That 14T 'Anthropic Mythos' leak on Hugging Face is RNG noise
7.72TB of name-seeded random numbers wore Anthropic's codename, and 5,846 people grabbed it. Plus: a 4B local Three.js coder and an AI-slop linter.
Copy markdown
43 tensors, one seed: the '14T leak' is pure RNG
A Hugging Face repo, burner22/numbat, posed as Anthropic's unreleased ~14T 'Mythos' (772 shards, 7.72TB, NVFP4 mixture-of-experts) and hit 5,846 downloads before reviewers proved every tensor is just SHA-256(name)-seeded noise — one check matched all 162,246,496 values across 770 of 772 shards, with zero training involved. You can falsify the entire 7.72TB with a ~90KB download, so don't burn the bandwidth, and never run the 'custom modeling code' a fake like this ships with.
A 4B model that writes Three.js scenes on your laptop
CCORE's Qwen3.5-4B-3D-Coder, an Apache-2.0 fine-tune of Qwen3.5-4B, turns a text prompt into a self-contained Three.js scene and is trending on Hugging Face now; it runs locally via `ollama run hf.co/ccore/Qwen3.5-4B-3D-Coder:Q4_K_M` or llama.cpp, with optional image-to-scene vision. It's an independent fine-tune trained on 4,146 scenes, not an official Qwen release, so expect the occasional retry (usually duplicate variable names) and disable multi-token prediction if you run it in LM Studio.
A linter that catches AI slop before it ships
pbakaus's impeccable, back on GitHub trending, pairs a 'design vocabulary' for steering coding agents with a linter that flags low-quality AI-generated artifacts before they reach your repo. If you let agents write UI, it's a cheap gate against output that passes tests but reads as slop — duplicate logic, dead props, half-wired handlers.