Just saw a few L2s going at each other again over TPS and subsidies—honestly, I’m a bit tired. Anyway, what I’ve been focusing on recently is the “humanized” part when AI agents execute on-chain—for example, whether they can sense subtle Gwei fluctuations, or whether they can read the opcodes in a transaction and tell if it’s bait. These parts really can’t be solved by simply throwing more compute at it.

In plain terms, AI can run strategies and test slippage, but when it encounters malicious traps or a block reorg, it still has to rely on humans to “watch the private key,” or to “pause” for a moment. The signal I care about isn’t the win rate or returns, but when it knows it should stop. For now, humans are still the backstop.

Of course, maybe I’m just being too conservative. Let’s leave it at that for now.