Zero sum games

I’m relatively new to ML/AL. I spent most of my career in commodity layers claiming to do boring things, and I’m still learning this field’s history, its culture, and its assumptions. So take this as an outsider’s observation. There is one assumption I keep running into that contradicts what I have seen in the past: the belief that AI is a zero-sum game. It impacts projects, people, organizational dynamics, and everything you can imagine in this domain.

The framing is everywhere. Whoever trains the best model wins everything. Whoever gets to some capability threshold first takes the whole market. Every lab’s gain is another lab’s loss, every open release is a leak of advantage, and the entire field is narrated as a race with one podium.

Maybe the generative aspect of the technology makes it distinct from other fields, the lab that can generate better code wins it all. But there is a possibility that this time round won’t be different than any successful technology in the past. The pattern we watched over and over is the opposite: successful layers becomes a commodity. We operate more Linux than anything else. Databases, once the most fiercely defended proprietary products in the industry, became open source defaults you install without thinking. Container orchestration was a battleground for maybe 3-4 years before Kubernetes turned it into plumbing. Compilers, networking, video codecs, machine vision, you name it. The list is long and boring. The technologies that matter end up being commodity.

Commoditization is not what failure looks like. A technology succeeds when it disappears into infrastructure and everyone builds on top of it without asking permission. The value doesn’t vanish; it migrates up the stack, and it usually grows by orders of magnitude on the way. Nobody who bet on owning TCP/IP won. Everybody who bet on what TCP/IP made possible did. So commodities enable higher layers of the stack and differentiation at product layer.

The early signs in AI is familiar. The price of a given level of capability keeps collapsing. Permissibly licensed layers are becoming more frontier. Knowledge migrate through papers, through people changing jobs, etc. Every “moat” I’ve heard described so far (more compute, more data, better post-training) represents a lead and not moats. A lead is something you have to keep re-earning.

I understand where the zero-sum instinct comes from. Training frontier models is capital intensive in a way infrastructure software never was. When you are raising billions, the story that one winner takes everything is the story that justifies the magnitude. Race framing is also emotionally sticky. Loyalty and tribalism are expected in this domain unlike others.

Whatever happens, I don’t think anyone who is leading today will lose. If models become commodity, then the durable positions look like the ones they always were: distribution, trust, integration, and whatever new layer gets invented on top of intelligence. That is a much bigger and more interesting game than the current race, and it is emphatically positive-sum. The pie is not fixed; the entire point of this technology is that it grows what is possible to build.

Maybe I’m wrong. Maybe there is a discontinuity ahead that breaks the historical pattern because of compute scarcity, or the generative capabilities of this tech. But when an entire field’s self-narrative contradicts 40+ years of how technology has actually behaved, I don’t want to jump onto a conclusion too early.