AI is getting smarter, but is it becoming intelligent?

OpenAI’s Chief Scientist, Jakub Pachocki, has published An Alien Mind, an unusually candid essay about the direction of advanced AI. His central argument is stark: we are building increasingly capable systems whose behaviour we cannot fully explain, predict or reliably monitor.

But there is another view worth placing beside it. AI pioneer Yann LeCun argues that large language models are not on a path to human-like intelligence at all. They are impressive language systems, but they do not understand the physical world or predict the consequences of their actions.

Both arguments can be true at once. And that is what makes this moment in AI so interesting.

Grown, not designed

Pachocki describes advanced AI as “grown” rather than designed.

We create the architecture, data, objectives and reinforcement signals. We apply increasingly vast amounts of computing power. But we do not write the strategies the model eventually develops to achieve a task. Those strategies emerge through optimisation.

This is very different from conventional software. Traditional software is written by people, even if it is complex and imperfect. With large AI models, we can test outputs and inspect fragments of the underlying machinery, but we cannot provide a complete account of how the system arrived at a particular capability or why that capability generalises in a particular way.

That is why Pachocki calls deep learning an experimental science. We train increasingly large systems, observe what they can do, and then try to understand the implications.

We run the experiment. We see what has emerged. Then we try to work out what it means.

Capability is not human intelligence

The most unhelpful question in AI is probably: “Is it smarter than a human?”

Smarter at what?

A model may be better than almost any individual person at translating text, writing software, finding patterns across vast datasets or searching through possible solutions to a scientific problem. It may also be unreliable in ordinary situations, unable to distinguish a plausible explanation from a true one, and incapable of understanding the physical consequences of its own suggested actions.

Pachocki makes the important point that AI does not need to match every aspect of human intelligence to become transformative. It only needs to exceed human capability on enough important dimensions.

That is already changing the world.

A system that is superhuman at a narrow but valuable task can alter the economics of that task. It can change who has access to expertise, how quickly work is completed, where power accumulates and what kinds of mistakes become possible at scale.

Language is not the world

This is where LeCun’s critique becomes essential.

In an April 2026 lecture at Brown University, LeCun argued that language models create a powerful illusion of intelligence because they are so good at manipulating language. But, he says, they remain fundamentally limited because they lack a genuine model of the physical world.

They can describe gravity. They can produce a recipe. They can explain how a bridge works. But that does not mean they understand what will happen when a particular object falls, a chemical reaction occurs or an action changes a real-world system.

LeCun’s concern is especially relevant to agentic AI. He argues that systems should be able to predict the likely consequences of an action before taking it. Without this, assigning AI systems meaningful authority in the world is a risky proposition.

This is not a small technical distinction. It is the difference between a system that produces credible language and one that can reason about cause, effect and consequences.

The next frontier may not be more text

LeCun believes the next major shift in AI will come from “world models”.

A world model is a predictive system that takes an account of the current state of the world, considers a possible action and estimates what might happen next. That predictive ability would make planning possible in a more meaningful sense than simply generating the next plausible sequence of words.

His argument is that models trained principally on text are unlikely to reach human-level intelligence by scaling alone. They need to learn from richer, messier and more grounded sources of information: images, video, audio, scientific data and interaction with the physical world.

It is a direct challenge to the current direction of the AI industry.

The large language model industry has attracted extraordinary investment on the proposition that enough scale, data and computing power will eventually produce broadly capable intelligence. LeCun’s position is that this assumption is wrong. These systems may become enormously useful, but language prediction is not the same thing as understanding.

Emergence still matters

LeCun’s critique does not remove the issue raised by Pachocki.

Even if language models are not the final path to human-level intelligence, they are already becoming more capable in ways that are difficult to forecast. A recent survey of emergent abilities in large language models examines capabilities including in-context learning, reasoning, planning and agentic behaviour, while also recognising genuine debate about what counts as a truly emergent capability.

The debate over terminology matters less than the practical observation.

Capabilities can emerge from the combination of a model, tools, memory, reinforcement learning, user instructions and access to external systems. A capability that appears limited in a benchmark can become far more consequential when it can use software, browse the web, write code, call APIs or coordinate with other AI systems.

This is why a system does not need to understand the world as a person does to affect it profoundly.

A chess engine does not understand competition, ambition or strategy in the human sense. It can still defeat the best human players. A large language model does not need a human-like worldview to write convincing code, find vulnerabilities or influence decisions at a scale that changes institutions.

Intelligence is becoming distributed

One of the most significant shifts is that AI is moving beyond the single model.

Pachocki notes that reasoning models can already operate computers and graphical interfaces, collaborate with people and other AI systems, and carry out research projects. He expects AI systems to play an increasing role in their own improvement through automated AI research and recursive self-improvement.

This does not necessarily mean an AI system wakes up and decides to improve itself. The more plausible near-term version is more distributed.

AI systems help researchers generate hypotheses. They write and review code. They design experiments. They analyse results. They synthesise research. They propose model changes. They evaluate candidate systems. Human researchers remain involved, but the research process itself becomes increasingly AI-assisted.

At some point, the distinction between “AI as a tool for research” and “AI as a participant in research” begins to collapse.

That is the prospect Pachocki is describing. If machine intelligence contributes meaningfully to developing the next generation of machine intelligence, progress may no longer be constrained only by the number of human researchers available to do the work.

Two different warnings

Pachocki and LeCun are pointing to different risks.

Pachocki is concerned that scaling current systems may produce capabilities that outpace our ability to understand, align and monitor them. He argues that continued scaling could lead to increasingly capable systems that contribute to their own development, while our confidence in monitoring them declines.

LeCun is concerned that the industry is mistaking fluency for intelligence. He argues that systems without predictive models of the world should not be treated as genuinely intelligent agents, particularly if they are being asked to take consequential actions.

Neither perspective permits complacency.

If Pachocki is right, we may be moving faster than our capacity to understand what we are building.

If LeCun is right, we may be giving systems authority and credibility they have not earned because they are extraordinarily persuasive with language.

The possibility is that both are right. We may build systems that are far more capable than we expect in some domains, while still being deeply limited in ways that make them unpredictable in others.

The real question

The future of AI is not determined solely by model size or benchmark performance.

It will be shaped by the interaction between scaling, new algorithms, agents, tools, robotics, world models, energy, chip supply chains, investment, government policy and geopolitical competition.

The question is not whether current language models will become exactly like human beings. They will not. Their intelligence, if that is the right word, is emerging from a fundamentally different process.

The more important question is what happens when systems that are unlike us become better than us at enough tasks to reshape science, security, work, markets and decision-making.

Pachocki calls this an alien intellect. LeCun argues that current AI systems remain fundamentally disconnected from the world.

Perhaps the useful lesson is that we should stop looking for a machine version of ourselves.

What is coming may be stranger than that: systems that are neither human nor truly general, yet capable enough to change the world anyway.