But why stop with datasets that induce languages with “grammars” that can be rendered legible to us? Could you make a “Large Solar Flares and Sunspots Model” (LSFASM) and learn to talk to the Sun and ask it where it might flare up next? How about a Large Oceanic Model that allows ships to talk to ocean currents? Or a Large History Model that works as a Prime Radiant for Asimovian psychohistory? Maybe a Large Climate Model constructed out of weather data can talk to us and supply strategies for climate change?

One reason it is hard is, once again, our tendency to mistake discoveries for inventions, or equivalently, cameras for engines. Instruments of discovery measure more than they are measured. Yes, there are a number of ways you can measure a telescope (mirror diameter or focal length for example), but the interesting measuring going on is what the telescope is doing to what it’s turned towards (the analogy to AI here is perhaps to things like floating-point precision — that’s closer to mirror diameter).

Lukas Berglund and others, led by Owain Evans, asked a simple, powerful, elegant question: can LLMs trained on A is B infer automatically that B is A? The shocking (yet, in historical context, see below, unsurprising) answer is no

I’m sharing my intro notes introducing ARRG!’s mission, which I’m excited to say was shared with a room that was at capacity (and, unfortunately, many had to be turned away).

The old web is dying, and the new web struggles to be born.