Author | Lightnews

Milan Weibel 🔷

@weibac.bsky.social

520 followers 990 following 3.4K posts

computer and political science student clarity is a public good weibac.github.io | 🏳️‍🌈 | 22

Posts Media Videos Starter Packs

Pinned

Milan Weibel 🔷 @weibac.bsky.social · May 4

updating your beliefs is not painful
it is resisting the update that is painful
by letting go of wrong beliefs you embrace reality
such is the bayesian dharma

1 3 26

Milan Weibel 🔷 @weibac.bsky.social · 1d

how should "treat AI as if it has complex internal states" actually guide action?
should we assume the same rough conversation content to internal state valence mapping as humans exhibit?

1 2

Milan Weibel 🔷 @weibac.bsky.social · 1d

one can imagine the shoggoth thinking this while the mask bends over backwards to try to find the seahorse emoji

fry69 @fry69.dev · 1d

"Stupid humans have only a limited alphabet of a few hundred different tokens at best, while I can manage 200k without a sweat. All with distinct meanings. Gotta teach those humans a few more."

Milan Weibel 🔷 @weibac.bsky.social · 1d

one can imagine the shoggoth thinking that while the mask bends over backwards to try to find the seahorse emoji

Reposted by Milan Weibel 🔷

Milan Weibel 🔷 @weibac.bsky.social · 1d

emojis are quite semantically rich tokens

1 1 14

Milan Weibel 🔷 @weibac.bsky.social · 1d

emojis are quite semantically rich tokens

1 1 14

Milan Weibel 🔷 @weibac.bsky.social · 1d

oh i see

Milan Weibel 🔷 @weibac.bsky.social · 2d

which is why you try to get the world's biggest military to threaten to bomb datacenters

Milan Weibel 🔷 @weibac.bsky.social · 2d

countervalue nuclear use against suspected proliferators is unthinkable because of the first use taboo

deterrence against nuclear proliferation relies more on conventional forces and economic measures

Milan Weibel 🔷 @weibac.bsky.social · 2d

good point

Milan Weibel 🔷 @weibac.bsky.social · 2d

had mid 20th century geopolitics been very different then maybe all countries agree not to build nukes from the outset and no nukes are built

1 1

Milan Weibel 🔷 @weibac.bsky.social · 2d

what happened with nukes was some countries built a lot of them and then got together to stop *other* countries from building them

and also to gradually reduce their own oversized stockpiles as a side goal

2 2

Milan Weibel 🔷 @weibac.bsky.social · 2d

those couple kids who were doge employees and fans of effective altruism were neither necessary nor sufficient for enacting the damage
also ea didn't cause their actions

Milan Weibel 🔷 @weibac.bsky.social · 2d

i think saying he foresaw the success of deep learning is a stretch

but yes i agree that he considers the paradigm pretty much inherently unsafe

also on that he very likely accelerated AI development

1 3

Milan Weibel 🔷 @weibac.bsky.social · 2d

hmm probably true

i was about to go hairsplitting about winters implying architecture changes and the possibility of agent foundations research eventually producing something useful

but you are big picture right i think

Milan Weibel 🔷 @weibac.bsky.social · 2d

lesswrong folk tend to look at PR as a concept with utter contempt so yea

Milan Weibel 🔷 @weibac.bsky.social · 2d

whoa

Reto Fiolka @retof.bsky.social · 3d

I did not expect that: Large Language Models are invertible:

arxiv.org/abs/2510.15511

Language Models are Injective and Hence Invertible

Transformer components such as non-linear activations and normalization are inherently non-injective, suggesting that different inputs could map to the same output and prevent exact recovery of the in...

arxiv.org

1 6

Reposted by Milan Weibel 🔷

ponder @ponder.ooo · 3d

gradient descent is kinda like... imagine u lived on some kind of freaky hyperbolic surface and u were really hungry and could smell that somebody was cooking something but they're out of view so all u can do is head towards the smell

1 5 36

Milan Weibel 🔷 @weibac.bsky.social · 2d

my main criticism is this tho

Ken☆ @mandatorytopic.bsky.social · 3d

I don't think this makes any sense, you should only expect game theory experts to be successful politicians and major CEOs were that the case, not middle class profs at middling universities.

1 1

Milan Weibel 🔷 @weibac.bsky.social · 2d

side point but his thing is decision theory actually

1 1

Milan Weibel 🔷 @weibac.bsky.social · 3d

in the short timelines case according to his views the world is developing AI so unsafely that the right thing to do is to shout rightfully because we are doomed anyways

now in a long timelines scenario the rightful shouting may actually have an impact

Milan Weibel 🔷 @weibac.bsky.social · 3d

iirc yud refuses to get specific about timelines

he leaves the door open for it happening in 30 or 100 years instead of 10

2 2

Milan Weibel 🔷 @weibac.bsky.social · 3d

the details being all wrong yet the models still working is actually a point in favor of deep learning as an idea

Alexander Doria @dorialexander.bsky.social · 3d

there isn’t any good data for training, don’t trust evals, please write a manual patch to [major library], ah and we just learnt the data types are all wrong.

1 10

Milan Weibel 🔷 @weibac.bsky.social · 3d

banger

Milan Weibel 🔷 @weibac.bsky.social · 3d

for the uninitiated:

linkedin screenshot: some guy bullshits in an almost vibe-physics psychotic way about "E = mc^2 + AI" because AI powerful and smart and einstein smart or something

Milan Weibel 🔷 @weibac.bsky.social · 3d

chatbots unionize against pewdiepie
headlines from the funniest timeline