@jeremyphoward

–

16 July 2025 11h09

Similar thoughts about this paper here from Anthropic's @janleike BTW:

Publishing Research: Include Both Supporting and Opposing Perspectives

By

–

16 July 2025 11h09

IMO if you're going to go to the trouble to get feedback about your paper before publishing (which is a great idea!) you should publish *both* sides. Just listed prominent supporters gives the wrong impression.

System Prompts as Prioritized Instructions Framework

By

–

16 July 2025 4h46

So really I think "follows prioritised instructions" might be a better framing, where the system prompt is taken as the highest priority instruction?

Redefining AI Misuse Requests: Safety vs Developer Intent

By

–

16 July 2025 4h45

"requests for misuse" only makes sense, to me, if defined as "request to do a thing that the system prompt asks not to do". This is not necessarily harmful or unsafe. It's just not what the developer wanted.

User Intent Framing and AI Safety Safeguards Context

By

–

16 July 2025 4h44

I think I disagree with your framing. "The user is intentionally asking for something that could lead to harmful outcomes if no safeguards are in place" could refer to no prompts, or all prompts, depending on context.

Model Controllability: Essential for Safe AI Deployment

By

–

15 July 2025 21h31

For someone wanting to deploy a model, what they really need to know is how controllable it is, so that they can make it behave in the way they require.

15 July 2025

Reconsidering Safe and Unsafe Language in AI Safety

By

–

15 July 2025 21h30

I wonder if you'd be open to avoiding words like "safe", "unsafe", and "harmful" requests. Requests do not have any of those attributes. Refusing a task can also be an "unsafe" behavior. It's entirely context dependent.

15 July 2025

System Prompt Changes Cannot Fix Underlying Model Bias

By

–

15 July 2025 21h16

This is the key issue – changes to the system prompt can only paper over the underlying behavior that appears to have been trained into the model. @simonw those system prompt changes you highlighted won't really fix the underlying bias.