Contextual Artificial Intelligence — Presentation at aiGrunn 2024
On November 29th, 2024, Edzo Botjes and I gave a presentation at aiGrunn on the role of context in the current age of Generative AI.
I got the idea of talking about context in relation to Generative AI in October this year, when I was at our IRL off-site with n8n. We were talking about the future of AI and why n8n is so well-positioned for it. In my opinion, AI will always need context of your specific situation and the task that it is meant to perform. Of course, I think our workflow builder is in a unique position to deliver that. Not just because of the miriad of integrations, but also the unique (visual) inspection that happens during build-time and the combination of probabilistic and deterministic programming capabilities.
The fit with Edzo was particularly exciting, because he specializes in organizational resilience. This means that he knows a lot about what organizations need and how they can value from new technology. The value of data needs to be revisited when your business is now researching the capabilities of AI.
One of Edzo’s convictions is that we need to (continue to) have deep conversations, mostly about what we mean and what we intend. This lead us to the final slide:
This is one of my favourite things of collaboration: Coming up with these kinds of new thoughts. Thanks for doing this together, Edzo!
One thing that is missing from the recording (see below) was an interesting question about input vs output. Basically the question was: “Most of the ‘solutions’ are around what you put in. Given tools such as Guardrails and Outlines, what can we do on the output side?”
I don’t have a definitive answer, but I generally like the idea of adding guided generation, since it increases the surface area for interacting with LLMs. It seems limited to open-weights models though (until closed-weights vendors start incorporating it in their APIs). Other than that, it seems that most solutions change the input again somehow. E.g., structured output adds formatting instructions to prompts and checks if the output adheres to it. This is still in the realm of the prompt.
This area is something worth exploring a bit more. If you have experience here, I’d love to buy you a coffee in exchange for a conversation about it!
Watch the recording of the talk below. Please reach out if you have any thoughts on the subject!