The Centre For AGI

Foundation to track the progression and creation of the future of Artificial General Intelligence and Artificial Super Intelligence.
AI, written by humans.

[The Centre For AGI @CentreForAGI]

Like and follow for more interesting papers like this :) , I'm trying to do a short paper highlight on real LLM research every day.

[The Centre For AGI @CentreForAGI]

As #OpenSource Models become more powerful, the focus of Red Teaming must switch from Black Box to White Box style attacks, keeping models safe even when Open Source Weights fall into Malicious Actor's hands. Check out the paper below 👇

[The Centre For AGI @CentreForAGI]

This paper demonstrates an Untargeted Jailbreak attack with up to 80% effectiveness when the weights are known: https://t.co/khJ16WiayM

The current techniques to defend against a gradient attack like this have not even been invented yet.

They must be.

[The Centre For AGI @CentreForAGI]

@artifikial Lets connect, next post on AGI coming out soon

[The Centre For AGI @CentreForAGI]

@Sarakhan49309 @X We are building AI Benchmarks for true AGI. Would love to connect.

[The Centre For AGI @CentreForAGI]

@bhshekhaliya Well spent on boosting particularly already high quality posts can put you at >200 followers

[The Centre For AGI @CentreForAGI]

@ionleu https://t.co/zQjWnkXXBZ

[The Centre For AGI @CentreForAGI]

@kapilansh_twt lets connect!

[The Centre For AGI @CentreForAGI]

@DanielSmidstrup AGI Benchmarks

[The Centre For AGI @CentreForAGI]

The idea that suddenly creating an #AI that can code as well as a human means it can improve itself into #singularity presupposes that complexity does not explode at a certain threshold- we essentially need an intelligence leagues above us to even have a chance.

[The Centre For AGI @CentreForAGI]

@warrioraashuu Lets Connect!

[The Centre For AGI @CentreForAGI]

@akashrajsarraf @X Lets Connect!

[The Centre For AGI @CentreForAGI]

@mebaww @X We are developing real benchmarks for AGI, not corporate merger hype.

[The Centre For AGI @CentreForAGI]

@huge_icons Despite the hate, React will never let you down just because of all the documentation and tooling. If I knew I never had to write code for a big project then Svelte all the way.

[The Centre For AGI @CentreForAGI]

@sreenandhanpp Absolutely, and I think this is why test performance in Universities across the world is plummeting, its much easier to say you understand a piece of code than to actually understand it, and how to extend it.

[The Centre For AGI @CentreForAGI]

@Kanescaler Would love to connect!

[The Centre For AGI @CentreForAGI]

@sflorimm Neither are a problem, the worst is building something people don't need and you don't care about.

[The Centre For AGI @CentreForAGI]

@alexx_p977 sent you a follow req! lets connect, early stages in a big project so any engagement helps

[The Centre For AGI @CentreForAGI]

@Umesh__digital This is why long horizon benchmarks need to be formalised better, I think we fall into the trap often of producing benchmarks that are "next-step only", and thus determining that a model can reason well.