A forensic investigation has concluded that the alleged "Mythos Class" model leak, which claimed to be a 14 trillion parameter model from Anthropic, is actually synthetic. While the packaging, tokenizer, and embedding matrix appear genuine, the model weights themselves were found to be generated. The investigation used two tests: one analyzing the relationship between the language model head and embedding matrix, and another checking for differences in input scaling between modules sharing an input tensor, both of which indicated the weights were not from a real Anthropic checkpoint. AI
IMPACT Investigative techniques for verifying AI model authenticity are becoming crucial as synthetic data and model leaks proliferate.
RANK_REASON Analysis of a purported model leak concluding it is fake.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
