Is the AI Singularity Concept a Mathematical Mirage?

There’s a curve in mathematics called an asymptote. It accelerates toward a line forever without ever reaching it, and the closer it gets the more effort each step requires. Nature is full of these patterns, really. Systems tend to grow fast when conditions are favourable, then slow down as friction builds in. Limits always assert themselves, because nature loves asymptotes.

I think AI progress might be behaving the same way.

The early breakthroughs in AI felt explosive because they were. Pattern recognition and language modelling were tractable problems once everything aligned, and they created the impression that intelligence itself was being “solved.” But look closely at the domains that fell first.

When Deep Blue beat Kasparov at Chess in the late 90s the press treated it as a turning point moment, man against machine, the future is here. AlphaGo did the same thing to Go world champ Lee Sedol a couple decades later and the same headlines ran. The interesting part is what happened next in each case, which was…very little. AI didn’t unlock infinite layers of either game. It approached optimal play within a bounded system and then improvement became marginal. Chess and Go are closed systems with defined end states, and once you’re near optimal inside those rules there isn’t much ceiling left. LLMs are another version of that. Vast training data with clean ways to measure outputs. The first layers to fall are always the ones most suited to the method. But that says more about what we attack first than how far the curve extends.

The last decade worked so well for AI because the conditions were unusually favourable. Compute scaled fast and cheap right when there was a huge backlog of human data sitting around to train on. The problems we attacked happened to suit the method, which is a separate piece of luck that doesn’t repeat just because we want it to.

What’s left next is messier. The remaining problems tend to have limited feedback and vague goals, and they sit in really subjective contexts. Throwing more compute at that helps less with each pass.

The data problem is real too. As models increasingly train on AI-generated content and recycled material, the signal gets thinner and thinner. Fresh high-quality data now requires deliberate human judgement and controlled environments to produce. Compute keeps improving, but the gains are mostly architectural now. Which means slower work that’s harder to predict. As AI moves into healthcare, law, finance and infrastructure, every layer of safety and governance adds real friction.

Human cognition is layered and emotional, shaped by experience in ways no benchmark catches. There isn’t a single dial to turn up. Systems can outperform humans in specific domains anyway, without understanding anything about what they’re doing.

GPT-4 scoring near the top of professional exams is impressive. But the result demonstrates pattern mastery inside a constrained environment, which is a very different kind of capability to consciousness. The distinction sounds philosophical, but in a boardroom it turns out to be one of the more practical things to understand.

In most boardrooms the conversation is about compliance, liability, and where the data lives. Machine consciousness doesn’t typically come up. Philosophy of mind makes for a great dinner conversation, but Procurement just wants to know if the vendor has SOC 2 certification.

If the singularity requires subjective intelligence (a system that experiences and integrates itself as a self) then more compute won’t get us there.

Successful organisations are looking for leverage, and for systems that scale judgement without scaling headcount. Sentient machines don’t factor in. The effort has moved to orchestration and the systems sitting around the models, which is where most of the real engineering work lives now.

I think the singularity, if we ever get there, will arrive looking very different to what was promised. It’ll be slower and more administrative and considerably less cinematic, and we’ll mostly only notice in retrospect.