Unauthoritative Pronouncements

I know, linking to myself is very Obama-awards-Obama meme. This post is relevant if you’ve been keeping tabs on what I’ve been writing about on my blog recently. Those posts here helped me see a pattern. I’m not sure if anyone else sees a pattern, or if I’ve just made a mess with red string and a pile of mashed potatoes that looks like Devil’s Tower, but I like to think it makes sense.

So here’s a thought for those who might suddenly find themselves in charge of Siri: Search is a foundational element of smart assistants, and the current state of Apple’s search technologies leaves much to be desired.

While all today’s web search engines are placing sparkly and unreliable AI-synthesized answers above everything else, they still generally deliver solid search results underneath. Refining Siri without bolstering the foundation is a recipe for disaster.

Everyone’s making the magic box on top of their results, and Apple’s trying to only make the magic box, sans search results. Let’s get some ye olde heuristics in here first.

Matt Birchler has a tangentially related blog post about LLMs that he put up yesterday while this was being edited. I’ll link to it here for this passage:

This is one example, but I’ve also seen people try to use ChatGPT as a calculator or Claude to give them the weather at their current location, and I sigh because these aren’t the right tools for the jobs. That said, the Google search box has been super powerful at training people to expect search similar fields to behave the same. It’s not exactly better when these LLM search boxes often look like Google and have “ask anything” as the placeholder text.

The simple fact of the matter is that not everything is best solved by an LLM, but LLM chatbots give the impression that they are good for everything. A further problem is that LLMs have a hard time saying they don’t know something, so they’ll always give you an answer whether it’s right or not. This is why I find Google to be in such a spectacular position to have the best of all worlds. It knows where you are (if you let it) and can tell you the weather right now, it scrapes the web constantly so it can give you news that literally broke a few minutes ago, it can show a calculator if you ask it to do some math, and yes, it can detect if an LLM would be best suited to respond to your query and use the LLM for that specific case. That LLM is also backed by the most powerful search engine ever and can parse real time data just like ChatGPT or Perplexity, but with even better search results.

Substitute LLM for Siri and you have something akin to what I’m saying, so I don’t feel like a total crank.

I’m proud of my post as I think it makes the case well (no string and mashed potatoes). The advantage of writing for Six Colors is that Jason Snell, as Editor Supreme, can tell me what’s not working, especially when it’s the whole thing.

I totally rewrote my piece based on his feedback on an earlier draft that was too scattered and didn’t make the point I was going for. He did a nip and tuck on that draft too. I kept getting bogged down with this example and that example, and it just didn’t need it. If you think I don’t have enough examples then, oh boy do I have a folder full of screenshots for you. I’m always glad to go through this process. Hopefully it’ll make me a better writer someday, and not just a guy who rants on the internet.