> There were too many open-ended options , a lot of people who were loud online. So, the developers decided to rent an Airbnb outside the city for a week so the

> There were too many open-ended options , a lot of people who were loud online. So, the developers decided to rent an Airbnb outside the city for a week so they could really focus, isolate, and ship some code. When they got together around a whiteboard, they frantically started researching what tools to build an LLM with .. This ostensibly post-modern article illustrates a few key things: * Gradient decent is indeed a key actor in the play, but it might be easier to explain e.g. classic TF-IDF search or Penn Treebank parsing and compare it with LLMs in order to grasp the ROI of modern techniques. The hype has the crowd confused about relative improvement versus magic. * Mozilla is definitely not going to build and launch an LLM-based search product that could rival Google. Extra fodder for the argument that Google has an unfair monopoly. * ML deployment in the past decade has failed to formalize things like the “reprex” (reproducible input/output pairs) and the “vibe check” (ignoring the test/validation accuracy and throwing subjective inputs through the inference path). The industry around LLMs is getting so “big” that these things aren’t just basic day-to-day skills but whole teams and maybe even dedicated start-ups. But will this hold in 3-5 years after LLMs have evolved?