Unverified50% confidenceEventExact time
The author's earlier article "We tested our own AI safety checks; they caught nothing" describes: after removing the answer from the text, the model borrowed another deadline from the same article (15 days instead of 30 days), and that 15 days belonged to the institution's reply deadline rather than the individual's complaint deadline, yet the check system still passed that answer.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/16/2026
First Seen
Valid until: 1/14/2027
Sources
Related Entities
Related Claims
UnverifiedAt approximately 115K context, the model requires about 21 minutes of waiting before it begins generating, decoding drops to approximately 14 tok/s, but all four integrity checks pass.74% similarUnverified对于参数量达数千亿的模型，单次检查点的写入可能需要数十分钟，整个故障恢复过程可能耗时30分钟到数小时不等72% similarUnverifiedOpenAI's monitoring system flagged the anomalous behavior within 15 minutes, and human review intervened 3 minutes later, but the training task was not terminated until two and a half hours later, and the automatic stop mechanism failed to work properly and had to be shut down manually.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1045689API
curl https://kongchang.com/api/v1/knowledge/claims/1045689MCP
get_claim(id=1045689)