4-chip Metis accelerator PCIe card coming soon:

PCIe and M.2 Metis boards available now:.- -

50+ models pre-configured in the model zoo are ready to run:

Blog post by A-Tang Fan and Doug Watt about 's Voyager SDK:

The rest of this post is on Bluesky:

I'm delighted to share that I joined the Axelera team this week to deliver the next generation AI compute platform.

The rest of this post is on BlueSky, right here:

Reinforcement Learning (RL) just landed a stellar breakthrough with reasoning language models. Yet, RL has a distinctly bad reputation. See “To RL or not to RL” ( on reddit. I'd like to revisit the basic math of RL to see why. Let's enter the dungeon!

The rest of this thread is on Bluesky:

Are you still using LoRA? 2024 has seen an explosion of new parameter-efficient fine tuning technique (PEFT), thanks to clever uses of the singular value decomposition (SVD). Let's dive into the alphabet soup: SVF, SVFT, MiLoRA, PiSSA, LoRA-XS 🤯:

The rest of this thread is on bsky:

s Transformer² paper features a cool new parameter-efficient fine-tuning (PEFT) technique that makes tuned models composable! Let’s dive in 💦. (They have stunning artwork on their website 🤩 too)

Twitter recently served me an ad, not a user's post, an ad, for an antisemitic podcast. The ad explicitly mentioned "burning Jews". It's high time to leave this place. I'm using #HelloQuitX to leave along with 19567 followers and followees to #BlueSky:

Personal update: I am no longer at Hugging Face. I will take some time to pursue personal projects and find my next adventure. DMs open. Feel free to ping me if you have an interesting AI/ML project to share!

You can play with this yourselves. In short:.- the system prompt is in the blog post: - and the Keras chatbot arena app is here:

End result here: This is a very simple exercice on purpose. I'm not looking for AGI, just workable human-computer interactions in plain English. Gemma 2 9B and Llama 3.1 8B did pretty well, but it goes downhill from there.

I built this small demo app with Gradio on Hugging Face Spaces. It loads 7 models at once on a TPUv5e. You can chat with 2 models at once, and swap them at any time. Click here to try:

This would be a great way to interact with computers, but does it work with current LLMs?.

The LLMs would make various mistakes and I'm testing if you can ask them to fix them. Simply by asking in English e.g. "you got the time wrong, 5PM is 17:00".

It's just a typical "assistant" conversation, asking the assistant to set or cancel meetings, using a bare-bones API with just two calls: add_calendar_entry(. ) and remove_calendar_entry(. ).