LessWrong 2.0 Reader

COMMENT THREAD

If you comment anywhere other than here, Nathan will delete your comment.

This all seems very teleological. Do you have thoughts on what the teleology of the universe could be under this model? 

One optimal predictor for this data would be to maintain a belief over which of the three HMM states we are in

As well as inferring the HMM itself from the data.

Features to benefit people accused of X may benefit mostly people who have been unjustly accused.  So looking at the value to the entire category "people accused of X" may be wrong.  You should look at the value to the subset that it was meant to protect.

Personally, I like mentally splitting the space into AI safety (emphasis on measurement and control), AI alignment (getting it to align to the operators purposes and actually do what the operators desire), and AI value-alignment (getting the AI to understand and care about what people need and want). Feels like a Venn diagram with a lot of overlap, and yet some distinct non-overlap spaces.

By my framing, Redwood research and METR are more centrally AI safety. ARC/Paul's research agenda more of a mix of AI safety and AI alignment. MIRI's work to fundamentally understand and shape Agents is a mix of AI alignment and AI value-alignment. Obviously success there would have the downstream effect of robustly improving AI safety (reducing the need for careful evals and control), but is a more difficult approach in general with less immediate applicability. I think we need all these things!

That is a fair summary.

It might be tempting to think you could use multivariate statistics like factor analysis to distill garbage information by identifying axes which give you unusually much information about the system. In my experience, that doesn't work well, and if you think about it for a bit, it becomes clear why: if the garbage information has a 50 000 : 1 ratio of garbage : blessed, then finding an axis which explains 10 variables worth of information still leaves you with a 5 000 : 1 ratio of garbage : blessed. The distillation you get with such techniques is simply not strong enough.[1][2]

That doesn't seem like it. In many contexts, a 10x saving is awesome and definitely a 'blessed' improvement if you can kill 90% of the noise in anything you have to work with. But you don't want to do that with logs. You can't distill information in advance of a bug (or anomaly, or attack) because a bug by definition is going to be breaking all of the past behavior & invariants governing normal behavior that any distillation was based on. If it didn't, it would usually be fixed already. ("We don't need to record variable X in the log because X cannot change." NARRATOR: "X changed.") The logs are for the exceptions - which are precisely what any non-end-to-end lossy compression (factor analysis or otherwise) will correctly throw out information about to compress as residuals to ignore in favor of the 'signal'. Which is why the best debugging systems like time-travel debugging or the shiny new Antithesis work hard to de facto save everything.

Hah, literally just what I did.

Thank you, I was looking for a post.

Of interest, Daniel Dennett | From Bacteria to Bach and Back | Talks at Google in 2017. It's worth reviewing his other notable ideas and views of philosophy that he explored, from his Wikipedia page. I look forward to reading other testimonies of his influence and the effects of his work.

The implication of ICL being implicit BI is that the model is locating concepts it already learned in its training data, so ICL is not a new form of learning that has not been seen before.

I'm not sure I follow this. Are you saying that, if ICL is BI, then a model could not learn a fundamentally new concept in context? Can some of the hypotheses not be unknown — e.g., the model's no-context priors are that it's doing wikipedia prediction (50%), chat bot roleplay (40%), or some unknown role (10%). And ICL seems like it could increase the weight on the unknown role. Meanwhile, actually figuring out how to do a good job in the previously-unknown role would require piecing together other knowledge the model has — and sufficiently strong building blocks would allow a lot of learning of new concepts.