Is the AI Singularity already here?

For decades, the idea of the AI singularity has been framed as a dramatic future event: a moment when artificial intelligence crosses some clear threshold, exceeds human capability, and changes the trajectory of civilisation overnight.

That framing makes for good science fiction. It may also be the wrong mental model.

The singularity, if it comes, may not arrive as a single visible event. It may arrive gradually, through product workflows, evaluation loops, and systems that become better at improving themselves.

I have been thinking about this while working on some new prototypes for a better AI tutor.

Today’s LLMs are not yet great tutors

Large language models can already be impressive in conversation. They can explain, encourage, rephrase, translate, quiz, and simulate dialogue. In the classroom, that alone is enough to create the impression that we are already close to truly intelligent tutoring systems.

But in my view, most of today’s LLMs are still surprisingly poor tutors.

They are often:

• fluent, but not reliably pedagogical
• engaging, but inconsistent in sequencing knowledge
• helpful in the moment, but weak at building long-term mastery
• plausible, but not tightly optimised for measurable learning progress

This is a critical distinction.

A model that can produce a convincing answer is not the same as a system that can reliably improve student outcomes.

In education, the goal is not “generate a good-looking response.” The goal is to help a learner move from not understanding to understanding; from hesitation to confidence; from fragile recall to durable mastery.

That requires something much more demanding than generic conversational ability.

A genuinely strong AI tutor should be able to:

• infer what the student currently knows
• identify the most likely misconception
• choose the next best intervention
• adapt difficulty and pacing in real time
• track whether learning is actually occurring
• improve its own tutoring strategy based on evidence

That last point is where things get especially interesting.

The real frontier is not smarter answers. It is smarter feedback loops.

What I have been exploring is not just how to make an AI tutor sound more intelligent, but how to make it become better at tutoring over time by learning from outcomes.

That means shifting the optimisation target.

Most current AI products are still implicitly optimised around signals like:

• response quality
• human preference
• task completion
• latency
• cost
• user satisfaction

Those matter. But for an AI tutor, they are not enough.

A truly effective AI tutor should also be optimised around:

• student progress
• retention over time
• reduction in repeated mistakes
• ability to transfer knowledge into new contexts
• confidence calibrated to actual competence
• speed of movement toward mastery

In other words, the key question is not:

Did the model say something that sounded helpful?

It is:

Did the interaction improve learning, and can the system use that information to improve the next interaction?

That is a very different technical problem.

It implies a tutoring architecture built around continuous measurement, evaluation, and adaptation:

• instrument every learner interaction
• estimate knowledge state dynamically
• test intervention quality against downstream learning signals
• compare tutoring strategies across similar learner profiles
• update prompts, policies, or model behaviour using the results
• repeat continuously

This is not magic. It is an engineering problem.

But it is an engineering problem with unusually profound implications.

Because the moment a system starts improving itself based on evidence of how well it is teaching, it stops being a static software feature and starts becoming a compounding learning system.

This is where the singularity question becomes less theoretical

Back in the 1960s, mathematician I. J. Good described the possibility of an “ultraintelligent machine” that could design even better machines, leading to what he called an intelligence explosion. (ScienceDirect)

Later, Vernor Vinge popularised the modern idea of the technological singularity: a point at which we create greater-than-human intelligence and the old assumptions about the future stop holding. (edoras.sdsu.edu)

For years, these ideas were discussed mainly in abstract, philosophical, or speculative terms.

But today, the mechanics are becoming more concrete.

Modern AI development is increasingly shaped by evaluation systems. OpenAI’s own documentation makes clear that robust evals are central to building reliable LLM applications, because traditional software testing is not enough for variable generative systems. (OpenAI Developers)

At the same time, frontier labs are also publicly describing AI systems that help accelerate research itself. Google’s AI co-scientist, for example, is presented as a multi-agent system that helps scientists generate and refine hypotheses and research proposals. (research.google)

This does not mean we have reached a fully autonomous, runaway intelligence explosion.

We have not.

Humans are still designing the objectives, the infrastructure, the reward signals, the training pipelines, and the deployment constraints.

But the loop has undeniably changed.

It is no longer simply:

humans build AI

It is increasingly:

humans build systems in which AI helps evaluate, refine, and accelerate the creation of better AI-enabled systems

That matters.

Because once intelligence becomes part of the machinery that improves intelligence, even in a constrained and human-supervised way, the dynamic starts to look different from traditional software development.

In education, this may matter earlier than in many other fields

Education is one of the clearest domains in which this shift could become both visible and valuable.

Why?

Because education has something many other domains lack: a comparatively rich stream of measurable feedback.

A tutoring system can observe:

• how long a student hesitates
• which distractors they choose
• where they repeatedly fail
• when they answer correctly after scaffolding
• what they retain days later
• which explanations work for which learner profiles
• whether confidence and performance are aligned

That means an AI tutor can, in principle, be connected to a far better improvement loop than a generic chatbot.

If designed properly, it should be possible to build tutoring systems that:

• learn which interventions work best for different types of learners
• improve their sequencing of concepts
• detect unproductive struggle earlier
• personalise support more precisely
• become more effective not just globally, but at the level of individual learning pathways

This is the deeper opportunity behind AI in education.

Not merely replacing worksheets with chat.

Not merely making revision feel more interactive.

But creating instructional systems that can continuously get better at the thing we actually care about: helping students learn more effectively.

That is where AI in education becomes strategically different from AI as a generic interface layer.

A static digital textbook does not improve itself.

A conventional educational app improves only when a product team ships an update.

But a well-designed AI tutor could improve continuously:

• across cohorts
• across lesson types
• across misconceptions
• across languages
• across age groups
• across millions of micro-interactions

That is a qualitatively different type of product.

The most important design challenge is what we choose to optimise

Of course, none of this is automatic.

A system can optimise for the wrong thing just as easily as the right one.

If an AI tutor is optimised for engagement alone, it may become entertaining but pedagogically shallow.

If it is optimised for short-term correctness, it may over-help and reduce long-term retention.

If it is optimised for satisfaction, it may avoid the productive difficulty that real learning often requires.

So the central design question is not simply whether AI can improve itself.

It is:

What objective is it improving itself against?

This is why education is such an important test case for the future of AI.

If we can build systems that are explicitly optimised for genuine student progress, then we are not just making AI more useful. We are creating a model for outcome-driven, continuously improving intelligence.

And if we get that wrong, we risk building systems that look impressive while quietly optimising for the wrong metrics.

The future of AI tutoring will be shaped less by model size alone, and more by:

• the quality of the feedback loop
• the clarity of the learning objective
• the granularity of the evaluation system
• the alignment between optimisation and educational value

In other words: the most important breakthroughs may not come from bigger models, but from better learning architectures around the model.

So is the singularity already here?

If by “the singularity” we mean a sudden, fully autonomous, irreversible intelligence explosion, then no, not yet.

But if we mean the beginning of a world in which intelligence is becoming part of the loop that systematically improves intelligence, then I think there is a serious case that the earliest form of it may already be emerging.

Not as a cinematic event.

As a design pattern.

As eval loops. As optimisation systems. As AI-assisted research. As products that improve their own performance by learning from outcomes.

And in education, that pattern may become especially powerful.

Because an AI tutor that can measure whether a student is genuinely learning, and then use that signal to become a better tutor, is not just another chatbot feature.

It is the beginning of a new category of educational technology: software that compounds pedagogical effectiveness over time.

This is also the core design principle behind what we are building at Marvely. We are not interested in AI that merely automates marking or produces polished-sounding responses; we are interested in systems that give learners immediate, contextual feedback, encourage them to try again straight away, and improve through repeated closed-loop interaction over time. In language learning especially, that matters because students rarely get enough chances to practice speaking in a psychologically safe environment, and an AI speaking partner can create far more room for iteration, confidence, and self-motivated improvement.

That is the idea I keep coming back to.

The singularity, if it arrives, may not first appear as a machine announcing its superiority.

It may first appear as systems that quietly, continuously, and measurably get better at improving human capability.

And if that is true, one of the most important places to build it is education.

Zzish is the company behind Quizalize, Marvely and Blockerzz.