github.com/deepseek-ai/...

No doubt, it would help people to understand that there are tangible ways to experience the bias of a LLM’s training data…that they probably won’t enjoy.

WALL-E here we come

Wait, you don’t do that? 😂

Am I the only one who doesn’t keep my heads of lettuce on a plate in my fridge?

Let’s hope it doesn’t take them long to find their way over here

I think it’s sometimes not considered that something that developers do WITH code is have fun. No code doesn’t tickle the brain the same way. This doesn’t mean it’s bad, it just makes it boring for devs imo 🤷‍♂️

Fascinating. Thanks for providing that excerpt. Seems that, at least for now, promoting technique remains a crucial component of a performant LLM system.

I didn’t get a chance to read the entire paper, but it would be interesting to see the difference in results if the scratchpad wasn’t provided. I.e. does the model only exhibit alignment faking if given the opportunity to reason?

Thanks for the reply! Super interesting to hear the inside rationale behind these decisions 🔥

Really curious to know why it was ∞ in the first place?