I'm still using Kagi for all my searches, I still like it. I've always had the thought rattling around in my head, though, that their approach to ML integration makes me wary.
Recently, they introduced a new pricing model and even more ML functionality. To quote @iliana on their previous ML functionality,
Kagi seems dead-set on adding [ML]-powered functionality to their site. […] I subscribed to Kagi because they didn't have [ML] in my face, and yet they intend to make one of the most stereotypically-unsustainable buzzword features one of their highest priorities because they think people want that, likely based on the subset of their customers that bothered to stumble into the Discord.
This, and similar thoughts I had already had, rattle around in my mind as these newer features are announced.
They already had ML-powered answer generation, which I have left on out of curiosity more than anything else. It's much less aggressive about showing probably-wrong answers than Google's is, in my experience. I don't hate it. It has to be very, very confident to appear, which is about right for this technology IMO.
The new features include a button to summarise the results for a given search, which is something I have fiddled with a few times, but I don't think I will use often. Okay, I can kind of see how this could be useful to someone, but I don't think I can trust like that.
Finally, they've produced what they're calling the Universal Summarizer, which (kind of impressively) can turn audio, YouTube videos, web pages, Twitter threads (lol) and more into a paragraph or three detailing what they mean. Again, I'm unlikely to use it, and I can vaguely see what this is useful for.
The big kicker with all these new features, though, is the pricing model. Each of them costs "interactions", which eat up the also relatively newly-introduced plan search limits1. This lays bare that these ML operations have a cost. So does search, of course, but it's much, much less intensive than the ML stuff.
This is maybe one of the more responsible applications of this I've seen; making it clear that this is a specialised thing, which has a real cost, makes it pretty clear that these operations are not a normal part of search.
Listening to ATP episode 5272 last night, this segment stuck out to me, among others:
you can’t just replace Siri with a large language model because language models are best thought of as search engines with an amazing summarizer. […] It is very much like a different form of search engine, which kind of makes sense that being in Google with its Bard thing would be using this. […] The result is not here’s a link. That’s web search. The result is here’s an answer. But that answer is informed by all the knowledge that it had. And the summarization, again, it’s not as simple as that. You can see all these articles about how the language models work with probability models or whatever. But the whole point is, there’s no intelligence there. There is no understanding. There’s no intelligence. There’s no credibility.
captions from catatp.fm, which ironically enough, uses the Whisper ML model to do this. I've copyedited some of its more egregious mistakes, though.
John's talking specifically about the topic that Siri feels antiquated by comparison to the advent of large language models, which of course is dismissing that Siri can be hooked up to actually useful functions, but that's kind of beside the point.
I think Kagi's current approach to machine learning passes a personal sniff test which is just now forming in my mind.
It's not taking over search. The search is still really good. It's not a thing I can just directly ask a question of and get an overly confident answer that's completely wrong. It's deployed automatically only when it's very highly confident, in a way that isn't obtrusive to me. I often will need to explicitly call on the ML model to do its more abstract work, and that work has a specific cost to me (because Kagi is not ad-supported). John's "amazing summariser" is exactly what Kagi is currently offering and from what I have seen so far, delivering. I don't think I will personally rely on it, but I am at least comfortable with it existing.
At least for now.
-
Kagi recently updated their pricing so instead of having a stated break-even number of searches, they are capped and overages are charged a small fee https://blog.kagi.com/update-kagi-search-pricing
-
Outing myself as an ATP listener here, please don't be mean to me.
