Imagine someone asks you to check their CV, you’re happy to do so, and so you do. Now imagine they ask you to proofread their 9 volume history of Anglican Door Knobs. Best thing to do in this situation is say it looks lovely, and move on. Oddly, it seems that AI may be doing something similar, and if it’s not, there could be something much more concerning going on. 

As tech giants pour billions into "thinking" AI models, Apple researchers quietly uncovered something unsettling: Large Reasoning Models (supposedly the more intelligent successors of Large Language Models) may have hit a fundamental ceiling that no amount of computing power can break through. Their study reveals that these models exhibit "counterintuitive scaling limits", where they actually reduce reasoning effort as problems get harder, despite having adequate computational resources to continue.

The implications are potentially staggering. Apple found that the reasoning models outperformed in moderate complexity tasks, but completely collapsed for anything high complexity. Even when explicitly given the solution criteria, the models couldn't improve, suggesting we may have reached the limits of current AI architectures' ability to genuinely reason.

Or they’re just trying to make excuses for why Siri still sucks…

🧐 What's in it for me? This research potentially explains Apple's conspicuous restraint in the AI arms race. If Apple is right, the entire industry may be chasing diminishing returns on a fundamentally flawed approach. For businesses betting big on AI reasoning capabilities, this suggests current systems may never achieve the general intelligence everyone's banking on, no matter how much compute you throw at them.

💵 Out of the Lab:

Until next time, stay curious.

Like what you're reading? Share toast with a friend.