Blog post argues LLMs break the deterministic abstraction ladder (binary→assembly→C→Python) because they output probability distributions, not fixed artifacts.
Key Takeaways
Each prior abstraction layer satisfied f(x) -> y: same source input always produces the same binary artifact.
LLMs produce f(x) -> P(y | z1 | z2 | ... zN): you may get what you asked for plus unasked-for, potentially dangerous additions.
Test suites checking only for the presence of y will silently pass even when harmful z artifacts (credential exposure, open FTP) are also present.
The author frames this not as a limitation to work around but as a categorical difference that invalidates the abstraction metaphor entirely.
Hacker News Comment Review
Commenters largely agreed on the probabilistic-vs-deterministic point but disputed the framing: LLMs with fixed seeds are deterministic, functioning as universal function approximators rather than truly stochastic systems.
The deeper disagreement is cultural: critics note that LLM-enthusiast builders often accept non-determinism as a tradeoff, making the technical argument miss its target audience.
The compiler analogy was challenged: C and Python across different compilers already produce varying machine code, so the f(x)->y purity claim for prior abstraction layers is overstated.
Notable Comments
@conorbergin: LLMs are deterministic under fixed conditions; randomness is injected deliberately, not inherent.
@bigstrat2003: “They are happy to hand off the thinking to a third party, even if it will give wrong answers they don’t notice.”
