My latest free upgrade just arrived 🚀

Today at Answer.AI we've got something new for you: FSDP/QDoRA. We've tested it with AI at Meta Llama3 and the results blow away anything we've seen before.

I believe that this combination is likely to create better task-specific models than anything else at any cost. 🧵

Congrats to AI at Meta on Llama 3 release!! 🎉
ai.meta.com/blog/meta-llam…
Notes:

Releasing 8B and 70B (both base and finetuned) models, strong-performing in their model class (but we'll see when the rankings come in @ lmsys.org :))
400B is still training, but already encroaching…

llama-3 is out! 🚀

- 8B and 70B
- 400B model is still training
- 8k context length

llama.meta.com/llama3/

RAG vs. Fine-Tuning debates drive me nuts. They are not fungible replacements for each other. Dan Becker and I are teaching a free lighting course on the topic

maven.com/p/787a2f/rag-a…

I read over 100 pitch decks in the last 24 hours for AI companies, most people need to 10x the level of their ambition.

A few weeks ago, I had an incredible conversation with Jeremy Howard

We chatted all things AnswerAI (his new startup), FastAI, and more. I hope you enjoy this as much as I did 🫶

youtu.be/OujUZnXf4J0?si…

There are a growing number of voices expressing disillusionment with fine-tuning.

I'm curious about the sentiment more generally. (I am withholding sharing my opinion rn).

Tweets below are from Emmanuel Ameisen anton Ethan Mollick

I've been coding on GitHub quite consistently for about 12 years, but I honestly never expected to find myself up there 😊

The incredible Radek Osmulski 🇺🇦 was kind enough to interview me 🙏

We talked about:

- My journey into programming
- How fastai changed my life
- Why I started my podcast series
- My approach to goal setting
- Reading LLM papers

Thanks again Radek!

m.youtube.com/watch?v=JrwO7f…

whichever one of you guys wrote this, thanks, you made my week😊

You need to put 2x the quantity as your goal, where x is the quantity you can achieve easily in a week.

Eg: you should aim daily 2 hours to read paper, if you can comfortably read 1 hour now.

- Sanyam Bhutani said this during an interview with Radek.

youtu.be/JrwO7f2C__U?si…

After watching the video
youtu.be/haRCu4hoI2A?si…

I have started reading this book.
I will be posting some ideas from this book.

Thanks Sanyam Bhutani and Radek Osmulski 🇺🇦

I sold all four of my 3090s today 🥹

Last week, Radek Osmulski 🇺🇦 and I got to hangout with the CEO of Answer.ai 🙏

Jeremy Howard has always been very kind to us with his time

Also, he owns a large beach in Brisbane so that makes it an awesome trip 😁

A perfect intro to open source LLMs! 🙏

The course by Amit Sangani is now my top recommendation for getting started with Large Language Models:

- Just enough theory for a whole picture

- Teaches prompting, special tokens and conversational agents

- Perfectly abstracts the…

After 3 years of learning, being mentored and building with this legend

I finally met Aman Arora in person! 🙏

New (2h13m 😅) lecture: 'Let's build the GPT Tokenizer'

Tokenizers are a completely separate stage of the LLM pipeline: they have their own training set, training algorithm (Byte Pair Encoding), and after training implement two functions: encode() from strings to tokens, and…

uv venv creates a virtual environment

✨ with a sensible default (.venv)
🤩 includes a .gitignore for it

chefkiss 👩‍🍳

uv pip compile and uv pip install are crazy fast 😮

I have to manually check the files to confirm it actually did something, it's annoyingly fast 🤯

LLM Fine-Tuning Benchmarks! 🙏

Super excited to finally publish this report comparing different GPUs and precisions:

- First, why do it and what is it?

- There are MANY GPU benchmarks but few specific to Large Language Models

- I wanted to compare their behaviour from…