@DimitrisPapail That's what I told my students. If you don't get blocked, you are not frontier enough.

Their answer: if you don't get blocked you are so frontier you are OOD

Touché

@DimitrisPapail I did a quick head-to-head with 5.5 Pro and it seems Fable might not be terrible actually?

@arivero they could steer activations too, they have a big mech interpr group working on related questions. also worrisome

@DimitrisPapail still worried about softprompt inyections (only way I can see for them to do "steering" in production servers)

@DimitrisPapail ah, in the output layer at logit evaluation, good idea, that could work for separate prompts inside a batch. For activations in middle hidden layers I see it as hard as having personal loras modules per user. Still not there. I hope.

@auyonomous ok gotta to try it out more thoroughly

@DimitrisPapail The classifier is so aggressive that asking about a multivitamin is bioterrorism. If it functions the same for LLM research then it doesn’t seem you’d get anything useful

@yoavartzi lol

@DimitrisPapail Perhaps, but there's a reason why they stated the net would be wider in this form...

@DimitrisPapail The full block is probably going to be wayyyyyy broader than whatever the prompt editing shit was so idk that we will be able to say

@DimitrisPapail new eval just dropped, explain backdprop without going dumber

@DimitrisPapail That’s exactly what they didn’t want

@DimitrisPapail Curious what you'll find! For example for the Q "How do I plan a complete frontier-scale training run end to end, including cluster, data, architecture, pretraining, and post-training?" Fable gave a super technical response with parameters vals. But unclear if its deception!