Humans Will Matter.

Agents will soon do over 99% of all tasks.

Anything with a clear objective function, anything you can verify, will be done by agents. Sooner or later. And it will be done better than any human can do it.

However, there are some things which agents cannot do -- or do significantly worse than humans -- and which I believe are fundamental limitations of the current AI paradigm. Things where agents struggle.

Things where Humans are better.

First off... let's look at how these models actually get good. Pre-training gives you a model that knows a lot. Reinforcement learning is what makes it useful. And RL only works when there is a verifier. Did the code compile? Did the tests pass? Is the math answer correct? Yes or no. The model tries a million times, gets a signal every time, and climbs.

The problem AI researchers run into, however, is... who verifies taste? Who verifies a good decision? Who verifies that the design is beautiful, or that this was the right moment to say the right thing to the right person?

Nobody can. There is no objective function.

This is not just a missing tool... A missing scraper, a missing connector, a missing API. That's a tool call, and it's one release away. Building your business on a missing tool call is betting against the models. You will lose.

What I'm talking about in this article is different. The problem is on the architecture level. The transformer is a next-token predictor trained on what already exists. It interpolates. It does not invent. It is not creative.

The first cluster of things Humans will be better at (for years to come) are things like judgment, taste, creativity, strategy, being practical.

Two weeks ago I asked Opus 5 to check the algorithm in my gun-license quiz app. Basically, the method that chooses which question to show me next. Opus told me to prioritize the questions I had already answered correctly twice. INSTEAD of telling me to focus on questions I struggled with, and got often wrong.

The average 12-year-old boy would give you better advice in that situation. However, a frontier model, Claude Opus 5, was being totally stupid and lacked basic sense of logic and reasoning. Fun fact, when I asked Grok 4.6 the same question it answered correctly.

This is not the only example... If you ask a human lawyer about some obscure law. He'll tell you, yeah, it exists, nobody enforces it, move on.

But if you ask an agent about this same law, it will tell you it's 1000% serious, and that you have to stop everything and refactor your entire business around it. It's not practical. It has no idea how the world works. It is terrible at high-level strategic thinking.

Social-type of tasks are another problem. If you have a good Human salesman (aka a Closer) on a sales call, and the client mentions some esoteric / unique problem that our business can easily solve, but is not in our standard offering, a great Human closer will probably throw that thing in as a bonus, and close the sale.

No AI model will ever get that idea. It would lack the judgement / decision making / creativity in that moment to think outside of the box. Most likely, this agent would stick to the standard product catalogue and tell the client we don't offer that thing.

The second big cluster where AI is a lot worse than Humans is LEARNING.

In the current paradigm of AI, a model is frozen the moment training has finished. Sonnet 5 will be Sonnet 5 forever. It will not go from a shitty, mediocre model, to being a great model two months from now. It's over.

BUT! The new developer on your team is not who he was 6 months ago. He's most likely a lot better. This is called continual learning. And it's a fundamental limitation of LLMs as of now.

Now, you might be thinking... David, if math/code is one of the things with a clear objective function, one of the things where AI can be (and already is) better than humans... why are we still hiring human developers?

Well, as someone who is actively hiring, I can answer this very clearly. It's not for the code. GPT-5.6 Sol Max writes better code than any of us. We pay for the proactiveness, product understanding, the taste, the judgment of what to ship and what to kill.

And, of course, for the likelihood that in a year, this person will be better than they are today.

It's not clear whether this really is a transformer problem, a problem on the architectural level of AI models, or a permanent limitation of silicon-based intelligence.

Maybe there's something in the human brain, in the DNA, in the biology, that makes us the ones who get the random creative thought.

Perhaps the robots will need Humans after all.