Search and LLM Tools
“AI BAD!”, IT'S LLM NOT AI!”, FURTHER ANGRY INTERNET NOISES!
Now that we have that out of the way...
So far my journey into the land of LLM has been relatively pedestrian, I think. I waited several months after OpenAI's ChatGPT hype peaked before I started tinkering with it. I signed up for a paid account for a couple of months and found it to be not very useful. In hindsight, I think this was a combination being stuck in my ways for information seeking methods and because it wasn't quite integrated into anything I was using.
Not too long after that, many of the developers at a the company I was working at started to use Github Copilot and there was a big fuss about the legality of code that it regurgitated. I spent a few minutes tinkering with it occasionally from inside VSCode. I found its answers useful for boilerplate and not much else so I abandoned its use outright.
Fast forward a year and some change and I discovered Codeium. What was interesting about Codeium to me is that it had some sort of context for my open files and my project structure. It was better at suggesting refactoring options and helping with syntax for languages I didn't often use. I had a paid account for some time and eventually abandoned it later this year because I just don't spend all my time slinging code. It was not particularly good at devoops tooling so I went back to my old ways.
My current job uses Google things for their email/docs/storage and I noticed that Gemini was available through this account. I tried it out for a while and it was relatively responsive for code and configuration help. I ultimately abandoned it because it wasn't integrated into my editor and deleting my history was really difficult. Google really wants that data.
During this timeline, I signed up for Kagi. I have been increasingly frustrated with search results coming out of DuckDuckGo and didn't find Searx or related tools to be very useful. I signed up for a trial of Kagi even though I thought paid search in the present was absurd. I was blown away by how great the results were and started with a paid account. It was also really interesting to see just how often I searched for things. Thousands of times per month!
Kagi began to add AI features and I feared it meant the enshittification of a lovely service. It was surprisingly useful. The Quick Answer feature was a nice way to get a quick view of the first few results and the Universal Summarizer reliably spit out bullet points for large articles. I still verify the output of these regularly but they have been consistently good.
Eventually they added an Assistant feature that provides access to a handful of LLMs (including Gemini and ChatGPT). This has been really nice for comparing results across models and shaking loose some information about documentation that I don't want to read in full. Having access to multiple models simultaneously with one subscription has been good for my wallet as well.
So in the present, my LLM usage seems to have fallen into two categories: summaries of search results/articles and code boilerplate/syntax. I'm not sure it's worth all the earth burning for those features but I think we can't unsqueeze the toothpaste now.