The Biggest Impact of AI Won't Come From Smarter Models

I have a theory. The biggest impact of AI will come from removing waiting, more than from smarter models.

I've been thinking about this since reading about NVIDIA's RTX Spark, mostly because of what the hardware points toward: AI that lives next to your work instead of somewhere else entirely. That shift, when it fully arrives, will change things more than any benchmark improvement.

Here's why.

The waiting problem

If you're a heavy AI user, think honestly about how much of your day is spent waiting.

Waiting for generations to complete.
Waiting for uploads to finish.
Waiting because you've hit a token limit.
Waiting because the AI lost the thread of what you were building.
Waiting because you had to start a new chat and explain everything from scratch.

We don't talk about this enough. We've quietly built our entire AI workflow around these constraints and adapted our thinking to fit the tool. We've become part creator, part AI babysitter.

Compress this. Split that. Remove these files. Rewrite the prompt. Start over.

Your focus shifts from the work to fitting the work into the AI.

Tokenmaxxing: the hidden tax on AI productivity

I've started calling this tokenmaxxing.

Tokenmaxxing is what you do when the constraints of an AI system start shaping your creative decisions. You trim context that's still relevant because you're running out of space. You restructure a project to fit inside a limit rather than to think more clearly. You spend more energy managing the AI's memory than doing the work.

Some people call it prompt engineering, but prompt engineering sounds like a skill. Tokenmaxxing is a workaround, a tax on every interaction.

And the more complex your work, the heavier the tax, because context windows were never built for ongoing projects with real depth.

Somewhere along the way, we stopped asking the AI to help us think. We started asking it to help us survive its own limitations.

What happens when AI sits next to your work

The shift from cloud AI to local AI changes the relationship between you and the tool.

Right now, cloud AI is like having a brilliant assistant who gets amnesia at the end of every meeting. You re-brief them every time. You carry the context yourself. Your files, your projects, your history: none of it transfers. You start from zero on every session.

Local AI changes that. Your context and work history are always there, and the AI knows where you left off without you having to explain it again. It stops being something you log into and starts being something that runs alongside everything you do.

When you remove that friction, you think differently. You stop protecting your ideas from the overhead of re-explaining them, and you move faster. You follow threads you would have abandoned before because starting felt too expensive.

That's the tokenmaxxing tax, removed.

When generation is fast, the cost of experimentation collapses

Speed changes what's possible.

Right now, generation latency shapes creative decisions in ways most people don't notice. You think of an idea, wait for the output, assess it, iterate. That cycle has a cost. Slow generation means you protect your decisions. You invest in fewer directions because each one takes time.

When generation becomes fast enough, that changes completely.

Instead of exploring five ideas, you explore five hundred.
Instead of testing three directions, you test thirty.
Instead of protecting every decision, you start following curiosity.

Creation stops being the bottleneck, because it becomes cheap, almost free.

Which means the bottleneck shifts somewhere else entirely.

The new bottleneck: judgment

When AI can generate almost anything, anyone can make more things.

The real skill is knowing which things are worth keeping.

What's worth building.
What's worth pursuing.
What's worth ignoring.

For most of creative and knowledge work, the bottleneck has always been production. Getting the thing out of your head and into the world. AI has already started moving that bottleneck. Fast, local, always-on AI will finish the job.

And when production is no longer the constraint, judgment takes its place: the ability to look at a hundred technically correct, competent outputs and know which one is right for this moment, this audience, this purpose.

That's a human skill, and AI makes it more valuable.

Taste is the skill nobody is training for

The most valuable people in the AI era might be the people with the best taste.

That would put them ahead of the best prompters and the people with the most powerful hardware.

Taste is the ability to recognize signal in an ocean of noise, to spot the one great idea inside hundreds of average ones, and to know when something has soul and when it's technically impressive but emotionally empty. That conviction is why I built a music discovery platform with no algorithm in it. Every record is hand-picked and written about, because the curator is the part a model cannot replace.

I've believed this for a long time. AI just makes it more urgent.

When everyone has access to the same generation tools, the same models, the same compute, the differentiator will be who can recognize what's worth producing.

AI does not remove the need for taste. It exposes the lack of it.

Taste isn't a technical skill. You can't prompt for it or fine-tune your way to it. It comes from deep exposure to great work, from strong opinions held and revised over time, from caring enough to develop real standards.

That takes years, so the window to build it is now, before everyone else realizes it's the thing that matters.

So what actually becomes valuable?

When AI can generate almost anything, instantly, locally and without friction, which human skill is most valuable?

My answer is taste, the ability to decide what's worth keeping from an endless stream of capable output.

But I'm curious what you think. Is it taste? Creativity? Judgment? Decision-making under uncertainty? Something else entirely?

We spend a lot of time talking about what AI can do. We don't spend nearly enough time asking what humans should get better at.

What's your answer?