In the TLDR IT newsletter I was reading today, there was an interesting link to Jet Xu’s blog post - The “AI Psychosis” Divide: Why Coders are Terrified and Everyone Else is Bored.

The underlying idea that Jet presents, inspired by an Andrej Karpathy twitter post, is that there is a substantial difference between the experience that coders have with using the latest generation of AI tools, and the experience that more general business users have with AI. He suggests that one contributing factor to this may be that the extensive structure, organisation, documentation and history that are present in the typical codebase provides a lot of context that allows AI agents to perform better. And, that because the actions of the AI agents are then incorporated into the codebase, the AI has “memory” that allows it to learn and improve from previous interactions.

He suggests that one of the challenges of realising the full benefit of AI in more general corporate work, and consulting work in particular, is the haphazard approach to structuring and storing information across all different forms of document and the transactional “chat” interactions with LLMs that doesn’t provide the depth of context and history available to coding agents.

An alternative approach that he discusses, with reference to Andrej Karpathy’s LLM Wiki idea, is to take a more structured approach to doing knowledge work and creating the PowerPoint decks, spreadsheets and other documents that are the outputs of this work. This approach then allows these knowledge workers to see some of the massive benefits of using AI agents that coders get excited about.

As an aside, as a knowledge worker wanting to get some hands on experience with the potential of these new coding agents, I have started dabbling in using Open Code at home and my first project has been using it to put together an agentic system for compiling and summarising the evidence on topics of interest to me from various papers. The approach is quite similar to what has been described by Karpathy and Xu, although not yet anywhere near as refined.

Another related thought is that this approach of using LLM to enhance knowledge structuring addresses the observation attributed to Kazuyuki Hiraoka, author of the howm package for note taking and organisation in Emacs:

If you don’t organize notes, reading feels like a hassle. If you try to organize notes, writing feels like a hassle. This trade-off is a dilemma.

With the appropriate structure, the LLM can assist in organising and compiling the notes for you so that you don’t have to deal with the “reading hassle” noted above.