IMO, THIS.
Joined September 2014
- Tweets 5,257
- Following 234
- Followers 1,584
- Likes 1,070
29 Photos and videos
Pinned Tweet
3 Mar 2020
New goal: learn the original language.
8
31
105
I think I'm really interested in local approximations of a distribution, where the approximation also goes to zero as you move away from the point of approximation.
Like clusters, or kernels, or proposal distributions.
36
The "you shouldn't have a favorite weapon" principle, but generalized to "you shouldn't have any likes or dislikes."
64
Proof that morality is fake, using analogies:
(i) North : South :: Good : Bad
(ii) North : South :: Bad : Good
Convert to latent space vector representation:
(i) North - South = Good - Bad
(ii) North - South = Bad - Good
Subtract:
(i - ii) 0 = 2 (Good - Bad)
So:
Good = Bad
2
105
Are there any good examples of something that's actually one thing, but has two different identities in the popular consciousness?
Like. As a hypothetical example, if the two Noam Chomskies were actually one guy.
1
3
121
I think it'll be a big deal when software engineering AIs are no longer dependent on human-created languages and frameworks.
Seems like the transition to new languages could both boost AI productivity and reduce human ability to review new code.
3
103
Robin gets it.
x.com/i/status/2059657886179…
May 27
If you wanted to worship something big, powerful, ancient, and nurturing, hard to see a better option than the Sun. So its actually a bit odd to see people ever worship anything else. Is it the Sun's constancy what they hold against it?
1
120
I wish "gamification" were about using game design principles to create mechanically satisfying tools.
3
103
I wonder if an LLM agent could exhibit much better introspection just by training on simple tasks related to its embodiment (e.g. "how many tokens do you see in this word?")
1
3
147
I'm interested in perspectives that have a clearly-defined origin.
1
84
I think if I ever write a prequel, I'll just abandon the original storyline and let the good guys win earlier.
2
101
Horror premise: the discovery that we are what you might call "typical observers" in the sense of the anthropic principle.
Like, typical within the multiverse.
2
12
849
I think human art is really crippled by not having an affordance to directly pipe large arrays of hallucinated imagery out of the brain.
1
3
118
If you organize something, does that mean you turned it into an organ?
6
172
I think some aspect of human motivation is likely encoded in protolinguistic "prompts" in our genes.
1
5
352
How would you name "the fate of one's soul" in naturalistic or humanistic terms? Like, you don't want to see someone you care about walk the path of destruction, even when the damage they cause is small.
2
128
People tend to underestimate the value of really tall towers. Because they tend to judge towers based on what they can see. But in order to know what's on top of a tower, you'd need to already be that tall yourself.
2
136
The future is more like the past than the present in the following sense:
Both the future and the past are billions of years long, while the present is but an instant.
1
154
Reinforcement learning for skill gain doesn't need to happen in post-training. It can happen in some totally different system that feeds into the pre-training dataset. For example, you can add a gigabyte of Go games played by KataGo to your data.
3
140
I have a neat idea for a language model. A bit like a Markov chain or HMM, but it's for sentences of a fixed size N. Then there are K clusters. For each cluster, you have some tree spanning the N word indices, so each word's distribution depends on a single parent.
2
171