If you haven’t started building solid, repeatable data models, organizing those data into knowledge graphs, or deploying open platforms that ensure AI can interface with your data, its not too late. 

It’s been a year since I last opined on the topic of AI, so I guess I’m due to update my thoughts on what is proving, for better or worse, to finally be a generational shift in computing.

OpenAI’s initial forecast of imminent AGI is no closer to reality than it was two years ago. What has changed is that we’re getting better at figuring out how to harness and apply AI, and limit some of its downsides. Funnily enough, the key to making AI more useful was to make it less general.

More focused AI “agents” and technologies like Retrieval Augmented Generation (RAG) that focus AI interactions on more specific information sources, have significantly reduced hallucinations, and made it more practical to use AI for particular tasks. Stringing these individual tasks together in the form of agentic orchestration, and leveraging different models for different sub-tasks, improves the odds that Large Language Models can be used in a way that return value to a user in specific problem domain. Two things are true about this evolution in how Generative AI is being understood and applied:

First, the Foundation Models that give agents their general understanding remain expensive, unwieldy, error prone, and poorly understood. OpenAI’s trend away from openness hasn’t resulted in more useful “general” intelligence, but Anthropic’s focused, scientific research and publications “psycho-analyzing” their Foundation Models has made the space less susceptible to accepting wild assertions from grifters like Sam Altman.

Continue reading on LinkedIn…