Tech Tree

OpenAI dropped GPT 5.6 today. Sol now has an “Ultra” setting for thinking effort. I saw that and my first thought wasn’t about the model itself. I was thinking about tech trees, like in an RTS game or the talent tree in an RPG.

In a tech tree, once a path gets paved, everyone piles onto it. Reasoning effort got paved first. More compute poured into the model at inference so it thinks longer and comes out smarter, and “Ultra” is just more points dumped into that same branch. Then agentic loops got paved, wrapping the model so it can act on its own, and everyone built agents. Every new model keeps pouring into the two branches that are already there.

What I keep noticing is how automatic the tendency is. People invest into the paved path whether or not it’s actually the right one. Not because anyone proved it’s the best direction, just because the road already exists and everyone else is walking it. The money and attention flow there, and that’s most of the reason the path keeps getting bigger.

So I can’t really tell if reasoning effort and agentic loops are the best branches, or just the first two that happened to get paved. Once a path gets all the investment, the other directions never get enough to become real paths. Maybe one of them was better and just never got the points. I run into the same thing in audio hardware. Ribbon drivers, the flat square kind, can sound better than the round drivers everyone buys, but the money went into the round ones and the square side stayed small enough that most people never heard the difference.

The two paved branches are clearly working and the models really are getting better. I just keep wondering if that’s because these are the right directions, or because all the resources went there and nothing else got a real chance.