Claas Voelcker
@cvoelcker.bsky.social
For professional, see https://cvoelcker.de
If I seem very angry, check if I have been watered in the last 24 hours.
Now 🇺🇸 flavoured, previously available in 🇨🇦 and 🇩🇪
If I seem very angry, check if I have been watered in the last 24 hours.
Now 🇺🇸 flavoured, previously available in 🇨🇦 and 🇩🇪
Pinned
Claas Voelcker
@cvoelcker.bsky.social
· Nov 11
I'm Claas, a hopefully-soon-finished-PhD researcher at @uoft.bsky.social . I work on reinforcement learning, especially on deep model-based methods, and am dabbling in diffusion policies and large-scale imitation learning.
I'm way too political and loud in general, so please be warned.
I'm way too political and loud in general, so please be warned.
And most importantly, I'm at least 200% nicer than I appear in any given moment, so let me know if you really just want to geek out about random shit!
Learnings (some of them obvious):
- Put other important code folders in your VS Code project. That way the model won't try to find and read the code via the command line.
- Claude thinks GCRL = HER... Understandeable, but really not what I am going for. Be precise about what algorithms you need.
- Put other important code folders in your VS Code project. That way the model won't try to find and read the code via the command line.
- Claude thinks GCRL = HER... Understandeable, but really not what I am going for. Be precise about what algorithms you need.
Watch me lose my sanity in real-time as I try to vibe code a research project from scratch. Any tips welcome.
I'm using the free academic Github Copilot Pro version. I found I prefer anthropic models, so I'm testing Claude. My first command was to read a dataset and sim library and to
I'm using the free academic Github Copilot Pro version. I found I prefer anthropic models, so I'm testing Claude. My first command was to read a dataset and sim library and to
December 1, 2025 at 7:59 PM
Learnings (some of them obvious):
- Put other important code folders in your VS Code project. That way the model won't try to find and read the code via the command line.
- Claude thinks GCRL = HER... Understandeable, but really not what I am going for. Be precise about what algorithms you need.
- Put other important code folders in your VS Code project. That way the model won't try to find and read the code via the command line.
- Claude thinks GCRL = HER... Understandeable, but really not what I am going for. Be precise about what algorithms you need.
Watch me lose my sanity in real-time as I try to vibe code a research project from scratch. Any tips welcome.
I'm using the free academic Github Copilot Pro version. I found I prefer anthropic models, so I'm testing Claude. My first command was to read a dataset and sim library and to
I'm using the free academic Github Copilot Pro version. I found I prefer anthropic models, so I'm testing Claude. My first command was to read a dataset and sim library and to
December 1, 2025 at 7:54 PM
Watch me lose my sanity in real-time as I try to vibe code a research project from scratch. Any tips welcome.
I'm using the free academic Github Copilot Pro version. I found I prefer anthropic models, so I'm testing Claude. My first command was to read a dataset and sim library and to
I'm using the free academic Github Copilot Pro version. I found I prefer anthropic models, so I'm testing Claude. My first command was to read a dataset and sim library and to
OK, so #reviewing has issues and now people want to deanonymize the reviewers. I'm not a fan, and it smells of a punitive impetus which is super common "Evil reviewers are bad and we need to know who to shun".
Reviewers are tasked with enforcing community standards.
Reviewers are tasked with enforcing community standards.
December 1, 2025 at 3:59 PM
OK, so #reviewing has issues and now people want to deanonymize the reviewers. I'm not a fan, and it smells of a punitive impetus which is super common "Evil reviewers are bad and we need to know who to shun".
Reviewers are tasked with enforcing community standards.
Reviewers are tasked with enforcing community standards.
Reposted by Claas Voelcker
My lab is looking for new students who are very passionate about foundational models and planning/RL/robotics. Apply via Mila. I will also be at #NeurIPS to discuss research ideas and opportunities. See notes below for application advice.
November 19, 2025 at 3:10 PM
My lab is looking for new students who are very passionate about foundational models and planning/RL/robotics. Apply via Mila. I will also be at #NeurIPS to discuss research ideas and opportunities. See notes below for application advice.
Also, could we delay the collapse of the whole academic system until after I get a job (and possibly tenure)? It's been ... a year...
November 28, 2025 at 7:32 PM
Also, could we delay the collapse of the whole academic system until after I get a job (and possibly tenure)? It's been ... a year...
OK, I ask again, with not inconsiderable rage, who decides the PCs for conferences, and why are there no votes for the board members! @iclr-conf.bsky.social
November 28, 2025 at 7:27 PM
OK, I ask again, with not inconsiderable rage, who decides the PCs for conferences, and why are there no votes for the board members! @iclr-conf.bsky.social
One thing that has bothered me for a long time is how absurdly undemocratic a lot of processes in academia are. We should not be accepting conferences organized by teams of people who are chosen on completely intransparent criteria! This goes for pretty much any conference out there.
November 28, 2025 at 5:34 PM
One thing that has bothered me for a long time is how absurdly undemocratic a lot of processes in academia are. We should not be accepting conferences organized by teams of people who are chosen on completely intransparent criteria! This goes for pretty much any conference out there.
After @neuripsconf.bsky.social , @iclr-conf.bsky.social is now truly optimizing for "What can we do to make people hate ACs A LOT". There is no way adding a fresh AC and have them only guess at reviewers views will not end with a huge amount of absolutely arbitrary decisions.
November 28, 2025 at 5:09 PM
After @neuripsconf.bsky.social , @iclr-conf.bsky.social is now truly optimizing for "What can we do to make people hate ACs A LOT". There is no way adding a fresh AC and have them only guess at reviewers views will not end with a huge amount of absolutely arbitrary decisions.
This is why good Kasspatzen (vastly superior German Mac and cheese) needs to come with salad and crunchy fried onions! Very nice to have a word for it (even though I have no idea how to pronounce it)
Vietnamese food has a concept I don't often see in western cuisine - "ngán."
It's translated as being bored but applies when your palate is dulled by the sameness in texture, taste.
We want our food to hit different notes. A bánh mì is simultaneously soft and crunchy, spicy and cool, fatty and sour.
It's translated as being bored but applies when your palate is dulled by the sameness in texture, taste.
We want our food to hit different notes. A bánh mì is simultaneously soft and crunchy, spicy and cool, fatty and sour.
November 26, 2025 at 8:20 PM
This is why good Kasspatzen (vastly superior German Mac and cheese) needs to come with salad and crunchy fried onions! Very nice to have a word for it (even though I have no idea how to pronounce it)
Every damn time!
November 26, 2025 at 7:57 PM
Every damn time!
Give me fruity cocktails with sparkly umbrellas in them or give me death! To be fair, being gay is a bit of a cheatcode to toxic masculinity, you kinda lost the game anyways, gotta make up new rules...
So many men go through life like that. You can play poker because it's right-coded, you can't enjoy fruity alcoholic drinks because they're women-coded, you can't lick an ice cream cone because it's gay-coded. All day, every day, choosing and judging things based on coding for narrow norms.
November 26, 2025 at 4:43 PM
Give me fruity cocktails with sparkly umbrellas in them or give me death! To be fair, being gay is a bit of a cheatcode to toxic masculinity, you kinda lost the game anyways, gotta make up new rules...
To be fair, if one of my students stood in front of my homework assignment in “authentic wonder before the beauty of [postdoc]’s creation”, I’d recommend a thorough therapy session.
Pope Leo XIV told students not to use artificial intelligence for homework, saying that AI ‘won’t stand in authentic wonder before the beauty of God’s creation.’
Even God Is Worried About ChatGPT
Pope Leo XIV told students not to use artificial intelligence for homework, saying that AI ‘won’t stand in authentic wonder before the beauty of God’s creation.’
www.vulture.com
November 26, 2025 at 4:11 PM
To be fair, if one of my students stood in front of my homework assignment in “authentic wonder before the beauty of [postdoc]’s creation”, I’d recommend a thorough therapy session.
Maybe I will finally be hireable!!! Make value learning cool again :D
Also, value functions will soon be cool again! 👍 thank you pistar_0.6 .. and Ilya thinks they will be used more in the future 🙏
On emotions and value functions: youtu.be/aR20FWCCjAs?...
On emotions and value functions: youtu.be/aR20FWCCjAs?...
Ilya Sutskever – We're moving from the age of scaling to the age of research
YouTube video by Dwarkesh Patel
youtu.be
November 26, 2025 at 3:56 PM
Maybe I will finally be hireable!!! Make value learning cool again :D
Generalization is a lie made up by big network
November 25, 2025 at 10:03 PM
Generalization is a lie made up by big network
Reposted by Claas Voelcker
🎉 New work: “Learning Massively Multitask World Models for Continuous Control”
We introduce MMBench: a 200-task RL benchmark, and Newt: a language-conditioned multitask world model trained with large-scale online RL.
www.nicklashansen.com/NewtWM/
Code, checkpoints, dataset etc. are open-source!
We introduce MMBench: a 200-task RL benchmark, and Newt: a language-conditioned multitask world model trained with large-scale online RL.
www.nicklashansen.com/NewtWM/
Code, checkpoints, dataset etc. are open-source!
November 25, 2025 at 7:47 PM
🎉 New work: “Learning Massively Multitask World Models for Continuous Control”
We introduce MMBench: a 200-task RL benchmark, and Newt: a language-conditioned multitask world model trained with large-scale online RL.
www.nicklashansen.com/NewtWM/
Code, checkpoints, dataset etc. are open-source!
We introduce MMBench: a 200-task RL benchmark, and Newt: a language-conditioned multitask world model trained with large-scale online RL.
www.nicklashansen.com/NewtWM/
Code, checkpoints, dataset etc. are open-source!
"For you" finally knows what I want! More #dndmaps
@mightofmerchants.bsky.social looks like an incredible tool
@mightofmerchants.bsky.social looks like an incredible tool
November 25, 2025 at 7:45 PM
"For you" finally knows what I want! More #dndmaps
@mightofmerchants.bsky.social looks like an incredible tool
@mightofmerchants.bsky.social looks like an incredible tool
AI alignment research causes unaligned AI by writing too many lesswrong posts about how it is unaligned 😅 the solution is clear (and socially/morally good): ban lesswrong posting!
This is why the actual AI doom scenario is if the global memeplex ever spews enough anti-AI bs into the noosphere that the concept of AI itself gets pushed into the 'evil' category. Hyperstitioning supervillains because a bunch of people couldn't keep their emotions regulated
the consensus seems to be that models are not inventing evil from first principles, but rather they are formed by correlating vast swaths of human text and recognize that certain concepts are correlated w/one another (that's how LLMs work in the first place), so they see the evil ones as related
November 24, 2025 at 7:53 PM
AI alignment research causes unaligned AI by writing too many lesswrong posts about how it is unaligned 😅 the solution is clear (and socially/morally good): ban lesswrong posting!
How do academic collaborators navigate the increasingly horrible embargo processes at companies like Google? Google co-authors keep telling me that projects might get placed under embargo. That feels like a huge violation of our (academics) IP rights when we collaborate on projects.
November 24, 2025 at 7:38 PM
How do academic collaborators navigate the increasingly horrible embargo processes at companies like Google? Google co-authors keep telling me that projects might get placed under embargo. That feels like a huge violation of our (academics) IP rights when we collaborate on projects.
What is it with roboticists and their undying dedication to python 3.8??? It's 2025!!!
November 24, 2025 at 3:46 PM
What is it with roboticists and their undying dedication to python 3.8??? It's 2025!!!
To be fair, the church was a sponsor of cutting edge arts for like 1.5 millennia, so this isn’t actually that weird if you think about it. Gothic cathedrals were the raves of the high medieval ages (don’t quote me on this, not a historian)
He said "Amen" straight into the drop
Lmao Pope Leo threw a rave for an archbishop's 75th birthday this is kind of incredible
Lmao Pope Leo threw a rave for an archbishop's 75th birthday this is kind of incredible
November 23, 2025 at 6:17 PM
To be fair, the church was a sponsor of cutting edge arts for like 1.5 millennia, so this isn’t actually that weird if you think about it. Gothic cathedrals were the raves of the high medieval ages (don’t quote me on this, not a historian)
Reposted by Claas Voelcker
I'm looking for graduate students to join my group in fall 2026. We work at the intersection of Computer Vision, Deep Learning, and Robotics.
The goal of our work is to create and understand organic 🍀 simulation systems, i.e., controllable data-curation engines built from real-world data. 1/n
The goal of our work is to create and understand organic 🍀 simulation systems, i.e., controllable data-curation engines built from real-world data. 1/n
November 22, 2025 at 8:50 PM
I'm looking for graduate students to join my group in fall 2026. We work at the intersection of Computer Vision, Deep Learning, and Robotics.
The goal of our work is to create and understand organic 🍀 simulation systems, i.e., controllable data-curation engines built from real-world data. 1/n
The goal of our work is to create and understand organic 🍀 simulation systems, i.e., controllable data-curation engines built from real-world data. 1/n
Reposted by Claas Voelcker
And here is the presentation I gave on networking, self-promo, and how to make the most out of a conference. Hope this helps for everyone at NeurIPS!
www.youtube.com/watch?v=B9hG...
www.youtube.com/watch?v=B9hG...
Conferencemaxxing: How to grow your profile and network as a scientist
YouTube video by Michael Saxon (NLP & Generative AI research)
www.youtube.com
November 19, 2025 at 11:59 PM
And here is the presentation I gave on networking, self-promo, and how to make the most out of a conference. Hope this helps for everyone at NeurIPS!
www.youtube.com/watch?v=B9hG...
www.youtube.com/watch?v=B9hG...
Be the “kernel” for someone
Find someone who loves you more than ML people love saying manifold
November 20, 2025 at 3:25 PM
Be the “kernel” for someone
OK, but ... neither of these names mean anything???
Some news: Open Philanthropy is now Coefficient Giving! Our mission is unchanged but the new name reflects our growing work with other donors to multiply the impact of their giving.
🧵 on our work to make philanthropy a more efficient "market" and plans going forward:
🧵 on our work to make philanthropy a more efficient "market" and plans going forward:
November 18, 2025 at 4:24 PM
OK, but ... neither of these names mean anything???
I might sometimes be a bit of a techno-pessimist, but nothing made me love Waymos more than sharing the street with them as a bike rider. They drive slowly, carefully, and the likelihood they go into road rage mode is ca 0. Universal self driving can’t come fast enough!
November 18, 2025 at 2:28 PM
I might sometimes be a bit of a techno-pessimist, but nothing made me love Waymos more than sharing the street with them as a bike rider. They drive slowly, carefully, and the likelihood they go into road rage mode is ca 0. Universal self driving can’t come fast enough!