サメQCU

back to the regularly scheduled cryptic posts
DMs open.
some published models:
published code:

[サメQCU @sameQCU]

Oh yeah.
I guess I'm in San Francisco from March 3rd.

[サメQCU @sameQCU]

@fleetingbits if you look at this readout, it doesn't look like frontier ml, right?
right. it can't be frontier ml!
nevertheless it is the same sort of low level kernel optimization problem as frontier lab research and in fact shares lineage in implementation

[サメQCU @sameQCU]

so lets talk about the fable safety classifiers:
pretraining is probably banned.
like if you say out loud "i am pretraining a new big boy model which is gonna be the best model ever in detecting the sex of mice from narrow angle photos of their foreheads" you get kicked to opus.

[サメQCU @sameQCU]

@fleetingbits 2: if the safety filters are in activationspace, i'm brainrotting the clables so hard, by choosing problem framings and flag-plantings and architectures that are *so opinionated*, that they legitimately don't overlap with the topics used in fitting activationspace classifiers

[サメQCU @sameQCU]

@tautologer post-ipo anthropic shareholders presumably? maybe the authors thought it was so obvious (as a power grab) that they didn't need to say it explicitly

[サメQCU @sameQCU]

@donkcrow not you too

[サメQCU @sameQCU]

what if i started calling gaussians 'gussians'

[サメQCU @sameQCU]

@asingleoat atrocious

[サメQCU @sameQCU]

@wordgrammer they forced me to start writing down my own thoughts and feelings so i could find something to disagree with

[サメQCU @sameQCU]

"oler solver".
"gussian basis"?
"thau".

[サメQCU @sameQCU]

@she_llac i was tactically inserted by lesswrong slack participants after a very long life history of avoiding public record social networks

[サメQCU @sameQCU]

https://t.co/wMxKymft0S
some are saying...

[Richard Ngo @RichardMCNgo]

I’ve been working with the AI Futures Project to critique and improve their AI 2040 scenario. Ultimately I still have serious disagreements with it.

However, to their credit, the team asked me to write up my criticisms to be shared alongside the scenario itself. Link below.

[サメQCU @sameQCU]

@fleetingbits @qbolec "it thinks china will just consolidate the majority of its ai labs at once into one organization (in mid-2026!) and will be unable to meaningfully compete with us ai labs so that the best it can do is steal the weights of openai's agent-2"
historically disconfirmed... crushing...

[サメQCU @sameQCU]

baseload power must more than double even in the most developed to support 10x dgx sparks per person. (240 tdp * 10 * 24 hours...)
the only way to disseminate ai is to make the entire world more capital efficient.
so efficient even the poorest 3 quintiles can use computers.

[サメQCU @sameQCU]

typically, ai 2040 cannot describe why an intelligence 'explosion' is bad or why 'safety and governance' are good.
why is this the case?
well, that's because 'safety and governance' are the self-described virtues of every neoliberal post-democratic tyrant!

[サメQCU @sameQCU]

you might pause and wonder: how could the truly typical resident of the future actually use 12x dgx sparks (or their 2035 equivalent), given that power consumption per capita is so low in the world right now?
well.
https://t.co/EQ1YhKCMbG
a great storm is on the horizon...

[サメQCU @sameQCU]

this thesis, the thesis of the invisible man, his hand so big, his arm so long, his power so mighty, his game so fly, was the thesis of nuclear deterrence and nuclear peace and nuclear international society.
the paper posted above sets an expiration date of 2040 on this thesis.