First-hand information for everyone
A new study shows that Claude Opus 5.5 can describe what small neural networks learn by doing local calculations on their weights, rather than stepping through every computation from input to output. The researchers fed the model the raw numbers of a 125‑k‑parameter convolutional network as plain text. No tools, no data, no code were provided. In a fresh chat, Claude answered questions about the network as reliably as if it had seen the dataset.
| Dataset | Classes | Train / val / test | Majority (%) | Networks | Test accuracy (%) |
|---|---|---|---|---|---|
| CIFAR-10 | 10 | 45,000 / 5,000 / 10,000 |
| 10.00 |
| M1 / M2 / R |
| 82.36 / 82.28 / 83.95 |
| KMNIST | 10 | 54,000 / 6,000 / 10,000 | 10.00 | M1 / M2 / R | 97.84 / 97.58 / 97.95 |
| Cats vs Dogs | 2 | 18,728 / 2,341 / 2,341 | 50.15 | M1 / M2 | 87.70 / 87.95 |
| Fast food | 10 | 2,400 / 300 / 300 | 10.00 | M1 / M2 / R | 54.67 / 52.67 / 58.00 |
| SAT-6 | 6 | 64,800 / 8,100 / 8,100 | 37.12 | M2 high / M2 low | 99.57 / 36.01 |
The key evidence came from two straightforward clues. First, the model inferred that the network had been trained for roughly three optimization steps, using only the batch‑normalization statistics that the weights encode. Second, it correctly identified the vehicle‑vs‑animal split of the CIFAR‑10 output layer, a hallmark of that dataset. The model’s confidence matches the known dataset structure and even remembered the order of class labels, showing it relies on dataset priors.
However, the study did not show that the model traces intermediate activations or performs a forward pass. Many attempts to reconstruct hidden‑layer evidence failed, and no experiment asked the model to compute a prediction on a test image.
The finding suggests that large language models can read weight files and produce useful high‑level summaries by combining arithmetic on the weight edges with memorized knowledge of common datasets. It also highlights the limits of current interpretability: the model does not traverse the full computational graph, and the results depend on the exact way weights are presented.
The study “Zero‑shot weight reading by Claude Opus 5.5 is better described as local arithmetic plus dataset priors” (arXiv 2611.08070) reports these results in detail.
Visual 1: Table of the small convolutional networks used.
Visual 2: Diagram showing the vehicle/animal split recovered by Claude on CIFAR‑10.