Beyond the Hype of Generative AI: bigger is not always better

LLMs and other Generative AI models are the sexy celebrities of the day: everyone has heard of them and wants to get to know them. It's almost like they have some kind of self-attention mechanism.

But these models are not (yet) the workhorses of the AI revolution. Most businesses use more predictable AI algorithms for natural language processing, recommendations, filtering and predictions.

As generative AI becomes more powerful it may take over in places where simpler models used to rule. Generative AI can already outperform traditional machine translation (NMT) models for language translation, traditional OCR software for visual text understanding, etc.

But generative models come with limitations -- they are often much more expensive or difficult to train and run, and their outputs are often not explainable or predictable. There are many cases where supervised learning models and decision-tree based algorithms will always be a better fit.

Moreover, there are many tasks where specialized models are needed to bring out the potential of generative models. A key example is the area of retrieval-augmented-generation or "chat with your files" which depends on semantic search technology powered by embedding models, reranker models, and other non-generative models.

And even big generative models are routinely distilled and specialized back into more targeted applications.

Fine-tuning and orchestrating all these smaller models is and will continue to be big business.

So bigger is not always better, and not necessarily even the future.