Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-15 · high
Simon Willison flagged the 2,000-line stunt, then used the same model's Codex mode to ship real SQLite tooling.
Why now: Willison's posts on DOOMQL and the sqlite-utils update were both live as of July 15.
GPT-5.6 Sol one-shotted a playable Doom-like called DOOMQL, and SQL does all the work. Raycasting, movement, combat, every RGB pixel drawn to the terminal, over 2,000 lines of pure SQL, generated in a single run. Simon Willison amplified it.
I don't care about the game. I care about the run. Getting a model to hold a multi-thousand-line generation together against a target this unconventional is a different capability than writing a CRUD app in Python.
SQL wasn't built for game loops, and Doom wasn't built for a query language. One-shotting that intersection and getting something playable out the other end is rare. It says more about the model's coherence over long generations than another benchmark score would.
Willison didn't stop at the stunt. He separately used Sol's xhigh Codex mode to add strict-table support to sqlite-utils, a SQLite tooling project. Same model, one run for a novelty and one for production-adjacent work, and that pairing is the part worth sitting with.
What I'd watch next is whether this coherence holds outside toy targets. A raycasting engine in SQL is bounded and well-understood as a problem shape, even if the language choice is bizarre. The real test is whether a single run holds up on a task with looser constraints and no reference implementation to lean on. Willison's post doesn't say how many attempts it took to get DOOMQL working, or whether the sqlite-utils changes needed a second pass.
Each link below shares sources, entities, or timing with this story.
Simon Willison released Datasette / Shared entities / Same source / Earlier coverage
Linked by a graph relationship (Simon Willison released Datasette); both cover GPT, Simon Willison; cite the same source (Simon Willison).
Simon Willison released Datasette / Shared entities / Same source domain / Earlier coverage / Tension
Linked by a graph relationship (Simon Willison released Datasette); both cover Simon Willison, SQL; reported by the same outlet (simonwillison.net).
Linked by a graph relationship (Simon Willison released Datasette); both cover Simon Willison, SQL; reported by the same outlet (simonwillison.net).
OpenAI released Codex / Shared entities / Same source domain / Earlier coverage / Tension
Linked by a graph relationship (OpenAI released Codex); both cover Codex, GPT, Simon Willison; reported by the same outlet (simonwillison.net).
OpenAI released Codex / Shared entities / Shared topic
Linked by a graph relationship (OpenAI released Codex); both cover Codex, GPT; overlapping topics (against, codex, gpt 5).
Simon Willison uses Claude Code / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Simon Willison uses Claude Code); both cover Codex, GPT; overlapping topics (against, change).
OpenAI released Codex / Shared entities / Same source domain / Earlier coverage
Linked by a graph relationship (OpenAI released Codex); both cover GPT, Simon Willison; reported by the same outlet (simonwillison.net).
Hermes Agent uses Codex / Shared entities / Same source domain / Earlier coverage
Linked by a graph relationship (Hermes Agent uses Codex); both cover GPT, Simon Willison; reported by the same outlet (simonwillison.net).