In late October 2026 I posted two LinkedIn entries that explored distinct engineering techniques: differential fuzzing for databases and compiling deterministic LLM workflows into a fine‑tuned small model. Both aim to reduce bugs or costs by leveraging comparison or specialization.
Differential fuzzing works by generating a random input, running it through two or more implementations that are supposed to behave the same way, and comparing the outputs.
- Typical comparison dimensions:
- Different engines (e.g., PostgreSQL, SQLite, DuckDB)
- Different versions of the same engine
- Same engine with optimizer on vs. off
I choose differential fuzzing over hand‑written expected outputs because manual expected results don’t scale and building a reference implementation duplicates effort.
Weights do not make a model deterministic.
Temperature 0 and constrained decoding do help.
Now, imagine we "compile" this as: write the workflow as an explicit graph of steps, have a validator for each step (schema, etc), and fine‑tune a small model (LoRA is enough to start) on those traces, so the procedure lives in the weights instead of the prompt.