> In almost every run, the GPT agents were much more aggressive .. riskier strategy vs. Claude’s more defensive posture that tried to win by kill count.
> A tool bug stopped block-breaking. GPT diagnosed it and shared workarounds; several Claudes concluded breaking was impossible.
I wonder if Claude's defensive strategy would have emerged had it believed breaking the bed was actually possible?
> A tool bug stopped block-breaking. GPT diagnosed it and shared workarounds; several Claudes concluded breaking was impossible.
I wonder if Claude's defensive strategy would have emerged had it believed breaking the bed was actually possible?