Twitxy Saved

Rohin Shah

@rohinmshah

AGI Safety & Alignment @ Google DeepMind
London, UK rohinshah.com Joined October 2017
96 Following 14K Followers 409 Posts

Latest posts 5

Rohin Shah @rohinmshah
For the inaugural essay collection of the DeepMind Institute, I’ve written with Anca about why chain of thought is a useful tool (not a silver bullet!), and how we can preserve it. I’d love to discuss more – it’s an important topic! bit.ly/the-case-for-r…
x.com/ancadianadraga…
institute.deepmind.com The case for reasoning transparency Today, we have a window into the thoughts of the most powerful AI models. We need to ensure it stays open.
Anca Dragan Verified account @ancadianadragan Sep 16
The idea that Chain of Thought is inevitably going to go away misses that we have agency in designing these models and can do something about that. @rohinmshah and I make the case that we should intentionally preserve monitorability (as part for of the launch of the DeepMind Institute) here: institute.deepmind.com/essays/the-cas…
View quoted post on X
8 replies 8 reposts 105 likes 11K views View on X
Rohin Shah @rohinmshah
My team at GDM is hiring once again!

Join us in the foothills of the singularity, developing the techniques we need during the ascent to the peak.

x.com/NeelNanda5/sta…
Neel Nanda Verified account @NeelNanda5 Jul 31
GDM AGI Safety is hiring! Open roles on all subteams, Lon/Bay/more

There's lots to do to reduce risks from AGI at Google, but we're bottlenecked on people. If you want to help, please apply!

I really love this team and my excellent colleagues. I'm excited to hire more!
View quoted post on X
5 replies 10 reposts 170 likes 15K views View on X
Rohin Shah @rohinmshah
Though the project itself is about the transparency of DiffusionGemma, I'm most excited about this as an example of what it could look like to assess the transparency of latent reasoning models, in a manner that lets us compare to autoregressive CoT

x.com/JoshAEngels/st…
Josh Engels Verified account @JoshAEngels Jun 19
Text diffusion models are fast, but are less transparent than today's LLMs because they do many forward passes before outputting text.

We audit the transparency of DiffusionGemma and find that the intermediates are interpretable. This recovers many of the benefits of CoT!

🧵
View quoted post on X
5 replies 7 reposts 63 likes 9.5K views View on X
Rohin Shah @rohinmshah
Excited to share our control roadmap!

There is a *lot* to learn from prior art in security against rogue human insiders. At the same time, AI does have several systematic differences, so we thought about how to adapt these approaches for AI.

x.com/MaryPhuong10/s…
Mary Phuong @MaryPhuong10 Jun 18
We're releasing the GDM AI Control Roadmap -- our plan for building internal security against potentially adversarial AI agents, as they grow harder to oversee and contain.

Paper: storage.googleapis.com/deepmind-media…
Blog: deepmind.google/blog/securing-…

🧵👇
View quoted post on X
3 replies 7 reposts 77 likes 8.9K views View on X