While waiting for Nvidia earnings on Wednesday to determine the short-term course of the market, let’s see where the technology will be going.
The first two years of the AI Era were about “Big AI” projects like ChatGPT, Gemini, Claude, and Co-Pilot, run by the biggest tech companies. They were wholly proprietary, depending on huge investments in Nvidia hardware.
But it’s clear that these are hitting a wall. As LLMs get bigger, they don’t get much better. There are diminishing returns. It’s inherent in the deep learning model. If you want the algorithm to think, you must learn how thought works.
AI development won’t stop. But it will change direction. The Nvidia of next year’s market may be Meta Platforms with its LLaMa tools.
There are already hundreds of small AIs out there. Some are horizontal, devoted to replacing artists in movies. Others are vertical, devoted to a specific industry or even a large company.
Small LLMs require less training data. They’re less expensive to create and run. They allow their creators to focus on getting their money out, something Big AI has failed to do.
While Big AI focuses on unanswerable questions, like control of data, the truth of output, and the nature of truth, small AIs solve real world problems, at prices people and companies can afford to pay.
What Comes Next
Om Malik’s latest project, Crazy Stupid Tech (CST), is following this trend.
CST is following projects like Citate.AI, which wants to deliver the equivalent of search engine optimization to AI. CST has also introduced me to Abridge, which is automating medical note taking.
You’re going to need a search engine to keep up with all this. It’s going to relaunch the computer press industry, and hundreds of analyst careers. (There will also be a subset of the industry covering the analysts, sussing out who has a Clue and who is Clueless.)
As this happens, remember the basics. AI is an outgrowth of database computing. The rule of Garbage In, Garbage Out (GIGO) still applies. It’s about collecting, massaging, and delivering new kinds of output from the “data vaults” and “data lakes” companies have been building throughout the Internet era.
The question I have yet to answer is whether all this “small AI” activity can compensate, in the market, for the collapse we’re going to see in “big AI.” Can these projects, together, create enough demand for cloud resources that the Czars will keep buying Nvidia chips with both hands to serve them?
