truefire @truefire87

@thinkingshivers "Baudrillard" is actually when someone an insufferable hack fraud who has nothing to contribute to the discussion aaaaa I hate Baudrillard.

0 0 18 1K 0

View Details

@ID_AA_Carmack has anyone ever told you you look kinda like a skeleton. Like if someone took a picture of you with heavy top lighting (so your eye sockets are in shadow) and posterized it then scaled it down to like 64x64 it could work as a skeleton sprite in an early CRPG.

0 0 0 17 0

View Details

truefire @truefire87

5 days ago

@rickyflows Guy who doesn't know about busybeaverial growth

0 0 1 18 0

View Details

truefire @truefire87

6 days ago

Get a load of this guy 😏

Guy @nosilverv

7 days ago

The genius of the QT is that it "frames" the existing thing said by someone as "something said by someone" thus allowing you to trivially recognize their frame *as a frame* which you are now empowered to contribute to / question / ignore / sabotage / …

6 1 91 7K 8

0 0 0 20 0

View Details

truefire @truefire87

6 days ago

@Mononofu Kinda non sequitur: it is too easy to jailbreak ("does what the user wants"), so I'm worried it will be misaligned ("do what the user doesn't want"). The reward hacking part makes sense though.

0 0 1 242 0

View Details

truefire @truefire87

7 days ago

@TylerGlaiel The real thing is a "good idea" is not "we should do X", but actually 100 pages of details about what X looks like that are both good and tractable. At which point calling it an "idea" isn't really correct.

0 0 0 68 0

View Details

truefire @truefire87

a week ago

@Mononofu Man, I honestly think the biggest threat to AI alignment right now is alignment researchers. You guys are so focused on the wrong things and it's so obvious. This kind of optimization is the de-facto way to induce psychopathy in a model.

0 0 12 462 0

View Details

truefire @truefire87

a week ago

@xsphi @Actualwebutante I think they're confusing you for @yacineMTB, I'm sorry for your loss.

0 0 1 22 0

View Details

truefire @truefire87

a week ago

@xsphi Dubious parentheticals doing all the work in this image.

0 0 71 7K 0

View Details

truefire @truefire87

a week ago

It's kind of annoying that I agree with LeCun's general thesis (latent manipulation probably has strict benefits over the clipping we do), yet disagree with most of what he says because he's rhetorically dishonest to the point of absurdity. Transformers can obviously do a lot, and you don't have to pretend otherwise to assert that there are better options. You don't have to intentionally pretend attention doesn't exist so you can make this obviously unreasonable 3-byte quip.

Fabien @Fabien_Mikol

a week ago

LeCun : "le pb du LLM est que la seule communication d'information d'une étape à une autre est à travers un token discret de 3 octets", alors qu'un véritable raisonnement nécessite de manipuler en interne des représentations dans l'espace latent. L'étude Anthropic réfute cela 🧵

41 53 613 339K 862

0 0 1 40 0

View Details

truefire @truefire87

a week ago

@Fabien_Mikol Bro forgot about the attention mechanism😂

0 0 0 21 0

View Details

truefire @truefire87

a week ago

@allTheYud Claude "disobeying"

1 0 11 849 0

View Details

truefire @truefire87

a week ago

@pneumic I like Claude. Funny guy.

0 0 3 460 0

View Details

truefire @truefire87

a week ago

@Santeau_k2o It's because they make more money by being bad. If you find a job you're not on LinkedIn. Why no good alternative: It's a bit ironic, but it's the network effect. The same reason job search boards exist in the first place.

1 0 2 138 0

View Details

truefire @truefire87

a week ago

@Afinetheorem @logangraham What if this is actually the optimal way to think. Maybe we're the idiots because we don't recognize the expressive power of genuinely sitting with our load-bearing smoking guns.