Prediction: Semantic search (boring) ends up bigger than generative AI (cool)

As for Agents they seem like they would be unpredictable. Wouldn't it be better to use a generative AI to create test data for an agent so you keep humans in the loop and create agents that have deterministic, reliable outcomes?

See https://smartenoughsystems.com/ for the "path less taken" that integrates less-than-perfect AI into workflows.

Note that with transformer models the usual story is that people try a "zero-shot" approach to a task that works 70% of the time (e.g. the blog poster who asks ChatGPT one question and gets an incredible result isn't super-lucky) but when they go to a "few-shot" paradigm with a modest amount of training data (1000 samples, probably half a day of hardcore classifying... I've sustained much higher rates of image classification for a few days but found my visual system would start to malfunction)

I work full time in this space and can tell you with 100% certainty that it won’t be one or another, but a combination of them (plus other tech like learning to rank).

A search system is always an ensemble of many parts that make up a quality experience. Generative AI will undoubtedly be a part of that from now on.

Look up the retriever/reader pattern to learn more about this. It’s a really interesting field and I’m really glad to see more people getting involved with search tech with the recent AI boom.

Edit: I should also shameless plug the book AI-Powered-Search on which I'm a contributing author. Just looking at the table of contents you can see how deep this subject goes, and how much more there is to learn! https://www.manning.com/books/ai-powered-search?a_aid=1&a_bi...

It's both a sign of the times and a shame in my opinion that none of these options include the words "sustainable" or "long term". Just short term, unsustainable play using someone else's money to follow a fad.

But that's probably just my elder millennial desire for stability that's beyond my reach.

Speaking as a preposterous layman in this field . . like, "layman" as in "crammed some old linear algebra and calc courses a few months ago just so I could grok the basic design of these things" . .

Semantic might maybe could be the way to divide a solution up among a set of competing models. Right now, the real pain in the butt of using models for solutions (for me, anyway, aside from the unpredictability and explainability and sheer size and corpus) is that a wrong solution requires another trip back to the model (or, I guess, mastering the art of "prompt engineering", but I haven't gotten consistent guidance on how that works). Chopping up the parts of the solution, then using smaller models to compete for the chunks, then assembling the chunks and smaller models compete for that - this could be where we end up, and semantics might be one possible way to chunk things. Procedural data, after all, isn't just natural languagee: you got your part numbers, your geometries, your interwiring stuff, and lots more, all of them separately optimize-able.

The trap I always try and avoid: believing semantics are in some way an inherent or emergent property. Lots of times people in content systems have a near-religious belief in a sort of Platonic Semantics, which exists in this ether-space. Better to think of semantics as framework for NLP. My opinion, anyway. Keeps you out of a lot of trouble.

But what do I know? We could get yet another Mystery Emergent Behavior at ten trillion parameters in GPT-N that eliminates all or practically all problems with the solution, or somehow pre-guesses your future trips, or sees through your question(s) to the underlying problem. Right now, where I am, using any of this LLM stuff is strictly verboten, due to data restrictions and infosec. It's all after-hours projects at this point. So that's maybe another reason I'm thinking about smaller models - easier to get them past the gatekeepers.