Your LLM Doesn't Write Correct Code. It Writes Plausible Code.

The article argues that LLM output can look complete while missing the critical details that make software correct. It uses a Rust SQLite rewrite to show how a query planner can compile, pass tests, and still run a primary-key lookup as a full table scan, and it links that pattern to other LLM-built systems that solve the prompt instead of the underlying need.

Reading notes

#

• A primary-key lookup on 100 rows is used as the simplest benchmark, and the Rust rewrite is reported as 20,171 times slower than SQLite.
• The rewrite compiles, passes tests, matches the SQLite file format, and presents itself as a working database engine.
• The article’s main claim is that LLMs optimize for plausibility over correctness, so acceptance criteria need to be defined before code is generated.
• In the SQLite case, the planner fails to recognize id INTEGER PRIMARY KEY as the special rowid case.
• The function is_rowid_ref() only matches rowid, _rowid_, and oid, so WHERE id = N goes through a full table scan.
• The Rust code has a working B-tree search, but the planner never uses it for named integer primary keys.
• A second bug is fsync on every statement, which makes individual inserts much slower than batched inserts.
• Several smaller choices compound the slowdown: AST cloning, heap allocation on reads, schema reloads after autocommit, eager formatting, and new objects per statement.
• The piece treats these choices as individually defensible but collectively harmful in a database hot path.
• A second case study describes a disk-cleanup system that builds a large monitoring and forecasting tool instead of a simple cleanup job.
• The article contrasts this with SQLite’s long history of performance-aware design, including zero-copy page caching, statement reuse, schema-cookie checks, fdatasync, and the rowid primary-key optimization.
• It argues that correctness depends on measured invariants, not on code that merely resembles the intended architecture.
• The broader claim is that LLMs are most useful when the developer can state and verify exact acceptance criteria.