Some thoughts on Apple's recent findings on limitations of current gen AI:

The work is interesting and probably useful.

The research doesn't actually say much of anything about whether current gen models are "reasoning" vs "just pattern matching". These are semantic debates not scientific ones. It does however show interesting limitations in their reasoning capabilities.

These limitations probably aren't fundamental, and within 6 months I expect we'll have models that overcome them.

Reasoning isn’t magical. When I form words into sentences that are logical, refine any illogical from them, and utter them to you— did I not reason? There’s no secret reasoning portion of the brain. We do it at test time, too.