Andrew Jardine

🤔 How do we make AI accessible for all? ...... Guess what, it already is 😉. This guide shows how anyone (without code 😮) can train their own #LLM today!

At Hugging Face we've been working to democratize access to ML with #nocode tools like AutoTrain, Inference Endpoints, Spaces and more. But awareness of these tools beyond engineers can be limited. I wrote this article in hope we can introduce ML to a wider audience. 👬 🌎 👭

The guide shows how anyone can:
1️⃣ Fine-tune an #opensource base LLM with your data
2️⃣ Deploy your new LLM to a #ChatGPT type web app

Tools used:
🪐 Spaces - Easy to deploy web hosted ML demos and applications
🚂 AutoTrain - A no-code UI for fine-tuning ML models (including LLMs)
💬 ChatUI - The open-source ChatGPT style UI that power #huggingchat

Thanks to 🚀 Abhishek Thakur for building AutoTrain and helping me to write this blog. 🙏

Yesterday Elon Musk's XAI released #Grok 1 under Apache 2.0 ...........What does this mean for the ML community?........Some initial thoughts 👇👇👇

If you missed the details of the Grok release, here's a recap:
✨ Base model only
✨ MoE architecture (8 experts, 2 active)
✨ 314B params (86B active)
✨ 73% on MMLU, 62.9% GMSK, 63.2% HumanEval
✨ 8K context
✨ RoPe embeddings enabling context extension
✨ Apache 2.0

1️⃣ At first, most won't be able to use it
Grok needs multiple H100s to run and more to fine-tune, so will remain the preserve of the GPU rich for now. But, over time the community will find ways to efficiently deploy and fine-tune Grok, cloud providers will then make it accessible opening up use.

2️⃣ Increases pressure on closed #LLM business model
With Grok now, and #Llama 3 soon there will be fewer reasons to pay a premium for #GPT4, #Claude3, #Gemini or #Mistral Large. Eventually businesses will use frontier models selectively when needed, and choose to privately host next best open models for most workloads.

3️⃣ Reduces reliance on Meta
Some have bemoaned the reliance on Meta as our sole champion among big tech for #opensource ML. While I disagree given the recent contributions from Tencent, and Google. The release of Grok illustrates there will be continued incentives to release open models and cements a bright future for open ML.