Melvin Sevi

PhD student @Inria Grenoble. RL finetuning/TTS of image gen models (autoregressive + flow matching). ex-ENS Saclay, ex- Sorbonne U .

[Stanislav Fort @stanislavfort]

We discovered a surprising, training-free way to generate images: no GANs or diffusion models, but a ✨secret third thing✨! Standard models like CLIP can already create images directly, with zero training. We just needed to find the right key to unlock this ability = DAS
1/11

[terminally onλine εngineer @tekbog]

i keep hearing how good LLMs are and how much people generate code with them yet every single time i try to do something relevant it's absolute garbage, are you people just writing simple CRUD apps all day? wtf

[ℏεsam @Hesamation]

32 year old @ylecun shows off the world's first Convolutional Network in 1993 for Text Recognition.

[Machine Learning Street Talk @MLStreetTalk]

Iman Mirzadeh from Apple @i_mirzadeh wrote the famous GSM-Symbolic paper a couple of months back which argued that LLMs are learning surface statistics and not genuinely reasoning due to their sensitivity to distractors and out of distribution examples.

[JFPuget 🇫🇷🇺🇦🇨🇦🇬🇱 @JFPuget]

I used to say LLM can't reason. But I am now convinced that tree search + LLMs can achieve results similar to reasoning. Like tree search can yield better chess players than reasoning players (a.k.a. humans)

[Melvin Sevi @melvin_sevi]

@JokeTmcp2 Without warning over rodeo is insane

[Melvin Sevi @melvin_sevi]

@TravisScott__FR @VOAngo Le lien fonctionne plus !

[brett goldstein @thatguybg]

Sora's video quality seems impossible so I dug into how it works under the hood

it uses both diffusion (starting with noise, refining towards a desired video) and transformer architectures (handling sequential video frames)

read on 🧵

[Doron Adler @Norod78]

A very interesting explanation of what LoRA is and how LoRA works by implementing one from scratch upon an existing DistilBERT model
https://t.co/YvqVq8YLlF

[Sylvain Filoni @fffiloni]

Whisper Speech X DreamTalk ? Yes