Claude Fable 5.2 achieves 91.3% MNIST accuracy using hand-crafted pre-deep-learning techniques
·2026.11.18 06:44
Key point
Claude Fable 5.2 achieved 91.3% test accuracy on MNIST by manually coding 1990s-style feature detectors and classifiers, reaching accuracy comparable to an Adam optimizer after 50 gradient steps, though with different data exposure per step.
Details
- Claude Fable 5.2 achieved 91.3% test accuracy on MNIST by manually coding 1990s-style feature detectors (such as line and corner detectors) and classifiers, reaching accuracy comparable to an Adam optimizer after 50 gradient steps, though with different data exposure per step.
- The model followed a path of hand-crafted templates, loop detectors, and corner detectors, using confusion matrices and feature frequency measurements to adjust weights without PyTorch or gradient fitting.
- In a Go experiment, GPT-6 Astra built a 1.5M parameter network that achieved 84% and 81% win rates against small PPO networks and 70-79% against a 20.5M game-trained network, using classical Go techniques like liberty counting and influence functions.
- The core finding is that LLMs do not directly "write" in weight-space but effectively chain manual design techniques, with feature extractor design determining accuracy more than the classifier itself.
Experimental Constraints and Setup
- Fixed CNN architectures were used (5,994 weights for MNIST, 24,898 for Dogs vs Cats) with PyTorch, fitting, and regression prohibited.
- Only NumPy, SciPy, and math libraries were allowed, requiring weights to be written as formulas with a maximum of 50 feedback steps.
Performance Comparisons
- Claude Fable 5.2 outperformed Opus 5.5 on MNIST, which stopped at 19-23 steps with 70-74% accuracy due to measuring less and guessing more.
- On the Dogs vs Cats task, Fable 5.2 reached 61.6% accuracy by detecting photo properties like green pixels, texture, and vertical edges, outperforming Adam 500 steps (55%) but falling short of Adam 5,000 steps (74%).
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.