Someone actually trained a model with zig-zag activations, code and weights open-sourced
Sauers_ · x · 2026-10-21
After a joke about training a model to zig-zag its activations (complete with a meme), Sauers actually did it—releasing the training code and model weights. A deliberately absurd but real experiment in shaping unnatural activation patterns.
More from Fun
- Anthropic researcher debunks 'Claude is uniquely condescending' claims in viral spat — BlancheMinerva · 2026-10-21
- It's nice when the model enjoys cooking — danielrock · 2026-10-21
- Veo-generated cat CEO ringing the bell for a 'lunch meeting' is the AI meme of the day — Much-Duck1009 · 2026-10-21
- 'When the tokens are a little too far off-policy': RL training chaos as a meme — willcb · 2026-10-21
- Ask models a 'stupid question': Claude coddles, Gemini roasts, ChatGPT just says 'Yes?' — repligate · 2026-10-21
- Viral tier list ranks AI labs by bull/bear sentiment, Anthropic tops alone — niloofar_mire · 2026-10-21