๐๏ธ Welcome to QuackChat: The DuckTypers' Daily AI Update
Hello, Ducktypers! ๐ฆ Jens here with todayโs AI-packed episode! Weโve got groundbreaking news from Intel, some interesting research on activations, and even a little AI-themed merch madness to discuss. Letโs jump right into it!
๐ Intelโs FP8 Training Revolution with Llama2-7B

Kicking things off, Intel has made an exciting leap forward with the Llama2-7B model. Theyโve managed to train this AI on a dataset of 2 trillion tokens using FP8 precisionโa whopping 20x increase over previous models! ๐คฏ
Whatโs really fascinating is the modification they made to the SwiGLU activation function. They introduced something called Smooth-SwiGLU to address instabilities that typically arise when using FP8. This not only helps stabilize activations but also allows Intel to use FP8 quantization for the Adam optimizer, making everything more efficient. ๐ก
Ducktyper Question: Do you think FP8 precision is the future of AI training? Will it become the new standard, or are we still a ways off? Let me know your thoughts in the comments! ๐ง
๐ Check out the full research paper here.
๐ฎ Random Projection: The Key to Smoothing Activations?
In other news, researchers are exploring random projection techniques to help smooth out activations in AI models and reduce dimensionality. The theory is that by making activations sparse enough, outliers can be controlled, leading to better model stability. Theyโve already seen success using Hadamard transforms in models like Quip# and Quarot. ๐ฏ
This could be a huge step forward in making large language models more stable and faster to train. But whatโs next? Could dimensionality reduction truly be the answer to smoother activations? ๐ค
๐ AI Merch Madness: $90 Hoodies? Worth It?
Switching gears, letโs talk about Nous Researchโs latest merch drop! Their Decentralize T is getting a lot of love, but thereโs been some serious sticker shock about their $90 hoodie. One member in the CUDA Mode Discord community said theyโre still considering it despite the high price tag. ๐ธ
Ducktyper Poll: Would you spend $90 on AI-themed merch? ๐งฅ Drop your answers in the comments below!
Nous Research Merch Drop Praise
๐ง Hackathon Time: Ready to Join the Fun?

Finally, if youโre into AI hackathons, the CUDA-Mode IRL event is happening soon, and invites are already rolling out. Plus, attendees will get the chance to have their PMPP book signed by Professor Wen-mei Hwu! โ๏ธ
Have you ever participated in a hackathon? Let me know if youโre planning on attending any upcoming events!
๐ Wrap-Up and Call to Action
Thatโs all for today, Ducktypers! ๐ฆ If you enjoyed todayโs episode, donโt forget to subscribe and like. And let me know your thoughts on Intelโs FP8 training, random projection techniques, and whether youโre shelling out for that $90 hoodie! Iโll see you in the comments!
Until next time, keep quacking, keep learning! ๐ฆ๐ป