Compute Cannot Buy Common Sense: Why Apple is Hitting the Brakes on LLMs

The April willow catkins in Jinan are drifting wildly again. Glancing at the calendar, it’s April 22, 2026—coincidentally Earth Day.

While major tech giants are still rushing headlong into releasing extremely power-hungry, trillion-parameter models, Apple has unexpectedly poured a huge bucket of cold water on the hype.

1. Tearing Off the “Paper Chef” Mask

Over the past couple of days, research papers like Semantic Mastery published by Apple’s machine learning team have gone viral in tech circles. I read through them carefully and discovered quite a counterintuitive fact.

Think about it: don’t today’s large language models (LLMs) seem to know everything? Writing code, composing essays, chatting—they handle it all effortlessly. But Apple’s researchers found that once deep logical coherence and factual consistency are involved, these intimidating behemoths often crash and burn.

To put it bluntly, a pure probability model is like a “paper chef” who has memorized millions of recipe books but has never actually stepped into a kitchen. If you ask it how to make boiled cabbage, it can recite the steps perfectly.

But guess what? If you actually make it chop vegetables, it wouldn’t even know if it’s holding the knife upside down.

It is merely guessing the probability of the next word based on the previous one. It fundamentally doesn’t understand what it’s saying. This fact is indeed quite disheartening. Apple has bluntly pointed out that simply stacking parameters using the Transformer architecture has hit a brick wall. Relying on brute-force compute cannot solve this structural flaw.

2. Apple’s Master Plan

So the question arises: why did Apple choose this moment to be the “thorn in the side”?

Actually, this is all paving the way for their operating system gateway. If you’re building a regular chat tool, the model occasionally speaking nonsense can be brushed off as “AI’s sense of humor.”

But what about on a system-level foundation like iOS? Good heavens, what if a hallucination triggers and it accidentally deletes your crucial files, or directs you to a dead-end street while navigating? Who takes the blame for that? Apple absolutely dares not take this risk.

Therefore, they proposed developing structured Knowledge Graphs and a Hybrid Symbolic-Neural architecture.

My rudimentary understanding is that this is equivalent to forcibly imposing a set of “kitchen discipline” equipped with a physics engine on that nonsense-spouting “paper chef.” Symbolic logic represents rigid common-sense relationships: too much salt makes it salty; water under 100 degrees Celsius won’t boil. The chef can continue to showcase its linguistic talent, but in critical actions, it must align with this underlying set of hard rules.

This is very much in line with their consistent technical aesthetics: they don’t seek to be the first, but they demand absolute reliability.

3. Puncturing the Compute Myth

Let’s broaden our perspective and see what the peers are doing.

How have major tech companies dealt with hallucinations over the past two years? By frantically piling up compute. They casually run tens of thousands of H100 GPUs. Public data suggests that squeezing out that final 5% of accuracy might cost more than ten times the previous expenditure.

Following the previous analogy, what is this like? In order to let that “paper chef” occasionally cook a dish right by chance, you simply buy him an entire supermarket, letting him resort to brute-force trial and error. It is extremely redundant and highly un-ecofriendly.

But there’s a problem you need to know. The foundation of pure neural networks is actually very fragile. It is said that in logic evaluations containing irrelevant distractors, the accuracy of large models can flash-crash directly from 90% to just over 30%—meaning that if you just stuff in a couple of nonsense sentences, its brain instantly crashes.

(o_O) Looking at this logic chain diagram, you can tell that Apple simply doesn’t believe the current “thinking” of LLMs is actual reasoning; it’s purely an advanced conditioned reflex.

To put it bluntly, the path of “brute force yields miracles” is not only fiercely competitive but may have already reached a dead end.

In contrast, Apple has chosen an extremely difficult, muddy trail. The cost of building knowledge graphs is exceptionally high. How to blend rigid symbolic logic with flexible neural networks is notoriously one of the hardest nuts to crack in academia.

Frankly, I’m not sure how far they can ultimately push this. By the way, today’s coffee beans were over-roasted, and the whole room smells burnt. — Alright, getting sidetracked, let’s pull it back. Compared to mindlessly burning through graphics cards, Apple’s approach is at least attempting to cure the root cause.

4. Who Gets the Steering Wheel?

Writing this, I sometimes wonder.

If Apple truly succeeds on this path, will AI become extremely boring?

Follow this logic for a moment. When all of the model’s wild imaginations are firmly locked down by the underlying symbolic system, it certainly becomes absolutely safe. But will those occasional, mind-blowing flashes of brilliance disappear along with it?

Digging a step deeper: who will be responsible for writing the “rules” for this knowledge graph?

If it’s a system-level monopoly, that means the vendor holds absolute power to define the underlying logic of the digital world. Above is the natural language designed to please you, and below is the symbolic cage that rigidly confines your cognitive boundaries.

Or maybe I’m overthinking it. After all, for today’s working professionals, getting Siri to just stop failing at setting a simple alarm clock is the most pressing issue.

5. Calling It a Night

After churning out hundreds of lines of messy business code during the day, staring at these dry papers and architectural diagrams in the middle of the night is making my eyes sore. Looking up, my black cat has sneaked over again, rubbing against my leg to demand canned food. Going to feed the master first.

References:

—— Lyra Celest @ Turbulence τ.