Even if it's "just" a text predictor, it already outperforms the average person in certain domains, particularly software engineering. With all the recent advances in agentic systems like Claude Code and OpenCLAW, these text predictors can iterate and debug faster than the average human. Looking ahead at the next decade, I totally agree with Sean's view here.

> I'm a daily user of the most recent Claude and while it's amazing at presenting other people's knowledge and reducing cognitive load.

'Presenting other people's knowledge' is enough to get the job done when that knowledge encompasses the entire internet.

it doesn't outperform on any code i ever threw at it. It added more bugs, and invented them when there were none

These are strong claims and I would want to see equally strong evidence for them.

My experience is that it's really darn good at producing text, but it's not a logic engine - it's not designed to be one and even the most recent versions make mistakes which indicate it's not actually thinking.

This is pretty strong evidence to me. https://simonwillison.net/2026/Mar/3/donald-knuth/

Well it’s not just a “text predictor” is it? You can pretend that today we still only have ChatGPT 2 and that there is only pre-training on a large corpus of information, but that simply isn’t true?

> Well it’s not just a “text predictor” is it?

It is, by definition, design and architecture a system that produces believable text.

Here's a task to give it which pulls the veil right off:

Ask it to add tests to a piece of code where code coverage is 100%, but it doesn't actually test functionality 100%. You'll start seeing nonsense sooner or later.

A leading theory in neuroscience is that human brains are fundamentally prediction machines too, constantly predicting sensory input, other people’s behavior, the next word in a sentence. “it’s just prediction” isn’t the gotcha you think it is. Prediction and attention turn out to be a surprisingly powerful foundation for intelligence.

The “just a text predictor” framing was fair a couple years ago but hasn’t kept up. Current models can genuinely identify untested edge cases even when coverage is 100%. You're definitely using the latest and greatest models?

The architecture started as next-token prediction, sure, and yes, human judgment is still required, but that judgment is being captured and integrated too. Every time millions of people use these models, their feedback feeds the next round of improvements.

Also, these models don’t need to replace your best engineers to be disruptive. They just need to outcompete the bottom of the bell curve. For a lot of junior-level work, we’re already getting close.

> You're definitely using the latest and greatest models?

Claude 4.6 opus high, specifically.

As for human brains: every self respecting neural networks 101 course is prefaced with "don't draw analogues to the human brain". And for good reason. Natural neural networks are fundamentally way more complex at every scale.

Also the brain indeed predicts, but also verifies and learns from the predictions. LLMs don't do that - not in real time at least.

You'll need to give it another metric to improve on.

I agree that the software industry should expect a major upheaval. Developers won’t be replaced by LLMs, but it’s extremely likely that software as a product will become less valuable. You can already vibe code solutions you would’ve otherwise had to pay for. As tools come out that take advantage of this, it’s going to just get easier to spin the app you want up instead of paying for someone else’s. Which is pretty cool if you aren’t already a software developer.

Yes, maybe in a decade all of the most popular LLM are dropped because lack of income. Only companies that learned how to use their own models and local optimized software using very specific hardware could still run them for very specific tasks