LARQL Turns Weights Into Queryable Index

Hot take: this is a legit research/prototype demo with an aggressive marketing gloss. It’s interesting, but the “LLMs are just graph databases” framing is more metaphor than established fact.

• –The repo itself says “The model IS the database” and documents LQL commands like `DESCRIBE`, `WALK`, `INSERT`, and `INFER`, so the core demo is real.
• –The repository also shows concrete support for model extraction into a queryable structure (`vindex`) and benchmark data, which makes it more than a vibes-only video.
• –The claim that inference is just graphwalking is too strong. The README still lists attention, logits, and FFN work in the inference path.
• –The “no matrix multiplication” / “run Gemma 4 31B on a laptop” interpretation looks like extrapolation, not a claim I could verify from the repo.
• –Best framing for experts: interesting interpretability/query-interface prototype, not evidence that transformers have been reduced to database traversal in the general case.