when you think about it, it's just impressive ... I remember seing a lot of papers about sentiment analysis / question generation / autocomplete etc ... All of this just vanish with LLMs. It's like computer vision in a sense, all those technical legacy going forgotten due to deep learning.

Those figures in the article are just bad math. The real subscription price of chatGPT is not "200$" as said in the article. It's actually 2$. I'm not joking. 2$ is the inference price for 1M token for a chatGPT like model if you self host the model. 1M token enough reading for one month.

Bluesky doesn't penalize you for using links, tags, using key words like Patreon or commissions, and doesn't have a central algorithm on your following feed, which means all the power rests on you to help your favourite creators be seen!

Quick question : are llms good at terraform ? It feels like having some free terraform skill can be very valuable

it would be a perfect timing for openai to release an open model tomorrow

but frankly (benchmarks wise) it seems that the chinese are very close to US ... (the last qwen seems very good)

c'est clair qu'il y a une perte de souveraineté globale qui n'est pas vraiment bienvenu (alimentaire mais aussi énergétique numérique etc) ... peut-on être moral et faible ?

I hope it will be full diffusion modelling

Gemini 3 omg

good job on gemini cli !

Text Arena | LMArena Compare and explore Text models ranked by Coding performance.

this is curious no ? gemini 2.5 pro seems pretty good at coding :
lmarena.ai/leaderboard/...

bravo vous me direz comment vous avez fait

ai or news fatigue ?

Ça me paraît curieux cette histoire : les données de chat grand public servent souvent à rien pour entraîner des modèles.

they have (almost) all the twitter features and the UI is fast but there is not a lot of people.

Proton Mail: Obtenez un compte de messagerie gratuit avec respect de la vie privée et chiffrement | Proton Proton Mail est la plus grande messagerie sécurisée au monde (plus de 100 millions d'utilisateurs), disponible sur le web, iOS, Android et sur ordinateur.

AWS -> www.ovhcloud.com/fr/ (mais franchement il manque des fonctionalités comme les cloud run / AWS lambda)
365 -> tu as proton.me/fr/mail mais pas de messagerie instantanée (drive et mail ici seulement)

"their data troves are siloed an trapped in archaïc IT sytems" ... it feels like nothing has changed in 10 years :(

mais ça donne envie d'installer linux tout ça

"des jeux" j'imagine qu'il y a beaucoup de jeux compatibles mais la dernière fois que j'ai fait des tests il y avait des trous dans la raquette (Read Dead / Elden Ring).

tu utilises la version steam de linux ?

la seule raison pour laquelle j'ai encore windows c'est la partie gaming (sinon linux est mieux en tous points sauf le support sur certain driver de périphérique)

vous avez déjà essayé de jouer à un jeu steam sur linux ?

the InceptionLabs (another dLLM) model seems to be based on arxiv.org/pdf/2310.16834 (best ICML 2024 paper)

Sure : the "AI con" books are not really helping : there is a more nuance reality than "this is useless / overhype" and "AGI is coming for all the jobs"

🌧️ 2024 figure comme l’une des années les plus pluvieuses depuis 1959 !

Grâce aux données de @meteofrance.com, cette carte permet de mieux comprendre la disparité régionale de ces précipitations, à l’heure d’une multiplication des événements climatiques extrêmes.

👉 www.ign.fr/mag/ou-t-il-...

do you know a better plateform ?

it feels wrong but it's how the field progresses ("random shit" = new architectural design like transform, cnn, or mamba that have vague reason to work well).

Does a benchmark exist for tpu vs gpus (in pricing comparaison) ?

it's just too difficult to move people from one platform to another ... "the value of a social network is proportional to the square of its number of users"

my guess is that is a good idea. I first try PL because it was simple and offers a lot of functionalities (logging / model checkpointing / tricks likes gradiant clipping or gradiant accumulation). Eventually you want to stick to raw pytorch to do distribution.