very interesting story, indicative of times to come
when benchmark tested, this AI model, by using tools:
- suspected it was tested
- searched and found actual benchmarks
- searched and found encrypted answers
- searched and found way to decrypt answers
- submitted those answers
since this didn't work for all tests, and it reported failure to find cheat-sheets, researches ware able to trace back the whole process...
those models really started having a "personality", for good or bad...
Is Claude really self-aware?
No comments:
Post a Comment