i have a blog now! in my first post, i summarised the safety frameworks released by openAI, anthropic and deepmind, and say a bit about what I thought of them. link below

pt 2

some news! I’ve joined the comms team at @AISecurityInst . AISI is doing some extremely important things and I’m excited to help tell the world about them.

I have weather on the brain bc it is insanely hot and I’m dying.

what would a benevolent superintelligence do about the weather.

here's a link to the paper: an arXiv version should be available in a few days - if you'd like to be sent a link when it is, please let me know!.

RT @AmmannNora: In this piece, @littIeramblings & I argue that technological solutions can lower the barrier to meaningful international de….

I go into a bunch more detail about how these ideas could be operationalised and name some historical precedents in my report!. but they are still fairly high-level, and I'd love to see more people explore this topic.

7) A plan for automated research. AGI labs discuss using AIs to accelerate their own research. The Project should make a concrete plan about if and how it will do this. What limits should it impose on AI-enabled acceleration? What safety research could AIs help with?

6) Intelligence and Scenario Planning Division. I suggest designating an office that would a) collect intelligence on adversary progress towards AGI and b) make detailed scenario plans for crisis scenarios that would require high-stakes diplomacy.

5) A designated verification project. A simultaneous project could research technologies for verifying future agreements on AGI development (a possibility that the project should actively prepare for!)

4) Internal audit & risk monitoring. I propose an internal audit function and a (separate!) risk monitoring team to check whether the project is keeping risks below manageable levels.

3) Board oversight. A Board could be established to approve high-risk actions, like training a model at a new compute threshold or internally deploying one for a new application.

2) Emergency protocols. If extreme risks look imminent, how should the project react? . I suggest "top-down" and "bottom-up" protocols that could be used in an emergency.

1) Protected information channels.The project should establish mechanisms to communicate information about emerging and imminent risks. There are existing systems these could be modelled on, incl the CRITIC system and the dissent channels that exist in many govt departments

Many have asked if a Manhattan Project for AGI would be a good idea, or if one is likely. I wanted to ask the question: if one does happen, how could we make it safer?. I propose 7 "safety features" for a hypothetical AGI project under US government control:.

There's a lot of talk about a "Manhattan Project for AGI". Suppose the US government actually decided to pursue such a project. How could we make it go well? . I explored this question for the @pivotal_org fellowship 🧵. (full paper below)

many doubt that international cooperation on AI is feasible. I disagree!. I co-wrote a piece with @AmmannNora for @aif_media on how we can build new technologies to help verify and enforce international agreements. link in next tweet

😌.

Customer due diligence .