jd_pressman: Formal verification stalled on labor costs, not complexity—LLMs may change that

In a multi-round discussion with lusichu, jdpressman laid out a systematic view on the practicality of formal verification: the claim that "complex systems can't be formally verified" is an outdated cached intuition. Since it was abandoned as a general-purpose approach, academia has seen steady progress in formal verification methods and tools, and formally verified operating system kernels already exist (e.g., work like seL4). The real bottleneck isn't computability but labor cost—it was historically too labor-intensive to be practical, a situation that LLMs may change.

Confirmed

• Discussion consensus: the difficulty of formal verification lies mainly in dependency structure, not computability itself. Enforcing verification from the ground up (the kernel) breaks compatibility with all downstream software; insisting on compatibility forces programming patterns that are hard to prove; pushing inward from the outside yields minimal benefit, and realizing real gains requires recursively rewriting all dependencies—so the practically verifiable domain remains tiny today.
• Even if some practical programming patterns can't be verified or replaced with verifiable equivalents, shrinking the search space still has value.
• Academia already has formally verified OS kernel results such as seL4.

Why it matters

• Addressing the "tweezers-and-toothpicks" idea of using LLMs to harden application-layer software piece by piece, jdpressman noted that application software is nearly disposable anyway, making this unreasonable—and the underlying dependencies are still C, so the protection gained is limited. The real leverage point sits at the root of the dependency chain, offering a framework for re-evaluating formal verification's value in the LLM era.

2026-10-23 ~ 2026-10-23 · 7 related posts

• Formal verification's barrier is labor, not complexity — verified OS kernels already exist — jd_pressman · 2026-10-23
• [source] Formal verification was abandoned for labor costs, not infeasibility — jd_pressman · 2026-10-23
• Formal verification debate: narrowing code search space to the verifiable 1% — jd_pressman · 2026-10-23
• Why LLM-driven software rewrites hit a wall: kernel breaks compat, compat blocks provability — jd_pressman · 2026-10-23
• jd_pressman: recursive dependency proofs make formal verification impractically narrow — jd_pressman · 2026-10-23
• LLMs make code cheap enough that recursive rewrites become tractable, argues researcher — jd_pressman · 2026-10-23
• [source] LLMs make code so cheap that recursive rewrites could make verified software tractable — jd_pressman · 2026-10-23