Hey there 👋 I would like to quickly plug a publication I am working on to help teams build better AI-enabled products. If you are building a software product and want to integrate LLMs or make sure you're ready for going to production, make sure to check it out!
In recent months, with the general availability of increasingly powerful large language models (LLMs) like GPT4 I’ve noticed a change in my work-related browsing behavior, away from blogs and forums like StackOverflow. The reason for this is simple: Whereas the good old internet required you to get really good at coming up with search queries to find sites describing roughly equal problems and somewhat feasible solutions, the new age of information retrieval is based on natural language as a tool for getting personalized answers. With GitHub Copilot, line-by-line code suggestions reduce lookups for specific libraries. With ChatGPT, understanding complex scenarios for individual cases is remarkably easy. While there is a chance of hallucinations or outright incorrect code being generated, in most cases, it just works.
I try to keep my predictions about the future very limited (mostly because things end up completely different from my wildest imaginations), but I am convinced that there is a definite trend towards mediation layers for the most essential knowledge work tasks and for software engineering, that seems to entail much more “collaboration” with NLP tools for working out solutions to problems. At that point, individual data sources like blogs and StackOverflow are becoming less relevant because they cannot keep up with a general-purpose reasoning engine, designed to combine a multitude of concepts picked up from the textual content of the internet (like the post you’re reading right now) and applied to a very individual problem.
Remaining edge cases where AI lacks capabilities to help include asking questions related to content after the knowledge cutoff of the respective models. Asking ChatGPT to implement a piece of code using a function or library it has never seen in training data will most probably yield a polite non-answer or something completely made up. When augmented with browsing capabilities or access to external data sources, models should be able to fill those knowledge gaps just fine, so there is no need to encode every piece of available information on the planet in a single model.
I’m not quite sure what all of this means for the future of this blog. Fewer of you will use this site to solve their problems directly, mostly because ChatGPT could solve your problem way quicker and better than my write-ups of personal experiences, and in part because you won’t be able to get the attribution to find my content, even if it was used to come up with the answer you received. If I was writing guides and posts for a living, I’d probably be quite agitated by the prospect of this future. Luckily for you (and me), I haven’t written one guide for (immediate) commercial upside, because this is more of a personal notes collection than it is a content marketing site. Just know that I am using the available tools myself, though, so it there may be less of a need for a certain kind of code-heavy guides in the future than up until now. Maybe that’s for the best if it helps timeless content stick around.