John Burden

AI Evaluation

[John Burden @JohnJBurden]

Check out the paper on arxiv!
https://t.co/XSlIJaYZ1K

[John Burden @JohnJBurden]

New preprint! 📜
We apply classic visual search paradigms from cognitive psychology to multimodal LLMs. Instead of just benchmarking outputs, we probe the human-likeness of their responses. How do models like GPT‑4o, Claude and Llama respond to simple vs. compound visual features? A thread 🧵

[John Burden @JohnJBurden]

We think this type of work is valuable --- rather than identifying whether a model *can* perform some task, we want to get a better picture of the computational/cognitive processes driving performance. This sits somewhere between benchmarking and Mechanistic Interpretability---we try to see what the model is doing at a higher level of abstraction.

[John Burden @JohnJBurden]

@TheZvi Hey now, it can't solve the 555th problem! Definitely can't solve that one yet.

[John Burden @JohnJBurden]

@dioscuri I admit I was surprised and saddened when I found out his level of analysis was really the only thing he had time to do. It's one of those things that's really simple in retrospect but that takes a special something to come up with--hallmark of a good idea.

[John Burden @JohnJBurden]

@sebkrier Yeah but in the meantime it's been renamed to "governance hacking" or something.

[IJCAIconf @IJCAIconf]

#IJCAI2025 John Burden, University of Cambridge, delivering their talk #8954 on Paradigms of AI Evaluation: Mapping Goals, Methodologies and Culture at the Machine Learning (1/4).

[John Burden @JohnJBurden]

@Jsevillamol This seems valid... But also somewhat sad. I don't want AI for business, I want it for general wellbeing and reducing the load on working people.

[John Burden @JohnJBurden]

"Just-In-Time" AI Evaluation: use the model for what you want, when you need it. If it works, use it. If it does t work, don't use it.

[John Burden @JohnJBurden]

@PrinceVogel I think that's fair.

[John Burden @JohnJBurden]

@PrinceVogel What if you're pro it because you think it could be the catalyst for interviews to ask more meaningful questions/ tasks?

[John Burden @JohnJBurden]

@fchollet Shameless self-plug here on this topic
https://t.co/srdu6C9iwl
We make a very similar argument in this paper from a legal perspective on the right to have access to unpolluted data.

[John Burden @JohnJBurden]

@BlackHC @OwainEvans_UK Moreover there's the whole " inference between the gaps" thing (another banger from @OwainEvans_UK ). What does an inferential-world look like without any traces of these papers/stories?

[John Burden @JohnJBurden]

@BlackHC @OwainEvans_UK I agree, but I'm not sure how you do this without guaranteeing some kind of screwup? Thinking about canary codes from big bench etc.

[John Burden @JohnJBurden]

@RosieCampbell Super interesting and a bit disturbing about human psychology. But what do we do about this a few years down the line when (presumably) this effect is much more widespread?

[John Burden @JohnJBurden]

@HumanHarlan Imagine growing up thinking any of this is vaguely normal or well understood

[John Burden @JohnJBurden]

@robinhanson I was waiting to see what term you were using CPR for and then have a good laugh at myself for think it was... Whatever CPR actually stands for. Man am I disappointed.