The Output Is Converging
Rip a CD to WAV and you got a perfect copy - typically huge files. It used to be that a few albums could swallow half your drive. That was the 90s. Then MP3 arrived and you could hear what was missing. With MP3s and aggressive compression, cymbals turn to static and busy parts turn into mush. The audiophiles understandably were furious.
But encoders got good and bandwidth got cheap. The typical bitrates went up, artifacts went down, and one day the difference just was not there anymore, not to anyone without lab gear. When did you last hear someone defend WAV? Lossy won the moment it got close enough that the gap stopped mattering.
TVs went the same way. If you walk past the wall of screens in any electronic store you will notice that one set has a better panel and faster processing, and on the spec sheet that gap is real, but if you step back it is gone. Laptops too. Almost anything you can buy now handles real work, and nobody has shopped for clock speed in years. Ask Apple. They sell you a feeling and a finish, not a number.
LLMs are next. A year ago you could pick a frontier model out of a lineup instantly. That tell is fading. These days I look at a piece of work and have no idea what made it. Frontier model? Open weights? Something small running on a laptop? Nobody knows.
Though the models are not equally smart. The gap on paper is something that can be measured, and it may even be widening. But intelligence and output are different things. Past a certain point the extra intelligence stops showing up as something that can be detected with a naked eye.
Frontier models are not going anywhere as far as I can tell. In the narrow places where a fraction of a percent swings the outcome, people will pay a fortune for that fraction. Everywhere else, nobody can tell which model did the work, and the models quietly become a commodity.
Once the model is a commodity, the only thing left worth anything is what you do with it. Turning raw capability into something useful is its own skill, and that is the one that does not commoditize. Electricity is a commodity too. What you do with it is what matters.
Soon or later how many billion or trillion parameters a model has will not matter at all for most practical purposes. What will matter is how that model is used and what it produces.
The output is converging.