Choosing an AI Model Has Become Its Own Damn Project

AI is supposed to make it easier for me to build things, but lately I’m spending an aggravating amount of time deciding which AI to use before I even get started. At some point, choosing the right tool starts feeling like more work than the actual job.

Another Decision Before the Work Even Begins

My biggest annoyance with artificial intelligence isn’t the technology itself. I actually love the technology. What drives me absolutely nuts is the ridiculous number of models, reasoning levels, and execution options I have to sort through every time I want to accomplish something.

I open ChatGPT with a project in mind. Maybe I’m working on Bonumark Stream, improving my WordPress theme, investigating a technical problem, or developing another piece of software. I already know what I want to accomplish, but before I can get to work, there’s another decision waiting for me.

Which model should I use? What reasoning level does this particular task require? Should I use the most powerful option or something faster? Is the extra reasoning actually going to improve the result, or am I just burning through my available usage for no meaningful benefit?

And if I choose the faster or more efficient option, am I going to regret that decision halfway through the job? Will I spend twice as much time fixing its mistakes?

I love having options, and I definitely want control over the tools I’m using. But there’s a point where the process of choosing the right AI becomes its own damn project. I’m trying to start work, not hold a meeting with myself about the best way to start work.

I Love What AI Makes Possible

I don’t want this mistaken for another article about how AI sucks or how everything was better before these tools existed. That isn’t my position, and it would be pretty damn dishonest considering how much I use them.

I’ve built websites, developed software, improved publishing systems, and worked through complicated technical problems with AI assistance. Bonumark Stream, my self-hosted microblogging platform, is a real application that I use. The custom WordPress theme running JimLunsford.com is another example. Only Empowerment is a different kind of project, built around helping people understand and apply the philosophies I’ve written about.

These aren’t theoretical experiments where I ask an AI to generate something interesting and then forget about it. They’re projects I actively develop, test, maintain, and improve. The decisions matter because the results have to work outside the conversation.

AI has made an enormous difference in what I can accomplish, especially when I’m directing work that requires technical knowledge, careful judgment, and repeated testing. I want these systems to keep getting better, and I want access to the stronger capabilities when the job requires them.

That’s what makes the frustration so ridiculous. The better these tools become, the more options I seem to have to evaluate before I can use them. I can appreciate the improvements and still be annoyed by the amount of homework they create.

The Wrong Choice Can Cost More Than It Saves

When people compare AI models, the discussion often centers on speed, capability, and usage. I care about those things, but I care even more about what happens after I make the choice.

Suppose I’m working on a technical assignment that involves changing several files, preserving existing behavior, and running tests. I choose a lighter model to conserve usage. That sounds reasonable until the work comes back with a misunderstood requirement, an unrelated change, or a failure that should have been caught.

Now I have to figure out what went wrong, explain the problem, prepare another assignment, run the correction, and verify the result again. The initial run may have been cheaper, but the total job wasn’t. Saving usage on one execution means very little if I have to spend more of it cleaning up afterward.

Of course, going straight to the most powerful model at the highest reasoning level isn’t automatically the answer either. I don’t need an enormous reasoning budget to correct a paragraph, update a known configuration value, or perform a straightforward formatting task. Throwing maximum effort at every little job is its own kind of waste.

An important task isn’t necessarily a difficult reasoning task. If I already know exactly what needs to change, the source material is available, and the acceptance criteria are clear, much of the hard thinking has already happened. Compare that with investigating a production failure, evaluating a security boundary, or trying to understand why a technical change broke something else. Those jobs demand a different level of investigation.

Then there’s another complication: model capability and reasoning effort aren’t the same thing. Turning up the reasoning on one model doesn’t guarantee it will outperform a different model at a lower setting. So I can’t just pick a model and use the reasoning slider as a universal measure of quality.

Wonderful. Two decisions before the first damn instruction goes out.

What I want is the configuration that gets the work done correctly with the least total rework. If a stronger model finishes a job in one verified pass, it may be a better use of my allowance than a lighter one that needs multiple corrections. If a faster model handles something simple correctly, I have no reason to waste the stronger option on it.

And the result still has to be verified. A confident completion report doesn’t prove that a website works or that a software change preserved what it was supposed to preserve. I’ve written about the way I manage that in How I Actually Build Software With AI. The model helps carry out the work, but its confidence doesn’t determine whether I accept the result.

I Actually Created a Standard for Choosing AI Models

Here’s the part that makes me laugh at myself. I’ve reached the point where I have a written operating standard specifically for choosing which AI model and reasoning effort to use for my projects.

Think about how ridiculous that is. I use AI to build websites and software. Then I use AI to analyze the available AI models. Then I develop rules for deciding which model should handle a particular assignment. When new models or reasoning options appear, I have to revisit those rules.

I’ve essentially created an entire process to manage the process of choosing the tool that manages my projects.

Holy shit.

The funny thing is that the standard actually serves a purpose. It keeps me from treating every important project as an excuse to select the highest setting. It reminds me to look at the actual work, consider the cost of corrections, and judge models by their performance on my projects rather than by marketing or somebody else’s benchmark.

It also helps me stop reopening the same questions unnecessarily. I can make a reasonable choice for the next assignment, do the work, examine the evidence, and adjust when something actually justifies a change. That’s useful, especially when I’m moving between software, writing, website work, and technical operations.

But there’s something absurd about needing a formal operating standard just to decide how to start working with a product that’s supposed to make complicated things easier.

I’m glad I have the system. I’m annoyed that I need it.

Every New Model Reopens Old Questions

Just when I start getting comfortable with a particular model, another option appears. Maybe it’s an updated version of one I already use. Maybe it’s a model intended for a different kind of work, another reasoning level, or a new execution mode.

Now I have to reconsider assumptions I only recently finished evaluating. Is the new model better at coding? Can it follow a long assignment without losing the important boundaries? Is it better at writing, design, or research? Is it actually more reliable for my projects, or does it just look impressive on paper?

Those aren’t meaningless questions. I work on very different things, and what works well for a manuscript might not be the best choice for a repository change or a production investigation. I’ve also learned not to assume a new model is automatically better for my particular workload just because it’s newer.

I don’t want to become the guy refusing to try something better because he’s comfortable with what he knows. But I also don’t want every new release to trigger another full evaluation of how I work. There’s a difference between learning a useful new capability and constantly reorganizing your workflow around a menu that won’t stop changing.

At some point, keeping up with the tools starts competing with the time I have to actually use them. That’s the part I resent.

Give Me Control Without Making It Another Job

I don’t want AI companies to eliminate these choices. That would frustrate me just as much. I want to decide when a task deserves a stronger model, more reasoning, or additional usage. I’ve spent too much time building things to be comfortable giving up that control.

But there’s a difference between giving someone control and forcing them to constantly manage unnecessary complexity. A powerful tool should let me get into the details when they matter without making every ordinary assignment a miniature research project.

Give me a sensible starting point. Tell me why it fits the job. Show me the important tradeoffs, and let me override the recommendation when I have a reason. I don’t need the system making every decision for me, but I shouldn’t have to study every available configuration just to get moving.

And what about someone who isn’t building software? Maybe they want to organize a spreadsheet, write an article, research a question, or prepare something for work. They may have no interest in model families, reasoning levels, and execution modes. They shouldn’t need that knowledge just to get useful work done.

People still need to understand their own assignments, give clear direction, and evaluate the results. AI doesn’t eliminate judgment. But understanding the job you’re trying to accomplish and understanding every technical difference between AI configurations are two very different things.

I want control without unnecessary friction. That isn’t asking for less powerful technology. It’s asking for the power to become easier to use.

I Just Want to Build Things

I’ll keep learning and adapting because these tools are valuable enough to justify it. I’ll pay attention when a new model actually helps my work, and I’ll change my standards when the results give me a reason to. I don’t expect every choice to be made for me.

What I don’t want is to spend more time selecting the AI than defining the work. I don’t want to conserve usage so aggressively that I buy myself another round of corrections, and I don’t want to burn through the strongest settings just to avoid making a decision.

I open ChatGPT because I have something I want to accomplish. I want to write, build software, improve websites, solve problems, and turn ideas into things that actually work. The models are tools for doing that work, not the reason I’m there.

I love having options. I want the control. I appreciate what the technology makes possible. But holy shit, I shouldn’t need a doctorate in model selection before I can get started.

I just want to build things.

New Here?

Start Here: Why I Build This Way