AI Fatigue and the Abundance Phase

I spent years turning coding into a skill: learning where a bug probably lives before opening the file, knowing which abstraction will hurt in six months, holding enough of a system in my head to change one thing without breaking three others. That skill now feels pointless. I can describe what I want and get most of it back, and the part of me that took pride in producing it by hand has nothing to do.

The same technology has made me faster than I have ever been. Things I would have scoped as a month of evenings now happen in an afternoon, and some of them happen at all instead of dying in the backlog because the activation energy was too high. Both of those are true for me every day and they do not resolve into a position.

So when people say they are tired of AI, I am not outside that. I feel it while riding the benefit. I want to write about the combination, because in what I read about AI fatigue the writer is usually standing on one side of it explaining the other side's error.

What the fatigue is made of

There is no single cause. It is a pile-up, and each part of the pile is reasonable on its own.

Start with the gap between progress and perceived benefit. Models keep getting better fast enough that I notice it month to month in work I know well. Most people who use AI as a chatbot notice any change at all. What they notice is that applications they already knew how to use have grown assistants they did not ask for, usually in the top right, usually with a sparkle on it. The company shipping it is measuring adoption. The user is measuring whether the thing they came to do got easier. From inside that experience, AI is a product decision being pushed at you by a lot of companies at once, and the pushing is the whole of the evidence.

Then jobs. This is the part where I have the least to offer and the most sympathy. Anthropic publishes research on AI's labor market impacts that tries to measure which tasks people actually hand to models and whether the pattern looks like automation or augmentation, which is the right question and a hard one to answer from usage data. I do not know the magnitude of what is coming. I do not think anyone does. But anyone whose employer has started asking which headcount the new tools make unnecessary does not need a magnitude estimate to be worried, and telling that person their fear is unquantified does not make it irrational.

Then the discourse. My feeds, the podcasts I listen to, the conferences I go to, the newsletters I subscribed to for other reasons, and most parties where someone finds out what I do. I like the technology and I am worn down by the volume of talk about it. Fatigue is a fair response to saturation, independent of the merits of the topic.

Then slop. Producing text and images and video got cheaper. Reading them did not. So the burden moved to the reader. Recruiter emails that were clearly generated, pull request descriptions longer than the diff, comment sections where you cannot tell which accounts belong to people, product pages assembled by a model that has never touched the product. Each instance is minor. Together they changed my default with anything that arrives: before reading it, I check whether a person bothered to write it.

And then skill degradation, which is the strand I feel personally. I read code less carefully than I used to. I catch myself asking for something rather than thinking about it for ninety seconds first, and I am fairly sure the ninety seconds was where the learning happened. Some projects have made the review cost explicit. Gentoo's council adopted a policy in April 2024 that expressly forbids contributing any content created with the assistance of natural language processing AI tools; the RFC that preceded it described the covered material as ebuilds, code, documentation, messages and bug reports. NetBSD's commit guidelines treat code generated by an LLM, naming Copilot and ChatGPT, as tainted code that cannot be committed to the source repository. Those rules read as unfriendly until you have been the person on the other end of a confident, plausible, wrong patch that takes an hour to disprove. The submitter spent two minutes. On the projects I have reviewed for, review time was the scarce resource.

None of these require anyone to be wrong about the technology. Someone can believe models are improving quickly and still be exhausted by all of the above. I do.

Two camps, both fine

Roughly, there are people who use the current tools seriously and people who refuse or ignore them, and they are working from different evidence.

If you use the models daily, you have a private record of progression. You remember the specific task that failed a year ago, failed with hand-holding six months ago, and now completes in one shot. That record is convincing, and it is the reason users tend to talk about the future with a confidence that sounds unhinged from outside. It is also unverifiable to anyone who has not run the same experiments. I cannot hand you my last two years of watching a model get better at a job I know well.

If you refuse, you are judging the technology by what is public: launch posts, keynote demos, ads, assistants inserted where nothing needed one, enthusiastic threads from people holding equity. My judgment is that most of that surface is marketing. Concluding from it that the whole thing is a bubble of talk is a defensible inference. It is what the evidence looks like when you have not run the experiments.

The asymmetry is that the users' evidence is real and private, and the refusers' evidence is real and public. Both groups are reasoning correctly from what they can see. My guess is that the capability is more real than the skeptics think and the current products are more noise than the enthusiasts admit.

My guess about the shape of it

Models will keep getting smarter, cheaper, and more efficient. Dario describes where that leads as a country of geniuses in a datacenter, millions of instances of something as capable as the best researchers in a field, running continuously, across every field at once. His later essay is more sober about the risks of getting there. I find the first one genuinely moving, and I think it is the most useful thing to hold in mind when you are tired of AI, because it is about biology, disease, mental health, and poverty rather than about productivity software. The argument is that a great deal of progress in medicine is bottlenecked on a small number of very smart people having time to think, and that this bottleneck is the kind a country of geniuses removes.

A deep learning model screening chemical space turned up halicin, a structurally novel antibiotic, in preclinical work published in Cell in 2020. Rentosertib, a TNIK inhibitor with an AI-discovered target and molecule, ran a Phase 2a trial in 71 patients with idiopathic pulmonary fibrosis and showed FVC of +98.4 mL against -62.3 mL for placebo, published in Nature Medicine in 2025. AlphaFold2 predicted structures now work about as well as experimental structures for prospectively discovering new ligands, per Lyu et al. in Science in 2025. These are small trials, early compounds, and narrow findings, which is also the shape the early part of a curve has.

The shape I expect is the gradual-then-sudden one. For a long stretch not much seems to happen from outside, because the results look like the three above, small enough that a reasonable person can file them under interesting and not important. Then several arrive close together, if the constraint really is the number of hours that highly trained people can spend thinking about a given problem and models add hours to that supply. I do not know how long the gradual part lasts. Among the people I read, the ones who use these systems daily guess short and the ones who do not guess never, and I have no way to score either group in advance.

The part I feel most confident about is smaller and nearer. The abundance phase we are in now, where consumer and workplace software gets an assistant regardless of fit, is temporary. It is what happens when a capability gets cheap faster than anyone works out where it belongs. My assumption: shipping an assistant is cheap, not shipping one looks like falling behind, and the sorting happens later, when the features nobody uses get quietly removed. Most of the fatigue I listed above is fatigue at this phase specifically, the pushing and the sparkles and the slop and the noise. That is the layer I expect to thin out.

What I expect underneath is more boring and more consequential. AI becomes a tool woven into ordinary work, the way version control or search is, mostly unremarked. Nobody has a position on whether they are a spreadsheet person. Meanwhile the capability keeps compounding in the places where it compounds, and the outputs arrive as specific things: a drug, a diagnosis that used to be missed, a protein structure. I expect AI to end up as infrastructure, with occasional discrete arrivals that matter enormously.