The article discusses the evolving trajectory of AI development, emphasizing that the future lies in the model itself rather than applications. It highlights challenges such as the stalling of generalist scaling, with compute costs skyrocketing, isolating model deployment without an affordable pricing strategy. Innovative training methods are yielding unexpected results, allowing smaller models to excel in complex tasks. With inference costs plummeting, the model providers need to transition higher up the value chain, impacting investors who focused on application layers that may soon be automated.
It's time to call it: the model is the product; all current factors in research and market development push in this direction.
Investors have been betting on the application layer, but this layer is likely to be the first automated and disrupted in AI's next evolution.
Generalist scaling is stalling; capacities are growing linearly while compute costs are on a geometric curve, making deployment of massive models economically unfeasible.
The recent optimizations mean the available GPUs could theoretically fulfill global token demand, rendering current token economies unworkable for model providers.
Collection
[
|
...
]