2026-11-02 · America/Los_Angeles · 社区动态 · #8

The 7.72TB Alleged Anthropic AI Model Is Fake — I Reconstructed Its Generated Weights

I investigated burner22/numbat, a 7.72TB checkpoint claiming to be an internal Anthropic Mythos-class model. My forensic findings show that the tested weights are deterministically generated, not learned through training. The checkpoint's claim to be a genuine trained model is not supported by this evidence. Here's the proof: 633 tensors: 11,374,080 values reconstructed 100% bit-for-bit using SHA-256 and NumPy PCG64. 373.6 million weight signs: 100% matched the predicted pseudorandom sequences. NVFP4 weights: 99.51% of 173 million held-out quantization codes reconstructed. 4.73B-parameter…

热度 42.6 / 100；排名与评分保留该期记录。