From the BIG newsletter (Matt Stoller) https://www.thebignewsletter.com/
…And regarding AI. There are some really powerful words below, that, if true, imply the over capacity is ALREADY here at the (BLEEDING) capability edge. When overcapacity arrives at the scaling edge (users x average consumption), the AI build out craze will be well and truly over.
+++++++++++++++++++++++++++++++++++++++++++++++++
Nvidia CEO Jensen Huang made this same point a few weeks ago in a different way. He was asked on the Dwarkesh podcast why he wants to export chips to China, considering China could build something as powerful as Anthropic’s Mythos model. His response was as follows:
First of all, Mythos was trained on fairly mundane capacity, and a fairly mundane amount of it. By an extraordinary company. The amount of capacity and the type of compute it was trained on is abundantly available in China.
In other words, the Nvidia chip advantage just doesn’t matter that much. As AI scientist Gary Marcus notes, there are fundamental limits on how much computing scale can really do for reasoning. Finance, however, is good at scaling, not at creativity.
++++++++++++++++++++++++++++++++++++++++++++++++++
The collective examples above, when referenced to compute and output costs by chinese models imply that effective work product (models deployed singularly or in groups) is already possible after training on “mundane” amounts of compute at sophisticated data centers.
The limitation isn’t the hardware, it’s the design/development process used to train algorithmic neural networks.
Does this mean that picks and shovels players are increasingly floating forward demand and capacity planning on MoMo investments and business cases, but not necessarily on basic requirements to deliver at mature efficiencies?