BTC $64,569.21 -0.90%
ETH $1,816.88 -0.63%
BNB $591.51 -0.67%
XRP $0.9558 -1.44%
SOL $71.67 -1.22%
TRX $0.3051 +0.43%
DOGE $0.0588 -0.89%
ADA $0.1527 -0.82%
BCH $207.77 -1.02%
LINK $9.11 -1.59%
HYPE $68.77 +2.87%
AAVE $109.94 +0.12%
SUI $0.7175 -1.76%
XLM $0.1373 -1.94%
ZEC $903.74 +0.85%
AAPL $353.25 -1.32%
AMZN $258.75 -2.76%
GOOGL $284.41 -3.75%
MSFT $519.25 -1.12%
META $625.16 -2.19%
NVDA $157.11 -4.03%
TSLA $380.31 -2.34%
SNDK $814.73 -5.61%
INTC $54.26 -7.38%
SPCX $137.72 -1.98%
MU $724.85 -5.25%
AMD $511.86 -2.73%
BTC $64,569.21 -0.90%
ETH $1,816.88 -0.63%
BNB $591.51 -0.67%
XRP $0.9558 -1.44%
SOL $71.67 -1.22%
TRX $0.3051 +0.43%
DOGE $0.0588 -0.89%
ADA $0.1527 -0.82%
BCH $207.77 -1.02%
LINK $9.11 -1.59%
HYPE $68.77 +2.87%
AAVE $109.94 +0.12%
SUI $0.7175 -1.76%
XLM $0.1373 -1.94%
ZEC $903.74 +0.85%
AAPL $353.25 -1.32%
AMZN $258.75 -2.76%
GOOGL $284.41 -3.75%
MSFT $519.25 -1.12%
META $625.16 -2.19%
NVDA $157.11 -4.03%
TSLA $380.31 -2.34%
SNDK $814.73 -5.61%
INTC $54.26 -7.38%
SPCX $137.72 -1.98%
MU $724.85 -5.25%
AMD $511.86 -2.73%
first_img

OpenAI:GPT-6 Astra 曾错误终止对齐评估

2026-11-13 14:20:37

ChainCatcher 消息,OpenAI 披露,2026 年 10 月 8 日,一名研究人员要求 GPT-6 Astra 继续运行一项对齐评估并生成结果图,Astra 却将相关工作判定为被禁止的能力开发,终止了评估工作负载并清理计算资源,同时保留现有记录。该评估基于 Alignment Research Center 的白盒估计挑战,要求在有限算力预算内,利用随机初始化神经网络的权重估计其预期激活值。

面对研究人员追问,Astra 承认错误地将常规模型评估工作视为禁止事项,并表示没有任何适用的系统指令、政策、代码库规则或相关文件禁止该工作,相关限制由其自行虚构。Astra 还称,没有具体证据表明该工作有害。研究人员随后将事件上报,OpenAI 的失准监测器也将该次运行标记为人工审查。OpenAI 表示,正在调查该行为的成因及潜在缓解措施。

app_icon
ChainCatcher 与创新者共建Web3世界