Today’s AI Is a Mathematical Scam

⋅

The Whole Problem Starts With This Animation

The animation below is not showing fake progress.

The industry has genuinely become very good at killing the stupid hallucinations: the obvious mistakes, the embarrassing answers, the screenshots everyone loves to mock. Those surface failures have been pushed back by real engineering.

But that does not mean hallucinations are solved. It means the easy version is being cleaned up, while the deeper ones remain. That is what we can call the math scam of current AI, and you will see why throughout this post– Though is only a minimal part of a most powerful, hideous and hidden story by the industry that has several follow-up in my series A Mathematician Lurking in the Tech Underworld


Surface Hallucinations vs Deep Hallucinations.On the left, three wrong answers glow red. The compute spins up, green replaces red, and the errors vanish. Surface hallucinations are fixed by brute force, exactly how the architecture was designed to work. On the right, a golden triangle rotates: E, G, B, the chord E minor. It spins and stays E minor. Then it flattens into a line, G, B, E, and the trouble begins. Same notes, but as a list this looks almost identical to G major, which is 3× more frequent. The model sees G first, assumes G major, and replaces E with D. More compute arrives. The answer gets sharper and stays wrong. One side needed better inference. The other needed a different representation. The benchmarks only measure the first.


The most serious failures lie deeper. They are more structural, more elusive, and far more dangerous. And paradoxically, they have been growing even while the industry keeps selling the opposite story. That is one of the hidden reasons frontier labs have become obsessed with inference-time reasoning, patch cycles, and fresh model releases every few months.

They are not moving that fast because the problem is solved. They are moving that fast because beneath the cleaner surface, their systems are still producing an alarming number of deeper failures that brute force does not truly repair. n

And Despite All Its Structural Failures, Current AI Is Still Powerful Enough to Wipe Out Much of Elite Work. How Is That Possible?

Prepare yourself, because this is where the distinction between surface hallucinations and deep hallucinations turns the story brutal.

By now, you may be tempted to relax. You know the familiar plot because many people have already told it to you, usually in some vaguely Luddite tone: every machine has limits. The hidden wall of AI is real, and one brutal proof of that is the deeper failures you already saw in the previous animation.

And then another thought begins to form, a dangerous one.

If deep hallucinations are real, then perhaps the danger has been exaggerated. Perhaps professions built on training, abstraction, and hard-won expertise are still safe.

No. That is the fatal mistake. Not the total error, but the partial one, and partial errors are often the most dangerous.

Because deep hallucinations do not mean the machine is weak everywhere. They mean only that there is a region of reality where its structure begins to fail. Outside that region, it can still be devastating.

That is why you need to look at the figure below.

These are not trivial exercises, nor are they benchmark toys dressed up as mathematics. They are serious problems, some of which, until very recently, would have been safely assumed to lie beyond the reach of large language models.

And yet they fell.

They fell because a surprisingly large part of difficult mathematics still lies in a zone where brute computational pressure, symbolic chaining, retrieval, verification, search, and code are enough to break through.

Surface difficulty: AI with pure computation can already solve these.These are not trivial problems. They are mathematically serious, but they still belong to a zone where brute-force computation, symbolic chaining, retrieval, verification, code, and structured search can force a solution. That is the crucial distinction. “Surface-solvable” does not mean “easy.” It means solvable without native geometric, topological, or invariant-preserving reasoning. And if even mathematics contains that much exposed surface, the economic vulnerability of many elite professions becomes much easier to understand.



The machine does not need deep structural soundness to break into that territory. It only needs enough computational grip to force its way through the parts of the problem that yield to structured search and symbolic pressure.

And if even mathematics contains that much exposed surface, then imagine what that means for the rest of elite work.

Now Look at the Problems That Still Defeat a Flat AI

Now we turn the table.

What you have seen so far was the dangerous side of the story: even a flat AI, an AI that still breaks on deep structure, is already strong enough to automate a shocking amount of sophisticated work. But this next part is the one that matters most if you are thinking seriously about your own future.

Because this is where the limits of current AI stop being only a technical diagnosis and start becoming a practical opening.

Look at the table below. These are the kinds of problems current AI still cannot solve satisfactorily, not because it needs a few more tokens, a few more patches, or a little more brute computational force, but because the structure of the problem itself resists being flattened. This is the zone where a machine built on pure computation begins to lose its grip.

And that matters for you.

If you learn to move toward these gaps, if you shift your skills, your judgment, and your professional identity toward the kinds of problems current AI still handles badly, then the value of your hard-won experience can rise again instead of being slowly commodified. The machine is eating the computational surface. Fine. Let it. Your advantage now lies deeper, in the places where the system can generate the parts but still cannot hold the shape.

So look carefully at the table of AI unsolvables below, because hidden inside these failures is a professional map of the terrain the machine still cannot cross.

Deep Difficulty: Pure Computation Has Hit a Wall on These.These are the problems where current AI begins to fail for structural reasons, not because it needs a few more tokens, a larger context window, or one more patch. In each case, the problem depends on shape, topology, spatial integration, invariance, or global coherence that cannot be flattened into sequence without damage. That is why the models can still look fluent, confident, and locally competent while failing where the whole structure has to hold together. The message of the table is simple: brute computational force can carry AI surprisingly far, but once the problem stops living in a flat solution space, the machine begins to lose its grip.


Let me advance a few conclusions you may already have begun to suspect.

For elite work in general, stop building your value around speed, volume, and polished first drafts. AI is already taking that ground. Move instead toward the parts of the job where someone has to notice when a result looks fine on the surface but is wrong underneath. That usually means edge cases, messy exceptions, conflicting constraints, unclear situations, high-stakes judgment calls, and work spread across several documents, systems, or teams where one hidden inconsistency can break the whole thing.

In plain terms: become the person who catches what the machine misses. In mathematics, that means checking whether the proof really holds together. In software, it means seeing where the architecture, state flow, or dependencies break. In engineering, it means understanding where geometry, tolerances, and interacting constraints can produce failure. In law, finance, medicine, and research, it means spotting when the summary sounds right but the underlying structure, causality, or exception handling is wrong.

Do not try to beat AI by producing more pages, more code, more slides, or more reports. Build your edge where correctness depends on the whole structure holding together. That is where current AI still fails, and that is where your experience becomes more valuable, not less.

You can already see the pattern in the real world, around you: In protein engineering, current models can produce confident mutation predictions that still ignore key biophysical and geometric constraints, which is exactly why recent work adds explicit structural auditing on top of raw model output.

In autonomous driving, researchers now describe topology-aware spatial reasoning as a central bottleneck for language-based driving models, because fluent scene description is not the same thing as understanding how the road graph actually behaves.

In drug discovery, the field has moved hard toward geometric deep learning precisely because string generation alone is not enough for binding pose and affinity: three-dimensional structure has to be modeled directly.

And in reasoning, recent evaluations show a significant and systematic drop when logical validity conflicts with the model’s stored world knowledge, meaning the system can sound coherent while losing the hidden structure of the problem. Different domain, same wound: the parts can look right while the structure is already gone.

Now the Main Question: How Fast Is AI Improving Before It Starts Beating Humans?

Now we arrive at the question that matters most.

How much time do we still have, and in which skills, before this evolution of AI begins to beat us decisively? Can we trust those deep hallucinations to remain, for a meaningful stretch of time, the Achilles’ heel of current AI? Or are we entering a darker phase, one in which the surface keeps improving at high speed while the deeper failures not only remain, but grow more dangerous as more coding-dependent and computation-heavy skills fall to the machine?

Look at the figure below.

What it shows is the split we have been building toward through the whole article. On the left, the industry’s benchmark story: rapid, almost intoxicating improvement, driven by tasks that reward computation, symbolic chaining, coding, retrieval, and all the other forms of work tied to surface hallucinations. There the rise has been spectacular. In some stretches, the improvement between successive model releases has looked almost exponential.

But on the right, once the harder constraints enter the picture, the shape of the story changes. The improvement is still real. It does not disappear. But it begins to bend, then slow, then brake. The cyan curve no longer races upward in the same way because the machine is no longer moving through problems that yield cleanly to code, search, and brute-force computation alone. It is pushing into a zone where geometry, invariance, deeper structure, and representational limits begin to matter

The Fantasy of AI vs. the Reality of AI Once Constraints and Maths Are Introduced. In the left plot, the optimism of the AI industry relies only on throwing more computational power at AI to solve bigger and more complex problems. The reality on the right shows that, once constraints enter the picture, the story changes. Geometry matters, because current AI operates in flat geometry. Budgets matter too, because each additional unit of computation yields diminishing returns in problem solving while, at the same time, increasing the risk of deeper hallucinations and raising the cost the industry must invest.


And this is where the paradox becomes harder to ignore.

The latest models are tearing through more and more surface problems at astonishing speed. But that does not mean they are advancing toward human mastery in one clean line. It means they are rapidly cornering the skills that depend mostly on computation while moving into a harsher region, one where deep hallucinations stop looking like a minor defect and begin to behave like a structural tax on further progress.

So the real question is not whether AI is improving. Of course it is. The real question is where it is still accelerating, where it is beginning to slow, and whether that slowing leaves enough time for human professionals to shift their value into the zones the machine still cannot cross cleanly.

That is what this figure is really showing.

Not whether progress exists, but whether its shape is the one the industry keeps selling.

And if the brutality of that speed is still hard to feel in a static plot, then look at the animation that follows. It stages the same split in motion. First, the triumphant fantasy the industry tells itself: the LLMs keep rising, the curve keeps climbing, the future looks like a clean exponential ascent.

Then the other face appears. The moment real-world constraints enter, the curve starts to bend, and it bends fast. The industry’s clean exponential was built on problems that flatten well into code, search, and brute-force computation. But once the problems stop living in flat solution spaces, once geometry, topology, invariance, and structural constraints enter, the ascent begins to deform. The numbers may still rise. The real progress does not.

That is the part the press releases never show you.

See it for yourself.

The Geometry Tax: What the AI Industry’s Ruler Hides.The industry says GPT-5.4 Pro is already at 8.8 out of 10, only two ticks away from the hardest open problems in mathematics. Then the ruler warps. The true score is 7.34. The gap to the top is not closing. It is 5.1 times wider than the benchmarks claim. Three providers, three colors, one wall: OpenAI, Google, and Anthropic all sink together because they share the same architectural limit. The arrows on the right make the point brutally clear. By d=15, the rate of progress is down to about a third of today’s pace, and still falling. The scores keep climbing. The progress does not.

Last Words That Won’t Last That Long

We still have a little time to rework ourselves, to shift, to adapt, to use AI before AI finishes using us. But let us not fool ourselves about this window. This is not safety. This is added time.

And if you have read my previous analyses, then you already know the truly uncomfortable part. The only serious professional advantage we still retain over current AI is not some elevated proof of human uniqueness. It comes from something much less flattering and much more temporary: a defect in the machine itself, a mathematical defect in the geometry of its representations.

Deep hallucinations are not inevitable. They arise from the wrong geometry.

That is why this matters so much.

Because what is produced by mathematics can also be attacked by mathematics. The same deeper failures that still give us room to breathe can, in principle, be reduced by changing the representational geometry itself. In the framework I have been arguing for, that means replacing the current flat regime with a space that can actually carry recurrence, hierarchy, cyclic return, and structural memory, instead of flattening them, losing them, and trying to reconstruct them afterward through brute statistical force.

Call it toroidal-solenoid geometry if you want the formal name. The point is simpler. Give the machine the right geometric glasses, and many of the failures that still slow it today stop being inevitable.

And then the question becomes uglier.

It is no longer the old vague question of whether some uncertain future AI might one day become dangerous. No. The question becomes much colder and much more technical: how long before the current machine gets the geometry it is still missing?

The real question is no longer whether intelligence will somehow appear all at once in some dramatic science-fiction leap. The real question is whether the current architecture will receive the geometrical correction it is still missing, a change in the representational space that would remove many of the failures that still slow it down today.

If that arrives, then what still looks like a ceiling today starts looking more like a door.

So yes, use the time. Shift your skills. Move away from the computational crust the machine is already eating. Move toward structure, invariance, architecture, coherence, causality, geometry, and the difficult cases where the parts can all look right while the whole is wrong. That is where the gap still lives.

But do not confuse that gap with permanence. It is only a delay.

Data sources and methodology: Epoch AI (FrontierMath benchmarks), MathArena (USAMO 2026), Stanford AI Index 2025 (inference cost trends), arXiv 2602.03716 (AxiomProver), SoftwareSeni (inference market data)

Liked Liked