Mmastodon TechnologyAI first seen 7 h ago, last 7 h ago, peak #4
Paris emerges at layer 28 of a 36-layer model
Original: Paris shows up at layer 28 of 36 https:// youtube.com/shorts/s8DcpY1IQXI # AI # LLM # Transformer # llamacpp
Machine learning practitioners are discussing which layer of a 36-layer transformer language model starts representing 'Paris', a phenomenon tracked in internal activations and popularized by interpretability work. The claim, shared alongside llama.cpp and transformer hashtags, is that the concept of Paris only becomes visible from layer 28 onward, illustrating how such models build concepts progressively across their layers.
Why now: unclear
Evidence
- Paris shows up at layer 28 of 36 https:// youtube.com/shorts/s8DcpY1IQXI # AI # LLM # Transformer # llamacpp · R4TSQ · 1
API: https://socialmediatrends-api.osmike.com/v1/trends/2033439