MagpAI retweeted
intelligence might have two main parts:
1. generalization capacity (g)
2. expertise requiring lots of practice
larger models seem to have higher g, but new capabilities seem to mostly require expertise via RLVR
this could explain much spikiness in both humans and LLMs
Maybe it speaks to my cultural milieu that i saw this shot, knew it was trying to be funny, but was confused because that's just what everyone does. Guy has a startup, of course he's in a wework.
My favorite visual gag in You Can See Everything is when Billy’s working inside a fancy looking office building and it cuts further out and it’s a WeWork
7
296
MagpAI retweeted
somewhat spoiler-y tentative hypothesis about elizabeth holmes’ deal after having watched you can see everything and not doing additional research, skip if you want to watch and come to your own conclusions first:
—-
so, first of all, she’s obviously some kind of autistic. she has sort of elfin neotenous facial features that remind me of some other autistic women i know and i assume there’s a connection there but idk what specifically
there’s a kind of rizz that autistic women in particular can develop, that involves being unguarded in a certain way, learning to drop energetic defenses that most other women have up most of the time. it’s compelling in a confusing way, and i think you can see nathan being simultaneously compelled and confused by liz doing this at several points
when it’s happening it feels like she is radiating childlike innocence, it doesn’t feel like you’re being charmed by a master manipulator or whatever. but, i think you can tell a plausible story about a type of woman who learns this is a button she can push to get men in particular to like her, who gets RLed into pushing this button, not necessarily consciously. strategies like this work best when you aren’t aware you have them
when you put that idea together with stuff nathan has said about the experience of being around liz - “we’d completely disagree on what something meant to us”, “feeling of cognitive dissonance”, “our brains were scrambled” - and with what’s in the doc itself, i think you can vibecode a picture of a reality distortion field powered by this sort of fundamental innocence. i think she fundamentally believes what she’s saying, and nathan picks up on this energy repeatedly. but that doesn’t mean she doesn’t have huge blindspots, which are on some level being intentionally maintained so that her life strategies don’t collapse
MagpAI retweeted
Many eons ago I was particiating in a NASA competition about how to best rotate the solar panels on the ISS for max solar throughput per orbit (got 10th place).
The evaluator was a ray tracer and we were given the ISS model down to every tiny shit antennae casting shadows.
The first place winner just deleted all the small details of the ISS and used a rough model, making his simulations immensely faster. That's it, it won him the comp.
Ref: NASA ISS Longeron Challenge.
MagpAI retweeted
Interesting how closely this mirrors the basic setup high throughput biology roughly does (eg. 10^8 HTP sort or selection, hits go into 10^3 plate based, hits go into 10^1 in situ / extensive charactisation).
Without getting into too much detail about the conversations I’ve had with my friends at Anthropic, it’s become obvious to me that their long-term ambition is to continue running these small-to-large experiments while scaling experiment throughput with AI agents.
Now, researchers already use smaller models to test architectures and optimizers. However, what I’m hearing is that they want to do this on a much bigger scale.
Think millions of smaller, cheaper experiments with models that have a billion parameters, with the most promising discoveries being validated at 10 billion, then 100 billion, and so on and so forth.
The idea is to have autonomous research agents explore this enormous hypothesis space and identify which architectures and training methods scale best.
Now, this probably sounds obvious to a lot of people already in the AI field. However, from what I’m hearing, there’s a lot of success happening right now, and they’re continuing to scale up this experimentation, which is quite exciting to hear.
MagpAI retweeted
Replying to @cackenbools
I think my point here is that in less-skilled hands it wouldn't have been produced at all. This isn't so much to Levent's credit — he does have access to models to which the public doesn't, after all, and he presumably has practically-unlimited token spend, so it's hard to say that he had any unique edge here over a similarly-exceptional mathematician — as it's meant to point out that, hysteria aside, we're not (yet) in the proof-button regime. We may get there, perhaps soon, but we're still in a centaur phase, despite rumors of the field's complete death.
There is zero evidence that writing essays is hard for LLMs
Replying to @HellenicVibes
Writing essays is easy for (some) humans and hard for LLMs. Solving Millennium Prize problems is hard for humans and easy for LLMs. There's no law of the universe that says that solving Millennium Prize problems is harder than writing essays
13
318
This is probably a decent time to completely lose your grip on reality, if you were considering it already.
3
19
391




















