“Some thoughts on GRAFT and follow-up work” by shawnghu

This is a link post. Everyone has some actions that come naturally to them, and some actions that are more unnatural, and some actions that wouldn’t even occur to them. For me, blogging comes naturally. Which is not to say that it's easy, but the challenge of it is one I readily set out on and feel confident in my ability to overcome. And for me it is a natural response to all kinds of situations. Unsatisfied with the going views on AI policy? Blog. Want some impact? Blog. Want to make an artistic statement? Blog. Want a boy to like you? Blog. Want a job? Blog. Want to change the norms of American society? Blog. Want to start a revolution? Blog. Another thing in this category for me is throwing parties. I organize social events at the drop of a hat, and feel pretty comfortable in some sense, even though the dance of it does involve various known discomforts. This is important to notice because you can turn other things into blogging and parties. Once I had agreed to give a keynote speech, and was seriously floundering constructing one. My friend suggested I instead write a blog post on [...] ---

Or if you prefer, "@Claude", "@Opus", "@Astra" @Grok is a popular Twitter/X feature where if you tag Grok in a tweet, Grok will respond in a comment. The stereotypical use case is "@Grok is this true?" But you can also ask things like "How does this compare to layoffs in 2025, 2024, and 2023?" and "How would you verify AI in the domain of AI research?" Notwithstanding that Grok is often recruited in a charged or condescending way to win arguments, I think it's a pretty good feature: It's requires no effort to invoke an LLM with the context of the tweet thread. It provides a neutral invocation where you're guaranteed that the person didn't insert other prior context to get the answer they want. It means you get the first output, and know it's not the case that someone kept rerunning the query until they get the response they wanted. In short, you get convenience and a neutral fact checker. In launching this feature on LessWrong, I'm not fully sure how people will use it and I'm hoping to see some novel valuable use-cases. Likely things like little experiments, retrieving useful content, and similar. Going beyond @Grok's [...] The original text contained 1 footnote which was omitted from this narration. ---

The initial promise of crypto was digital money without a middleman. Censorship resistance. Financial sovereignty. Anonymity. Satoshi Nakamoto's 2008 whitepaper describes a payment system with no bank, no processor, and no government standing between you and the person you're paying.

tl;dr We fine-tune 12 open-weight models on correct, obviously incorrect, and subtly incorrect advice across 8 narrow domains We replicate the OpenAI result and find that subtly incorrect data produces more emergent misalignment than obviously incorrect data +10.6 points pooled over all model-domain pairs (95% CI [9.1, 12.0]) We also extend this analysis to Petri evals and find uplift in fabrication and concerning behavior Fabrication: 2.6 → 6.7 (obvious) and 8.7 (subtle) on the 10-point judge scale Concerning: 3.7 → 7.5 (obvious) and 8.2 (subtle)

The full text is finally here, so we are giving it the full RTFB of model card treatment.