By Skye VernonTechnology has always elevated mediocrity. Cars make people think they are better drivers than they are. Team games give everybody enough ambiguity to believe they understand the strategy better than everyone else. Modern tools compensate for mistakes so effectively that people can confuse the capability of the tool with their own capability. AI is doing that to creation.
For a long time, crappy ideas had a natural filter: lack of skill. You could have a terrible app idea, but if you could not code and could not afford a developer, it stayed an idea. You could have a ridiculous game concept, but if you could not make games, nobody ever had to play it. Now AI fills that gap. Now you can make the crappy idea, and yes, we are going to get a fucking mountain of them.
This should not be controversial. Lower the cost of creating something and more things get created. We are already seeing people deploy absurd amounts of intelligence and compute to solve incredibly basic problems. Someone proudly tells you they had an advanced agent check the weather. Okay. My phone can check the weather. A useful agentic system would notice that tomorrow’s weather affects something I actually care about, determine what needs to change, handle whatever it has permission to handle, and bring me the exception it could not resolve.
Instead, I keep seeing these huge agent setups where one agent manages several sub-agents. One researches, one plans, one critiques, one verifies, and all of them pass information around, generate reports, maintain state, and deliberate over the answer. Then you find out what the system actually does. It checks the weather. That is not a flex. You have built a Rube Goldberg machine.
There are absolutely problems where agentic orchestration makes sense. Different processes can need different tools, permissions, context, or expertise. Breaking work apart can be powerful. But complexity is not capability. Every additional agent creates another handoff, another inference, another opportunity to misunderstand context, another thing to debug, more latency, and more compute. Sometimes the most sophisticated architecture is one model, one tool, done.
I think AI is temporarily making this worse because the machinery itself is still novel. People want to demonstrate that they can use agents, so they find problems for agents to solve. That is backwards. The problem should determine the architecture. My personal hierarchy for building things is brutally simple: does it fucking work, is it fucking useful, and is it pretty? Architecture does not even make the top three. Nobody using your product gives a shit how impressive the diagram behind it looks.
But crappy ideas are not new. The overengineered person with too much confidence is not new either. What changed is that the person can finally build their thing. Previously, technical difficulty acted as an accidental filter. It was not a particularly good one. It did not ask whether an idea was useful. It asked whether you knew how to code, had enough money to hire somebody, or could convince enough customers that the idea justified development.
Those questions killed terrible ideas. They also killed brilliant ones.
Giving someone powerful AI does not magically give them taste. It does not make them a good product designer or make them understand users. It does not make them good at identifying problems. It does not make their complicated architecture necessary. AI dramatically increases execution capacity without necessarily increasing judgment, which means mediocre ideas can suddenly become functional products.
We are going to get applications nobody asked for. We are going to get beautiful interfaces wrapped around nonexistent problems. We are going to get people proudly explaining architectures that could have been one function. We are going to get the twelve-agent weather checker. Fine. I think we have to get through that period because people need time to learn that being able to build something was never proof it deserved to be built.
Eventually, the novelty wears off. Nobody is impressed today because you made a website. The website either works, helps someone, looks good, or it does not. AI-made software is heading for the same reality. The question will become what did you make, does it fucking work, is it useful, and is it good? Once everybody has access to enormous execution capacity, execution capacity itself stops being the differentiator.
That is where I become much more optimistic than most of the AI discourse I see.
I keep seeing people actually make things. A guy builds an incredibly detailed app for tracking allergens in his area. Someone makes a companion app for kids. Someone makes little reading buddies. People make their own games. People make tools specifically for themselves. They are not necessarily building the next billion-dollar startup, and that is what makes this interesting.
The old software industry asks whether enough people can pay for something to justify building it. AI-assisted personal software can ask whether it would make my life better enough that I want to spend Saturday building it. That is an enormous change. For decades, most software available to ordinary people has been shaped by commercial incentives. The company needs growth, the app needs subscriptions, the platform needs engagement, and the product needs a sufficiently large addressable market.
So we adapt ourselves to whatever generalized product somebody decided was profitable enough to make. But what happens when I can make the thing I specifically want? Not because 400,000 other people need it. Because I need it. Maybe 30 other people need it too. Maybe nobody else does. Who cares? If it solves the problem, it has value.
That opens an enormous design space for niche tools, accessibility software, community tools, personal automation, experimental interfaces, mods, tiny games, and software designed around extremely specific lives. Things that commercial software companies would never bother creating because the market is too small or too weird or too personal to interest them.
Games might be one of the clearest examples. Modding already showed us what happens when people are given enough access to change games themselves. People rebuild interfaces, create campaigns, fix mechanics they hate, add characters, change difficulty, and create entirely new genres inside existing games. Generative AI pushes that accessibility much further.
“I wish this game did X” can increasingly become “Fine. I’ll make X.”
Someone who understands games but never learned programming can start experimenting with mechanics. An artist can prototype their own game. A writer can make an interactive world. A player can make the weird mod only they and 200 other people wanted. Some of it will be awful. Some of it will be fucking incredible.
The environmental argument needs the same nuance. There are legitimate questions surrounding AI. Energy consumption matters. Water use matters. Hardware manufacturing matters. Where data centers are built matters. Who pays for infrastructure matters. Those deserve serious scrutiny.
But I increasingly see a much weaker argument mixed into those legitimate concerns: AI is consuming all these resources to make things nobody needs. That is an extraordinarily strange standard to suddenly apply to computing. How much infrastructure supports streaming video, gaming, advertising, social media, cloud photo storage, cryptocurrency, and millions of hours of short-form video? How much compute is currently being used to host somebody’s useless opinion about how other people’s computing is useless?
If we are going to have a serious conversation about technological resource consumption, fine. Let’s have it. But “I personally do not value what somebody created” cannot suddenly become our measurement for whether computation was justified.
I do not think AI automatically gives us better ideas. It gives us more realized ideas. First, we get the slop. Then we get the overengineered slop. We get people who can finally build a bad idea, and people who can finally build an absurd system around their bad idea. Then the novelty wears off and people get better at asking whether the thing is actually useful, whether it works, who it benefits, and whether they are making life easier for the person who has to use it.
The same lowered barrier that makes the twelve-agent weather checker possible also makes the brilliant allergen tracker possible. It makes the reading companion, the accessibility tool, the weird game, the useful mod, the tiny community project, and the thing one person has needed for years but could not justify spending years learning an unrelated technical skill to create.
That is the part I am excited about. Not another corporation generating infinite personalized content. Not another agent architecture diagram. Not replacing every human activity with a chatbot. I mean millions of little inventors making weird tools, personal software, accessibility devices, games, mods, experiments, and things designed for communities too small for corporations to care about.
Things designed for one fucking person.
Technology has dramatically reduced the distance between “I wish this existed” and “I made it.” We are currently staring at all the garbage coming through that newly opened door. I think we are underestimating what else is about to walk through it.