I think a huge number of devs will eventually have an itch to build their own coding agent that works just the way they want -- similar to web frameworks, bug trackers, etc. When a piece of tech becomes so central to your work, even small differences of opinion magnify.

There is some nuance. If someone used AI to help brainstorm and then thoroughly edited and refined the result to clearly communicate, it's fine. But when I see hallmarks of AI in writing I immediately pay less attention and start skimming because I assume it's bloviating.

I'm a dev but I'm way more annoyed by AI-generated writing than code. The purpose of code is its function - if the app works or the tests pass, I'm good. The purpose of writing is communicating - I don't want to spend time reading what you didn't spend time writing.

Imagining a sci-fi story where a far-future alien civilization tries to reconstruct what humanity was like based only on a thumb drive containing the weights to an LLM.

Tiny soapbox: "SWE-bench" is frustratingly misnamed as it consists *only* of Python code, and I genuinely think the name misleads people into thinking it's a general-purpose software engineering benchmark.

What I would much rather see is a $10M indie movie with the creative vision of an auteur but the production quality of a $250M blockbuster. I wish creatives would get over AI hate and use it as a tool and I wish suits would stop trying to turn the world into a slop factory.

Using AI to enhance and scale human creativity makes sense to me, but fully autonomous media makes no sense to me. It's like saying the ultimate restaurant would be one where you can order whatever you want and a mediocre chef will microwave it from frozen ingredients.

When I realized that LangChain will automatically create new tables and populate data into a Postgres database without any kind of flag or request to do so, my jaw hit the floor. WTF?. It feels like a cultural problem with "how things are done" in Python that I can't get over.

Python has a problem with black magic -- libs hide critical details behind black box behaviors. Downloading models is the most obvious one. Models are often multi-gigabyte files and Python libs just auto-download them without so much as a Y/n prompt. Deeply unserious practice.

It's fascinating to me how OnTrac exists given they are materially worse at delivering packages than every other provider. Consistently late, consistently not following delivery instructions, and this has been true in multiple places I've lived.

Prediction: most services will largely migrate to remote HTTP MCP servers, but developer-oriented services will continue to ship stdio MCP servers for the foreseeable future. There's just too much value in having access to the local filesystem and shell.

Popular JS project: this has been tested on all LTS versions of Node.js, Bun, Deno, and 3 browser runtimes. It works with ESM and CJS and all popular bundlers. Popular ML project: this has worked exactly once on my machine running on a Python version released in 2022. Good luck!.

One practice I learned in Go that I use everywhere now is to strongly prefer "table tests" where each test case is just a struct with various inputs and expected output states. It lowers the bar to adding additional tests as corner cases arise, and it's more readable.

Okay at least it's not just me:

This concludes my grumpy rant about JSON prompting. Now get off my lawn ya darn kids.

My suspicion is that "JSON prompting" is hacking the person prompting, not the LLM, into being more precise in defining the desired result. Nobody building those "AI apps that always work" uses a raw JSON prompt. Also AI apps that always work don't exist.

JSON prompting is a cargo cult. Saying "give me three five-word exercise descriptions as a JSON array" will be just as effective as a JSON prompt for same. You do need to be explicit about desired output format, number of examples, etc.

If you want actual advice, go to the labs that build the models and see what they say. Here's Google's:

This seems to be a growing "thing" in non technical circles and I just want to clarify that it has no basis in reality. A *small* JSON prompt is better than an awful prose prompt, sometimes. A *large* JSON prompt will nearly always be worse than prose.

Saddened to report that AI is no better at figuring out how to cleanly handle this merge conflict than I am. Long-lived branches are sometimes necessary but always painful.