(Photo by Greyson Joralemon on Unsplash)

Thought for the week:

With all of the fanfare around genAI, it may surprise you to know that ... some companies still build their own ML/AI models. I don't just mean "fine-tuning LLMs." I mean "training a small, focused model based on their internal data." Like they've been doing for the past several years.

Why would they do this? It's because there are some good reasons to skip the genAI/LLM ride.

Building your own models in-house:

• gives you control over what data went into the model
• provides data privacy, because requests to those models never leave your four walls (and you can remove training data if needed)
• scales better, cost-wise (you're not paying an OpenAI or Anthropic per request or per token)
• insulates you from the drama or problems inside the major genAI providers
• means that you're working with tried-and-true technology (the world has a better understanding of where ML/AI can go wrong; with genAI, a lot of the problems are obscured by the excitement)

... and so much more.

Granted, training your own models requires a greater up-front cost of time and money. And you'll also need specialized skills around to train and maintain those models. But for many companies this is a fair tradeoff for all of the benefits I mentioned.

What about you? Does your company still train its own models? And after road-testing genAI, are they giving ML/AI a second look?