@alth.fr - Al_th

21 Mar 2025 10:43 +0000

https://www.lemonde.fr/m-perso/article/2025/03/21/le-running-c-est-pas-du-flan-quoique_6584124_4497916.html "Le flan n’est pas prétentieux comme un macaron ; il est « terroir », mais pas snob comme le pâté en croûte. " Je vous jure, je suis trigger de fou.

25 Feb 2025 06:58 +0000

A bit frustrated by how arXiv accounts are integrated in the #MLSky feed. Endless scrolling of links without context is uninformative, and just leads to me to ignore them all. I can block but is this really a good route…

28 Feb 2025 12:36 +0000

Impressive. I really like the fact that you can interrupt. It's always difficult to speak with an AI algorithm, or even do speech to text as the moment you stop talking, it's the algorithm's "turn". IRL we do not play turn-based, it's much more subtle than that. https://www.sesame.com/research/crossing_the_uncanny_valley_of_voice#demo

25 Feb 2025 06:58 +0000

24 Feb 2025 07:02 +0000

Il faut les comprendre. Le MO coûte cher ! 😂

22 Feb 2025 09:31 +0000

Il y a une discussion entre deux personnes dans cette salle. Celle qui porte la culotte, si vous me pardonnez cette expression un peu datée, n’est pas celle que vous croyez. Go girl ! [contains quote post or other embedded content]

07 Feb 2025 12:46 +0000

An arguably "easy to read" simple GRPO implementation, for teaching purpose #MLSky [contains quote post or other embedded content]

06 Feb 2025 17:07 +0000

It’s really funny to me that the hottest RL algorithm in town is just a simplification (z-score normalization for advantage calculation) of a simplification (KL penalization over hard KL constraint). GRPO is quite intuitive, although I guess the devil is in the details and « convergence » speeds

05 Feb 2025 17:20 +0000

I implemented GRPO from scratch to RL a tiny toy LLM and it works surprisingly well. Rule base reward inspired by @dorialexander.bsky.social to make my Shakespeare shout more. I went for Outcome Supervision as both OS and PS we’re kind of close in DeepseekMath paper…

26 Jan 2025 08:19 +0000

Je suis mort de rire : un médecin retraité >>>>n’exerçant plus aucune activité médicale rémunérée <<<< doit toujours filer ses 100 balles a l’ordre des médecins 😂

24 Jan 2025 09:15 +0000

AGI milestone passed

23 Jan 2025 10:55 +0000

J’aime cet article parce qu’il utilise le prétexte d’un fait divers pour éduquer. Mieux, il éduque à la fois sur la sémantique et sur la prise en charge des cancers. [contains quote post or other embedded content]

22 Jan 2025 17:30 +0000

I must be using DeepSeek’s distilled r1 model wrong but I’m sincerely unimpressed by the reasoning bit. Reasoning often gets circular, and just gets ignored at the end by the model. My inputs might be too vague for current capabilities probably…

22 Jan 2025 16:32 +0000

Y’en avait un peu marre de la frangipane, alors j’ai œuvré avec une galette cochonne, comme dirait Verot… [contains quote post or other embedded content]

21 Jan 2025 13:43 +0000

I unfortunately don't have access to the underlying article. That being said, I remember vividly being under critics on X, when I discussed the data on organic food health benefits and argued for a cautious stance (see second skeet below) 1/2 https://www.science.org/content/article/people-are-bad-reporting-what-they-eat-s-problem-dietary-research

21 Jan 2025 07:18 +0000

Pecan biscuit, apple and dates, whipped cream 🍽️

21 Jan 2025 07:11 +0000

Does the NLP community know if there is a context size threshold under which KV caching does not make sense ? I guess for small sizes, matmul may be cheaper than slicing of a PyTorch tensor ? #NLP #LLM

15 Jan 2025 08:26 +0000

Sur le réseau social du démon, j'ai à N=3 liens (record : d'habitude l'algo me pousse de l'ED en direct), une vidéo de Bardella qui """"défend la liberté d'expression"""" et prend en exemple les élections Roumaines. Rappel que @flefgraph.bsky.social fait un boulot d'utilité publique... [contains quote post or other embedded content]

08 Jan 2025 18:52 +0000

Savory butternut pie / Tarte au butternut 🍽️

07 Jan 2025 15:31 +0000

Zuckerberg's take on content moderation is political. No doubt. More puzzling is that he frames the message by spitting on the performance of Meta's content moderation algorithms. "The issue with complex systems is that they make mistakes [...] and we reached a point where it's too many mistakes"