A truly open source LLM and what it can do for your business and the open community

Recently (as of writing this), I got an opportunity to speak at AllThingsOpen.AI. The event was the first of hopefully many more focused on the Open Source ecosystem that is coming together around AI. This is unbelievably important: there are too many walled gardens for AI, and AllThingsOpen.AI is doing its best to break down these walls. It was an excellent privilege for me to talk at this event.

In my talk, "A truly community-based Open Source LLM and what it can do for your business and the open community as a whole (InstructLab)," I spent 30 minutes in front of over 600 people building a narrative around the power of open source AI and InstructLab. (They recorded me, and you can watch my talk out on YouTube.)

Video will open in new tab or window

I discussed one of the definitions of open source AI and the core problems of the lack of pedigree for these open source models, and I gave the answers to enterprises and developers. Before I did that, though, I talked about InstructLab and the core value add that InstructLab brings to the opinionated workflow for fine-tuning.

After this, I took an aside to AI agents and agentic workflows, discussing where they fit in this budding ecosystem. I mentioned IBM donating BeeAI to the Linux Foundation, and called out one of the best agentic tutorials I've found with CrewAI. (It really is excellent...it built my foundational knowledge about how agents actually work).

Bringing the conversation back to the main goal of understanding the power and privilege of Open Source AI, I pivoted the conversation to the IBM Granite models. With over 1.4 million open source models on Hugging Face, the conversation is overwhelming on what to pick for your base model. Ideally, though, with IBM backing the Granite ecosystem, companies can trust them.

There are six different Granite models, and I spent some time walking through each of them and their main use cases, focusing on the Granite-Time Series models, which has some of the most unique and interesting engineering use cases.

This brings me to probably the thing I'm most proud about at this event: We did some mob programming leveraging the Granite-Time Series model and my personal data on my diabetes management. We pulled (at the event) from nothing a working Jupyter notebook that was able to project some data from my CSVs.

All in all, AllThingsOpen.AI was unbelievably amazing, especially for the first year. The audience was correct, the speakers were correct, and the venue was the perfect size. I look forward to seeing the awesomeness there next year.