Claude Fable 5.2 set all 5,994 weights of a small digit-recognition network by writing code, and the method it chose looks like computer vision from the 1990s. Lex Toumbourou, an independent researcher, described the experiment on his blog, notesbylex.com, on November 18. It is one person’s test, run three times, and it comes with a caveat the author states himself: the harder problem he tried afterward did not go nearly as well.
A “weight” is one of the numbers inside a neural network that decides how much a given input signal counts toward the output. Training normally finds those numbers automatically. Here, Claude had to choose every one itself, with rules that barred fitting, regression, and search. Sketching its own filters and templates was permitted.
The setup borrows from a recent paper on Weight-Space Programming, where Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab had GPT-6 Astra write a Go-playing network. Toumbourou swapped in a classic benchmark. MNIST is a set of handwritten digits, 0 to 9, to be labeled. The fixed network was a small convolutional design. Claude could import only NumPy, SciPy, and math, with no PyTorch and no practice games.
Each run allowed up to 50 steps. After every attempt, Claude saw only its training loss, accuracy, and a confusion matrix (a table of which digits it mixed up). It trained on 10,000 MNIST images and was never shown the official held-out test set of another 10,000 images. All three runs used every step and landed around 90 percent on the test set, according to Toumbourou, which is near what a run of Adam, a standard training method, reaches after 50 gradient steps. The best run scored 91.3 percent.
The route is the interesting part. Claude began by writing line detectors from formulas and drawing each digit as a few strokes, then scored images against those templates. Toumbourou ties this to template matching, an old staple of character recognition, and to LeCun’s 1998 paper, which noted that earlier systems relied on hand-crafted feature extractors. Its first network scored 71 percent. Redrawing the digits the confusion matrix flagged lifted it to 77 percent.
Then Claude started measuring. At step 4 it stopped classifying and tested a loop detector alone, rewiring the output so each row of the confusion matrix showed how often the detector fired for one digit. Toumbourou compares this to a unit test, and notes the code never sees an image, so the test scores stay honest. By step 18 the network reached 83 percent. At step 31 it discarded the drawn templates and rebuilt the classifier from rates measured on the training set, reaching 89 percent, then 90 percent after tuning global settings.
Measurement seems to separate the models. Toumbourou says Fable’s runs devoted between 22 and 27 steps to measuring. Opus 5.5 spent far fewer, kept guessing, stopped after 19 to 23 steps, and scored 70 to 74 percent. The comparison is not perfectly even either: Adam’s 50 steps saw 3,200 images in total, while Claude got scores over all 10,000 at every step.
Cats versus dogs, a Kaggle photo task at 64x64 pixels with a 24,898-weight network, went worse. Fable built a bank of detectors for pointed ears, facial geometry, and grass. The ear idea pointed the wrong way. Its best network, at step 9, used photo properties instead: green, texture, and vertical edges. It scored 61.6 percent against a 50 percent coin flip. That beat 500 steps of Adam (55 percent), but Adam reached 74 percent after 5,000 steps, and the 17-step run cost $8.08.
Toumbourou then reread the Go result from the original paper through the same lens. The 1.5 million parameter network, he says, amounts to a rule-based Go program driven by 26 hand-set numbers. He admits he knows little about Go and consulted Opus 5.5 for that section, so treat it as the weakest part of the post. His conclusion is that these models are chaining classical hand-built techniques more than writing in some new weight space.
For teams weighing weight writing as an alternative to training, the evidence here is one blog and one easy benchmark: it matches a short training run on digits, and on photos it needed $8.08 to beat a coin flip by 11.6 points.
Lex Toumbourou, notesbylex.com, November 18, 2026.