Brace for GenAI letting the early adopters of this technology down

livemint.com:

Subscribe to enjoy similar stories. When I write about Generative Artificial Intelligence (GenAI) in IT Matters, it is usually about its latest advances and the nuances of its seemingly relentless march. I try to discuss both its triumphs and its travails, especially technological aspects that aren’t readily apparent to most.

Today, however, I would like to cautiously join the mug’s game of trying to predict its near-term future. GenAI is impressive and its purveyors promise the moon, but I believe its hype exceeds reality and that we will soon have a reckoning. First, there is the problem of GenAI hallucination.

Since GenAI works by predicting the next words (or pixels) in a sequence, it will become exceedingly difficult to have these sentence-level predictions always grounded in fact. GenAI engines are simply trying to predict the next words in a sentence and not providing constructive or predictive analysis at a larger level. And so they make mistakes.

Attempts to fix such errors seem primarily focused on making their training models bigger, but there is research to show that this approach does not work, since the fundamental method is flawed (arxiv.org/pdf/2203.02155). At a more real-world level, most firms seem confused about integrating GenAI into their operations. Most use cases in large enterprises remain in their proof-of-concept (POC) stages.

Several critical factors must be addressed as IT service providers look to harness this technology to create sustainable revenue streams. The journey from POC to deployment of GenAI solutions is fraught with challenges. POCs are typically small-scale experiments designed to test feasibility and potential impact.