A criticism of Kagi's LLMs

NotAHugeFanOfLLMs

I subscribed to Kagi, because I loved the idea of having a search engine with custom results, and paying for it with money instead of my data.

Unfortunately, I am very uncomfortable seeing how much energy is being put into doing everything with LLMs. The LLM "quick answers" I get through Kagi are wrong more often than they are right. Even in the own example from Kagi, the performance of LLMs is very poor, 7 out of the 10 "key points" are just straight up wrong.

The changelog from Kagi claims that "No doubt [LLMs] will get better in the future", but I see no evidence to show that is true. ChatGPT is doing exactly what it was designed to do: spit out the next text that seems to fit. The model inherently has no idea what is true, and what isn't.

I use Kagi because I want a search engine that shows me the pages I'm looking for. I'm not using Kagi as a ChatGPT interface. If I wanted to use ChatGPT, I would use ChatGPT.

This isn't an "I hate all AI" post - I genuinely believe that AI can help us in many areas. I just think using an AI that is designed to spew out plausible-looking nonsense to "answer" my search queries, isn't a great idea.

I know that I can turn off many of the AI features - I just find it a shame that I'm still paying for a technology that I fundamentally disagree with. And my only other choice, is to go back to an ad-funded search engine.

KagiForMe

NotAHugeFanOfLLMs To be fair, you won't escape the AI-ism at Google or Bing or whatever. But yeah, AI created content (such as Quick Peek) should be more prominently labelled as such.

Vlad

NotAHugeFanOfLLMs

Kagi is very careful with LLMs. Check out our AI ntegration philosophy.
https://blog.kagi.com/kagi-ai-search

No feature in Kagi triggers LLM automatically. Everything is on-demand, including quick answer. Meaning you can use Kagi without touching an LLM just fine.

I am very uncomfortable seeing how much energy is being put into doing everything with LLMs.

I think you are just being unfair here. 90% of deployed work each release as seen in release notes has nothing to do with LLMs. Search is at the core of what we do.

Even in the own example from Kagi, the performance of LLMs is very poor, 7 out of the 10 "key points" are just straight up wrong.

The image sent was low fidelity it has no issues with high fidelity. And the message in our release notes was exactly anti "AI hype", so I struggle to understand your concern.

LLMs will have their use in search, for example sumamrization or the ability to ask questions about document. These are the things that personally excite me as it can speed up information retrieval dramatically, and finding stuff is the purpose of search.

CrunchyFritos

I agree with the idea that it can be problematic to mix up LLM results with search results in an ambiguous way, since these sources have different levels of verifiability and are thus often make up different parts of the research workflow. That said, Kagi's FastGPT implementation seems to strike a very good balance by trying to ground truth LLM answers with actual search results. I wonder if some of these concerns could be addressed in a similar way by giving good footnotes / references to LLM-assisted responses?

WoundedMoose

CrunchyFritos This is, I believe, akin to my previous comment about being able to adjust from where the LLM pulls from to give answers. IIRC, the answer was that this was going to be able to be done through the Assistant. I would very much like this, because it can improve the accuracy of results.

Dustin

@NotAHugeFanOfLLMs have you tried using the LLM features (like Universal Summarizer) to summarize search results, PDFs or articles and if so, what do you think of that?

I'm also not all that interested in my search engine having a "chat bot". However, I think these kinds of LLM tools are incredibly powerful when they are focused on a single data source.

I think it's easy to get distracted by the AI mania but I feel that the Kagi team are putting these kinds of tools to good use to enhance search in a really meaningful way.

KagiForMe

Vlad No feature in Kagi triggers LLM automatically.

I feel like making the automatic LLM quick answer for certain queries opt-in, or at least opt-out would have made the community a bit more happy with this feature. I quite like it, but I think this might be worth keeping in mind for the future.

chris_20017

One of Kagi‘s advances is the ability to customize your search experience. While I occasionally use my OpenAI API key elsewhere, I don‘t use AI output with Kagi and Kagi does not force me to see AI generated results. That is how I like it! Given the advent of LLMs, I would be concerned if Kagi would not integrate it into their product, as users will demand it more and more.

Igak

It is the Valley's signature style: churn things out quick, stay on top of trends, "move fast and break things".

You know, I would happily pay 10 bucks a month for the peace of mind of a rock-solid, highly dependable search experience with sub-second loading times, and where the basics are taken care of obsessively well. I would love to see a slow and steady development process where you don't get a totally untested, thoroughly botched iOS app release between Christmas and New Year's eve (next time, just chill!) but instead advanced, new features take years to materialize, as they should, and 2FA is fully functional from the get-go, as it should. Few things done OCD style.

Sadly the general trend in the industry goes in the opposite direction.

CrunchyFritos

Igak I mean, the quality of Kagi's basic search is already very good and in general has only been improving despite the dynamic nature of maintaining search quality in the face of adverse selection.

And the Kagi team's feature development has so far been pretty spot on: I look forward to every new release because the vast majority of features tend to directly improve functionality. So FWIW, I think the Kagi team has done a good job of balancing quality with trying new things, and their track record of improving the functionality and stickiness of their service in my workflows has been very good.

CrunchyFritos

KagiForMe being able to opt out makes sense to me. I personally really like the feature and imagine the marginal user probably wouldn’t want to turn it off but might not discover it if it were not default on.

Edit: seems like this is being addressed https://kagifeedback.org/d/3185-option-to-disable-quick-answers/6