I don’t mean AI development should stop forever, or that today’s models are perfect. They absolutely aren’t. There’s still plenty to improve: reliability, hallucinations, efficiency, security, interpretability, memory, agent behavior, and a hundred other things.
But those improvements don’t necessarily require this constant race toward making AI dramatically more capable at everything.
At some point, I think we have to ask what exactly we’re racing toward.
We already have models that can write software, dig through huge amounts of information, generate images and video, research topics, use tools, tutor people, automate workflows, and handle some seriously complicated tasks.
But the question I keep asking is:
Do we really need AI to do everything for us?
We’re already struggling to keep up with how quickly this technology is moving. Companies are already figuring out which jobs can be automated and which parts of people’s work can be handed over to AI. Entry-level roles are changing. Entire industries are trying to work out which skills are still going to matter. Schools barely have time to decide how students should be using these tools before another generation of models shows up.
And there’s a physical cost to all of this too.
More capable AI usually means more demand for compute. More compute means more data centers, more electricity, more cooling, more water, and more land.
Those data centers aren’t floating around in some magical cloud. They’re built in actual communities, and the people living in those communities are the ones dealing with whatever comes with them.
That doesn’t mean data centers are automatically bad, or that AI shouldn’t exist. Obviously I don’t think that. It just means there’s more to this race than benchmark scores and flashy model launches.
And while all of that is happening, the attitude still seems to be:
Bigger. Smarter. Faster. More autonomous.
Why?
There’s another option.
We could spend some time getting really good at the AI we already have.
Make it cheaper. Make it more reliable. Make it easier to verify. Make it more energy-efficient. Build better tools around it. Figure out what reasonable rules around this stuff should actually look like. Help workers adapt. Learn where AI actually improves something and where we’re just shoving it in because we can.
I especially feel this in software development.
AI-assisted development is already insanely fast when the developer is still thinking, reviewing, testing, and making the decisions. I don’t need AI to replace that part.
I like making things.
I want a ridiculously capable assistant that lets me build more of them.
I don’t want the end goal to be pressing a button and having AI do absolutely everything for me.
Maybe progress doesn’t always have to mean removing more humans from the process.
And maybe when technology starts advancing faster than society can actually absorb the consequences, slowing down isn’t being anti-progress.
Maybe it’s just being responsible.
I don’t want weaker AI.
I just think we need to spend some time figuring out what the hell to do with the incredibly capable AI we already have before we keep flooring the accelerator.
Top comments (3)
The friction you are describing is the widening gap between raw model capability and organizational absorption capacity. Frontier labs are locked in a capital race where the primary asset is benchmark leadership, but for actual software teams, the bottleneck has shifted from model IQ to verification cost. When an agent produces code in two minutes, the human time spent auditing edge cases, security assumptions, and architectural fit does not compress at the same rate. Until workflows catch up with determinism, local routing, and better observability, dumping more raw parameters into the loop yields diminishing economic returns while driving infrastructure capex through the roof.
Yeah, this is pretty much what I’m getting at. Generating code is already incredibly fast, but that doesn’t mean reviewing it, testing it, and making sure it actually fits the project suddenly takes two minutes too. I think there’s a lot of room to improve the workflows around the models we already have instead of assuming the answer is always another jump in capability. I’m not against better models, I just don’t think bigger and more autonomous should automatically be the goal.
Some comments may only be visible to logged-in visitors. Sign in to view all comments.