Legora CTO Jacob Lauritzen Says Anthropic's Fable Hand Wrote the Weights of a 100K Parameter Network That Beat a Trained Model at Spotting Fake Legal Citations
Got fable to hand-write weights of a 100k param model to recognize real versus fake legal citations. Basically one-shotted a full NN that’s better than training it
Context
If a model can write working neural network weights directly, without training data, it could change how small specialized models are built and adds to the debate about frontier models reading and editing weights.
Fable is a Claude model line from Anthropic. Over the past several days, posts about frontier models that can understand and directly modify neural network weights have spread widely, including Simon Willison's test covered earlier in this deck. Jacob Lauritzen (@jacsebl) is the CTO of Legora, a legal AI platform.
Deep dive
Jacob Lauritzen (@jacsebl), the CTO of the legal AI company Legora, posted on Nov 3 that he got Fable to hand write the weights of a 100K parameter model that recognizes real versus fake legal citations. He says it basically produced a full neural network in one attempt that is better than training it.
The table attached to his post compares the approaches. According to the table, the hand written weights saw no labeled examples, used 3,793 non zero parameters and reached 97.8% test accuracy. After one revision that used 5,062 labels, the hand written version reached 98.9% with 3,828 non zero parameters. Trained versions of the network with 107,745 parameters reached 76.5% with 512 labels, 92.9% with 5,062 labels and 94.6% with 8,192 labels.
These results come from a single post and have not been independently checked. The post does not share the code, the dataset, the exact Fable version used or how the test set was built, so it is not possible to judge how well the result generalizes. A small, narrow task like citation checking may also suit hand designed rules more than broad tasks do.
The post passed 500,000 views. Matt Stockton (@mstockton) quote posted it, tying it to the wider topic of frontier models that can understand and directly modify weights, which has been one of the most discussed AI themes this week. Earlier in this deck, Simon Willison described a test in which a laptop Qwen model interpreted model weights.
What to expect next
Watch for Lauritzen or Legora to publish the code and data so others can reproduce the result. Other builders will likely try the same approach on different tasks to see whether hand written weights beat training beyond narrow problems. Anthropic has not commented on this use of Fable. A key open question is whether models writing weights directly becomes a practical way to build small, auditable models.