A new system called werr claims it can handle simple automated decisions without a single neural network weight, using the geometry of the Mandelbrot set instead of a trained model.
The system, built on top of prior work called Mandelbrot Fractal Neural Synthesis and offered through the answerr.me platform, turns each request into a 24-byte coordinate seed and reads how that point escapes the Mandelbrot set to produce a yes/no answer, a category, or a score. It runs with zero stored weight tensors and reportedly zero bytes of VRAM. On JevBench, a 231-example benchmark, the authors report 81.65% calibrated accuracy and a 7.08ms median latency. A separate built-in filter for catching prompt injection attacks is faster still, reaching 3.31ms throughput latency after a 45.8% cut in processing steps, and the authors say it blocked 100% of injection attempts in testing - though their own 95% confidence interval on that number spans from 0% to 27.8%, meaning the real bypass rate is far from settled.
That distinction matters because a lot of what gets marketed as an AI decision - route this ticket, flag this transaction, rank these results - is really just a fast classification call, not a task that needs a large language model's reasoning. Companies already chase that gap with smaller distilled models and rule engines; werr's pitch is to skip stored weights altogether, and its authors even demonstrate it running as an on-chain oracle on EVM-compatible smart contracts for about 21,438 gas.
A 231-example test set and a confidence interval wide enough to swallow its own headline claim suggest the fractal framing is doing more marketing work than the results can back up yet.