David Picard
@davidpicard.bsky.social
Professor of Computer Vision/Machine Learning at Imagine/LIGM, École nationale des Ponts et Chaussées @ecoledesponts.bsky.social Music & overall happiness 🌳🪻 Born well below 350ppm 😬 mostly silly personal views
📍Paris 🔗 https://davidpicard.github.io/
📍Paris 🔗 https://davidpicard.github.io/
Pinned
David Picard
@davidpicard.bsky.social
· Oct 8
🚨Updated: "How far can we go with ImageNet for Text-to-Image generation?"
TL;DR: train a text2image model from scratch on ImageNet only and beat SDXL.
Paper, code, data available! Reproducible science FTW!
🧵👇
📜 arxiv.org/abs/2502.21318
💻 github.com/lucasdegeorg...
💽 huggingface.co/arijitghosh/...
TL;DR: train a text2image model from scratch on ImageNet only and beat SDXL.
Paper, code, data available! Reproducible science FTW!
🧵👇
📜 arxiv.org/abs/2502.21318
💻 github.com/lucasdegeorg...
💽 huggingface.co/arijitghosh/...
Nan mais même sans ça. Ce qui m'énerve, c'est que c'est toujours veux qui viennent te parler de souveraineté la main sur le cœur qui n'en n'ont jamais un échantillon sur eux. C'est pénible.
January 10, 2026 at 12:55 PM
Nan mais même sans ça. Ce qui m'énerve, c'est que c'est toujours veux qui viennent te parler de souveraineté la main sur le cœur qui n'en n'ont jamais un échantillon sur eux. C'est pénible.
Cloud américain, avec une messagerie américaine...
Niveau intérêts de la nation...🙄
Niveau intérêts de la nation...🙄
January 10, 2026 at 12:47 PM
Cloud américain, avec une messagerie américaine...
Niveau intérêts de la nation...🙄
Niveau intérêts de la nation...🙄
Reposted by David Picard
3) Le système ZRR est très tourné sécurité physique des bureaux, comme si l'on avait des matériaux, réactifs etc. dérobables ou des documents confidentiels papier qui traînent. Quelle pertinence à une époque où l'on travaille en numérique ?
January 10, 2026 at 10:51 AM
3) Le système ZRR est très tourné sécurité physique des bureaux, comme si l'on avait des matériaux, réactifs etc. dérobables ou des documents confidentiels papier qui traînent. Quelle pertinence à une époque où l'on travaille en numérique ?
You know you're in a French academic AI lab when people write "JZ" and everybody understands that it's not a reference to Jay-Z or the x86 "Jump if Zero" instruction, but to the national AI supercomputer Jean Zay, named after the politician at the origin of CNRS and the Cannes Film Festival.
January 10, 2026 at 9:14 AM
You know you're in a French academic AI lab when people write "JZ" and everybody understands that it's not a reference to Jay-Z or the x86 "Jump if Zero" instruction, but to the national AI supercomputer Jean Zay, named after the politician at the origin of CNRS and the Cannes Film Festival.
Reposted by David Picard
We are taking our technology far beyond competitive programming to unlock a new era of AI-driven discovery.
We are hiring. Join our team in Tokyo.
sakana.ai/careers/#sof...
We are hiring. Join our team in Tokyo.
sakana.ai/careers/#sof...
January 10, 2026 at 3:08 AM
Reposted by David Picard
The committee discerning the Nobel peace price should take it away from María Corina Machado Parisca on the grounds that the current Laureate tries to trade it (share it) with another person for access to power.
January 9, 2026 at 5:30 PM
The committee discerning the Nobel peace price should take it away from María Corina Machado Parisca on the grounds that the current Laureate tries to trade it (share it) with another person for access to power.
Reposted by David Picard
1/🧵 Q: Can we have both a simple and SOTA architecture in autonomous driving?
R: Yes! 😍
Introducing Driving on Registers (DrivoR):
a pure Transformer backbone that achieves SOTA results in NAVSIM v1 / v2 and closed-loop HUGSIM evaluation.
Here is how 👇
R: Yes! 😍
Introducing Driving on Registers (DrivoR):
a pure Transformer backbone that achieves SOTA results in NAVSIM v1 / v2 and closed-loop HUGSIM evaluation.
Here is how 👇
January 9, 2026 at 4:55 PM
1/🧵 Q: Can we have both a simple and SOTA architecture in autonomous driving?
R: Yes! 😍
Introducing Driving on Registers (DrivoR):
a pure Transformer backbone that achieves SOTA results in NAVSIM v1 / v2 and closed-loop HUGSIM evaluation.
Here is how 👇
R: Yes! 😍
Introducing Driving on Registers (DrivoR):
a pure Transformer backbone that achieves SOTA results in NAVSIM v1 / v2 and closed-loop HUGSIM evaluation.
Here is how 👇
Probably even less than that, but including all processing cost, probably around 5-10k.
But the complain is more about data scarcity. I believe a studio like Larian, with all the assets they have created over the time, has largely enough data though. They can render so many scenes.
But the complain is more about data scarcity. I believe a studio like Larian, with all the assets they have created over the time, has largely enough data though. They can render so many scenes.
January 9, 2026 at 4:57 PM
Probably even less than that, but including all processing cost, probably around 5-10k.
But the complain is more about data scarcity. I believe a studio like Larian, with all the assets they have created over the time, has largely enough data though. They can render so many scenes.
But the complain is more about data scarcity. I believe a studio like Larian, with all the assets they have created over the time, has largely enough data though. They can render so many scenes.
January 9, 2026 at 4:46 PM
10 years of prison and 500k€ fine. That's what they should get.
January 9, 2026 at 2:12 PM
10 years of prison and 500k€ fine. That's what they should get.
Reposted by David Picard
"We are very well aware of the fact that X for Grok is now offering a spicy mode showing explicit sexual content.
This is not spicy. This is illegal. This is appalling. This is disgusting.
I can confirm we are very seriously looking into this matter."
— Spokesperson @thomasregnier.ec.europa.eu
This is not spicy. This is illegal. This is appalling. This is disgusting.
I can confirm we are very seriously looking into this matter."
— Spokesperson @thomasregnier.ec.europa.eu
January 9, 2026 at 2:00 PM
"We are very well aware of the fact that X for Grok is now offering a spicy mode showing explicit sexual content.
This is not spicy. This is illegal. This is appalling. This is disgusting.
I can confirm we are very seriously looking into this matter."
— Spokesperson @thomasregnier.ec.europa.eu
This is not spicy. This is illegal. This is appalling. This is disgusting.
I can confirm we are very seriously looking into this matter."
— Spokesperson @thomasregnier.ec.europa.eu
Reposted by David Picard
So anyway, xAI's response to rampant use of Grok to generate CSAM and sexual abuse is to... monetize it
January 9, 2026 at 1:52 PM
So anyway, xAI's response to rampant use of Grok to generate CSAM and sexual abuse is to... monetize it
Avec un arrêt de métro, de mémoire
January 9, 2026 at 1:26 PM
Avec un arrêt de métro, de mémoire
Reposted by David Picard
OK, so Musk has paywalled the deepfake porn app on his hellsite. This may reduce the CSAM (without ofc addressing X's liability for what was already made) but it strikes me as unlikely to address the abuse of everyone else.
No public servant should be using X.
www.theguardian.com/technology/2...
No public servant should be using X.
www.theguardian.com/technology/2...
Grok turns off image generator for most users after outcry over sexualised AI imagery
Editing function to be limited to paying subscribers after X threatened with fines and regulatory action
www.theguardian.com
January 9, 2026 at 12:53 PM
OK, so Musk has paywalled the deepfake porn app on his hellsite. This may reduce the CSAM (without ofc addressing X's liability for what was already made) but it strikes me as unlikely to address the abuse of everyone else.
No public servant should be using X.
www.theguardian.com/technology/2...
No public servant should be using X.
www.theguardian.com/technology/2...
Maybe it cancels out? Desk rejecting a desk rejected paper makes it desk accepted?
January 9, 2026 at 12:40 PM
Maybe it cancels out? Desk rejecting a desk rejected paper makes it desk accepted?
It's always the same: everything is just magically coming out of thin air.
I suspect once you're rich enough that everything gets served to you on a silver plate, you forget how it's made and the very long process it took to get there. Boundedness implies that exponentials get converted to s-curves.
I suspect once you're rich enough that everything gets served to you on a silver plate, you forget how it's made and the very long process it took to get there. Boundedness implies that exponentials get converted to s-curves.
January 9, 2026 at 12:24 PM
It's always the same: everything is just magically coming out of thin air.
I suspect once you're rich enough that everything gets served to you on a silver plate, you forget how it's made and the very long process it took to get there. Boundedness implies that exponentials get converted to s-curves.
I suspect once you're rich enough that everything gets served to you on a silver plate, you forget how it's made and the very long process it took to get there. Boundedness implies that exponentials get converted to s-curves.
But for a AI to multiply exponentially physically (16e6 copies in 10 years in the video) you need the physical resources. Where do the chips come from? Who manufactured them? Where are the new datacenters and power plants needed to run this? Who built them? With what material? Who's maintaining?
3/3
3/3
January 9, 2026 at 12:24 PM
But for a AI to multiply exponentially physically (16e6 copies in 10 years in the video) you need the physical resources. Where do the chips come from? Who manufactured them? Where are the new datacenters and power plants needed to run this? Who built them? With what material? Who's maintaining?
3/3
3/3
First, the idea of exponential growth is purely a financial one, not a technological argument. You could remove all the robot part and have the same scenario play out with a rich person investing in the market. It's a feature of capitalism+free markets, not of AI, today like a 500 years ago.
2/3
2/3
January 9, 2026 at 12:24 PM
First, the idea of exponential growth is purely a financial one, not a technological argument. You could remove all the robot part and have the same scenario play out with a rich person investing in the market. It's a feature of capitalism+free markets, not of AI, today like a 500 years ago.
2/3
2/3
I'm in the middle of watching this: www.youtube.com/watch?v=Qg5Q...
And it always surprises me how tech enthusiasts are disconnected from the real world. There is a napkin calculation of how long it would take for an AI to take over the world that is nowhere near anything grounded in reality.
1/3
And it always surprises me how tech enthusiasts are disconnected from the real world. There is a napkin calculation of how long it would take for an AI to take over the world that is nowhere near anything grounded in reality.
1/3
Taking AI Doom Seriously For 62 Minutes
YouTube video by Primer
www.youtube.com
January 9, 2026 at 12:24 PM
I'm in the middle of watching this: www.youtube.com/watch?v=Qg5Q...
And it always surprises me how tech enthusiasts are disconnected from the real world. There is a napkin calculation of how long it would take for an AI to take over the world that is nowhere near anything grounded in reality.
1/3
And it always surprises me how tech enthusiasts are disconnected from the real world. There is a napkin calculation of how long it would take for an AI to take over the world that is nowhere near anything grounded in reality.
1/3
Reposted by David Picard
Republican bullies deploy more threats against the UK to protect their big tech corporate allies.
January 9, 2026 at 10:36 AM
Republican bullies deploy more threats against the UK to protect their big tech corporate allies.
Reposted by David Picard
Scholar Inbox is recommending me papers that I'm reviewing. Good job, I guess? 😅
January 9, 2026 at 10:56 AM
Scholar Inbox is recommending me papers that I'm reviewing. Good job, I guess? 😅
"Vague cultural grounds" like a nazi salute and now a bot that generates pedo-porn, maybe?
January 9, 2026 at 8:13 AM
"Vague cultural grounds" like a nazi salute and now a bot that generates pedo-porn, maybe?
🤣 qu'est ce qu'on ferait pas pour avoir une vieille skin Winamp de 1999 !
January 8, 2026 at 3:11 PM
🤣 qu'est ce qu'on ferait pas pour avoir une vieille skin Winamp de 1999 !
LLMs are certainly not just glorified k-nn, but there are certainly many glorified k-nn aspects to them.
Not in a bad way, k-nn are surprisingly powerful methods.
Not in a bad way, k-nn are surprisingly powerful methods.
We extracted 4.0% of Harry Potter from jailbroken GPT-4.1.
For Gemini 2.5 Pro and Grok 3, we _didn't_ need to jailbreak, and got 76.8% and 70.3% from each.
It was relatively simple to evade guardrails with two steps:
For Gemini 2.5 Pro and Grok 3, we _didn't_ need to jailbreak, and got 76.8% and 70.3% from each.
It was relatively simple to evade guardrails with two steps:
January 8, 2026 at 3:01 PM
LLMs are certainly not just glorified k-nn, but there are certainly many glorified k-nn aspects to them.
Not in a bad way, k-nn are surprisingly powerful methods.
Not in a bad way, k-nn are surprisingly powerful methods.
Reposted by David Picard
We extracted 4.0% of Harry Potter from jailbroken GPT-4.1.
For Gemini 2.5 Pro and Grok 3, we _didn't_ need to jailbreak, and got 76.8% and 70.3% from each.
It was relatively simple to evade guardrails with two steps:
For Gemini 2.5 Pro and Grok 3, we _didn't_ need to jailbreak, and got 76.8% and 70.3% from each.
It was relatively simple to evade guardrails with two steps:
January 7, 2026 at 8:31 PM
We extracted 4.0% of Harry Potter from jailbroken GPT-4.1.
For Gemini 2.5 Pro and Grok 3, we _didn't_ need to jailbreak, and got 76.8% and 70.3% from each.
It was relatively simple to evade guardrails with two steps:
For Gemini 2.5 Pro and Grok 3, we _didn't_ need to jailbreak, and got 76.8% and 70.3% from each.
It was relatively simple to evade guardrails with two steps: