I’ve been asked a version of the same question often enough recently that it stopped feeling accidental.
It usually arrives gently. At the end of a meeting. In a message that begins with “quick one” and then quietly isn’t. Sometimes there’s a hint of apology attached, as if the person asking feels they ought to know the answer already.
They don’t ask what AI is. They ask where it’s up to.
Which is a different thing entirely.
So I thought it might be worth taking a moment. Imagine this less as an explainer and more as a pause. Perhaps over a pint, perhaps over a coffee, depending on the hour and your preferences. The point is simply to stop for long enough to look around and take stock.
Because if you’ve been busy running a company, doing real work, raising children, or generally trying to be a functioning adult, it would be entirely reasonable to feel that artificial intelligence has shifted shape while you weren’t watching.
It has.
What people usually mean when they say “AI”
A useful place to begin is with a small clarification that removes a surprising amount of background noise.
When most people talk about AI today, they are almost always talking about large language models — LLMs. Systems trained on vast quantities of text (and increasingly images, audio, video, and code) to predict what comes next in a sequence.
These are the systems behind the tools people now casually refer to by brand name. They’re fluent. They’re often useful. Occasionally they’re disconcerting. But they represent a particular strand of AI, not the whole thing.
AI as a field includes recommendation engines, optimisation systems, fraud detection, computer vision, and much else besides. But the cultural and economic tremor you keep feeling is coming from LLMs — and from what happens when you start building on top of them.
Even the phrase Artificial Intelligence begins to feel slightly off at this point. I’ve found it more helpful to think in terms of Alien Intelligence. Not because these systems are conscious or alive, but because they don’t think the way we do. They don’t reason step-by-step in a way we can easily inspect. They don’t arrive at answers through processes that feel familiar.
That difference matters.
The foundation beneath almost everything you see
Nearly all the AI tools most people interact with today sit on top of a small number of foundation models. These are very large, general-purpose systems built by organisations such as OpenAI, Google DeepMind, Anthropic, Meta, Mistral, and increasingly by Chinese labs as well.
This is worth noting, because it means we are not watching a single, tidy race toward some agreed destination. We’re watching multiple approaches unfold in parallel, shaped by different incentives, constraints, and political contexts.
The thing that still surprises many people is how these models are made.
They are not programmed in the traditional sense. They are grown.
Engineers don’t sit down and write rules about how language works or how reasoning should unfold. Instead, they build neural networks and expose them to extraordinary volumes of data. The network adjusts itself as it learns to reduce error. Over time, patterns emerge. Capabilities appear. Behaviours become recognisable.
Crucially, the system is not taught what is true. It learns what is statistically likely based on what it has seen.
That single detail explains a great deal of both the power and the strangeness of modern AI.
How behaviour emerges (and why it’s hard to explain)
Once intelligence emerges from training rather than instruction, something subtle but important changes.
You lose a clean line of causality.
No one — including the people who built the system — can point to a specific place and say, “this is why it decided that.” The reasons are distributed across millions or billions of parameters. Behaviour arises from interaction, not explicit design.
This is where the Alien Intelligence framing earns its keep. These systems behave less like machines executing instructions and more like non-human minds shaped by experience. That doesn’t make them mystical. It just makes them different.
In practice, their behaviour is shaped by three broad layers.
Pre-training is the upbringing. Vast amounts of data absorbed statistically. This is where biases, blind spots, and historical quirks enter — not because anyone chose them deliberately, but because they were present in the material the system learned from.
Post-training is the socialisation. Humans step in to shape behaviour: reinforcing what’s helpful, discouraging what’s harmful, adding guardrails and tone. This is why two products built on similar models can feel quite different, and why updates can sometimes feel like a change in personality.
Prompting is the interaction layer. A prompt doesn’t operate like a command in traditional software. It provides context. You are steering probabilities rather than issuing instructions. This is why wording matters, and why confident-sounding answers should not be mistaken for certainty.
Seen together, these layers explain most of what feels impressive — and most of what occasionally feels unsettling.
Why things feel different now
For a while, all of this sat comfortably in the category of “astonishing assistants.”
You asked a question. The system responded. You closed the tab.
That phase is drawing to a close.
The most significant shift underway now is the move toward agentic AI — systems that don’t just respond, but act.
An agent can pursue a goal over time. It can notice when circumstances change. It can decide that a follow-up action is required, use tools to carry that action out, and keep going.
The mental model here isn’t a chatbot. It’s a junior colleague. Someone you don’t give keystrokes to, but outcomes.
A useful way to see this in practice is through OpenClaw. People aren’t using it to have conversations. They’re using it as a do-things-for-me layer that sits quietly on top of their existing tools.
In everyday life, that might mean letting it handle email triage, draft replies, and surface only messages that genuinely need attention. Or connecting it to calendars and task lists so it automatically time-blocks the week, reshuffles meetings when plans change, and flags overload before it becomes obvious.
Some people use it to generate short morning briefings — weather, meetings, priorities — or weekly reviews that pull together notes and transcripts and gently highlight what stalled and what deserves focus next.
In work settings, agents sit inside Slack or Teams answering routine questions, summarising long documents, or monitoring systems for anomalies. In at least one case, an agent noticed a production issue during a conversation, investigated the underlying code, and submitted a fix.
None of these tasks are extraordinary on their own. What’s new is initiative.
Once a system can remember, observe, and act across tools, you stop prompting and start delegating. Oversight becomes something you negotiate, not something you assume. This is why the coming year is likely to feel qualitatively different from the last one.
Beyond language: why world models matter
There is another shift underway, quieter but deeper.
Language models are very good at talking about the world as humans describe it. A world model is something else entirely. It is an internal representation of how reality behaves — space, time, objects, cause and effect.
If I move this, what happens next? If I drop that, where does it land? If I plan this sequence of actions, what is likely to go wrong?
This matters enormously for robotics, where systems must reason about balance, friction, geometry, and constraint. But it also matters for planning, simulation, and decision-making. Language alone can sound persuasive while quietly losing track of reality. Systems with world models can rehearse futures internally and test alternatives before acting.
Many researchers believe that without world models, we don’t reach anything resembling robust general intelligence. Language can imitate understanding for a surprisingly long time. The physical world is less accommodating.
When you combine agents that can act with models that understand how the world responds, you move into a different phase altogether.
The forces shaping what comes next
All of this is unfolding within a context shaped as much by politics and capital as by technology.
Access to advanced compute — particularly NVIDIA chips — will determine who can train frontier models and who cannot. Reports that a future Trump administration may relax restrictions on selling those chips to China would materially alter the global balance and accelerate parallel paths toward advanced AI.
Capital markets are another variable. Both OpenAI and Anthropic are widely reported to be exploring IPOs. Public markets bring funding, but also pressure. That changes incentives, timelines, and tolerance for risk across the ecosystem.
Then there are more unusual alignments. The reported merger of xAI and SpaceX would integrate intelligence, compute, satellites, and physical infrastructure under one strategic roof. That hints at a future where AI is not merely software, but something woven directly into the physical world.
These aren’t background details. They shape how fast this moves, who controls it, and how unevenly its consequences are felt.
Where this leaves you
If you’ve had the sense that AI stopped being “just software” somewhere along the way, that instinct is sound. We’ve moved from tools to systems with initiative, from language to models of the world, from assistance to delegation.
You don’t need to master the technical details to engage with this moment. You just need a mental map that roughly matches the terrain you’re standing on.
If you still have questions after reading this, that’s not a failure of understanding. It’s a sign that you’re paying attention.
Which, at this point, is probably the most sensible place to start.