Research scientist @valeoai | Teaching @Polytechnique @ENS_ULM | Alumni @upb1818 @Mines_Paris @Inria @ENS_ULM | Feedback: https://xitter.cf/t.co/MHAm0ClYvh
Paris, France
Joined November 2008
- Tweets10K
- Following1.4K
- Followers8.5K
- Likes19.4K
Pinned Tweet
The unreasonable magic of simplicity!
Meet DrivoR (Driving on Registers): our latest end2end autonomous driving model.
We teared down complex dependencies & modules from current models to
obtain a pure Transformer-based SOTA driving agent (NAVSIM v1 & v2, HUGSIM).
Find out more👇
Andrei Bursuc retweeted
Why are French university students lining up for lunch at 11am when the cafeteria opens at noon?
I took this picture on my way to work this morning because it is a good lesson in “shadow prices.”
In May, France extended €1 university meals to all students. Previously, those not eligible for assistance paid €3.30.
The result was not cheaper meals, but shortages.
Demand surged, and students now queue an hour or more in advance to secure their lunch. What they used to pay with money, they now pay with time.
And this is a lesson in shadow prices. It tells us that students are willing to sacrifice an additional hour to save €2.30, so they are implicitly valuing their time at €2.30 an hour or less.
That's awful. 11 am is one of the best times to study. You are alert, but not yet tired. Even if a fraction of them studied at this time, there could be more value than the one saved by the “cheaper” meal.
While the government did provide €50 million in additional funding for 2026, for a system serving more than 3 million higher-education students (about 8 cents per student per day), the problem illustrates something deeper about France.
Don't get me wrong. I love living here, and I work hard every day to contribute to this country. But too many French institutions operate as though people's time were free.
It is a system where enormous resources flow through the state, and simple transactions are replaced by forms, permissions, and waiting. People grow accustomed to this dependency, sometimes even demanding more of it.
The implicit message of the system is that your time is cheap, your initiative is unwelcome, and your ability to decide for yourself is something to be managed.
It is a sad situation because I genuinely believe France can be one of the best places in the world to live.
But people are tired. And I'm not talking about those who looted sporting goods stores and vandalized public buildings over the weekend. I'm talking about the butchers, bakers, and produce sellers I meet every week at the market. The small antique dealers who worry that a mob could destroy the little capital they have. I talked to many of them yesterday. They are scared and they know there is nothing for them to do other than hide.
A country that values its people should value their time.
Vive la France!
Well articulated post by @arthurmensch on why Europe should aim for frontier AI as well & be more optimistic.
Don't know if Europe will catch up with this immediate wave, but working towards ambitious aims will eventually lead to all sorts of cool stuff whether aimed or sidetrack
Andrei Bursuc retweeted
Today we’re releasing Muse Spark Open, the most capable US open-weight model to date.
Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
Key Model Specs:
→ 816B total / 42B active parameters
→ 1M token context window
→ Natively multimodal across text, image, video, audio
Andrei Bursuc retweeted
We are thrilled to announce that we have joined forces with @kesai_labs towards building generalist autonomous driving systems together.
We aim high and plan to release models and code to push forward research boundaries.
valeo.com/en/valeo-and-kesai…
Thrilled about the partnership between @Valeo_Group and @kesai_labs towards open and high performing generalist autonomous driving systems.
An ambitious agenda ahead of us together with a dream team of researchers.
Let’s go! 🚗
We are excited to announce that we have joined forces with Valeo to launch a new partnership aimed at building generalist autonomous driving systems. Together, we will openly release models, code, and weights to advance physical AI research.
kesai.eu/blog/2026-10-13-val…
Andrei Bursuc retweeted
A glimpse into #computervision history.
The inaugural #ECCV1990 was held over 4 days in Antibes, France.
• ~40 attendees
• 63 accepted papers
• 60 oral presentations
• 20 posters
• 13 sessions
Fast forward 36 years. How times have changed!
Big fan of such works ever since ASIF (Coupled Data Turns Unimodal Models to Multimodal Without Training) came out a while ago, though without much traction until now arxiv.org/abs/2210.01738
Nice work!
DINOv2 has never seen a caption, and Qwen3 has never seen an image. We still aligned their embedding spaces without a single image-caption pair.
It even works when the images and the captions come from different datasets.
Project page: dominik-schnaus.github.io/un…
⬇️
Andrei Bursuc retweeted
For everybody sharing their papers here, if you have a bit of scientific decency left, please add your sota table and a few ablation results at least at the bottom of your thread…🧵 if not, please consider labeling your post as advertisement.
Andrei Bursuc retweeted
1/n Our paper REPA-G is accepted at #NeurIPS2026! 🎉 I'll present it at the NeurIPS Paris site 🇫🇷.
We introduce test-time conditioning of REPA-style diffusion models with visual features, with no extra training.
Andrei Bursuc retweeted
Besides scalability, self-supervised learning is attractive because we have a good real-life example of humans/mammals learning strong visual systems.
BUT: they do not learn from shuffled image collections. They learn from (essentially) a single long visual stream.
This is what we tackle in our new @NeurIPSConf paper, led by Ivan, together with @LukasKnobel1 and me. 1/
arxiv.org/abs/2609.40333
The much awaited @stateofai report 2026 is now up.
At this crazy level of things happening in AI, this is becoming the one tool to keep some sanity.
Thanks @nathanbenaich for the amazing diligence.
welcome to the 9th annual @stateofai report!
ai building better ai, world models, physical ai, geopolitics, cyber defense and 12-month predictions.
read it at stateof(dot)ai or watch my editor’s cut here.
a few of this year’s biggest findings follow:
Meet 𝚆𝚑𝚎𝚛𝚎-𝙾𝙿𝙳: a simple strategy to generate synthetic data and tasks for MLLM on-policy self-distillation to improve visual perception.
Counting is a super useful skill to have and giving the teacher hints about where to look makes this work.
Andrei Bursuc retweeted
In LLM post-training, policy gradient is "the RL loss".
When it fails, we blame the RL: exploration, credit assignment, sampling noise.
Image classification has none of these, and the policy gradient is exact. Following it is terrible...
ImageNet, 4% vs 62% for cross-entropy.
Andrei Bursuc retweeted
1/n 🚀 New paper 𝚆𝚑𝚎𝚛𝚎-𝙾𝙿𝙳
In on-policy self-distillation, the student learns from a privileged version of itself.
In recent methods, privilege often comes from a better view of the image.
What if, instead of a better view, the teacher knew where to look? 👀
Andrei Bursuc retweeted
We let an agent self-improve by exploring and internalizing ”on this sort of task, keep this sort of thing in mind“ bits of self-feedback. It learned to solve tasks where the original policy failed 128 attempts in a row and GRPO flatlined.
Paper: arxiv.org/abs/2609.37633
🧵1/8
Andrei Bursuc retweeted
Maybe computer vision conferences need something like an "AI methods" track. This would not be a traditional paper -- it would describe how you leveraged a frontier model to solve an outstanding problem.
The publication would describe the problem, give the prompts, resulting code, demo, etc. This would allow the field to say "ok, that problem is solved sufficiently well" and move on.
This acknowledges something that's been bothering reviewers for a long time. Reviewers feel papers require technical novelty. If one uses a frontier model, there may not be anything technically interesting. But the fact that an open problem is solvable, and how you solve it, is interesting.
I would not have a full review process for such things. An abstract and a link to a demo might be enough. People would get a poster or demo space to show what they did.
I can imagine that practical sessions where people get insights about how to solve problems today could be interesting.
Just a thought.
Andrei Bursuc retweeted
I’ll take the opposite point, that for the vast majority of people, it probably doesn’t make sense to do a PhD in AI right now, and is better to join industry. I say this with love for Phillip, whose work and philosophy I do admire, and respect for putting this out there!
Everything in AI is, generally speaking, working well. Most things come from stacking wins and good ideas over and over again, and if you’re in industry, you have a much better opportunity to learn what is actually working and what doesn’t work, and then push the forefront of what’s possible at a much larger scale. Internships also don’t really feel the same because you’re certainly not working on the most important projects for the company and often come in with a high risk agenda and work alone. In industry, you learn a lot about how to manage budgets, work in large teams, and push projects that require a lot of effort and work to get done.
Academic project budgets are generally quite minuscule, maybe being O($1 - $100K) for many projects and you’re often working alone. If you really want to push novel ideas, another route I don’t think most people consider is also a startup, where you may be able to raise 1-100x more money to do the same project you may do in a PhD, with more resources and the ability to hire other people. And, if it goes well, the ability to scale it dramatically.
When it comes to hiring, often times, it’s easier to get an offer having worked well with other people in industry and then having them change jobs and bring you over. Doing a PhD can get your foot in the door, but often it’s still for entry level positions after doing a 5-year PhD. In industry too, you also will make a lot of strong connections, similar to those you’d make in academia. Though it might be the case that it’s good to do 1 year of a PhD and then drop out to work in industry at this point to get the benefit of both.
Doing a PhD right now is on extremely hard mode, where not much work ends up mattering and even if it works, there’s high cost to change the ways of industry to adapt something that might provide incremental gains when everything generally is working well with enough resourcing and effort.
Students sometimes ask me if it still makes sense, in this accelerating age, to pursue a PhD in AI.
Perhaps counterintuitively, I think it's a great time to do so.
I wrote up some thoughts on this here: web.mit.edu/phillipi/www/wri…