← sources

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

source: reddit (reddit:ClaudeAI)published 2026-11-02 08:50ingested 2026-11-03 02:38

retained text

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 truncated model:** Successfully assembled and executed, but generated repetitive punctuation rather than meaningful language.
* **Embeddings:** Some semantic structure exists, but that does not make the reconstructed weight families genuine trained weights.

**Conclusion: This is a fake model checkpoint containing demonstrably generated weights.** The analysis does not establish who created it or whether it has any connection to Anthropic.

Full forensic research, reconstruction scripts, and evidence:

[https://huggingface.co/adityawakharkar/numbat-forensics](https://huggingface.co/adityawakharkar/numbat-forensics)

All findings are open for independent verification.

document fetched 2026-11-03 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xls4cj/the_772tb_alleged_anthropic_ai_model_is_fake_i