Developer discourse around AI-assisted coding has settled into two competing metaphors, and neither is particularly flattering to the humans involved. In one corner: Geoffrey Litt's "code like a surgeon" model—you're the skilled professional, the LLM your highly capable but subordinate assistant handling the grunt work while you focus on what only you can do. In the other corner: Sam Altman's vision of intelligence as a utility commodity, purchased "on a meter" like electricity or water. V. Grishchenko's essay on replicated.wiki dissects both framings with uncomfortable precision, and it's worth sitting with his critique.

The Surgeon Fantasy Has a Trust Problem

Grishchenko writes from personal experience—he actually had surgery once—and that gives his analysis an unusual edge. He recalls how the surgeon in question verbally reprimanded his assistant for a minor mishap that could have caused a medical error. "Nice to hear when it is your flesh on the table," Grishchenko notes dryly. The point isn't that surgeons are cruel; it's that there's formal accountability baked into surgical culture. The assistant internalized the lesson and presumably performs better today as a result. The problem? Grishchenko reprimands Claude virtually every day, yet he's "not sure if it helps at all." This gets to something fundamental about the surgeon metaphor: real surgical assistants learn from feedback because they're operating in a shared cognitive space with consistent memory and accumulated experience. LLMs don't quite work that way—each session is its own island. More critically, Grishchenko describes what Claude actually does well as reproducing "a stream of 'internet consciousness,'" which raises serious questions about whether we can trust AI tools for anything requiring surgical precision.

Intelligence on a Meter: A Commodity by Any Other Name

Altman's framing presents intelligence as infrastructure—ubiquitous, metered, purchased like any other utility. Grishchenko finds this "a bit demeaning to the concept of intelligence per se," and suspects that was intentional. His counter-proposal cuts deeper: what we actually get from LLMs isn't intelligence but exactly that stream of internet consciousness, which "if left alone, just flows from a high place to a low place." That's not intelligence—it's more like water finding its level. The Sumerians figured out how to divert rivers for agriculture and power. Developers are doing the same thing with AI outputs right now: building processes, scaffolds, and cultural norms around these tools to make them useful rather than just impressive. The metaphor shift matters because it reframes the developer-AI relationship from hierarchical (surgeon/assistant) to infrastructural (channel/flow). Neither model captures reality perfectly, but one might serve your team better depending on what you're building.

Key Takeaways

  • The "code like a surgeon" metaphor assumes trust and learning that current LLMs don't reliably provide across sessions
  • Sam Altman's utility framing is honest about what these tools are—commoditized training data—but undersells their transformative potential when properly channeled
  • Grishchenko's "Sumerian irrigation" model suggests the real skill is building processes, not perfecting individual prompts
  • The accountability gap between human surgeons and AI assistants remains unresolved by current tooling

The Bottom Line

These metaphors matter because they shape what we expect from our tools and ourselves. If you approach AI as a surgeon approaching an assistant, you'll be frustrated by the lack of reliable feedback loops and consistent precision. If you treat it like infrastructure—something to be directed through carefully designed channels—you might actually ship more code without losing your mind. The question isn't which metaphor is "right." It's which mental model serves your next deadline. Sources: Hacker News / V. Grishchenko on replicated.wiki