Apple’s “Illusion of Thinking” Whitepaper

Apple, last week, entered the AI debate in a very interesting way. They published a Whitepaper that basically challenges the AI industry’s core assumptions on machine learning and reasoning…

Apple’s research suggests that large reasoning models (LRMs) that form the basis for AI’s complex problem solving and large language models (LLMs) don’t truly “reason” as claimed but rather rely on sophisticated pattern matching. Basically implying that artificial generative intelligence (AGI) today may be significantly further from reality than what existing industry “hype” suggests.

My thoughts…

Firstly, I think this is a very bold claim.

It appears Apple’s researchers have committed to what psychologists would call the “essentialist reasoning trap”. In this case it’s making the assumption that reasoning must possess some mystical, non-pattern-based quality to be considered “genuine.” This conclusion is philosophically bankrupt especially if you’d consider how human mathematical reasoning operates in solving more a mid-level complex maths problem such as 265 × 83. You wouldn’t be channeling some ethereal mathematical essence but you’d definitely apply learned patterns. Think multiplication tables, place value systems, algorithmic procedures, etc… All these are acquired through repetitive pattern recognition.

So in essence, the breakthrough isn’t that LRMs use “sophisticated pattern matching” but rather that we can say they’ve achieved sufficient sophistication in their pattern integration to approximate the emergent properties that we currently refer to as “reasoning”. 

My main critique of the Apple researchers’ approach is around their focus on reasoning as a discrete feature rather than an emergent property. Who can say with definite surety that human consciousness, creativity and reasoning exist as separate modules?

One is reminded of a similar scenario in the last century, when for many years before the Wright brothers became successful with their experiment, humans had looked at birds and concluded that “true flight” required biological wings, hollow bones, and feathers. When the Wright brothers achieved powered flight in 1903, their critics dismissed it as “mere mechanical locomotion” rather than “genuine flight.” The argument went: “Birds don’t need engines or propellers as they achieve flight through natural, biological means. Therefore, airplanes don’t truly fly but are rather only sophisticated falling machines with temporary lift.” I guess we all know how that ended.

My point: This is precisely the same type of argument Apple’s researchers are currently making with machine reasoning.

It’s not hard to see that the same inconsistencies that apply to machine reasoning also characterise our own human reasoning. We don’t solve every mathematical puzzle with perfect algorithmic precision. We use heuristics, make errors, and sometims even arrive at correct conclusions through flawed reasoning paths (cue The Man Who Knew Infinity).

Apple’s focus on the LRMs failing to achieve “real” reasoning with the human data they’re trained on also misses the point, I feel,  that these machines are successfully modelling the actual messiness of our own human cognitive processes. These inconsistencies aren’t bugs. They’re actually features that prove that the systems are capturing genuine aspects of how intelligence (as we know it) operates.

The Apple researchers are of course a team of very smart people and perhaps the real question their Whitepaper is raising is a more philosophical one, i.e. What criteria should we use to distinguish authentic reasoning across the entire spectrum of intelligent systems, whether biological, artificial, or even yet unimagined.

It’s a question that becomes quite compelling when we consider its uncomfortable consequence: Are we, as humans, prepared to confront the possibility that reasoning itself may be fundamentally more algorithmic, more mechanistic, more reducible to computational processes than our deeply held convictions would have us believe?

So the real question isn’t whether machines can think like humans, but whether we’re intellectually courageous enough to accept that human thought itself may be more machine-like than we’ve so far dared to admit. Some food for thought.