molli

Clickable links:

Code: https://github.com/moritztng/hacker-league

Discord: https://discord.gg/bD2saQqdTG

X: https://x.com/moritzthuening

Hinton: Studied psychology, became cognitive scientist.

Schmidhuber: Studied cs and math, developed algorithmic theory of the computable universe.

Hinton wins Nobel price in physics.

:D

Yup, it's tiny (~100kb)

Yup, let's do it! The binary is ~100kb :D Join the Discord server that I posted here!

Thanks a lot for the feedback! I built it in ~2 weeks and it's really just a prototype at this stage. Some of the points I did on purpose to finish a prototype quickly (Eigen + GLM, one file, not syncing renderer and physics, hardcoding shaders etc.). But some things I didn't know - thanks for that :)

You're welcome. The binary is only ~100kb and no game engine etc. is used.

Because I use it only for the physics. For Vulkan I used glm. Ideally, we should only use one library.

Thanks! I'll keep it way simpler than RL - especially the graphics as you can see. But we can make the physics and multiplayer great, since we build everything from scratch.

Maybe you will be good at Hacker League:)

But I don't use textures :P

Thanks!

When a game is not popular anymore some developers just turn off their servers destroying the multiplayer. That won't happen with decentralized open source servers.

Some very interesting thoughts here. Imo open source software is mostly higher quality than proprietary one.

And apparently the industry is not so interested in Linux. Proton from Valve is interesting but crashes on my machine all the time. So I guess we have to build ourselves and have some fun in the process.

I've created a Discord server for Hacker League: https://discord.gg/BbNH27st. Everyone is welcome!

I'll also keep building in public on X: https://x.com/moritzthuening.

I love you all.

Similar: https://github.com/moritztng/fltr

Like grep but for natural language questions. Based on Mistral LLMs.

When implementations of the Transformer's self-attention layer utilize SRAM instead of DRAM, they can achieve significant speedups. The Tenstorrent Grayskull architecture provides a large SRAM, distributed across a grid of cores. This work presents a fused kernel for Grayskull, that exclusively utilizes its large SRAM by combining matrix multiplication, attention score scaling and Softmax operations. Additionally, a dedicated Softmax kernel utilizing the SRAM and a CPU implementation serving as a baseline are presented. The Softmax operation consumes most of the runtime in the computation of attention weights from queries and keys on Grayskull. The speedup of the dedicated Softmax kernel compared to the CPU implementation is up to 10x, and the Softmax implementation inside the fused kernel is approximately 1.8x faster than the dedicated Softmax kernel. The time and memory complexity of all implementations is quadratic in sequence length. Currently, the Grayskull e150 is approximately 30x cheaper for the general public than an Nvidia H100 PCIe (a state-of-the-art GPU) and offers approximately 1.5x more SRAM.

Code: https://github.com/moritztng/grayskull-attention

Can't reproduce it right now, but I'll look into it. Normally, it would show a button to sign in with google.

Did you click start and nothing happened? The service is up and running..

These are noble values! But how to build a tool for easy transcription and summarization then?

I know what you mean. But I can't run the transcription and llm locally on the user's machine. So it has to be in the cloud. And most video calls and transcription/translation services are in the cloud anyway. So yes, you have to trust that the backend is secure, and it is.

Coming soon

You watch a YouTube video or participate in a video call etc. where humans are talking. You click on the extension. You select a language. You click start. Every 30 seconds the audio gets transcribed and translated into short bullet points. The transcripts are saved. Finally, you can ask the transcript any question or transform it into a poem by Arnold Schwarzenegger - one of the great Austrian thinkers.

hi, it's cpu, makes full use of simd on multiple cores and i got 3 tok/s. it's also memory mapped, so it stays in ram once it's loaded. the weights are 8 bit quantized. it can run as a server and is the backend of https://x.com/paulpapers - a x bot, that reads all llm arxiv papers and shares the ones that are about running llms more efficiently.