HN Greets LamBench With Curiosity, Then Starts Arguing About One-Shot Scoring

Hacker News liked the idea before it trusted the number. LamBench, a new benchmark built around 120 pure lambda-calculus programming problems, landed with a live leaderboard and an immediate promise: maybe this is the kind of evaluation that is harder for frontier models to memorize. On the source page, the top entry on April 24, 2026 was openai/gpt-5.4 at 110 correct answers out of 120, with the rest of the top tier packed closely enough to make the benchmark feel competitive rather than settled.

That freshness is what caught people's attention. Several commenters said new, unbenched problem sets are now one of the few ways to get signal out of model comparisons, because older coding evals are so overworked that every launch turns into the same recycled scoreboard. Lambda calculus also has a nice property for this crowd: it is compact, formal, and unforgiving. A model either gets the transformation right or it does not. That makes LamBench feel less like prompt theater and more like a clean stress test for symbolic reasoning.

But the HN thread moved just as quickly into method complaints. The biggest one was that LamBench is scored in a single attempt per problem. Commenters argued that this clashes with how strong coding models are actually used in practice: with retries, test feedback, and iteration. One reply went further and said a non-deterministic model needs repeated runs before a benchmark can say anything stable. Others pushed back a bit, saying one-shot evaluation still has value when the goal is to find problems that labs have not already optimized against.

The more interesting part of the discussion was not who won. It was the sense that the community is now auditing benchmarks as aggressively as models. HN seems happy to see fresh eval ideas, but not willing to treat any new leaderboard as self-evidently meaningful. That tension may be the real story here: benchmarks still drive attention, yet the audience wants to know exactly what kind of intelligence is being counted before it gives the scoreboard much respect. The original benchmark is at LamBench, and the community debate unfolded in the HN thread.