Recently, I learned about Perplexity, a new and trending AI search engine. It gives you both the natural conversation style chat, and the ‘referencability’ of a search engine. The company seems to be a startup and, I assume, trying to beat big companies in the search game, which is a tough feat, but not impossible.
In playing with Perplexity a bit, I found that search engines and chatbots are entirely different beasts.
Nit Picks for Perplexity
I found that search engines today have a few things I took for granted that I lost when using perplexity.
- Editing past searches - sure you can edit past prompts, but with this new ChatGPT style, it is not as fast.
- The chat style is a bother - I would rather have a clean slate when searching than trying to steer the AI towards what I want. The ‘follow-up question’ style, statelessness is not first class.
Infoboxes and Trust
Ultimately, this AI search stuff is kindof like a summarizing tool. You search for stuff, and a tool (AI or not) shows you the summary. A good comparison for these AI search tools is an infobox, a summary of an article written as a list of important facts in a table. An example is Google Knowledge Graph’s infoboxes.
The fundamental problem for infoboxes is that it contains incomplete information, and thus results in a problem of trust. How do you know if the information you are getting is correct? In my experience, reading google’s infobox is only good for surface-level information and things that can be understood at-a-glance. Simple facts, for example. It also sometimes gives unrelated or blatantly incorrect information. This may be because of people gaming the SEO system, and google is summarizing the top N results. But the result is that, I feel the need to click on results rather than reading the infobox even for simple facts. For the case of AI search summarizing, I feel that it may be the same story for me. I can’t trust the results, so I can’t trust the AI to give correct results.
Search Techniques
I feel that search engines can be better, so here are some techniques I use for searching that I feel should be integrated into searching tools. Especially with AI and its ability to understand semantic information.
First using dates to serve better results. Newer articles may be better for new things, but information about the past or fundamental information you would likely need older text. By using date information or subtle context cues from the text (like version numbers, use of certain depracated software features), you can get better articles.
Second, cross-checking information with eachother and picking relavant parts of the text. This is important for sifting forums. Some some replies are partially wrong, and you may need parts of different replies of different forums to get relavant information. Summarizing together texts gives a ‘blend’ of information, generalizing everything. But what I want is ‘picking’ important points, not blending points. One merges which gives general, surface level information, the other shows important and specific points, often times found in a single source out of a bunch of sources. The former AI and summary bots can do. The latter I think only humans can. I think will need a more intelligent approach to this.
Third is showing partially wrong information and relavant questions. AI and summarizing tools give facts about data, things that are true (often generalizing, see previous point). But partial truths can be just as important as complete truths. Partial truths can be tested, and questions can be asked about these truths so they can be expanded upon. Wrong information can also have value too like being wrong but showing you the right mindset. This is something search does not show yet.
Fourth is opinions and alternative usecases. Some things are not black and white, and even if they are, the people using them may not be. My point is searching tools should show alternative uses and differing contexts for you to choose what is right.
AI is the Future
AI search is the future, and I think it will improve a lot in the coming years. Rough edges will be smoothed out, and wildly new features will be added. Either way, I’ll stick to a normal search engine for now.