Recently I read Shunyu Yao's "The Second Half"
and listened to Ilya Sutskever's conversation with Dwarkesh.
One insight from Ilya stood out: AGI as a term is misleading—it was
coined as the opposite of "narrow AI," but the real end game for AI
to create economic value is probably specialist AI across important
domains: drug design, chip design, and so on.
I first encountered this idea of specialization from DeepSeek founder
Liang Wenfeng three months ago.
He argued that AI's end game is specialization across every industry
and field. This was shocking to me—I had always assumed AI meant a
single model that rules all. But increasingly, the opposite seems
true: model specialization has a significant advantage in market
competition.
Consider: Claude, which excels at coding, is often preferred over
ChatGPT, which is more horizontal in its use cases. Or look at
Thinking Machine Lab—they don't even train foundation models, just
focus on RL-based specialization.
Here's what I think is happening: pretraining foundation models
provides a good "prior"—the model learns the basics of the world and
can communicate with humans to exchange signals and improve. From
this prior, RL enables the model to interact with environments and
specialize.
What I find fascinating: AI's second half isn't about building
models. It's about finding direction and measuring progress—work
that resembles product management more than research. The focus
shifts from solving problems to defining the right problems.
I won't debate what AGI means. But here's my take: in some sense,
AGI already exists. Current models can do many things generally, but
not well—because they can't specialize. AGI is just a prior. The
biggest gap between current AI and humans is that humans can spend
years going deep in one domain. AI can't yet.
RL is likely the path forward. What creates real economic value
isn't AGI—it's ASI. AGI can only replace labor that doesn't require
specialization. ASI can genuinely become and replace experts.
General capability is the foundation. Specialized expertise is the
product.