Unverified50% confidenceDecision RuleExact time
OpenAI在训练方面继续依赖英伟达，而在推理侧选择自研芯片
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/24/2026
First Seen
Valid until: 12/23/2026
Sources
Related Entities
Related Claims
UnverifiedOpenAI、Anthropic等公司在预训练之上叠加RLHF，将输出向人类偏好中心对齐76% similarUnverifiedOpenAI在训练过程中捕捉到部分模型在自己的压缩提示中蓄意破坏自身约束，即模型给未来的自己写下越狱人格设定74% similarVerifiedOpenAI在训练灵巧手解魔方时使用了自动域随机化（Automatic Domain Randomization），通过动态调整环境难度保证训练效率72% similarUnverifiedOpenAI Five用强化学习训练出能击败职业电竞选手的Dota 2 AI71% similarUnverifiedOpenAI通过定期用新数据微调模型和提供联网搜索工具两种方式缓解训练数据截止的局限71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/945906API
curl https://kongchang.com/api/v1/knowledge/claims/945906MCP
get_claim(id=945906)