Comment by kylecazar
8 hours ago
A very small group of companies have the infrastructure required to train and serve a frontier model. They have a stranglehold on the compute/resources required for a new participant to break in.
And an overwhelming majority of enterprises running AI in production do so via one of two providers.
We'll like have an open-weight opus 5.5 class model we can run on (fancy) local hardware in a year or two at full perf, especially with AI optimized M7 macs. That's pretty distributed IMO.
Which brings us to the other decentralization issue we're running into: a 16GB stick of RAM might be $1500 by then.