The rapture and the reckoning
Putting aside the fact we already have these roads—they’re called train tracks—at our current pace of improving two miles of road every six years, we’d be able to replace all four million miles of roads in the United States by the year 12,002,023. 8 Still, Balaji’s tweet highlights a useful trick. To solve a problem, we can solve it outright—or we can change the problem. Rather than asking self-driving cars to navigate hopelessly complex terrain, we could bulldoze the terrain into something more manageable
To be clear, Narrator’s activity schema is meant to be illustrative, not prescriptive. I haven’t used it in meaningful ways, and I have no idea how ChatGPT would actually interact with it. The point, though, is that it seems unlikely that the schema design patterns that we created for people are optimized for LLMs. As the skeptics of the chatbot analyst correctly pointed out, it’s the relational nuances that confuse AI models. But we could flip that. If AI-powered analytical tools have enough potential—which, seeing what ChatGPT can do, it seems like they do—would it not make sense to favor schema design patterns that enable the big-brained computers more than they enable us feeble-brained apes?
Because the model wouldn’t need to be semantically expressive—LLMs find patterns, not meaning—that spec would likely also be relatively simple and consistent. Our transformation layers wouldn’t be inventive kitchens, figuring how to design bespoke dishes for individual customers. They’d instead become factories, stamping out the same standardized and machine-readable perforations that everyone else does
The implications of this would be profound. For decades, we’ve thought about building data tools and processes for people. They’ve been our best analysts, our best query writers, and our best decision makers. But I’m not so convinced that’s true anymore—and the only thing that’s keeping us on top is we’ve rigged the game in our favor. If
orta do this. The difference is that these tools either require a semantic model to work, or are tripped up by the same problems that confound today’s ChatGPT bots. The point here is to back up a step: Can we structure our data such that these tools work on complex questions without predefined semantic configurations
