Why I Remain an AI Optimist

Georges Seurat built his paintings one dot at a time. Thousands of tiny, precisely placed points of colour — each one deliberate, exact, and correct. Up close, they’re just dots. Step back and a cathedral emerges from a riverbank - a Sunday afternoon in a park. You’ve almost certainly seen one of his paintings without knowing his name or contemplating the technique.

What’s remarkable about Pointillism — and what makes it the best analogy I’ve found for how AI actually works — is the painter’s particular relationship with the picture they’re making. Intimately focused on the immediate. Increasingly blind to the whole.

The Dot, Not the Canvas

When I work with Anthropic’s Opus 4.6 on a development problem, it is extraordinarily good at the point we are working on. It works through the immediate problem carefully. It anticipates and identifies edge cases. It produces structured, coherent, confident-sounding output. The work at the point of focus is genuinely impressive.

But ask it to hold the whole canvas in view — to work across a large, complex, evolving codebase while tracking how each small decision affects the whole — and something shifts. It’s sometimes referred to as the “lost in the middle” problem. It loses the thread. It circles back. It recommends solutions that contradict earlier decisions it made itself, or in an extreme case in my experience, suggests fully delusional solutions deprecated over a decade ago.

To be clear, I’m describing my experience in software development where decisions compound and architectural coherence matters enormously. There are many sophisticated people out there doing genuinely impressive things with LLMs at scale: scanning hundreds of documents for patterns, validating assumptions across large bodies of research, surfacing connections that would take a human analyst weeks to find. Those are real capabilities that satisfy real problems, and I don’t want to dismiss the impact of generative AI’s contributions in those applications.

But I think there’s a meaningful distinction between processing broad context and comprehending it. LLMs are incredibly adroit working with specific details. They are good at evaluating massive amounts of data that humans couldn’t possibly process in a single sitting.

But whether they truly grasp how the dots form a realistic, representative image — which pattern matters and which is noise, which implication changes everything — is a different question. In my experience, as context broadens, the potential for confident misunderstanding grows with it.

What’s more, I don’t think that this is a bug - I am pretty sure it’s the nature of the technique.

What an LLM is actually doing, at a fundamental level, is statistical: producing outputs that are highly probable given the inputs. It is unnervingly good at this. The outputs are often so good that they’re easy to mistake for something else — for understanding, for judgment, for genuine comprehension.

But they are not that. They are very pretty dots.

Two Ways to Think About Where This Goes

There are essentially two framings for where AI capability heads from here, and they lead to very different futures.

The first — call it the pessimistic view from a humanist perspective — holds that with sufficient training data and computational power, LLMs will get good enough at simulating human judgment that the distinction stops mattering. Hundreds of millions of daily interactions are feeding these models constantly. The training data gets richer. The simulations get more convincing. Extrapolate two or three years and it’s easy to imagine models that have absorbed enough human reasoning to functionally replicate it.

The second — the optimistic view, from the same humanist perspective — holds that because LLMs are built on a foundation of statistical probability modelling, their reasoning can never grow beyond their probabilistic outputs. Each new generation of model may be more expansive and more capable than the last, but fundamentally it will be similar in kind. Like an infinitely expanding Mandelbrot set, every zoom reveals more seemingly-endless complexity, but all of it sharing the same fundamental character. More sophisticated self-validation is not the same thing as genuine understanding.

I find myself increasingly in the optimist camp.

The more I work with the latest LLMs, the more I believe that what improves is the simulation of authentically human responses — not the evidence of actual judgment behind them. Context handling has certainly improved and grown. Real-time retrieval of current information helps to fill in specific knowledge and training gaps. But at a fundamental level, there is no comprehension there, only a very convincing simulation of it. And that simulation remains prone to errors that genuine understanding would not make.

A Direct Challenge to the Optimist Position

I want to be honest about what shook my confidence in this framing.

In February, Matt Shumer — a six-year AI startup founder and investor who lives inside this technology daily — published a post called Something Big Is Happening that went viral for good reason. His central claim is that the latest models have crossed a threshold. He now describes what he wants built, walks away from his computer for four hours, and returns to find finished, production-ready work requiring no corrections. That something in the newest models feels, for the first time, like genuine judgment.

I read it and I’ll admit it shook me.

And then I went back to my desk, and my AI coding partner recommended a deprecated solution from a decade ago.

My experience has been that ambitious, autonomous, and expansive agentic workflows — even with meticulous prompt engineering, detailed architectural guardrails, and comprehensive success criteria — have been consistently less successful than guided, small-scale, collaborative work. The more I step back and let the AI drive independently, the more the canvas problem asserts itself. Not less.

This may be a function of my specific domain — SwiftUI development is architecturally unforgiving in ways that document analysis isn’t. It may be a function of the specific models we’re each using. It may be that Matt’s workflows are genuinely more sophisticated than mine. He may be a better developer than I am.

But it may also be that the experience of watching a capable AI “one-shot” a prototype makes it easy to mistake it for something more than it is. The dots, placed quickly and confidently across a large canvas, can look very much like comprehension — right up until the moment you notice what’s missing.

If the optimistic model is right — if LLMs will plateau (in the way that Moore’s Law has seemed to) at increasingly sophisticated probabilistic simulation rather than crossing into genuine judgment — then after a few more years of incremental improvement, I think the world will broadly recognize AI for what it actually is: a remarkable and transformative digital tool, in the same category as web search, but not the obsolescence of human expertise.

That’s not a dismissal. The world wide web, Google’s web search, and the iPhone all changed the world irrevocably. So will this.

An Old Fear in New Clothes

But I want to be careful not to let the optimistic framing become its own kind of complacency. Because the threat I find more genuinely concerning isn’t LLMs surpassing human judgment on their own. It’s LLMs serving as the instrument through which humans — or human-AI teams — build something that does. The path to AGI may not involve LLMs becoming AGI. It may run through LLMs dramatically accelerating the development of whatever comes next. A million developers, each working with an AI collaborator that compresses years of work into months, covering ground that would have taken generations.

That’s a different kind of risk. Less cinematic than a robot uprising. Harder to see coming.

People have been telling this story for a very long time. Prometheus stealing fire from the gods. Frankenstein’s creature turning on its maker. The golem that cannot be stopped once set in motion. The atomic bomb. The monster we build that destroys us is one of our oldest and most persistent fears — not because we keep imagining it, but because we keep getting closer to it. Every generation inherits the fear and a more powerful set of tools.

The conversation right now is swinging between two unhelpful poles. On one end, experienced AI technologists are literally cashing out and studying poetry with their fortunes, convinced that the end of human relevance is imminent. On the other, economists are reassuring us that the panic will subside and everything will normalize. Both responses are understandable. Neither is particularly useful.

The more honest position is this: the tools we have today are still just tools — extraordinary ones, but tools. The risk is real but it is a future risk, located not in what AI is now but in what AI might help us build next. Treating that risk as present and inevitable wastes the considerable value available right now. Ignoring it entirely because it isn’t here yet is how you end up surprised.

Why This Matters for How We Build

The painter who hands the brush to an assistant and walks away hasn’t made a painting - they’ve delegated one. It’s easy with AI tools to mistake their confidence and fluency as evidence that the whole canvas is in view and that nothing has been missed. And that is exactly when the canvas problem is most likely to occur.

Build with AI. Let it carry what it can carry. But pay careful attention to the big picture so that you can see what’s forming — and what isn’t.