Lilian Delaveau (h/acc)
1.7K posts
Lilian Delaveau (h/acc)
@LilianDelaveau
CTO & cofounder | AI accelerationist | Shitposting about math, alignment & tech 💜 | Husband | Paris, France
Paris, France. Katılım Temmuz 2011
251 Takip Edilen342 Takipçiler
@LilianDelaveau Je shitpost beaucoup trop, pas sur que ce soit très intéressant ptdr
Français
@data_ccelerator C’est Luc Julia qui m’a tout appris. Tu savais que c’était le fondateur de Siri ?
Français
@data_ccelerator Les paramètres c’est le nombre de page que le LLM a dans sa base de données c’est ça ?
(Je tente de speedrun ton infarctus)
Français
@LilianDelaveau C'est un verbiage qui au final ne dit rien. Il y a des pelo qui fine-tune des modèles sans même savoir ce qu'est que le nombre de paramètre d'un modèle
Français
Ah oui ! Je me rends compte qu’à part 2/3 personnes dont toi j’ai aucun FR sur ML & co dans mon feed. Ils sont tous aussi nuls en fR?
Pas eu le temps de mater TurboQuant encore mais il me semble que c’est l’application d’un truc connu depuis des mois non? Anyway si c’est juste de la compressibilité c’est cool mais probablement pas de quoi casser 3 pattes à un connard.
Also je vais te follow j’aurai ptet plus de français cools
Français
@LilianDelaveau Pour ceux qui parle de deep learning, LLM, IA en bref : Aucune compétence technique réelle, on dirais qu'ils sortent tous d'école de commerces. C'est cringe au possible.
Ils sont tous entrain de bander sur TurboQuant sans même comprendre le fonctionnement.
Français
@scaling01 Not comparable, and I get and agree with the shitshow around the release, however this is a good way to measure against humans.
English
The Scoring of ARC-AGI-3 doesn't tell you how many levels the models completed but how efficiently they completed them compared to humans
actually using squared efficiency
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
@shaggysurvives Hard yes, and also depends on the rules and if the task is an infinitely quick hyper task - I did a small breakdown iyw x.com/liliandelaveau…
Lilian Delaveau (h/acc)@LilianDelaveau
This situation depends on being 100% sure everyone down the line is making the good choice but not only that. Someone could misunderstand, or simply trip on the lever. And let’s be clear: it’ll happen. A psychopath, an elderly, someone intoxicated or just having a bad day. You should minimize suffering by pulling the lever. While it doesn’t change the outcome, the reasoning also depends on the game getting back to 1 after choosing to kill or the game stopping. In the second case, kill ASAP, it’s the minimum person to be killed. In the first, it’s actually a choice of linear versus exponential killing series, so if choice and situation is to be made in fixed time, always kill even one person (only the 1 -> 2 transition is free so if we want to lift the burden of culpability you can skip one and kill 2 and it’s still linear). So, in the worst case, you should still pull and hope many people pull. If time isn’t a factor here, the daemon putting us to choice just wants everyone dead so do whatever you want, linear or not, it’s a genocidal hyper task.
English
@brainrotpostig À small breakdown i think is interesting. Depends on game rules and hyper task definition. x.com/liliandelaveau…
Lilian Delaveau (h/acc)@LilianDelaveau
This situation depends on being 100% sure everyone down the line is making the good choice but not only that. Someone could misunderstand, or simply trip on the lever. And let’s be clear: it’ll happen. A psychopath, an elderly, someone intoxicated or just having a bad day. You should minimize suffering by pulling the lever. While it doesn’t change the outcome, the reasoning also depends on the game getting back to 1 after choosing to kill or the game stopping. In the second case, kill ASAP, it’s the minimum person to be killed. In the first, it’s actually a choice of linear versus exponential killing series, so if choice and situation is to be made in fixed time, always kill even one person (only the 1 -> 2 transition is free so if we want to lift the burden of culpability you can skip one and kill 2 and it’s still linear). So, in the worst case, you should still pull and hope many people pull. If time isn’t a factor here, the daemon putting us to choice just wants everyone dead so do whatever you want, linear or not, it’s a genocidal hyper task.
English
This situation depends on being 100% sure everyone down the line is making the good choice but not only that. Someone could misunderstand, or simply trip on the lever.
And let’s be clear: it’ll happen. A psychopath, an elderly, someone intoxicated or just having a bad day.
You should minimize suffering by pulling the lever.
While it doesn’t change the outcome, the reasoning also depends on the game getting back to 1 after choosing to kill or the game stopping.
In the second case, kill ASAP, it’s the minimum person to be killed.
In the first, it’s actually a choice of linear versus exponential killing series, so if choice and situation is to be made in fixed time, always kill even one person (only the 1 -> 2 transition is free so if we want to lift the burden of culpability you can skip one and kill 2 and it’s still linear).
So, in the worst case, you should still pull and hope many people pull.
If time isn’t a factor here, the daemon putting us to choice just wants everyone dead so do whatever you want, linear or not, it’s a genocidal hyper task.
brainrot dimension@brainrotpostig
English
@nellymacbeal Stop bousculeshaming en fait
Plus sérieusement je me souviens quand j’étais gosse que des qu’il y a un peu de monde personne te calcule
Français
Mon fils, 8 ans, se plaint souvent du fait que dans la rue il est bousculé par les adultes, comme s’il n’existait pas. J’ai donc observé. Et bien il a raison…
Français
@levelsio I couldn’t ;-; and last time I checked I’m not a bot (or I’m a water proof one cuz I took a bath yesterday).
Anyway I feel sad
English
Okay and who of you couldn't reply to this? Please please only if you couldn't okay?
@levelsio@levelsio
Okay let's see who can reply to this
English
Man I’m thinking back about KAN at least once a month. I have reasonable ideas, mostly fueled by the intuitive (maybe wrong) feeling of some class of decoupling here that could benefit LLMs or ML in general - even if not being models per se.
I’m however not in a research position and have nobody to chat to about this - only a few prototypes here and there.
Would you mind us discussing this? Would be happy to read your paper if you have the ref btw.
English
I wrote a paper on it , it's non parellalizable nature is it's biggest bottleneck. But yes , i believe it's promising , the core feels really strong to me ... Even though i believe a large enough MLP can reach the similar results , but the core thesis of KAN makes me believe that it is really valuable in real world scenarios, as the splines/ any activation function we use , can become a really good way to steer the models, or understand internal representations.
I have few ideas regarding it which i wanna explore as well.
Btw the paper was in Driverless Vehicle path planning and Control Systems domain.
English
Whatever happened to Kolmogorov-Arnold Networks? Did it ever turn out to be useful??
English
Either it doesn’t work or it’s just isomorph to classic weights or my genius idea I never worked on to constrain the f_n families of decomposition for a specific class of easy-to-compute/easy-to-LUT while allowing only g(·) to be arbitrarily complex and recursively collapsed to known function, wasn’t ever worked correctly
English
@memoart_dadrien Pas le temps demain j’ai randonnée, hyper tôt en plus.
Français
Dites en commentaire que vous avez reconnu cet écrivain sans n’écrire ni son nom ni son prénom !
Vous devez donner des indices ⬇️
Français
@Flarelda @YoonnsOfficiel J’en arrive à, malheureusement, utiliser l’expression maudite. Jeudi en 8.
Français
@LilianDelaveau @YoonnsOfficiel Enfin jeudi avec n'importe quel autre jour ou week-end évidemment.
Français
@NumaBuilds POV il a optimisé ta vie t’es marié et t’as une assurance vie
Français
Auto-mode tu reviens 5min plus tard il t'as vidé ton wallet et il commencé a draguer ta meuf avec un alt acc
Claude@claudeai
New in Claude Code: auto mode. Instead of approving every file write and bash command, or skipping permissions entirely, auto mode lets Claude make permission decisions on your behalf. Safeguards check each action before it runs.
Français
@Saamodeus Bonus pour l’image générée par ia (je m’en tape) qui mets des lignes de flux d’air n’importe comment (je m’en fous pas je bouille)
Français
Y a pas de génie automobile qui permet aux insectes de vesqui les voitures. Y a moins d'insectes sur les pare-brise parce que y a un déclin des populations d’insectes lié notamment à la destruction de leurs habitats et une surutilisation des pesticides
Captain Louis Renault@luizhfeliciano
Français