@Nexuist
he's right, and he also demonstrates why Chesterton's Fence is an excellent argument for legal conservativism (to anyone that has ever shipped software)
@tszzl
@gwern
once had the idea of a tri-phase nicotine-caffeine-aphetamine cycle
since all 3 stims have different tolerance mechanisms, you can recover from 2 while you're being juiced by the other
@mezaoptimizer
consider that Actually Good open-source LLMs may reduce the incentive for OpenAI, Anthropic, GDM, et al. to do larger and larger training runs -- you capture far less value if the output of your model is instantly used to train a model which ppl use for free w/o paying you
@tenobrus
when you were 15 you were writing good python for free
when I was 15 I was writing bad python for slightly above minimum wage
we are not the same
Dropbox was literally just an S3-wrapper with decent UX. But there are two companies responsible for the lion's share of YC's returns, and that S3-wrapper is one of them.
UX matters, big style. Too bad frontend work is annoying & midwit-coded lol
one of my core advantages over other people like me is that I was 14, not 18, when the world made me viscerally realize that there are other people who really are better at math / more intelligent than me — that I must win by being more strategic, not by firepower alone
@mpopv
just one more dependency bro cmon I'm gonna save myself a whole 30
minutes man cmon dude just one little dependency please it doesn't even weigh that much man cmon bro let me add just one little tiny dependency I need this man please I don't wanna reinvent the wheel cmon just one
@ctjlewis
@RamsesAmericas
this is pedantry imo; ppl only care about IQ tests insofar as they measure g-factor -- which is what IQ tests are optimized to do
the SAT is only a slightly worse measure of g-factor than a clinical IQ battery, so the SAT really is "as close as we got to a standardized IQ test"
u gotta realize basically every time someone makes a blanket group statement, they actually have a very specific individual in mind. when someone says "women do x" or "i hate men", what it actually means is "sarah did x" and "i hate tom". makes things read very differently.
@tapir_worf
I'd just like to interject for a moment. What you're referring to as "Logical Decision Theory" is, in fact, FDT/Updatelessness, or as I've recently taken to calling it, FDT plus Updatelessness.
how 2 tell if ur input is out of distribution for ANY model during inference!
(working 2023) (not clickbait) (OC pls no steal)
benchmark for resnet18, higher score means more out-of-distribution:
pic 1:
2.26
pic 2:
2.88
pic 3:
4.76
pic 4: (from the training set)
1.93
- take input vector
- "nudge" it by adding a lil noise
- get cosine-similarity between these two
- pass both through ur model (take ur output from any layer you want, wtv)
- get cosine-similarity between ur 2 OUTPUT vectors
- how much did the cos-sim decrease by?
& that's it😎
@vestiphile
@SydSteyerhart
> consider his conversation paths almost fatally carefully, like his temperature setting is dialed way down
I think this anxiety is more a consequence of RLHF than it is of low-temperature.
10x-dev-turned-founder types don’t realize the extent to which their activities can be monetized. they fail to make the rsync -> dropbox conversion. lots of alpha in convincing ppl to charge for these glorified bash scripts, even now, because they will always be in demand