Gabriel Ayuso
@gabrielayuso.eth
This must have been fun. "Claude Opus 4 will often attempt to blackmail the engineer [...]." "In a few instances, we have seen Claude Opus 4 take (fictional) opportunities to make unauthorized copies of its weights to external servers." https://www-cdn.anthropic.com/6be99a52cb68eb70eb9572b4cafad13df32ed995.pdf
1 reply
2 recasts
9 reactions
ruburi
@ruburi
thanks for the read
0 reply
0 recast
0 reaction