknesses in an online system, and the danger is very real that some major AI-triggered cybersecurity incident could occur in the near future which would be far more significant event in world history than, say, the production of a solution to all remaining Millennium prize problems. I am certainly supportive of efforts to slow down the development and release of even more capable models, and to impose genuine regulation on the industry, although the political environment currently is not exactly conducive to such policies, to put it mildly." terrytao.wordpress.com/2026/…

Finite time blowup with smooth forcing term for the incompressible porous medium, Boussinesq, and...

There’s some exciting very recent work by Alpöge and Buckmaster, building upon prior work by Córdoba and Martínez-Zoroa, in the general topic around the infamous global regularity problem for…

terrytao.wordpress.com

Sep 10, 2026 · 4:48 PM UTC

4
42
229
40,980
Sort replies: Relevant Recent Liked
Replying to @LechMazur
They honestly cant my friend, all of that will soon become clear.
Roy
@roydherbert
Sep 9
This is dated 31 Jan this year and represents the formation of a temporal kernel, which is what GPT 6.0 almost reached a few days ago. This isn't perfect as this was early days for Grok Imagine. #NavierStokes
3
1,818
Replying to @LechMazur
Yes, they are dual use. They can solve math AND advance science. I hope that Terry won't joint to choir of AI doomers aiming to sabotage the future of humanity out of fear. We need to get together, yes, patch and align models, and then use them to advance sicence!
1
12
1,373
Replying to @LechMazur
Does this imply mathematicians are dual use and we should restrict them talking?
1
242
Replying to @LechMazur
Agree with the capability point, disagree with the remedy. The correlation between mathematical ability and exploit-finding ability is real. But it cuts both ways. The model that finds the hole finds it for whoever runs it first.... that can be the person(s) wearing the white hat. Dont forget, we;ve been here before. The 1990s attempt to control cryptography via export restrictions failed. What made systems safer was open algorithms, open scrutiny and a security community with the same tools the attackers had. Security by obscurity lost back and I see no reason it wins now. Slowing frontier releases assumes the capability stays where a regulator can see it. IT WONT. A state or criminal group is the last actor a licensing system affects. Restrictions impact more those people who follow the rules. My position is not let it rip. It is that safety comes from equitable access plus governance at the point of use: measure what a system actually does when it acts, hold the vendor/deployer accountable for it, and make that measurement open so anyone can check it. That is a regulation you can enforce.
225