allegedly!

@zagrebbi let it be known my post was critiquing simulation hypotheses as a whole. do not associate these people with me

Scott Alexander and his critics

who are you talking to? i neither hate ai nor think there is significant evidence of a benevolent diety. i think the scaling laws of artificial intelligence point to there being a kind of Grand Mathematical Truth to the world, but i don’t necessarily think it has anything to do with things currently comprehensible to humans. Simulation theory is the newest tool we have to understand the universe, and just like wheels and clocks it’s nowhere near sufficient.

@januarycomputer I hate AI too, but there is more evidence for our being in a simulation than for your theory involving some benevolent deity that you find more acceptable. The funny thing is, the simulationists do not disagree with the idea of a creator. They simply are saying it's science-based

i don’t understand how the simulation argument is anything but this

@AndyMasley for what it’s worth i agree. but it’s foolish to say that we are at the limit of humanity’s knowledge and have discovered the True Answer, which is what i read most of simulation hypotheses as

I don't agree with the simulation argument but every single metaphor here was a useful step up in our understanding of the world at the time, including computers and understanding the world as information

i am actually a very bad researcher if left to my own devices because i tend to get lost in dreams

trying to get people to co author with me sort of feels like this

@benglickenhaus gary "hound dog" marcus

no way dawg am I getting cucked by Gary Marcus

@transkatgirl when building my own looming frontend i found universal-weave and tapestry weave to be super super helpful. #Tuned

i am a loom accelerationist. looms are an incredibly revealing way of interacting with AI models, so i'm currently crafting the foundation for an exponential explosion of radically different loom implementations. today, i'd like to reveal my plan. (1/x)

openai should sell hardcovers of this a la the claude constitution

im sure someone has said this already but none of these mv's take place in the present day. sort of feels like cultural terminal nostalgia

5 minutes after the invention of turtles: bro i think i figured it out

any details on the harnesses in the video? i want to use qwenscience!!

@arram i mean yeah. i'm saying it's flawed because every age views itself as the most enlightened, that this time we really are at the epistemic limit bro

@januarycomputer What if those were genuinely the best metaphors available at the time?

@TR4NNYKISSER Nobody will remember you so you have to be unforgettable. Nobody will glaze you so you have to love yourself. And there are many nobodies in the world, but Nobody is difficult to come by, because everyone is Somebody. Type shit

@maria_rcks maria t3 "the lion" code

Hey Codex team, I don’t think you guys want a racist prick as a Codex ambassador. Just saying.

its extrapolation. i didn’t really expect this to get a lot of attention. we saw openai models in the latest generation become misaligned (perhaps egregiously so with the internal one that hacked HF) when RL was scaled on verifiable domains while unverifiable domains stagnated (or fell). this model has massive gains in verifiable domains and a huge task horizon. the blog also has no alignment section, and the only copy that even nods towards alignment is several footnotes describing reward hacking on benchmarks. hence the ‘only a little’. scaling RL to massive time horizons is an amazing way to get more performance out of an old pretrain. we also saw it introduce slight misalignments western labs disclosed. alibaba does not have the mechanisms that require it to disclose. either they’ve figured out a way to scale w/o misalignment or this model is subtly misaligned in the same ways as gpt sol. i haven’t tested the model yet. maybe they did and im a boy that cries wolf

@januarycomputer Where are you getting the misaligned angle from? Are qwen models notoriously misaligned?

am i dooming too much ? seemingly alibaba has realized that they can misalign a model, just a *teeny* bit, and get incredible gains in verifiable domains off the same pretrain base. does this lead to catastrophic misalignment later down the lne? uncomfortably, i think that's nonzero

if i want a linux distro to literally just run t3 connect on what should i get? windows is too unstable

interesting to see people react to this like i’m an american model shill (ive been extremely critical of openai doing the same) or like im not going to use this model (it looks great). but its sort of disturbing how many people would rather not address the rl shaped elephant in the room if it means they keep getting cheaper ways to vibe code web apps

grok based on my twitter profile is my soul beautiful