2026-06-03 15:06 UTC
Replies (17)
-
@reykjalin@social.treehouse.systems 2026-06-03 15:11
@b0rk@social.jvns.ca I’m personally excited for a future where tiny but hyper specialized models you can run on your own computer are available. I don’t think I ever need these generally-“good”-at-everything style models. But something tiny that’s really good at React? Or testing strategies? Or finding articles written by people? I think I’d love that
-
@xpomul@norden.social 2026-06-03 15:14
@b0rk@social.jvns.ca I think there might be. The maths behind searching is creating a matrix of occurrences of words and compressing that matrix (Latent Semantic Indexing). In a very simplified way, LSI is like an LLM with a context of 0. So I guess LLMs which are also only matrices might be used in a similar way… (I am not very deep into the maths and theory of LLMs, but aren’t embeddings something in the direction you are looking for?)
-
@tom@labyrinth.social 2026-06-03 15:14
@b0rk@social.jvns.ca I can definitely imagine a few ways LLMs could make a search experience "better" but idk if the amount of "better" is worth the resources
-
@d6@merveilles.town 2026-06-03 15:19
@b0rk@social.jvns.ca one of the really sad things about this moment is that it seems like LLMs work best (fsvo "work" and "best") when interfacing between a person and a predictable/reliable/deterministic thing (such as a programming language type system for example) but instead people seem to be trying to replace all the predictable/reliable/deterministic things in our world with LLMs (which are not predictable, reliable, or deterministic). i feel this is especially true for search engines.
-
@webology@mastodon.social 2026-06-03 15:20
@b0rk@social.jvns.ca the best I have search + llm website I have seen is https://www.statmuse.com I suspect they are doing a ton of vector work behind the scenes but even if you do not sportsball, it's solid tech.
-
@magsol@quinnwitz.house 2026-06-03 15:32
@b0rk@social.jvns.ca A bit tangential, but that’s disappointing to hear that kagi hasn’t been a great experience. That was one of the top Google alternatives I was considering.
-
@aaronb@social.ridetrans.it 2026-06-03 15:41
@b0rk@social.jvns.ca How do you (with a computer program) decide what was written by a human? Google has had a problem for a while that even their search results mostly lead to autogenerated crap. Internet search is harder than ever if you’re trying to find the humans or find true knowledge. Marginalia is the best attempt I’ve seen at a “find human information” search
-
@tmcfarlane@toot.community 2026-06-03 16:24
@b0rk@social.jvns.ca what bugs me is that one thing these things /could/ be good for is fact checking. Never going to be 100% accurate, but it's the kind of thing they are OK at. It turns out no one is interested in fact checking though.
-
@Orb2069@mastodon.online 2026-06-03 16:26
@b0rk@social.jvns.ca I thought the whole point of the Gemini pivot was that Google couldn't sufficiently monetize search to their liking? https://www.impactlab.com/2025/07/19/the-death-of-google-search/
-
@evana@hachyderm.io 2026-06-03 16:31
@b0rk@social.jvns.ca it's particularly ironic because LLMs came out of search (specifically, the attention mechanism from Google), but Google seems to be less on the ball about actual search quality than they used to be. Some of this is improvements in the content-generation tooling from folks looking to make a buck off content ads (and maybe some conflict-of-interest in that), but it feels like Google used to be more effective at surfacing the high-quality content. Phrasing it in this way, I wonder whether the cost of producing good content has risen above the revenue stream of content ads, so the "free" internet content is being subsumed by either paywalled resources or slop.
-
@mala@fosstodon.org 2026-06-03 16:47
@b0rk@social.jvns.ca > (please don't tell me that you can ask generative LLMs to do searches) I won’t tell you but… now I am curious about why 🙃 Is it because you already do it? Or because you don’t want to use them this way? And if you don’t want to, why is that? I wouldn’t ask Claude. But I like using open weights models to help me find material, links between things, and provide me with URLs and whatever extra info I need. And I found a combo of pi, local models and SearXNG to be pretty good at this.
-
@joandaro@mstdn.games 2026-06-03 17:15
@b0rk@social.jvns.ca It's one thing to use a tool but it's another thing to have it shoved to your face I think seeing a summary of things instead of actually seeing the things you want is always annoying, LLM or not. It doesn't matter if it's a LLM summary or if it's a hard advertisement, it's just annoying when the first thing that pops up in a search isn't a search result. Same goes to everything else. It's just annoying when something that *isn't* what you're looking for pretends it is. No matter if it's an ad, a clickbait, or some generated nonsense. It's not about the AI, it's about how people are using it to shove shit to people's face but people keep saying it's an AI problem. AI is a tool that has a lot of potential and the most obvious one is that it makes shoving shits to people way easier than before. But the problem is still people shoving shits to your face, not the tool.
-
@archagon@mastodon.social 2026-06-03 19:19
@b0rk@social.jvns.ca I don’t like LLMs, but I’ve used Gemini for fuzzy searches that would be very difficult with traditional Google semantics. For example, “find articles discussing obscure law X that also mention potential caveats Y and Z.” It’s not a panacea but I’ve found leads that I would have definitely otherwise missed.
-
@pjg1@mastodon.social 2026-06-03 21:59
@b0rk@social.jvns.ca My initial, half-baked thought goes towards "Stack Overflow, but people post the answers they ask LLMs instead of questions". They answers need to be backed up with actual sources, and the person posting could edit the answer to include additional gotchas/add their own two cents, leading to a hybrid of human-LLM responses.
-
@emilsit@discuss.systems 2026-06-03 23:09
@b0rk@social.jvns.ca what has been your experience with Kagi and DuckDuckGo? I have enjoyed Kagi for its UX and unremarkable search results. I appreciate conceptually that they have Kagi Small Web and SlopStop.
-
@paulehoffman@infosec.exchange 2026-06-04 03:37
@b0rk@social.jvns.ca hat parts of Kagi have not been great for you? I'm quite smitten with it, and feel it's easily worth the monthly cost.
-
@jhaas@a2mi.social 2026-06-06 16:32
@b0rk@social.jvns.ca Pattern recognition is what the tools are (theoretically) good for. Augmented search should be great. Maybe they'll even figure out when the sites are all same site in a dozen SEO trenchcoats. But that's not the problem. The forced intermediation is the issue.