Part 5/9:
The findings revealed that while additional thinking time, in the form of extra tokens, was beneficial for medium complexity puzzles, performance declined sharply for models facing high complexity challenges. This "performance cliff" is where the internet became particularly reactive, leading to the conclusion that AI systems have limitations without tools or the capacity for inference.
However, it is essential to recognize that Apple researchers did not assert that AI was "useless" due to these limitations. Rather, they delivered a nuanced finding regarding how models perform under specific constraints, inviting a broader conversation about the implications of such limitations.