Claas Voelcker

Since there is a decent amount of "Do we know how LLMs work?" discourse flowing around, I would be really interested to hear what people would accept as "understanding how LLMs work". What kind of knowledge are we speaking about here?

Sometimes D&D adventures are hilariously badly written:

Players hear a voice telling them to bring NPC X. They have to bring NPC X to open the door. If NPC X died earlier in the adventure, X will wait for them behind the door.

So if X dies it is behind the door that can ONLY be opened by X???

A lot of folks in the AI space are like people who insist it’s safe and smart to get drunk in the driver’s seat of their Tesla on a country road and a lot of their opponents are like people saying there’s no such thing as an automatic transmission

I don't understand how the Canadian AI/ML ecosystem wants to attract and retain talents when they are offering less then half the salary on representative roles...

Put on arxiv before acceptance, yes or no?

Really great rant!

Enjoying the For You feed? Give it a like ♡ to help more people discover it: bsky.app/profile/did:...

The more people use it -> the more feedback we get -> the better we can make it for you.

A totally unrelated question: does anybody know how to make long equations work on mobile with math jax and Jekyll 😅🙈

Thank you ❤️

a close up of a sad cat with the words pleeeaasse written below it ALT: a close up of a sad cat with the words pleeeaasse written below it

cvoelcker.de/blog/2025/re...

I finally gave in and made a nice blog post about my most recent paper. This was a surprising amount of work, so please be nice and go read it!

cvoelcker.de/blog/2025/re...

Here ya go!

Congrats! May your GPU and space access live long and prosper!

RL rant time after reading another LLM paper about whether "RL sharpens the distribution or discovers new knowledge.": RL is not magic. If your exploration policy takes an action with 0 probability, it can't explore that action! It trivially just affects the distribution of supported actions.

Stanford and Berkeley are functionally equivalent places and I refuse to treat them as separate entities.

But those are WOOOOOOORK :D

Happy guy sad guy meme with sad text: USE PPO AND TUNE HYPERPARAMETER FOR WEEKS and happy text: USE REPPO AND GET A POLICY

I have been told I need to get more modern in my paper promotion! github.com/cvoelcker/reppo / arxiv.org/abs/2507.11019 @marcelhussing.bsky.social

My grad school salary advise: find a loving partner before grad school, get them a work visa and a well-paid job in your school location, have them wildly out-earn you because they are brilliant and tada... sugar-partner! Tested for your convenience!

I pitched this ~8 years ago! It was going to provide the services "AI consultancy" and "blockchain consultancy". Our highly trained consultants would say "no" when you ask about AI or blockchain, and then just give you a normal database and some working SQL.

I know I'm like... 3 years late to the party, but wow, custom preamble prompts make chatgpt so much more useful.

Sometimes I wonder how certain researchers get famous when _none_ of their results are replicable, even with their own published code?!

You may chose if I mean you in this rant...

We've hired some *fantastic* researchers but our startup is still looking for 2-3 more people with skills in ML/RL/LLMs. If you'd like to work on some transformative applied problems, hit me up. We'll be launching publicly soon too...

Happy Rosh Hashanah. May this year be better than the last one!

Reminds me of my favourite high-dim stat tidbit: the more dimensions you measure, the less likely it is to be close to the center (or average across all of them). High-dim Gaussian is a ball.

The only thing I want in life is a math textbook that tells me _why_ we need a thing and what we need it to look like, before it rigorously defines it. Why is it always the other way around???

@tmlrorg.bsky.social I keep getting assigned to review papers where I know close to nothing about the subject area. Is there a way to change the paper matching algorithm (e.g. exclude some of my works) or refuse review?

Proud to announce that Meta-World+ was accepted to NeurIPs, Datasets and Benchmarks! Meta-World is a common benchmark for multi-task and meta-RL research! However, it was very difficult to do effective science with Meta-World as different versions produce different results.

Math and ML? Just looking to expand my reading list so I can feel guilty about not reading more books even more 😂

Huge shoutout to 👑 @axelbrunnbauer.bsky.social 👑 who took the lead on developing our Atari integration while I was off getting married and chilling for the summer.

Big if true 🤫: #REPPO works on Atari as well 😱 👾 🚀

Some tuning is still needed, but we are seeing results roughly on par with #PQN.

If you want to test out #REPPO (atari is not integrated due to issues with envpool and jax version), check out github.com/cvoelcker/re...

#reinforcementlearning

What’s your favorite textbooks?