Max Weichart

@AlexWeichart @yacineMTB @pangramlabs slop?

@yacineMTB i am judging your perfect punctuation and spelling on every post you make btw

If you post even slightly LLM generated trash I will block you

Suggestion for @perplexity_ai: make top links show up instantly (< 200ms, like Google), while LLM output is loading. That way, one could use it as a full Google replacement, currently loading time for top links is just too long to be competitive (> 1000ms)

Interested in RL? I'm planning to assemble a new online meetup, focused on reinforcement learning paper discussions. You can sign up, and as soon as enough people are interested, you'll get an invitation. More information and registration: max-we.github.io/R1/

@paddlepaddle_ No problem. Could you please create a GitHub issue with a problem description (if possible reproducible) and I will help you out! github.com/Max-We/Tetris-…

@MaxWeichart Hi Max, i am using your tetris environment for rl study. The problem of the grouped action space is that it misses some actions. In this attached example, it missed three legitimate cases. Vertically put in the left most (twice for two rotations), and in the second left most once

🎉Today marks the first release of Tetris Gymnasium! If you're an RL researcher or just somebody who wants to get into it, give it a look! You can start with ~5 lines of code and maybe create the next big RL algorithm! pip install tetris-gymnasium github.com/Max-We/Tetris-…

You can find me on bsky.app/profile/mweich…

@YouJiacheng Good catch! I updated the diagram here. The way the paper phrased exploring several approaches made it unclear if all or only some of the tricks were used for the cold start data. But probably the R1-Zero outputs were indeed used.

@iScienceLuvr I lost count how many papers are called "xyz Is All You Need" by now

Tensor Product Attention Is All You Need Proposes Tensor Product Attention (TPA), a mechanism that factorizes Q, K, and V activations using contextual tensor decompositions to achieve 10x or more reduction in inference-time KV cache size relative to standard attention mechanism with improved performance compared to previous methods such as MHA, MQA, GQA, and MLA.

This is a really thought-provoking view on LLMs that I'd like more people to talk about!

@Spideraxe30 I feel like the problem with this rune is that it has to be balanced around the 1% best ults in the game and won't be viable for anyone else

Axiom Arcanist buff (this was a change on Friday I missed): - Bonus ult effect increased from 12% to 14% - AoE ult damage increased from 8% to 9%

@daniela_muntyan Thanks for sharing!

@MaxWeichart I think it's this clock; it's a very simple one: amzn.eu/d/4eSrmnl

Check out Craft 3's new styling options! You can now customize the color of text and paper, save your custom styles, and set them as the default for all new documents. 💫

Tetris Gymnasium x Jax JIT is coming along well...

Sorting an array with a neural network isn't as trivial as one might expect! maximilian-weichart.de/posts/set-to-s…

@ylecun @karpathy Boltzmann left the chat

@karpathy Hot take: "because entropy is in the eye of the beholder. One observer's entropy is another observer's information."

The Last Question by Asimov is relevant today! users.ece.cmu.edu/~gamvrosi/thel… """ "How can the net amount of entropy of the universe be massively decreased?" Multivac fell dead and silent. The slow flashing of lights ceased, the distant sounds of clicking relays ended. Then, just as the frightened technicians felt they could hold their breath no longer, there was a sudden springing to life of the teletype attached to that portion of Multivac. Five words were printed: INSUFFICIENT DATA FOR MEANINGFUL ANSWER. "No bet," whispered Lupov. They left hurriedly. """ o1-mini, Sep 2024: chatgpt.com/share/66e38baf…

@RaphaelWimmer The real CoT is hidden from the user, and we only see a model-generated summary. Ofc there are no details on how this process works exactly, but I can imagine there are multiple moving parts that can lead to weird output like this... openai.com/index/learning…

Been playing around with #o1preview for some time now and ... does it just simulate thinking deeply? The chain of thought does not necessarily fit the final output. Pretty obvious when asked to tell a joke (same classic joke returned to English and German questions, by the way).

Students at Stanford have built alphaXiv, an open discussion forum for arXiv papers. Think of it as 𝕏 for research.

Got my first PR + bounty for @__tinygrad__ accepted today 🎉 I recommend giving it a try, maybe you can make a contribution too github.com/tinygrad/tinyg…

@MKBHD Interesting take to just close the eyes on what's happening with GenAI instead of doing something about it

Bookmark this. Such a fascinating announcement Procreate CEO gets on camera to make it clear he HATES generative AI, and they will not be integrating it ever into any of their products. Artists and users on social media celebrate. TAKE NOTES, ADOBE (buuuuut technically this is committing to never offering any of features in their products, no matter how good/useful they may get in the future. An announcement to not add features. I wonder if they ever bend this rule someday)

NOW you have my attention