Claude Is Not a Compiler

LLMs vs Compilers

• Many argue LLMs are not compilers: compilers are deterministic, map a well-defined source language to a well-defined target language, and mostly preserve operational semantics.
• LLMs generate code probabilistically from underspecified natural language, can change behavior drastically with tiny prompt tweaks, and may not preserve intended meaning.
• Some still find the “LLM as compiler” analogy useful in a loose, historical sense: both change how software is produced, and both can transform higher-level descriptions into lower-level artifacts.
• Others suggest LLMs more closely resemble interpreters, transpilers, or “cargo cult developers” than compilers.

Determinism, Chaos, and Algorithms

• Determinism is highlighted as a key differentiator: compilers produce the same output for the same input; LLMs typically do not.
• Even if made deterministic, LLMs remain “chaotic”: small prompt changes can yield completely different programs, unlike most source changes with compilers.
• There is debate over what an “algorithm” is; some note that in a broad computability sense, both compilers and LLMs are algorithms, but compilers are deductive and LLMs inductive/heuristic.

Specs, Prompts, and Development Process

• Several reject the idea that a long spec “just exists” and is then fed to an LLM; specs and code co-evolve through feedback loops.
• Prompts are often seen as underspecified specs; LLMs are good at filling gaps but can make hidden, consequential design choices.
• Some envision futures where higher-level specs are central, with deterministic compilation to code and lockfile-like guarantees, possibly with LLMs helping draft specs and tests.
• Others counter that specs aren’t truly “higher level” than code, likening them to rough doodles versus a finished painting; implementation decisions remain crucial.

Vibe Coding and Code Quality

• Supporters praise “vibe coding” for fast prototyping and low-friction infrastructure setup, especially for small or low-stakes projects.
• Critics are wary of “don’t read the code” attitudes, report poor experiences with vibe-coded software, and worry about subtle, long-term bugs and missing domain expertise.
• There is tension between reviewing all AI-generated code (losing speed) versus relying only on tests (risking incomplete or wrong systems).

DNS and Distributed Systems Concerns

• A significant subthread analyzes the custom distributed DNS system described:
  • Discussion of DNS propagation, caching, TTLs, and negative caching (NXDOMAIN).
  • Some think the approach is reasonable if TTLs are tuned; others say upstream caches still dominate and question what problem is actually solved.
  • Alternatives like wildcard records, anycast, and existing DNS software are suggested; it’s unclear from the thread how much real-world benefit the bespoke system delivers.

Probabilistic Tools, Verification, and Risk

• LLMs are compared to probabilistic Turing machines or Las Vegas algorithms: nondeterministic but with an accuracy distribution.
• Nondeterminism is seen by some as a feature (like UDP, SGD, speculative execution) if paired with strong verifiers: tests, benchmarks, formal methods.
• Others emphasize that current LLM hype downplays error rates and that robust risk assessment and correction mechanisms are still immature.

Ecosystem, Energy, and Hype Skepticism

• Some comments view the blog as “picks and shovels” marketing for the AI boom, framing the story as a polished internal success anecdote.
• Questions are raised about how much time was truly saved, how hidden AI-made product decisions will age, and how to measure long-term costs.
• There is brief disagreement over energy use: one side criticizes “burning dinosaur blood” to avoid human collaboration; another dismisses “energy bad” arguments as anti-human and favors more energy as a sign of progress.