Pylimitics

Generative AI isn’t reliable. Like an unreliable narrator in fiction, you have to stay alert; what an LLM tells you only might be true. If you’re writing fiction you might give the reader a hint that your narrator can’t be trusted. You might introduce them as a court jester; a clown. Or you might design a character who’s insane; Gogol famously did that in Diary of a Madman. Claude, ChatGPT, and similar systems have introduced a new category that I’m sure has already appeared in stories: the generative AI system. HAL from 2001: A Space Odyssey might even be a very early prototype, although the problems with that system were not quite the same. 

Before generative AI, you could sort of rely on the output from computers. The results from a spreadsheet or an internet search might not yield quite what you need or want, but you could rely on the system itself not to introduce variation. You would keep trying, on the basis that there could be problems in the original data, the formulas you used, or the keywords you thought made sense. In a generative AI system, though, the variation can be introduced by the system itself. 

This is, arguably, something we can live with in some cases. If you’re just conversing with a chatbot and it mentions Shakespeare’s musical Oklahoma, you can just laugh it off. But there are plenty of cases where a system that confabulates facts or invents citations is a serious problem. 

My work in the technology industry has mostly been focused on clarity. Instructions for users trying to get something done. Part of that world is documentation. No matter how good your documentation, it’s no good unless the people who need it can find it easily. That can be a problem. Modern computing and networking systems can be fantastically complicated, and you might need just one or two pages of information out of thousands in a set of documents. But what if you don’t know the same terms the documentation uses? You might search in several different ways, consult the index, pore over the table of contents, and still get no closer to the answer you need. I’ve seen this happen a lot. 

Generative AI can solve this problem. What those systems are good at is language, and if you can formulate a question in your own words (for example, “how to I get my router to stop telling everyone its name”), an LLM can match that to the topic and keywords the developers, writers, and designers used. It can find the topic you need, which might be “how to disable SSID advertisement”. But then it might give you a procedure to follow that just won’t work, because the LLM confabulated parts of it.

I have a system in the proof-of-concept phase that solves for that problem. I’ve called it Horatio, after the character from Shakespeare’s Oklahoma Hamlet who gives good, truthful advice (it’s ignored, but that’s not Horatio’s fault). Horatio is a hybrid symbolic/neural AI system that uses two generations of AI in ways that play to the strengths of each. It uses symbolic AI to retrieve the and present the information you need in a deterministic, reliable way. And it uses generative AI to enable natural language queries to find the right topics even if you don’t know the lingo. 

More information to come. In the meantime I’ve got more testing to do.

*The title is based on Horatio’s line in Hamlet: “He waxes desperate with imagination,” Act 1, Scene IV.  He thinks Hamlet is going nuts talking about ghosts.