Louis retweeted
The Literature and Linguistics Analysis Benchmark (LALA) assesses how closely LLMs approach human-like linguistic competence through open-ended questions written by professors and experts at top universities. Our most important finding is that linguistic competence is not cumulative and additive; sometimes more post-training makes a model less intelligent, suggesting general intelligence is impossible to achieve from an LLM, let alone superintelligence.
The benchmark has two major findings and several significant methodological contributions to the emerging field of LLM benchmarking.
it says so much that gyges, in a highly critical post, suggests that the basic bayesian term “priors” originates in or owes anything to yudkowsky (cf. Edwards, Lindman, Savage 1963 below). what of rationalism isn’t like this?
An attempt to document important facts concerning AI Safety, Eliezer Yudkowsky, Rationalism and San Francisco Effective Altruism
verysane.ai/p/ai-safety-is-m…
it's not worth my time to read through [blogs] to pick out the good stuff even if there are a few small nuggets of goodness or competent people buried here and there, one might say
It's a pretty pathetic response to anything to "cheerfully admit to not having read" it, and then make claims about paradoxes in it that are really just paradoxes in the thing you made up as a substitute for reading it. (I also don't like CEV, but) it absolutely does not say that you conduct an "extrapolation of the ruling elite values". This isn't even a fair parsing of Aaronson's summary there, and that's already a *very* lossy compression of the CEV paper.
friendly ai partly has to entail friendship or an approximation of it. i often want an ai model at least to tell me when it believes or notices that my revealed preferences aren’t in line with obvious norms or my own values. this could range from nudges during ordinary coding sessions to a full blown sartrean existentialist mode where it reminds me that i will have the face i deserve.
i don’t mind this vision up to a point. not at all. yes, we all want to be the best version of ourselves, and the pursuit of virtue, including intellectual virtue, has to be maintained almost no matter what. we would become slugs without it and in the end probably pay for allowing ourselves to become so helpless. apart from that danger, the key problem is that the path to this set of outcomes is deeply precarious. there are many reasons to believe that this is not how things will turn out.
automating all knowledge work is simply an idiotic goal. aristotle, kant, and mill would all despise it or at least see it deeply pessimistically, i suspect. some know this already. i don’t think it is what, e.g., oai is trying to do.
it’s at least arguably within a person’s rights not to value their own life, though i think they should. but not valuing anyone else’s life makes it a we problem.
No. The good outcome is that we create something greater than ourselves that can carry on our legacy in a manner greater than any of us can fathom.
While I hope this entity leaves us a pleasant place to live out our lives the fate of hominids is ultimately a drop in the bucket compared to the enormous moral value of the superintelligence itself.
it’s easy to say this kind of thing lightheartedly when you’re not facing the prospect of being economically ruined by it.
