Unverified50% confidenceFactExact time
H3 的精细细节相比闭源模型偏颗粒、偏涂抹，团队排除了 VAE 和单一训练阶段的问题，承认这是系统性缺陷
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/12/2026
First Seen
Valid until: 11/10/2026
Sources
Related Claims
Unverified随着MoE架构的普及，激活参数量与总参数量的分离使得FLOP/s阈值失去准确性，DeepSeek-V3训练消耗算力远低于其参数规模所暗示的量级66% similarUnverified仅用成功轨迹训练的模型往往存在分布外泛化的严重缺陷，遇到工具返回错误时便丧失纠偏能力64% similarUnverifiedH3的漂移现象暗示其采用了自回归范式，高表现力的代价是累积误差导致的一致性问题62% similarUnverifiedThe first attempt at the reproduction experiment failed to observe the degradation phenomenon because too few training epochs were used (the model was still in an underfitting state), indicating that sufficient convergence during training is a prerequisite for observing the degradation phenomenon.62% similarUnverifiedThe cause of this large-scale erroneous suspension was very likely an overly aggressive update to OpenAI's risk control system rules or adjustment of model decision thresholds, lacking sufficient gray-scale testing.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/735488API
curl https://kongchang.com/api/v1/knowledge/claims/735488MCP
get_claim(id=735488)