..

Talking about LLMs

People don’t really know what you do. Some long time friends far from science or the tech industry thought I was simply doing some kind of maths with data and computers. Not clear what exactly, surely I can help them connect to their printer or retrieve their forgotten passwords1.

Then ChatGPT came along and Artificial Intelligence became so trendy that everyone (even my lawyer of a mother) started to wonder what they can do with chatGPT or genAI, in their jobs or personal lives. And suddenly they ask “Have you seen this new thing called ChatGPT? Do you know anyone who could help implement it for my company?”… only to realise I could very well be this person. It’s partially on me, I talked about Machine Learning, never about Artificial Intelligence, how could they make the connection?

And that’s how I’ve started giving AI consulting services for a few friends and connections, from technical audits to educational talks. For instance, I recently gave this high-level talk at a company’s kick-off with a broad audience, from sales and marketing to devs. Being so high-level proved to be challenging, there is so many things to say and yet we cannot say everything. Having a diverse audience makes for an interesting challenge as we need to balance “cool and fun applications” and slightly more technical points2.

A few weeks ago, in a totally different context, I was invited by a VC friend (hello Marie 👋) to give a talk at the first edition of Binary Stars: the AI x SWE conference. With a great line-up of speakers (shout out to Harm - which I realised mid talk he was the author of this great blog post on the implications of scaling laws and smaller LLMs), the audience was more than tech-savyy. The objective of the talk was to report on one year of LLMs in the enterprise world as I saw it from Dataiku. I tried to discuss the discrepancy between what people talk about and what they should be concerned with instead. Indeed, the two most recurring topics were fine-tuning and data privacy, which I believe are false problems 99% of the time. While no one cares to think about evaluating LLMs (wink wink no free lunch theorem for LLMs).

  1. Mandatory reference to What People Think I Do / What I Really Do. ↩

  2. While researching on the history of the fear of “Ai taking over our jobs”, I found this great article from MIT’s Technology Review, which links to older editions of the review of 1962 and 1938… which are really worth a read ! ↩