Guillaume Sautière 🇵🇸
@gsautiere.bsky.social
machine learning researcher in video and audio compression at qualcomm ai research amsterdam.
co-founder of WP Frequencies
organizing raves in and out of amsterdam
naifrec.github.io
https://www.youtube.com/@beincheekym8
https://linktr.ee/wp_frequencies
co-founder of WP Frequencies
organizing raves in and out of amsterdam
naifrec.github.io
https://www.youtube.com/@beincheekym8
https://linktr.ee/wp_frequencies
such a great read from @basicchanel.bsky.social on the "low end psychedelia" french music scene, with great interviews interweaved with awesome tracks. I cannot recommend enough if you want to discover where the electronic music scene that excites me is at! www.untitled909.com/untangling-t...
Untangling the web of the French low end psychedelia scene
Let’s start with how this feature came about. Over the last few years, my obsession with digging on Bandcamp has grown even deeper in relation to my [...]
www.untitled909.com
December 14, 2025 at 1:16 PM
such a great read from @basicchanel.bsky.social on the "low end psychedelia" french music scene, with great interviews interweaved with awesome tracks. I cannot recommend enough if you want to discover where the electronic music scene that excites me is at! www.untitled909.com/untangling-t...
"OpenAI – or other frontier labs – have no fundamental advantage in their inference and infrastructure stacks. The only way to gain an advantage is by having slightly better rack-level hardware optim or data-center-level hardware optim. But these will also run out quickly – maybe 2026, maybe 2027."
“Why AGI Will Not Happen” by Tim Dettmers.
timdettmers.com/2025/12/10/w...
This essay is worth reading. Discusses diminishing returns (and risks) of scaling. The contrast between West and East: “Winner takes all” approach of building the biggest thing vs a long-term focus on practicality.
timdettmers.com/2025/12/10/w...
This essay is worth reading. Discusses diminishing returns (and risks) of scaling. The contrast between West and East: “Winner takes all” approach of building the biggest thing vs a long-term focus on practicality.
December 14, 2025 at 7:26 AM
"OpenAI – or other frontier labs – have no fundamental advantage in their inference and infrastructure stacks. The only way to gain an advantage is by having slightly better rack-level hardware optim or data-center-level hardware optim. But these will also run out quickly – maybe 2026, maybe 2027."
Reposted by Guillaume Sautière 🇵🇸
All travelers from ESTA countries (yes the ones on visa waiver programs) will have to disclose 5 years of social media + huge amounts of personal data to enter the US now. All US academic associations should now meet outside the US if we want to meet our international colleagues.
December 10, 2025 at 4:27 AM
All travelers from ESTA countries (yes the ones on visa waiver programs) will have to disclose 5 years of social media + huge amounts of personal data to enter the US now. All US academic associations should now meet outside the US if we want to meet our international colleagues.
Reposted by Guillaume Sautière 🇵🇸
I'm so exhausted. 🤕
December 9, 2025 at 6:30 PM
I'm so exhausted. 🤕
Reposted by Guillaume Sautière 🇵🇸
I complain a lot about RL lately, and here we go again.
The CS view of RL is wrong in how it thinks about rewards, already at the setup level. Briefly, the reward computation should be part of the agent, not part of the environment.
More at length here:
gist.github.com/yoavg/3eb3e7...
The CS view of RL is wrong in how it thinks about rewards, already at the setup level. Briefly, the reward computation should be part of the agent, not part of the environment.
More at length here:
gist.github.com/yoavg/3eb3e7...
rl-wrong-about-rewards.md
GitHub Gist: instantly share code, notes, and snippets.
gist.github.com
December 5, 2025 at 11:37 PM
I complain a lot about RL lately, and here we go again.
The CS view of RL is wrong in how it thinks about rewards, already at the setup level. Briefly, the reward computation should be part of the agent, not part of the environment.
More at length here:
gist.github.com/yoavg/3eb3e7...
The CS view of RL is wrong in how it thinks about rewards, already at the setup level. Briefly, the reward computation should be part of the agent, not part of the environment.
More at length here:
gist.github.com/yoavg/3eb3e7...
Regulatory and educational seems l part of the solution. I put a lot of hope in enforcing AI forensics tools that would tag suspicious media and force (through regulation) platforms to use such tools. This would put the barrier to entry higher as off the shelf models would be detectable.
The other sort of unhelpful response is to think that we can regulate this away. Models are already out there and the amount of surveillance it would take to stop their use is wildly invasive.
November 27, 2025 at 8:54 AM
Regulatory and educational seems l part of the solution. I put a lot of hope in enforcing AI forensics tools that would tag suspicious media and force (through regulation) platforms to use such tools. This would put the barrier to entry higher as off the shelf models would be detectable.
Reposted by Guillaume Sautière 🇵🇸
Hidden cameras. Surveillance networks. Paranoid delusions.
My new video is about the feeling that you're being watched.
youtu.be/gStjfje6Mq4
My new video is about the feeling that you're being watched.
youtu.be/gStjfje6Mq4
November 23, 2025 at 7:31 PM
Hidden cameras. Surveillance networks. Paranoid delusions.
My new video is about the feeling that you're being watched.
youtu.be/gStjfje6Mq4
My new video is about the feeling that you're being watched.
youtu.be/gStjfje6Mq4
Reposted by Guillaume Sautière 🇵🇸
Cette interview est à écouter. Erner tout à son obsession d'établir "le parti pris, donc l'antisémitisme, donc l'illégitimité" de Francesca Albanese — idéologie médiatique dominante — déroule sa charge anti palestinienne abjecte d'autant plus grossièrement qu'Albanese est parfaite dans ses réponses.
Guillaume Erner face à Francesca Albanese : suspicion et délégitimation www.acrimed.org/Guillaume-Er...
Guillaume Erner face à Francesca Albanese : suspicion et délégitimation
France Culture, 18 novembre 2025.
www.acrimed.org
November 23, 2025 at 12:38 PM
Cette interview est à écouter. Erner tout à son obsession d'établir "le parti pris, donc l'antisémitisme, donc l'illégitimité" de Francesca Albanese — idéologie médiatique dominante — déroule sa charge anti palestinienne abjecte d'autant plus grossièrement qu'Albanese est parfaite dans ses réponses.
Reposted by Guillaume Sautière 🇵🇸
i love writing about crypto crime because these dudes do this constantly
![However, the brothers' Google search history in the weeks after the theft for phrases including "how to wash crypto”, “top crypto lawyers”, “wire fraud statute / wire fraud statue [sic] of limitations”, and “prison or jail worse” suggest that perhaps the brothers had some clue that their actions could be interpreted criminally, as do the brothers' own descriptions of their plan as “sus”.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:exrxvyu6bpoym6mbnctke5tn/bafkreicpwu5pjoddgp2ovmaka2kfu6d6uc7uijrld5a4hovz5qpaqlkk5u@jpeg)
November 20, 2025 at 3:19 AM
i love writing about crypto crime because these dudes do this constantly
Still horrified that this was the outcome of the lawsuit, sets a horrible precedent imo. Flat fee, no royalties or licensing, and also you have to realize you're in the dataset and claim?
Hey authors! Check to see if Anthropic stole your book to train their slop generator on. You’re entitled to $1500 per stolen Work.
Look up your work, and if you’re in the database, file a claim
secure.anthropiccopyrightsettlement.com/lookup/
Look up your work, and if you’re in the database, file a claim
secure.anthropiccopyrightsettlement.com/lookup/
Submit a Claim
secure.anthropiccopyrightsettlement.com
November 19, 2025 at 7:22 AM
Still horrified that this was the outcome of the lawsuit, sets a horrible precedent imo. Flat fee, no royalties or licensing, and also you have to realize you're in the dataset and claim?
Reposted by Guillaume Sautière 🇵🇸
POV: you are a young woman celebrating a recent academic success
November 17, 2025 at 7:20 PM
POV: you are a young woman celebrating a recent academic success
Reposted by Guillaume Sautière 🇵🇸
if anything ever happened to wikipedia i would lose my mind
November 17, 2025 at 7:08 PM
if anything ever happened to wikipedia i would lose my mind
Flock's CEO crashing out and calling deflock "antifa / agent of chaos" is so predictable. Amazing work by Benn.
Such ANPR systems are also deployed outside the US... need to research who's behind them and if the (allegedly) "privacy-preserving" EU has been more careful in the rollout.
Such ANPR systems are also deployed outside the US... need to research who's behind them and if the (allegedly) "privacy-preserving" EU has been more careful in the rollout.
November 17, 2025 at 12:39 PM
Flock's CEO crashing out and calling deflock "antifa / agent of chaos" is so predictable. Amazing work by Benn.
Such ANPR systems are also deployed outside the US... need to research who's behind them and if the (allegedly) "privacy-preserving" EU has been more careful in the rollout.
Such ANPR systems are also deployed outside the US... need to research who's behind them and if the (allegedly) "privacy-preserving" EU has been more careful in the rollout.
November 17, 2025 at 6:56 AM
Reposted by Guillaume Sautière 🇵🇸
very funny to read these emails and then remember summers wondering whether the leaky pipeline in stem was because women are inherently bad at math
The emails have Summers reporting to Epstein about his attempts to date a Harvard economics student & to hit on her during a seminar she was giving.
November 16, 2025 at 1:56 AM
very funny to read these emails and then remember summers wondering whether the leaky pipeline in stem was because women are inherently bad at math
Learning about Alice Wong through these posts and also learning about the ancestors... beautiful and heartbreaking, I will add her work to my reading list and I hope she rests in peace.
Our beloved Alice Wong has joined the ancestors. It was one of the great honors of my life to call Alice my friend, co-author & co-conspirator. She was a true genius, a force of nature the likes of which the world has never seen before. I love you, Alice, and am equal parts grateful and devastated
November 15, 2025 at 11:06 AM
Learning about Alice Wong through these posts and also learning about the ancestors... beautiful and heartbreaking, I will add her work to my reading list and I hope she rests in peace.
Reposted by Guillaume Sautière 🇵🇸
it seems important to note in this moment that all the rich guys who are pushing the adoption of generative AI would benefit greatly from being able to dismiss any evidence of them doing pedo shit as generative AI
November 12, 2025 at 11:46 PM
it seems important to note in this moment that all the rich guys who are pushing the adoption of generative AI would benefit greatly from being able to dismiss any evidence of them doing pedo shit as generative AI
Reposted by Guillaume Sautière 🇵🇸
November 12, 2025 at 3:29 AM
From Guangxuan Xiao
"Its a common belief that L SWA layers (size W) yield L×W receptive field.
[This] post shows why the effective range is limited to O(W) regardless of depth. The reasons are information dilution and the exponential barrier from residual connections" guangxuanx.com/blog/stackin...
"Its a common belief that L SWA layers (size W) yield L×W receptive field.
[This] post shows why the effective range is limited to O(W) regardless of depth. The reasons are information dilution and the exponential barrier from residual connections" guangxuanx.com/blog/stackin...
Why Stacking Sliding Windows Can't See Very Far
Modern LLMs use sliding window attention for efficiency, but why can't stacking sliding windows see as far as theory suggests? A mathematical exploration of information dilution and the exponential ba...
guangxuanx.com
November 10, 2025 at 4:00 PM
From Guangxuan Xiao
"Its a common belief that L SWA layers (size W) yield L×W receptive field.
[This] post shows why the effective range is limited to O(W) regardless of depth. The reasons are information dilution and the exponential barrier from residual connections" guangxuanx.com/blog/stackin...
"Its a common belief that L SWA layers (size W) yield L×W receptive field.
[This] post shows why the effective range is limited to O(W) regardless of depth. The reasons are information dilution and the exponential barrier from residual connections" guangxuanx.com/blog/stackin...
Reposted by Guillaume Sautière 🇵🇸
Replace Variational Autoencoder (VAE) with pretrained representation encoders (e.g., DINO, SigLIP, MAE) paired with trained decoders, which they terms as Representation Autoencoders (RAE).
October 15, 2025 at 3:49 AM
Replace Variational Autoencoder (VAE) with pretrained representation encoders (e.g., DINO, SigLIP, MAE) paired with trained decoders, which they terms as Representation Autoencoders (RAE).
Reposted by Guillaume Sautière 🇵🇸
"The Principles of Diffusion Models" by Chieh-Hsin Lai, Yang Song, Dongjun Kim, Yuki Mitsufuji, Stefano Ermon. arxiv.org/abs/2510.21890
It might not be the easiest intro to diffusion models, but this monograph is an amazing deep dive into the math behind them and all the nuances
It might not be the easiest intro to diffusion models, but this monograph is an amazing deep dive into the math behind them and all the nuances
The Principles of Diffusion Models
This monograph presents the core principles that have guided the development of diffusion models, tracing their origins and showing how diverse formulations arise from shared mathematical ideas. Diffu...
arxiv.org
October 28, 2025 at 8:35 AM
"The Principles of Diffusion Models" by Chieh-Hsin Lai, Yang Song, Dongjun Kim, Yuki Mitsufuji, Stefano Ermon. arxiv.org/abs/2510.21890
It might not be the easiest intro to diffusion models, but this monograph is an amazing deep dive into the math behind them and all the nuances
It might not be the easiest intro to diffusion models, but this monograph is an amazing deep dive into the math behind them and all the nuances
Reposted by Guillaume Sautière 🇵🇸
Elon's new pay package is ironically the same amount required to cover SNAP benefits for the entire country throughout its 10-year term.
November 8, 2025 at 4:42 AM
Elon's new pay package is ironically the same amount required to cover SNAP benefits for the entire country throughout its 10-year term.
Reposted by Guillaume Sautière 🇵🇸
Really great and relatable talk from old mate Jim here about why creative's should make they want to make.
Don't let an algorithm decide what you do, don't let an audience decide what you do, don't let "industry advice" decide what you do, just make shit you think is cool and fun to make
Don't let an algorithm decide what you do, don't let an audience decide what you do, don't let "industry advice" decide what you do, just make shit you think is cool and fun to make
November 6, 2025 at 11:43 PM
Really great and relatable talk from old mate Jim here about why creative's should make they want to make.
Don't let an algorithm decide what you do, don't let an audience decide what you do, don't let "industry advice" decide what you do, just make shit you think is cool and fun to make
Don't let an algorithm decide what you do, don't let an audience decide what you do, don't let "industry advice" decide what you do, just make shit you think is cool and fun to make
All time fav of mine youtu.be/Z-wNgrXgCFs?...
November 7, 2025 at 8:03 AM
All time fav of mine youtu.be/Z-wNgrXgCFs?...
... i mean, he stopped trying a long time ago but this is somehow a new low.
Newsletter: Faced with blowback over his pardon of Binance founder Changpeng Zhao, President Trump has offered a curious defense: he doesn’t even know the guy.
Trump says he has “no idea” who he just pardoned
President Trump reacts to condemnations of his recent pardon of Binance founder Changpeng Zhao by claiming he doesn’t know who he is.
www.citationneeded.news
November 4, 2025 at 11:49 AM
... i mean, he stopped trying a long time ago but this is somehow a new low.