For a few tokens more

At my current job several of my colleagues use Cursor. I downloaded it but never got around to installing it.

I abandoned Visual Studio Code a long time ago and I wasn’t going back to a clone just to bring AI into my workflow. As I already told you, I was using Neovim, and I didn’t want to move from there, but the plugins didn’t convince me, and neither did the idea of a copilot writing code for me. I tried Avante for a while, and that’s when I configured it to use the API keys for ChatGPT and Claude.

And the worries began. Apart from the Avante plugin having several glitches, I kept running into rate limits, and also into the fact that the bill was starting to climb.

Maybe in the future the economy will be based on tokens as the unit of exchange, or processing units, or something that reflects the use of AI (this on the assumption that this technology will be central to what we do).

So I suspect we’re going to see alternative token markets, stolen API keys, or people sharing their Claude Code accounts the way Netflix accounts get shared.

It’s already happening, and in fact this week the people at Anthropic sent an email saying they’re going to put weekly quotas on their flagship product.

The token is the new dollar, at least for programmers.

I started out with Claude Code Pro, which costs about 20 dollars, and I already find myself experimenting with the Max versions. Because it’s like a drug.

But in my case, apart from how fun and challenging it is to deal with AI as an assistant, what interests me is investigating its impact on our profession, and measuring in the field how useful it really is.

For now, I think that if companies assign a budget of 100/200 dollars to their engineers to use these tools, they expect an impact.

Bottom line, we want to know what the ROI of AI is. Can it be known?

I have a few ideas for measuring it, and that’s partly the reason I’m doing this research.

Shut up and take my money

The Pro mode and the Max mode of Claude Code look like a suitable way to manage this expense at the company level. You know you can assign 1,200 to 2,400 dollars a year per engineer, which is something like fifty cents to a dollar extra per hour of work.

In the United States and some parts of Europe, the total cost to the company of a senior engineer is well above a hundred thousand dollars a year, so the extra cost to the company is under 1%.

In Latam the highest salaries don’t top 100,000 dollars on average; they’re probably closer to 80,000 for seniors, and down from there. I have no data, and I may be wrong; I haven’t done any research for this post either, this is just a back-of-the-napkin calculation.

Giving each engineer the Claude Code Max usage plan has an impact of 2.5%, assuming the numbers I just mentioned.

But if we spend that as a company, are we going to get a return that justifies the allocation?

What do you think?

Help me, AI, you’re my only hope

I’ll tell you about my own case. My current position is split between team leadership, support for the sales side, technical direction and product development. For that last part I have to develop software, but it isn’t 100% of my time; right now I’d say it’s between 60 and 70% of my time.

I decided to take on a medium-to-large project, and for weeks I made very little progress, at a turtle’s pace. I started using Claude Code and got to an average of three pull requests a day during the last week, which let me finish the project, late, but not as late as it could have been.

Unlike the personal projects, here there was a much larger Elixir codebase. This particular project was a new service that was about 70% developed, and I had to finish it. I took advantage of a new tool, Tidewave, which was developed by the Phoenix team itself, led by that beast José Valim.

Tidewave is an MCP that runs in your code. Here’s the demo so you can see what you can do with this tool:

With that new superpower I was able to improve and re-architect several things, and even rethink the operating flow the product would have.

And I’m going to keep using this technology at work. I think it helps me a lot and yes, it makes me more productive.

But careful: given that I have many years of experience, I understand how AI works, and I don’t swallow everything the snake-oil salesmen say, I’m not so sure the result will be the same for everyone.

I think you need a method for using AI, and I’m working on that, and that’s what I’ll talk about in the next post in this series.

As always, thank you for making it this far, and for sharing, and even better if you subscribe to my newsletter (in Spanish).