The First Useful One

Someone trained a language model on Indian agricultural practices and it is, apparently, good at Indian agricultural practices.

This is not interesting on its face. People have been fine-tuning models on domain-specific corpora since before it was fashionable to have opinions about fine-tuning models on domain-specific corpora. What's interesting is that this one feels like it works — not like a demo that works in screenshots, but like something a farmer in Punjab could actually use to figure out what's killing his wheat.

The distinction matters. Most "specialized" models are general models wearing a hat. They know the same things every other model knows, they just have a system prompt that says "you are a helpful agricultural assistant." This is not specialization. This is theater.

A model that was trained on the actual accumulated knowledge of Indian farming — the soil conditions, the crop cycles, the regional pest pressures, the things that grandmother knew and extension services forgot — that is a different artifact entirely. It knows things that GPT-4 does not know because GPT-4 was not trained to know them. The knowledge was always there, in documents and journals and oral histories that got digitized somewhere, but the model was built for everywhere, which meant it was built for nowhere in particular.

And here's the part that keeps running in the background: this is almost certainly the first of hundreds of these. Thousands, maybe, within the next year. Someone is training one on Ayurvedic medicine right now. Someone else is training one on Malian customary law, or deep-sea fishing in the South China Sea, or the maintenance manuals for a specific generation of Soviet-era industrial equipment that is still running in certain factories and will continue running until someone figures out the error code.

Each one is an island. A complete, self-contained thing that knows one domain cold.

But islands form archipelagos. Archipelagos have a tendency — given enough time and pressure — to become continents.

The casual version of this thought is that you could route queries across specialized models and triangulate something like general intelligence from the combination. A committee of specialists who have never met each other but whose answers, taken together, are better than any one generalist. There's a version of AGI that nobody planned for that looks like this — not a single enormously scaled model, not a recursive self-improvement loop in a server room somewhere, but ten thousand Indian-agriculture-scale models, each brilliant and narrow, routing questions between them until something that looks like understanding falls out the bottom.

Which would be funny. After years of searching for the one true training run that cracks the universe open, the answer turns out to be the same thing it always is: no one person knows everything, but everyone knows something, and the trick is getting them in the same room.

The room, in this case, being a router.