Some recent applications of AI

Trevor McKendrick asked for favorite recent applications of AI, and Nate Meyvis shared his own, including a system that extracts facts from a corpus, generates embeddings, and clusters them to find frequent facts; using Claude to manage school documents and update his calendar; and using the Zippyflash API via LLM for flashcard reviews. Meyvis notes that cheaper, faster models have made AI reliable for gathering and checking unstructured real-world data, though he still double-checks important items and finds AI lacking for high-judgment lifestyle questions.

Trevor McKendrick asked https://ckarchive.com/b/xmuph6hp4en3kurnppvnqc20gw8kranhd6g7 for favorite recent applications of AI. Here are some of mine: - I have a large corpus in which I want to find the most frequently occurring facts. I've built a system that i attempts to extract the facts from items in the corpus, ii generates their embeddings, and iii runs a clustering algorithm over the result to estimate which of the extracted facts should count as the same. - I'm keeping all important school-related documents in a Claude project. It is reliable enough https://www.natemeyvis.com/civic-applications-of-generative-ai/ at extracting and interpreting practical information that I can ask it to, e.g., update my personal calendar with all the relevant dates. - Speaking of "now good enough at juggling nontrivial real-world information," I'm finding it not only amusing but actually worthwhile to ask an email- and calendar-connected LLM something like "please find all the information related to my trip to X next week and make sure everything seems on track." - I'm now reliant on using the Zippyflash https://www.zippyflash.com/ API via LLM for i allowing me to do some of my flashcard reviews with pen and paper by asking it to generate PDFs of paper-friendly due cards and then asking it to grade my work, confirming any nontrivial right-wrong decisions with me , ii allowing me to customize my study sessions in ad hoc ways that would be very hard to anticipate in advance, and iii allowing me to reschedule low-priority cards in smart ways when I run up a backlog. Some patterns, here and in my other recent AI usage: - There's a shape of task that prompts a "OK, I need to gather a bunch of mostly-unstructured real-world data and check stuff" reaction. The school- and trip-information cases are like this. AI has graduated from "can see the data and be sort of useful" to "can reliably handle these tasks, though I double-check the important stuff." This is a huge step up. Cheaper, faster models https://www.natemeyvis.com/try-the-very-fast-models/ are good enough that I don't feel much friction when I find a question that might be addressable with AI. The fact-extraction example is one of several like this. There's a big difference, for me at least, between spending three hours and $50 on something and spending 30 minutes of active time and $5 on it. - I also find myself deferring some interesting, low-urgency projects in anticipation of their being much cheaper a few months from now. This involves both prioritization and also some guesswork about which kinds of tasks AI will improve at the most. - I don't find it much better, compared to a few months ago, with questions like "what books should be on my list but aren't?" and other high-judgment lifestyle questions. As AI gets vastly better at the basic triage and annoying, lower-judgment data management that I imagine a personal assistant might do, it's nowhere near replacing the highest-judgment aspects of a PA's job. N.B.: I have no first-hand experience of personal assistants.