hakesson

I like the idea - but I have two concerns.

Often an AI reads info that might be outdated and then assumes it is right. It doesn't really have the concept of "this is old info, maybe I shouldn't trust it". How about creating a hook for when an agent reads one of the journal files that checks the file mtime and outputs a "this content is x days old, be sure to verify it is not outdated" as additional info.

Secondly, I don't think an AI agent has any way to "skim" content. I mean it either reads all or parts of the content or it greps it. Why not do some index.md file that lists all the journal files together with a short description (inspired by the Open Knowledge Format)?

I have a friend who was drugged (date-rape style drug) and convinced to empty his account in an ATM. Woke up in a hospital without any memory of it. Some EU law forces banks to reimburse you when you didn't get the money of your free will - the LLM informed him about it and helped him to fill in all the forms etc and go through the full procedure, which included appealing the bank's negation to the "banco de España". He got his money. I think there are many cases where the LLM can help people that would never have the money to pay lawyers.

I've had to live with shipping software with really poor quality that had loads of small bugs. My Product Manager forced me to. However, at the end, the customers seemed happy, and I have to admit that I was partially wrong.

The QA guys flushed out the big bugs and many of the small ones. Without any proof, I imagine something like 99% of the customers would hit only a few bugs, that they are used to compensate for (reloading pages etc), leading to a view of the product as having good quality. I was wrong: the sw quality is poor, but the percieved quality is good. I would of course prefer to do sw of much better quality.

For the "harder tech" part: I played around with a "vibe code" project to see how far it gets. It's a cargo workspace linter that does things like find unused feature flags, unused pub, etc. Guiding it only on high-level feature-set requirements, it ended up doing workspace-wide lints backed by rustc's name resolution, not a parser. Something I would never have done myself. I picked that project because a missed lint costs nothing - I still wouldn't work this way on anything where a wrong answer hurts - yet.

Yeah, I'm a great fan of both HTML and CSS "can do that". I find more useful things on the CSS side like :has, keyframes and container queries. Why duplicate code when you have it for free :)

How do you detect AI slop today? I mean some 6 months ago it wasn't that hard but today, especially if you give it some simple instructions to be brief and honest, I can't really see the difference... And I am someone that spends my entire day communicating with LLMs as a developer.