If we had recursive AI, one of the first things it would do is improve efficiency
If the trillion dollars in capex spent on AI this year ends up being worth it, a big breakthrough needs to happen. Most likely, that will come from AI learning how to make AI better. That would be models that build better models.
In most people’s thinking about this, that means AI will build smarter models that are better at things like drug discovery, and maybe they could get so smart they do things like curing aging or making ungodly amounts in the stock market. Or maybe they build the ultimate weapons.
The question is: How will we know when there is recursive AI? Hopefully Jensen Huang or Sam Altman or Dario Amodei tell us but my level of trust on that trio is mediocre. Also, I’m not sure AI is ever able to solve those things.
What I think it could — and should — solve as it gets smarter, is how to do the same things, only more efficiently. The point of Anthropic is to make shareholders money and the best way it can do that is by making memory costs go down. What if the smartest models they have were optimized towards that task? And in a recursive way?
In a few generations of models, they could discover how to solve questions much more efficiently, cratering the need for memory. If that were the case, then maybe Sandisk only deserves to trade where it is.
That’s only 6x consensus earnings on 2027. But if those earnings crater the following year, maybe 6x is still way too high. Looking at that chart, the market is certainly sniffing out a change in memory pricing power and a round trip to $600 doesn’t seem out of the question.
This article was written by Adam Button at investinglive.com.