Unverified50% confidenceFactExact time
Ethayarajh's 2019 research found that the contextual embeddings of BERT and GPT-2 occupy a narrow cone.
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedResearch in 2024 showed that models such as GPT-4 and Claude, assisted by auxiliary tools, are already capable of independently solving approximately 30%–40% of medium-difficulty CTF challenges.59% similarUnverified2023年研究者在 GPT-2 中识别出负责间接宾语识别的完整电路，涉及不足1%的注意力头58% similarUnverified机制可解释性入门可从小模型（如 GPT-2 small）的可解释性分析做起，再逐步挑战更复杂的机制57% similarUnverifiedThe inference of GPT-2 is essentially a series of matrix multiplications, additions, GELU activations, LayerNorm, and Softmax operations.57% similarUnverifiedThe perplexity of the GPT series of models on Wikipedia text has decreased from approximately 29 for GPT-2 to an even lower level for GPT-4.56% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/998007API
curl https://kongchang.com/api/v1/knowledge/claims/998007MCP
get_claim(id=998007)