Bertrand Charpentier

Happy to announce NatPN at #ICLR2022 (Spotlight) ! - It predicts uncertainty for many supervised tasks like classification & regression. - It guarantees high uncertainty for far OOD. - It only needs one forward pass at testing time. - It does not need OOD data for training.

Fruits are very appreciated by @PrunaAI, and specially avocados!😁🥑

Cute and adorable,I created an animation video using P-Video-Animation from Pruna AI. Prepare a cartoon image and a reference video. The result is high quality and true to the cartoon character. @PrunaAI #madebypruna

With the @PrunaAI playground, you can (1) generate content in seconds and for free with the fastest AI models, and (2) take inspiration from high quality examples!

@Liedangi @PrunaAI Wow, very cool video!

Just made this video with P-Video-Avatar It is very quick to use,simply provide an image and a prompt. The video is ready to be generated and here is the result. offering quality, speed and a user friendly experience. @PrunaAI #madebypruna

@Aicreativecraft @PRUNA @PrunaAI That is super cool! Do not forget to submit it: app.gimi.co/en/campaigns/m… 😀

Very excited about this challenge. People are super creative with @PrunaAI models! I will try to reshare my favorite creations ;)

@OiiOii_AI Really cool AI movie competition!

OiiOii AI's Into the Paradox Forge Contest is now LIVE! 🔥 Create an AI film, have it reviewed by industry judges, and compete for up to $5,000 in prizes. 📅 July 1 – July 31 Ready to bend reality? Join and win the BIG prize!

P-Image is still the most efficient model on the Pareto front.

@sampatil1010 Thanks @sampatil1010!

@Bertrand_Charp You guys are on fire 🔥

P-Image-Try-On is now the fastest model for virtual try on. We keep releasing the most efficient models for each use case!

Super excited to now have all @PrunaAI models on @CloudflareDev!

Happy to support AI founders to build with the most efficient AI models!

@LoicBerthelot No one can keep up with them Mission impossible

trying to figure out Anthropic's next move like:

@tibo_maker I remember Tweet Hunter being super good in the past, but trying to use it again recently and it was practically unusable

if you're selling your startup - don't I sold my startups for $8m & I'll never sell again quick context: I sold Tweet Hunter and Taplio for $8m in 2024. $2m upfront, the rest as an earnout based on performance - on paper, a dream exit in reality, 3 things broke me first, I gave up my baby. the products I spent years building, the vision I had in my head all of it went to someone else. and now I sit here watching the new owners ship things I'd never ship, kill features I loved, and just make the product shitty looking at Tweet Hunter like this hurts more than I expected it to 😞 -- second, the earnout was kind of a 2-year prison we wanted a high multiple, so we agreed to performance-based milestones that meant 2 years of waking up every single day knowing the only thing that mattered was hitting aggressive revenue targets I was technically still running the company but I had a boss now: the contract 🥲 hitting milestones used to feel like winning but with an earnout it flips. every win is just dodging a loss and missing one doesn't feel like falling short - it feels like being a loser -- third, the money itself breaks you - nobody warns you about it you have no idea what to do with that kind of money you don't get the time to adjust to it and you start taking bad decisions. you stop identifying as the same person you were before and you go a little crazy, faster than you think and then comes the worst part after all this, when I got back to building, I realized everyone around me already saw me as the "successful guy who exited" and that froze me for months I couldn't ship anything new because I was scared of breaking the perception I've talked to a lot of founders who sold their companies. every single one of them feels this but I got back to it anyway - building is what I love and today I'm building a SaaS portfolio that crossed $1m MRR this year so if you're sitting on a term sheet right now, do me a favor don't read what the buyer is offering read what you'd be giving up and then make a decision

@magnific This is very cool. I love the 3D playground

Edit any image, layer by layer, in Magnific No more rebuilding a design from scratch over one typo Auto Layers fixes that: → Drop any image → Get text, subject, background as layers → Creative control, down to the detail Try it in Designer, Spaces, and Magnific MCP

@anulagarwal running ads is the new meta haha

i get better $$$ returns from running ads on my mobile game than investing in the stock market i invest $100 -> get back $150 by end of the month +50% in a month pretty crazy scaling it up slowly now, wish me luck 🫡

P-Video-Animate is next level 🚀 Upload a static image and reference video, then the image will be flawlessly animated with the motion of the video. Highest quality out there and it won’t break your bank. Try it in the playground: t.co/kUV7WzUXyd Everywhere else you can try the model: @eachlabs, eachlabs.ai/pruna/p-video/… @replicate, replicate.com/prunaai/p-vide… @runware, runware.ai/models/prunaai… @scenario_gg, scenario.com/models/p-video… @wavespeed_ai, wavespeed.ai/models/pruna-a… @wiroai, wiro.ai/models/pruna-p… @TellersAI, buff.ly/RBmFEco Check out all our @PrunaAI models: pruna.ai/all-models

Amazing AI workflow for podcast creation!

ICLR 2026, UC Berkeley: “Which module types should you adapt during finetuning?” The original LoRA paper recommended attention modules. He et al. (2021) later found MLP modules worked better for some models. Neither answer is universal, and inserting LoRA into every module type costs more parameters than needed while slowing down training. The key idea in PLoP (Precise LoRA Placement) is not all modules deserve the same finetuning. The authors propose an alignment score that ranks each module type by how much room it has to adapt to a given task, and the modules with the lowest alignment are the ones LoRA should target. The score is cheap to compute, it only requires a single forward pass with batch size 200 and no gradient computation. The placement decision can be merged with the first step of LoRA finetuning and adds no meaningful overhead. Their results contradict the conventional defaults. Across Llama, Qwen, and Gemma the modules with the highest alignment scores tend to be Query and Key, and they turn out to be the worst LoRA placement targets when tested empirically. PLoP consistently picks combinations involving Value, Output, and Down-projection modules instead. Empirically, PLoP beats both attention-only and MLP-only placement on Qwen3-1.7B for GRP math reasoning, hitting 74..52% on GSM8K versus 72.13% for attention-only at a similar budget. It also beats inserting LoRA into every module type while using a fraction of the trainable parameters. The main takeaway for anyone running LoRA finetuning at scale is that the placement decision is cheap to optimize and the gains compound across runs. You can check out the paper here: openreview.net/pdf?id=3lGkVgN…

KVPress is now in @PrunaAI package to easily compress the KVCache!