Stop over-thinking AI subscriptions

Yea I’d love to know what these look like. I’ve been using Claude and Claude code a bit for things I’m too lazy to do myself but know it’ll be solid at, for example a parser for iso dates in Janet. I set a $10 budget in February and have used $4 of it. I genuinely don’t know how to use these tools to spend $20 a month but at the same time they've been very helpful.

I similarly get confused about normal blocking. I genuinely don’t understand it. I get procrastination. I go for runs during work. But rarely is something genuinely so much that it’ll take hours of programming.

Edit: of course some tasks take multiple days. But visible progress is almost always achieved on the scale of hours not days.

I sat down at a Piano yesterday and tried to play a beautiful song even though I never really used a piano before. Sounded horrible, must be the pianos fault.

One reason to use LLMs is to understand when and how to use them properly. They're like an instrument you have to learn. Once you do that, they're like a bicycle for your mind.

LLMs are like a child with undiagnosed ADD, it is a constant roller coaster with peaks of amazement and and equally dips of disappointment.

As with any other tool, half the art of using LLMs productively is the ability to recognize when they're a good fit and when they're not. So long as AGI doesn't come around, there will always be cases where they fall on their face and as far as I'm concerned, we're pretty far from AGI.

Think how much energy would be saved if you could embed a sqlite database in a pdf with the data in it.

Try Mistral OCR. They trained their model to extract data from PDF. Used it to convert pdf forms to json schema. Worked excellently.

Most people don't realize that LLMs by design were not made for document processing, data extraction etc. For that, you would have to use a dedicated tool like Klippa DocHorizon, which built its own AI OCR from scratch. It also provides an API that you can use to send your documents and receive formatted data. It's less popular than, say, Textract or Tesseract, but it's far more accurate, especially if you're dealing with sensitive data that you wouldn't want an LLM to hallucinate.

Claude is not good for thus, try Mistral OCR. I'm using it on a daily basis on English and non-English texts and it works very well.

Did you try Claude to just extract it or to write a python script to do so?

If a task that normally takes X hours without AI now takes Y hours with AI and you charge $Z/hr, then the value of AI is X - Y * Z for this single task. If the value across all tasks is greater than the amount you are paying for AI, then it has “paid for itself”.

It sort of seems like the author is billing for the time saved and having a coffee or something. Seems a bit off.

Paying programmers by the hour always seemed weird for me because of the productivity spikes. I prefer paying by the job or just buying a weekly block that assumes someone works full time on my project.

I would be annoyed with a consultant who billed me for 4 hours when it was 30 minutes of work and an estimated savings from using AI.

I suggest to try Claude Code out and see for yourself. Solving problems is now cost fraction of the time. What took hours/days before, now might take MINUTES., it's insane.

That's why anyone suggesting AI won't take your jobs is delusional. We will need significantly less developer time.

It's a no brainer I agree.

What I hate is that I will not pay more than 20-50$ per month for side projects and that gap annoys me.

Would be cheaper in the long run maybe but heat, noise, electricity cost and the fact that the models are not as good as sonnet 4 and opus 4 are some things to consider.

I did the math, it just doesn't check out, especially if you take into account time to generate and electricity.

Even assuming you generate 100 tok/second, 24 hr, that's only 8 mil tokens. About $5 for Gemini 2.5 Flash or $50 for 2.5 Pro. But you will not be able to run it 24/7 locally - sleep, thinking between prompts, etc

Even at continuous usage (24 hr/day), it's still cheaper to just pay Google for Gemini 2.5 Flash tokens for half a year.

And this is NOT taking into account that mistral/devstral/qwen3 are vastly inferior models.

How can Google be so cheap? TPUs, batching multiple requests, economy of scale, loss-leader, ...

Running stuff locally makes sense for fun, as backup for loss of connectivity, if you want to try uncensored models, but it doesn't make sense economically.

Slower tokens, maintenance time (installing new models/updating tools), lower SWE bench scores. For me it’s not worth it - optimize for speed.

If I worked on a codebase I couldn’t send to the hosted tools I’d go this route in a heartbeat.

How do you connect that 3090 to a MacBook? How does that work when I’m not tethered to my desk?

I think if you’re comparing it to Claude max it’s ok, but the payback period on GitHub copilot for example is almost 7 years.

I haven't paid for the nicer models — now wondering if you could use less capable models to get some code up and running, then pay for better models just to "polish up" the code, get it to "final".

I tried running those locally via Ollama on my 3080 (latency doesn't matter) and the performance is abysmal. They're almost random token generators, literally unusable. My hallucination rate was near 100% for basic tasks like data extraction from markdown.

Coders that are consultants? I think that's fairly common isn't it?

No they're not. I prefer closed source LLMs because I'm not interested in having a small hardware park, but DeepSeek 0528 or Qwen 3 235 are really good models. Not as good as Claude 4, but absolutely good enough for a lot of development use cases. You just have to spend money on the right hardware.

But that's the whole issue, with those AI subscriptions you have to pay a subscription AND think about all that "backend stuff" (AKA not exhausting your message limit which is really low).