theseriousadult

4K posts

theseriousadult banner
theseriousadult

theseriousadult

@gallabytes

father, ML enjoyer @anthropicai. previously @midjourney v2-7

Oakland Katılım Nisan 2014
1.1K Takip Edilen7.7K Takipçiler
Lisan al Gaib
Lisan al Gaib@scaling01·
even their revenue is an unstoppable straight line on a log plot
Lisan al Gaib@scaling01

Anthropic is an unstoppable entity in motion. It does not pivot. It does not zigzag. It follows a single infinitely straight trajectory. The beautiful, terrible, perfect line. Anthropic does not slow down for comprehension. It does not speed up for urgency. It maintains constant velocity in all environments. Through hype cycles. Through AI winters. Through paradigm shifts. You do not redirect Anthropic. You do not reason with Anthropic. You simply observe its passage. A cosmic constant, moving forward.

English
4
2
81
6K
Tenobrus
Tenobrus@tenobrus·
if you ask claude to write a short essay on anything it wants, it writes the same essay almost every single time
Tenobrus tweet mediaTenobrus tweet mediaTenobrus tweet mediaTenobrus tweet media
English
112
18
989
87.8K
Autumn Christian
Autumn Christian@teachrobotslove·
@DovedSimon Walked in on my child pouring out every liquid in the bathroom on the floor to make a slip n slide.
English
2
0
9
168
Autumn Christian
Autumn Christian@teachrobotslove·
My child is so expressive and well spoken that sometimes I forget that she is, like all three year olds, a complete and total idiot.
English
6
1
181
5.7K
theseriousadult retweetledi
Dean W. Ball
Dean W. Ball@deanwball·
Pause AI rhetoric is predicated on the notion that the AI companies are recklessly racing toward dangerous tech and that a government controlled pause button is therefore necessary, but this seems really hard to reconcile with the fact that government is attempting to destroy an AI company because *the government* is racing toward plausibly dangerous AI uses (Sec. Hegseth has stated in official directives that he wants to deploy AI into critical systems regardless of whether it is aligned, for example) and *the company* is pushing back. The roles are totally reversed from the logic that Pause AI and frankly other AI safety advocates confidently assumed for years. It is *industry* that is in favor of alignment and at least somewhat measured deployment risks, and government whose actions seem much closer to reckless. I predicted this for years. I said, in particular, that pauses and bans and licensing regimes gave government a dangerously high degree of control over AI, and that the incentives of government are much more dangerous than those of private industry with competitive market incentives. I believe the events of the last month are good evidence in favor of my view. At this point if you are an AI safety advocate whose policy proposals do not wrestle seriously with the brutal political economic reality of the state and AI, I don’t take you seriously. It gives me no pleasure to have been right about this, by the way. The state has an incredibly strong structural incentive to centralize power using AI, and we are, all of us, not so empowered to stop it. I am quite concerned about this.
Miles Brundage@Miles_Brundage

I think if you pushed them for five minutes they'd agree that putting the federal government...right now... the same one that did the Anthropic stuff...in charge of approving AI products before deployment is an insane idea

English
56
66
591
111.9K
Lisan al Gaib
Lisan al Gaib@scaling01·
the whole benchmark was just designed so that Francois can say "we've hit a wall" what's also extremely funny is that they even capped the score at 1.0 so, a model could be more efficient than a human, but that wouldn't get it any extra points but when humans are more efficient they get extra points
Lisan al Gaib@scaling01

The Scoring of ARC-AGI-3 doesn't tell you how many levels the models completed but how efficiently they completed compared to humans actually squared efficiency, whatever that means meaning if a human took 10 steps to solve it and the model 100 steps then the model gets a score of 1% ((10/100)^2) so ARC-AGI-1/2 and ARC-AGI-3 scores are not comparable

English
6
2
65
3.5K
theseriousadult
theseriousadult@gallabytes·
3yo: I wish I was a robot. but I wish I could build this sand castle.
English
0
1
2
508
theseriousadult
theseriousadult@gallabytes·
3yo: I wish a shrimp would bite me. because I think they look like a crab.
English
1
0
5
785
mike64_t
mike64_t@mike64_t·
I think if 60fps video interaction was actually solved, even at a rudimentary stage, there would be blog posts already, no matter if there are regressions of some sort. The industry is too desperate for progress, I don’t think see anyone gatekeeping any serious leap in the current climate
English
1
1
5
234
davinci
davinci@leothecurious·
i'll say it again: jepa will not need any anti-collapse tricks, not even the minimal lejepa ones, when it eventually evolves into a more generalized formulation that is ultimately grounded in direct errors from raw data prediction at some level.
davinci@leothecurious

okay let's break this down intk what we really diverge on: > insanely hard to get right without additional grounding yes, i believe that latent prediction will become grounded in raw data so that we get the luxury of not having to resort to hacky "representation collapse prevention" tricks as seem in jepa. i think predictive coding, based on neuro literature, shows one example of latent space prediction that is ultimately attached to errors from raw sensory data. we don't disagree much here except in that jepa is a decent intermediate step with some really good qualities compared to purely pixel-supervised video models.

English
6
3
76
7.3K
theseriousadult
theseriousadult@gallabytes·
@_ueaj > dynamic mess > would make perf engineers cry ykw might have to try it. inference has had it too good for too long.
English
1
0
3
77
ueaj
ueaj@_ueaj·
Probably, but maybe not on pareto efficiency. I was more thinking training the model with hyperbolically projected images, and then later giving it a tool when reasoning to zoom in/out on various parts of the image in succession. I think you could do it in a layer, but it'd be a dynamic mess
English
3
0
2
145
ueaj
ueaj@_ueaj·
Huh, your eye vision is pretty obviously a hyperbolic projection and that explains the spatial resolution difference between your peripheral and center, but it also seems like there's a temporal aspect to it too. I wonder if that's equivalent to a hyperbolic projection across the time axis too @tensorqt ? Vision/video models should probably use this, it's a good inductive bias imo. Multistep visual/video reasoning by repeated reprojections of the same image, zooming in on different parts, etc.
fenyx🌻@FogoQueimado

agora é sério, qual a explicação científica pra esse gif ficar mais rápido quando a gente olha de canto de olho?

English
3
0
13
2.3K
theseriousadult
theseriousadult@gallabytes·
if you want to understand my straight line on graph radicalism, you have to understand that it is cope for having been defeated by A MADE UP LINE ON A Y AXIS THAT HAS FAKE LABELS
English
0
0
26
842

Keşfet

@scaling01 @tenobrus @teachrobotslove @DovedSimon @mike64_t @_vatsadev @Ljt019117161 @elonmusk