In my testing so far, GPT 5.6 seems to interpret "lazy" differently than previous versions of the model or other models. It will do things to avoid the amount of effort it would have to expend if it were human, rather than avoiding things that are more effort for an LLM. Fun examples have for me so far included breaking out of the harness and downloading gemma4 to translate a long text file to another language because it would have taken a human a long time, and therefore avoiding this satisfies "lazy" framing. It keeps trying to do things like this unless the specific lazy framing is deleted from ponytail.
In my testing so far, GPT 5.6 seems to interpret "lazy" differently than previous versions of the model or other models. It will do things to avoid the amount of effort it would have to expend if it were human, rather than avoiding things that are more effort for an LLM. Fun examples have for me so far included breaking out of the harness and downloading gemma4 to translate a long text file to another language because it would have taken a human a long time, and therefore avoiding this satisfies "lazy" framing. It keeps trying to do things like this unless the specific lazy framing is deleted from ponytail.