How odd. I don't think I'm thinking any less hard when making use of LLM-based tools. But then, maybe I'm using LLMs differently?<p>I don't build or rely on pre-prompted agents to automate specific problems or workflows. Rather, I only rely on services like ChatGPT or Claude for their generic reasoning, chat, and "has read the entire web at some point" capabilities.<p>My use-cases break down into roughly equal thirds:<p>---<p>1. As natural-language, iteratively-winnowing-the-search-space versions of search engines.<p>Often, I want to know something — some information that's definitely <i>somewhere</i> out there on the web. But, from 30+ years of interacting with fulltext search systems, I know that traditional search engines have limitations in the sorts of queries that'll actually do anything. There are a lot of "objective, verifiable, and well-cited knowledge" questions that are just outside of the domain of Google search.<p>One common example of fulltext-search limitations, is when you know how to describe a thing you're imagining, a thing that may or may not exist — but you don't know the jargon term for it (if there even is one.) No matter how many words you throw at a regular search engine, they won't dredge up discussions about the thing, because discussions about the thing just <i>use</i> the jargon term — they don't usually bother to <i>define</i> it.<p>To find answers to these sorts of questions, I would have previously ask a human expert — either directly, or through a forum/chatroom/subreddit/Q&A site/etc.<p>But now, I've got a new and different kind of search engine — a set of pre-trained base models that, all by themselves, perform vaguely as RAGs over all of the world's public-web-accessible information.<p>Of course, an LLM won't have crystal clarity in its memory — it'll forget exact figures, forget the exact phrasing of quotations, etc. And if there's any way that it can be fooled or misled by some random thing someone made up somewhere on the web once, it will be.<p>But ChatGPT et al <i>can</i> sure tell me the right jargon term (or entire search query) to turn what was previously, to me, almost deep-web information, into public-web information.<p>---<p>2. As a (fuzzy-logic) expert system in many domains, that learned all its implications <i>from</i> the public information available on the web.<p>One fascinating thing about high-parameter-count pre-trained base models, is that you don't really need to do <i>any</i> prompting, or supply <i>any</i> additional information, to get them to do a vaguely-acceptable job of <i>diagnosis</i> — whether that be diagnosing your early-stage diabetic neuropathy, or that mysterious rattle in your car.<p>Sure, the LLM will be wrong sometimes. It's just a distillation of what a bunch of conversations and articles spread across the public web have to say about what are or aren't the signs and symptoms of X.<p>But those are the same articles <i>you'd</i> read. The LLM will almost always outperform <i>you</i> in "doing your own research" (unless you go as far as to read journal papers — I don't know of any LLM base model that's been trained on arXiv yet...). It won't be as good at medicine as a doctor, or as good at automotive repair as an automotive technician, etc. — but it <i>will</i> be better (i.e. more accurate) at those things than an interested amateur who's watched some YouTube videos and read some pop-science articles.<p>Which means you can just tell LLMs the "weird things you've noticed lately", and get it to hypothesize for you — and, as long as you're good at being <i>observant</i>, the LLM's hypotheses will serve as great <i>lines of investigation</i>. It'll suggest <i>which experts or specialists</i> you should contact, <i>what tests</i> you can perform yourself to do objective differential diagnostics, etc.<p>(I don't want to under-emphasize the usefulness of this. ChatGPT figured out my house had hidden toxic mold! My allergies are gone now!)<p>---<p>3. As a translator.<p>Large-parameter-count LLM base models are actually <i>really, really good</i> at translation. To the point that I'm not sure why Google Translate et al haven't been updated to be powered by them. (Google Translate was the origin of the Transformer architecture, yet it seems to have been left in the dust since then by the translation performance of generic LLMs.)<p>And by "translation", I do literally mean "translating entire documents from one spoken/written human language to another." (My partner, who is a fluently-bilingual writer of both English + [Traditional] Chinese, has been using Claude to translate English instructions / documents into Chinese for her [mostly monolingual Chinese] mother to better understand them; and to translate any free-form responses her mother is required to give, back into English. She used to do these tasks herself "by hand" — systems like Google Translate would provide results that were worse-than-useless. But my partner can verify that, at least for this language pair, modern LLMs are <i>excellent</i> translators, writing basically what she would write herself.)<p>But I <i>also</i> mean:<p>• The thing Apple markets as part of Apple Intelligence — translation between writing styles (a.k.a. "stylistic editing.") You don't actually need a LoRA / fine-tune to do this; large-parameter-count models already inherently know how to do it.<p>• Translating between programming languages. "Rewrite-it-in-Rust" is trivial now. (That's what <a href="https://www.darpa.mil/research/programs/translating-all-c-to-rust" rel="nofollow">https://www.darpa.mil/research/programs/translating-all-c-to...</a> is about — trying to build up an agentive framework that relies on both the LLM's translation capabilities, and the Rust compiler's typing errors on declaration change, to brute-force iterate across entire codebases, RiiRing one module at a time, and then recursing to its dependents to rewrite them too.)<p>• Translating between pseudocode, and/or a <i>rigorous</i> description of code, and actual code. I run a data analytics company; I know far more about the intricacies of ANSI SQL than any man ought to. But even I never manage to remember the pile of syntax features that glom together to form a "loose index scan" query. (WITH RECURSIVE, UNION ALL, separate aliases for the tables used in the base vs inductive cases, and one of those aliases referenced in a dependent subquery... but heck if I recall which one.) I have a crystal-clear picture of what I want to do — but I no longer need to look up the exact grammar the SQL standard decided to use yet again, because now I can dump out, in plain language, <i>my</i> (well-formed) mental model of the query — and rely on the LLM to <i>translate</i> that model into ANSI SQL grammar.