Yeah, I just don't think people should trust Tesla. Example: 3 years ago Car and Driver did a comparison of tested EV ranges vs. EPA claims, and Tesla was the 3rd worst about overstating their ranges. The only 2 that did worse only produced one EV model, compared to Tesla's 6.

Show Post

Blog post: “Sympathy for both sides of the egregious misalignment debate” www.alignmentforum.org/posts/DZaZ3f...

Show Post

One thing I have revised my point of view on a bit is the nature of hallucination. I used to be fond of saying that hallucination and non-hallucination are not actually distinct, that from the model's perspective, so to speak, it's doing the same thing whether it's hallucinating or not. 🧵

Show Post

Moreover, I underrated how powerful it is to give it a kind of escape hatch by post-training the tendency to say something like "I think I made a mistake. Let me go back." When it's in the unstable hallucination region, this becomes a strong attractor which can often point it back towards truth.

Show Post

I think this is a sense in which hallucination and non-hallucination are qualitatively different, at least in some domains. P(nonsense | nonsense context) >> P(nonsense | sense context). Hallucination begets hallucination, whereas truth is highly attractive, especially in domains with a lot of facts

Show Post