Sebastian Szyller
@sebszyller.com
Assistant Professor @csaalto.bsky.social
Trustworthy & Adversarial Machine Learning
sebszyller.com & taclab.aalto.fi
Trustworthy & Adversarial Machine Learning
sebszyller.com & taclab.aalto.fi
Anthropic acquiring Bun has not been on my bingo card for this year: bun.com/blog/bun-joi...
Maybe they will splurge some money on developing Zig 🤷♂️🤞
Maybe they will splurge some money on developing Zig 🤷♂️🤞
Bun is joining Anthropic
Bun has been acquired by Anthropic. Anthropic is betting on Bun as the infrastructure powering Claude Code, Claude Agent SDK, and future AI coding products & tools.
bun.com
December 2, 2025 at 6:21 PM
Anthropic acquiring Bun has not been on my bingo card for this year: bun.com/blog/bun-joi...
Maybe they will splurge some money on developing Zig 🤷♂️🤞
Maybe they will splurge some money on developing Zig 🤷♂️🤞
Reposted by Sebastian Szyller
For context: the policy is now official. bsky.app/profile/ange...
blog.arxiv.org/2025/10/31/a...
FYI the blog post for the updated policy is out. Our llm future is dire:/
FYI the blog post for the updated policy is out. Our llm future is dire:/
October 31, 2025 at 8:18 PM
For context: the policy is now official. bsky.app/profile/ange...
Imo, the most disappointing part of the Neurips drama is not even the rejections but the blatant disrespect of people’s time.
Volunteers review the submitted manuscripts, and authors have a short window to work on the rebuttal. It all goes down the drain when the ACs are asked to overrule.
Volunteers review the submitted manuscripts, and authors have a short window to work on the rebuttal. It all goes down the drain when the ACs are asked to overrule.
September 20, 2025 at 7:50 PM
Imo, the most disappointing part of the Neurips drama is not even the rejections but the blatant disrespect of people’s time.
Volunteers review the submitted manuscripts, and authors have a short window to work on the rebuttal. It all goes down the drain when the ACs are asked to overrule.
Volunteers review the submitted manuscripts, and authors have a short window to work on the rebuttal. It all goes down the drain when the ACs are asked to overrule.
Looks like Germany is not going to support chat control. This means that right now, there're enough countries to block it.
Tbf, this proposal, in its various forms, comes back like a boomerang. I'm fully expecting another attempt to get it through in a couple of months.
fightchatcontrol.eu
Tbf, this proposal, in its various forms, comes back like a boomerang. I'm fully expecting another attempt to get it through in a couple of months.
fightchatcontrol.eu
Fight Chat Control - Protect Digital Privacy in the EU
Learn about the EU Chat Control proposal and contact your representatives to protect digital privacy and encryption.
fightchatcontrol.eu
September 11, 2025 at 10:38 AM
Looks like Germany is not going to support chat control. This means that right now, there're enough countries to block it.
Tbf, this proposal, in its various forms, comes back like a boomerang. I'm fully expecting another attempt to get it through in a couple of months.
fightchatcontrol.eu
Tbf, this proposal, in its various forms, comes back like a boomerang. I'm fully expecting another attempt to get it through in a couple of months.
fightchatcontrol.eu
In a positive spin on the privacy disaster that chat control is, both Finland and Poland are planning to oppose it.
fightchatcontrol.eu
fightchatcontrol.eu
Fight Chat Control - Protect Digital Privacy in the EU
Learn about the EU Chat Control proposal and contact your representatives to protect digital privacy and encryption.
fightchatcontrol.eu
August 30, 2025 at 11:44 AM
In a positive spin on the privacy disaster that chat control is, both Finland and Poland are planning to oppose it.
fightchatcontrol.eu
fightchatcontrol.eu
There's a new HIIT call for postdocs and research fellows (deadline Aug 31s): www.hiit.fi/hiit-postdoc...
If you wanted to do research on trustworthy machine learning in the Helsinki capital region but couldn't find a group/professor that specialises in it, I'll be there in September 🫡
If you wanted to do research on trustworthy machine learning in the Helsinki capital region but couldn't find a group/professor that specialises in it, I'll be there in September 🫡
LinkedIn
This link will take you to a page that’s not on LinkedIn
lnkd.in
July 12, 2025 at 3:22 PM
There's a new HIIT call for postdocs and research fellows (deadline Aug 31s): www.hiit.fi/hiit-postdoc...
If you wanted to do research on trustworthy machine learning in the Helsinki capital region but couldn't find a group/professor that specialises in it, I'll be there in September 🫡
If you wanted to do research on trustworthy machine learning in the Helsinki capital region but couldn't find a group/professor that specialises in it, I'll be there in September 🫡
For the EU folks:
quarterly reminder to contact your MEP and voice your concerns regarding regulatory proposals that will undermine private communications.
If someone has a more informative link, please post it in the comments
edri.org/our-work/pro...
quarterly reminder to contact your MEP and voice your concerns regarding regulatory proposals that will undermine private communications.
If someone has a more informative link, please post it in the comments
edri.org/our-work/pro...
‘ProtectEU’ security strategy - European Digital Rights (EDRi)
The European Commission presented an internal security strategy that would undermine digital rights and even increase security threats.
edri.org
June 3, 2025 at 8:08 AM
For the EU folks:
quarterly reminder to contact your MEP and voice your concerns regarding regulatory proposals that will undermine private communications.
If someone has a more informative link, please post it in the comments
edri.org/our-work/pro...
quarterly reminder to contact your MEP and voice your concerns regarding regulatory proposals that will undermine private communications.
If someone has a more informative link, please post it in the comments
edri.org/our-work/pro...
This is more important than most realise. We’ve been pursuing this space for a couple of years now: arxiv.org/abs/2207.01991 arxiv.org/abs/2312.04542
April 11, 2025 at 7:30 AM
This is more important than most realise. We’ve been pursuing this space for a couple of years now: arxiv.org/abs/2207.01991 arxiv.org/abs/2312.04542
Claude is having one of those days. I asked it to translate my shell script from bash to fish, and it went bonkers with a made up fish language 😅
claude.ai/share/3a2488...
claude.ai/share/3a2488...
Claude
Talk with Claude, an AI assistant from Anthropic
claude.ai
April 2, 2025 at 9:40 AM
Claude is having one of those days. I asked it to translate my shell script from bash to fish, and it went bonkers with a made up fish language 😅
claude.ai/share/3a2488...
claude.ai/share/3a2488...
Reposted by Sebastian Szyller
Would you present your next NeurIPS paper in Europe instead of traveling to San Diego (US) if this was an option? Søren Hauberg (DTU) and I would love to hear the answer through this poll: (1/6)
NeurIPS participation in Europe
We seek to understand if there is interest in being able to attend NeurIPS in Europe, i.e. without travelling to San Diego, US. In the following, assume that it is possible to present accepted papers ...
docs.google.com
March 30, 2025 at 6:04 PM
Would you present your next NeurIPS paper in Europe instead of traveling to San Diego (US) if this was an option? Søren Hauberg (DTU) and I would love to hear the answer through this poll: (1/6)
People in my broader research circles have been saying that we should finally stop wasting resources on fair (lEfTiSt) machine learning research. To me, it’s a Simone Biles level of mental gymnastics, so let’s unpack 🤓
sebszyller.com/blog/2025/fa...
sebszyller.com/blog/2025/fa...
sebszyller.com
sebszyller.com
March 11, 2025 at 3:13 PM
People in my broader research circles have been saying that we should finally stop wasting resources on fair (lEfTiSt) machine learning research. To me, it’s a Simone Biles level of mental gymnastics, so let’s unpack 🤓
sebszyller.com/blog/2025/fa...
sebszyller.com/blog/2025/fa...
Reposted by Sebastian Szyller
My new paper "Deep Learning is Not So Mysterious or Different": arxiv.org/abs/2503.02113. Generalization behaviours in deep learning can be intuitively understood through a notion of soft inductive biases, and formally characterized with countable hypothesis bounds! 1/12
March 5, 2025 at 3:38 PM
My new paper "Deep Learning is Not So Mysterious or Different": arxiv.org/abs/2503.02113. Generalization behaviours in deep learning can be intuitively understood through a notion of soft inductive biases, and formally characterized with countable hypothesis bounds! 1/12
Reposted by Sebastian Szyller
🏹 Job alert: @eng.cam.ac.uk is seeking a Uni Asst. Professor in #MachineLearning in the Computational and Biological Learning Lab (CBL) to lead a research programme and teach.
📍 Cambridge, England
📅 Apply by 16 March 2025
🔗 Get details at bit.ly/4khlntY
📍 Cambridge, England
📅 Apply by 16 March 2025
🔗 Get details at bit.ly/4khlntY
University Assistant Professor in Machine Learning - Job Opportunities - University of Cambridge
University Assistant Professor in Machine Learning in the Department of Engineering at the University of Cambridge.
bit.ly
March 4, 2025 at 9:16 AM
🏹 Job alert: @eng.cam.ac.uk is seeking a Uni Asst. Professor in #MachineLearning in the Computational and Biological Learning Lab (CBL) to lead a research programme and teach.
📍 Cambridge, England
📅 Apply by 16 March 2025
🔗 Get details at bit.ly/4khlntY
📍 Cambridge, England
📅 Apply by 16 March 2025
🔗 Get details at bit.ly/4khlntY
Reposted by Sebastian Szyller
Za 20 złotych można kupić sobie podróbkę mDowodu. Robią to dzieci, żeby kupować alkohol i energetyki.
Ale na lewy mDowód można też załatwić wiele innych spraw, np. zagłosować, bo praktycznie
❌ NIKT nie weryfikuje mDowodów tak, jak powinno się to robić, czyli 1/4
Ale na lewy mDowód można też załatwić wiele innych spraw, np. zagłosować, bo praktycznie
❌ NIKT nie weryfikuje mDowodów tak, jak powinno się to robić, czyli 1/4
February 4, 2025 at 4:26 PM
Za 20 złotych można kupić sobie podróbkę mDowodu. Robią to dzieci, żeby kupować alkohol i energetyki.
Ale na lewy mDowód można też załatwić wiele innych spraw, np. zagłosować, bo praktycznie
❌ NIKT nie weryfikuje mDowodów tak, jak powinno się to robić, czyli 1/4
Ale na lewy mDowód można też załatwić wiele innych spraw, np. zagłosować, bo praktycznie
❌ NIKT nie weryfikuje mDowodów tak, jak powinno się to robić, czyli 1/4
Ricing my terminal/nvim config is my man cave. So I like to check out other people's dotfiles. Though the number of times people have committed their API keys is crazy😅
February 4, 2025 at 10:16 AM
Ricing my terminal/nvim config is my man cave. So I like to check out other people's dotfiles. Though the number of times people have committed their API keys is crazy😅
Reposted by Sebastian Szyller
AI made in 🇪🇺
OpenEuroLLM, the first family of open source Large Language Models covering all EU languages, has earned the first STEP Seal for its excellence.
It brings together EU startups, research labs and supercomputing hosts to train AI on European supercomputers ↓
OpenEuroLLM, the first family of open source Large Language Models covering all EU languages, has earned the first STEP Seal for its excellence.
It brings together EU startups, research labs and supercomputing hosts to train AI on European supercomputers ↓
February 3, 2025 at 3:01 PM
AI made in 🇪🇺
OpenEuroLLM, the first family of open source Large Language Models covering all EU languages, has earned the first STEP Seal for its excellence.
It brings together EU startups, research labs and supercomputing hosts to train AI on European supercomputers ↓
OpenEuroLLM, the first family of open source Large Language Models covering all EU languages, has earned the first STEP Seal for its excellence.
It brings together EU startups, research labs and supercomputing hosts to train AI on European supercomputers ↓
DeepSeek has been accused of distilling ChatGPT. Now, I don't know or care if it's true or not but apparently OpenAI has some proof 🤔
How? OpenAI could have been watermarking their LLMs and/or their outputs; but it's tricky to get it right. So let's look into it 🤓
sebszyller.com/blog/2025/de...
How? OpenAI could have been watermarking their LLMs and/or their outputs; but it's tricky to get it right. So let's look into it 🤓
sebszyller.com/blog/2025/de...
sebszyller.com
sebszyller.com
January 31, 2025 at 7:39 PM
DeepSeek has been accused of distilling ChatGPT. Now, I don't know or care if it's true or not but apparently OpenAI has some proof 🤔
How? OpenAI could have been watermarking their LLMs and/or their outputs; but it's tricky to get it right. So let's look into it 🤓
sebszyller.com/blog/2025/de...
How? OpenAI could have been watermarking their LLMs and/or their outputs; but it's tricky to get it right. So let's look into it 🤓
sebszyller.com/blog/2025/de...
The hill I will die on
"Adversarial attacks" betray either laziness or carelessness. "Adaptive attacker" is now one who is unaware of the defense it is trying to defeat (Kerckhoff spinning in his grave), and not one who adapts the attack strategy based on the responses of the target system.
January 25, 2025 at 10:59 AM
The hill I will die on
Reposted by Sebastian Szyller
I discovered a fatal flaw in a paper by @floriantramer.bsky.social et al claiming to break our Ensemble Everything Everywhere defense. Due to a coding error they used attacks 20x above the standard 8/255. They confirmed this but the paper is already out & quoted on OpenReview. What should we do now?
December 12, 2024 at 4:29 PM
I discovered a fatal flaw in a paper by @floriantramer.bsky.social et al claiming to break our Ensemble Everything Everywhere defense. Due to a coding error they used attacks 20x above the standard 8/255. They confirmed this but the paper is already out & quoted on OpenReview. What should we do now?
Learning Rust with LLMs goes on 🔥
In this post, I'm going to recap my initial impressions. The good, the bad, and what's next to come.
sebszyller.com/blog/2024/ru...
In this post, I'm going to recap my initial impressions. The good, the bad, and what's next to come.
sebszyller.com/blog/2024/ru...
March 3, 2024 at 8:14 PM
Learning Rust with LLMs goes on 🔥
In this post, I'm going to recap my initial impressions. The good, the bad, and what's next to come.
sebszyller.com/blog/2024/ru...
In this post, I'm going to recap my initial impressions. The good, the bad, and what's next to come.
sebszyller.com/blog/2024/ru...
I've been writing about data marketplaces for a couple of years now. However, and I and others have been using the "food market" analogy to describe the relationship between the buyers and the sellers.
Honestly, it isn't a good analogy -- can we find a better one? 🤔
sebszyller.com/blog/2024/ma...
Honestly, it isn't a good analogy -- can we find a better one? 🤔
sebszyller.com/blog/2024/ma...
sebszyller.com
sebszyller.com
January 21, 2024 at 2:56 PM
I've been writing about data marketplaces for a couple of years now. However, and I and others have been using the "food market" analogy to describe the relationship between the buyers and the sellers.
Honestly, it isn't a good analogy -- can we find a better one? 🤔
sebszyller.com/blog/2024/ma...
Honestly, it isn't a good analogy -- can we find a better one? 🤔
sebszyller.com/blog/2024/ma...
This whole thing about “Sleeper Agents in LLMs” — it’s just model backdooring.
ML Security community has been showing for years that you can put backdoors into all kinds of models, and make them do all kinds of things.
Due diligence before you drum things up.
ML Security community has been showing for years that you can put backdoors into all kinds of models, and make them do all kinds of things.
Due diligence before you drum things up.
January 13, 2024 at 8:17 AM
This whole thing about “Sleeper Agents in LLMs” — it’s just model backdooring.
ML Security community has been showing for years that you can put backdoors into all kinds of models, and make them do all kinds of things.
Due diligence before you drum things up.
ML Security community has been showing for years that you can put backdoors into all kinds of models, and make them do all kinds of things.
Due diligence before you drum things up.
I tried learning Rust about five years ago but I didn’t have time to properly commit to it so I shortly dropped it. Given the hype around LLMs and tutoring, I'm going to learn Rust using ChatGPT as my main source of information: sebszyller.com/blog/2024/ru...
January 2, 2024 at 1:34 PM
I tried learning Rust about five years ago but I didn’t have time to properly commit to it so I shortly dropped it. Given the hype around LLMs and tutoring, I'm going to learn Rust using ChatGPT as my main source of information: sebszyller.com/blog/2024/ru...