Adam Scherlis
357 posts
Adam Scherlis
@ascherlis
Physicist → AI safety researcher @d_model_ai
Berkeley, CA Katılım Ağustos 2010
81 Takip Edilen188 Takipçiler
Naming their next model after Cthulhu makes it hard to take Anthropic seriously as the good guys. It's fun at any other software company, not one that actually is flirting with extinction.
Jimmy Apples 🍎/acc@apples_jimmy
“ A draft blog post that was available in an unsecured and publicly-searchable data store prior to Thursday evening said the new model is called “Claude Mythos” and that the company believes it poses unprecedented cybersecurity risks. “
English
Adam Scherlis retweetledi
you've joked about a spherical cow, but have you thought about higher multipoles of a cow, thus the tetrahexacontapole cow, and the octacosahectapole cow?
(figure showing 1, 2, 4, 8, 16, 32, 64 and 128-pole cow, and the full cow)
English
@liquiditygoblin Are the linear features NN embeddings or just one-hot tokens directly?
English
In an effort to try stop seeing so much slop I've been trying to train my own AI detection model.
Found something incredibly interesting.
for the most part LLM generated text and human written text are linearly separable.
English
i'm starting a collection of things like this
English
@Seb_81_ @Chad_Athena1371 @AfricaFirsts no force in heaven or on earth can stop the Gall–Peters from looking weird
English
@Chad_Athena1371 @AfricaFirsts It was for navigation because it helps make that easier. And now it's what the map looks like so peters looks weird for example.
English
@EVBetsNY @angutyoo @AfricaFirsts (ok that's not quite his name but I was enjoying the typo too much to spoil things)
English
@EVBetsNY @angutyoo @AfricaFirsts The Watermelon Butterfly, designed by Steve Watermelon based on a map by Cahill. A friend has a hand-painted copy in gold ink on one of her jackets.
English
@BronsonSchoen @DanielCHTan97 @ihsgnef You have to request it but I think they're pretty generous with it.
English
@DanielCHTan97 @ihsgnef Opus 3 is alive and well! (I’m not sure how api access works, might need to request it or something similar)
English
@eshear @QuetzalThoughts @nickcammarata I would've guessed mostly heuristics and slowly-learned behaviors rather than beliefs updating in a Bayesian way or on a short timescale. Emergent problem solving like with boids.
English
@ascherlis @QuetzalThoughts @nickcammarata What else do you think is driving ant behavior other than ant beliefs?
English
I don’t actually know how to process this, what’s a reasonable explanation of what is going on here
The Figen@TheFigen_
They are ants solving a geometric problem and it is mind-blowingly colorful.
English
@ascherlis @QuetzalThoughts @nickcammarata The momentum is in belief-space. arxiv.org/abs/2205.11543
English
@eshear @QuetzalThoughts @nickcammarata What does it mean for momentum to be "spread across frequencies" if you're using it to mean spatial frequency? Can you spell out your model a little more?
I did in fact think you meant physical momentum.
English
@ascherlis @QuetzalThoughts @nickcammarata They’re only operating in that regime. Unless you think I mean the ants have literal physical momentum?
English
@eshear @QuetzalThoughts @nickcammarata Reciprocal position space is only related to Newtonian momentum in a quantum setting (action ~ hbar). Ants are generally speaking not operating in that regime.
English
@QuetzalThoughts @nickcammarata @Mathematically, the duality between position and momentum is an example of Pontryagin duality. In particular, if a function is given in position space, f(r), then its Fourier transform obtains the function in momentum space, φ(p). “
English
@kristinatastic "pluralia tantum" is one (but isn't a plural tantum)
English
I've been working on this important list for a couple of years now. What am I missing?
English
Adam Scherlis retweetledi
Longtermists: We need to engineer a breed of cats which will change color upon exposure to radiation, release them to the wild, then compose a ditty about these cats that will be so catchy as to be handed down for 10,000 years at least
Musicians: on it boss
Geneticists: what
𝖓𝖎𝖓𝖊 🕯@atlanticesque
@MRichster Don’t change color kitty Keep your color kitty Stay that pretty grey Don’t change color kitty Keep your color kitty Keep sickness away Don’t change color kitty Keep your color kitty Please cause if you do Or glow your luminescent eyes We’re all gonna have to move :(
English
@reconfigurthing from an SSC comment by Roxolan in 2015: #comment-177111" target="_blank" rel="nofollow noopener">slatestarcodex.com/2015/01/25/a-p…
(oops, would've credited this if I knew it was gonna go ~viral)
English
@reconfigurthing The waitress brings Newcomb his coffee. “Did you spit in it?” asks Newcomb. “Are you going to tip me?” answers the waitress.
English
@Brilliand__ @qbolec (this is also why I'm ultimately a thirder -- I don't think my victories and failures are, in practice, diluted by those of distant copies of me.)
English
@Brilliand__ @qbolec I agree that scoring mechanism is important.
I think there's an important class of linear/local ones: you know how many points you got today without needing to know what you did / will do on other days. These point to SIA.
e.g., Claude gets a cookie for being right.
English
"I flipped a coin. If heads, I planned to give one Claude instance this prompt. If tails, two instances. What probability do you assign to the proposition that the coin came up heads?"
Robert Long@rgblong
the anthropic principle
English