Claas Voelcker
banner
cvoelcker.bsky.social
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 🇩🇪
Party is over, food* was not delivered

*review scores
November 11, 2025 at 10:16 PM
You know, it's chain rule, but then you also save the inputs, and the matrix convention is the other way around so you have to add a transpose here, and ah no, this framework doesn't do that, I don't know there is a transpose somewhere, oh and also the gradient of softmax and cross entropy kinda ...
a man in a blue shirt and tie is pointing at a board with papers on it .
ALT: a man in a blue shirt and tie is pointing at a board with papers on it .
media.tenor.com
November 7, 2025 at 5:24 PM
It is quite pretty and the CS and ECE facilities are very very nice, they even hold up well compared to Vector
November 6, 2025 at 12:11 AM
here = I moved to the US today :D
November 3, 2025 at 3:17 AM
Not perfect, but would the two of you be up for organizing an RLC workshop on this?
October 27, 2025 at 7:44 PM
I think @rl-conference.bsky.social tried that last time, but I haven't checked out the results!
October 27, 2025 at 6:00 PM
Maybe we just need to rename it from Intro to Summary? Also, I think the page limit is relevant: make the summary explicitly one page and rate it separately. It gets less tempting to squeeze it to get more formal stuff in.
Worst offenders here are theory papers ime, zero appreciation for intution.
October 27, 2025 at 6:00 PM
I think the paper is a relic from a past when we had far more robust and slow systems for moving from "fundamental research -> real world" not multi-million startups based on arxiv preprints 😁
October 27, 2025 at 5:27 PM
Oh yeah, I totally agree. I think we should give one extra page for an executive summary at the minimum. But there is a huge crowd that thinks "science == maximum precision to the point of obfuscation" (and an equally large amount that thinks "science == competitive engineering").
October 27, 2025 at 5:27 PM
Also, there is zero reason papers shouldn't have the tweet level summary added at the beginning, we should probably incentivize that. One extra page for executive summary?
October 27, 2025 at 5:23 PM
The boring answer is: Because a thread only makes you understand the surface level info and none of the nuances, but most of us are fine with it. The paper is for people who want to know all.
October 27, 2025 at 5:22 PM
There is a deeper issue with the original hindsight paper: It assumes that any trajectory you observe that moves from state a to state b took an optimal path. But that is obviously not true. Many offline GCRL algos actually inherit this weirdness a bit. Don't know if this is discussed somewhere.
October 27, 2025 at 5:19 PM
REPPO with hindsight experience when?
October 27, 2025 at 2:46 PM
Yes, I know, I could chose to just ignore the bullshit, but there is soooooooo much of it. How do people live like this, constantly protecting their vulnerable mental state? I guess we are where we are because they don't?
October 27, 2025 at 2:46 PM