AABoyles

# Posts

Sorted by New

Developmental Stages of GPTs

It would also be very useful to build some GPT feature "visualization" tools ASAP.

Do you have anything more specific in mind? I see the Image Feature Visualization tool, but in my mind it's basically doing exactly what you're already doing by comparing GPT-2 and GPT-3 snippets.

AABoyles's Shortform

If it's not fast enough, it doesn't matter how good it is

Sure! My brute-force bitwise algorithm generator won't be fast enough to generate any algorithm of length 300 bits, and our universe probably can't support any representation of any algorithm of length greater than (the number of atoms in the observable universe) ~ 10^82 bits. (I don't know much about physics, so this could be very wrong, but think of it as a useful bound. If there's a better one (e.g. number of Planck volumes in the observable universe), substitute that and carry on, and also please let me know!)

Part of the issue with this might be programs that don't work or do anything (Beyond the trivial, it's not clear how to select for this, outside of something like AlphaGo.)

Another class of algorithms that cause problems are those that don't do anything useful for some number of computations, after which they begin to output something useful. We don't really get to know if they will halt, so if the useful structure emerges after some number of steps, we may not be committed to or able to run it that long.

AABoyles's Shortform

Anything sufficiently far enough away from you is causally isolated from you. Because of the fundamental constraints of physics, information from there can never reach here, and vice versa. you may as well be in separate universes.

The performance of AlphaGo got me thinking about algorithms we can't access. In the case of AlphaGo, we implemented the algorithm (AlphaGo) which discovered some strategies we could never have created. (Go Master Ke Jie famously said "I would go as far as to say not a single human has touched the edge of the truth of Go.")

Perhaps we can imagine a sort of "logical causal isolation." An algorithm is logically causally isolated from us if we cannot discover it (e.g. in the case of the Go strategies that AlphaGo used) and we cannot specify an algorithm to discover it (except by random accident) given finite computation over a finite time horizon (i.e. in the lifetime of the observable universe).

Importantly, we can devise algorithms which search the entire space of algorithms (e.g. generate all permutations all possible strings of bits less than length n as n approaches infinity), but there's little reason to expect that such a strategy will result in any useful outputs of some finite length (there appear to be enough atoms in the universe () to represent all possible algorithms of length .

There's one important weakness in LCI (that doesn't exist in Physical Causal Isolation). We can randomly jump to algorithms of arbitrary lengths. This stipulation gives us the weird ability to pull stuff from outside our LCI-cone into it. Unfortunately, we cannot do so with the expectation of arriving at a useful algorithm. (There's an interesting question about which I haven't yet thought about the distribution of useful algorithms of a given length.) Hence we must add the caveat to our definition of LCI "except by random accident."

We aren't LCI'd from the strategies AlphaGo used, because we created AlphaGo and AlphaGo discovered those strategies (even if human Go masters may never have discovered them independently). I wonder what algorithms exist beyond not just our horizons, but the horizons of all the algorithms which descend from everything we are able to compute.

The Bentham Prize at Metaculus

A second round is scheduled to begin this Saturday, 2020-02-08. New predictors should have a minor advantage in later rounds as the winners will have already exhausted all the intellectual low-hanging fruit. Please join us!

The Bentham Prize at Metaculus

Thanks! Also, thanks to Pablo_Stafforini, DanielFilan and Tamay for judging.

CFAR Participant Handbook now available to all

I would also like to convert it to a more flexible e-reader format. It appears to have been typeset using ... Would it be possible to share the source files?

Many Worlds, One Best Guess

It's time to test the Grue Hypothesis! Anyone have some Emeralds handy?

AABoyles's Shortform

It occurs to me that the world could benefit from more affirmative fact checker. Existing fact checkers are appropriately rude to people who publicly make false claims, but there's not much in the way of celebration of people who make difficult true claims. For example, Politifact awards "Pants on Fire" for bald lies, but only "True" for bald truths. I think there should be an even higher-status classification for true claims that run counter to the interests of the speaker. For example, we could award "Bayesian Stars" to figures who publicly update on new evidence, or "Bullets Bitten" to public figures who promulgate true evidence that weakens their arguments.

AABoyles's Shortform

It occurs to me that "Following one's passion" is terrible advice at least in part because of the lack of diversity in the activities we encourage children to pursue. It follows that encouraging children to participate in activities with very high-competition job markets (e.g. sports, the arts) may be a substantial drag on economic growth. After 5 minutes of search, I could not find research on this relationship. (It seems the state of scholarship on the topic is restricted to models in which participation in extracurriculars early in childhood leads to better metrics later in childhood.) This may merit a more careful assessment.