Compilers are (generally speaking) pure functions - you get the same output for the same input. If it gets something wrong, it's getting it wrong for everyone given the same circumstances. They're fundamentally not comparable.

I think through aggressive nesting of AGENTS.md (or equivalent) files you could leave enough breadcrumbs (in a complex codebase) to assist an agent for easy issues. However, seeing how much context they burn you'd also have to break work up into microscopic tasks (burning even more tokens over time)

But what about less experienced Devs? They'll either accept blindly or not even catch the debts they're creating; and they'll have no incentive to learn about them if the code looks "good enough".

We need better and I'm not sure LLMs are _it_.

The tools can get smarter, but context windows are only so-long and only so-stable within that window. Furthermore (and more relevant to my concerns), they don't solve the problem of upskilling developers. It's all well and good having a senior guiding agents; they know the pitfalls of decisions.

Like, it's a godsend for throwing together new projects, but those come so rarely and the initial setup is never the time-sink. In the time Claude/OpenCode et al have fumbled their way to an incorrect solution for the mature codebase I've already opened a PR.

People are allowed to be fucked off by spam, especially if it's spam that has no actual purpose. Use your own rubric and move on without commenting if his response annoys you.

Browser interaction beyond just curl/wget would also be :chef-kiss:

To steal a term from AI-maximalists, tools are a "force multiplier", not a (work)force replacement. What makes me laugh is that these people would drop the relatively-speaking cheap labour to save money in getting the same productivity. Make it make sense.

This was my path as well. I've tried CC, and the Copilot cli - neither clicked as well as OpenCode. I don't think I'm happy with the results I'm getting with the model I'm testing with, but I want to push it to understand it before I compare/contrast. At least the UI/UX is on point.

Lorem ipsum forces you to replace text with first drafts that are intentional. Generating placeholders written by LLMs becomes the first draft. These are not the same.

Well, you get the LLM to give you a result from a well-known list of good results, and you validate its response against the list before storing it in the DB. I call this the "what I should have done in the first place + LLM to burn my company's money for no reason" algo.

Even if we wanted a free text field for convenience, we have Levenshtein and other algos that are stupid fast and capable especially small data sets. Throwing an LLM into the mix is like using a steamroller to push your lounge across the room

Be assured, if it was naked women they could goon to they'd be out in droves defending it as art. Ugly men with dongs on show? <I sleep>

Just Meta Things

3. Less optimistically, but still a happy ending:
An ethical company that needs a top security researcher to make their services robust hires Jon and puts his vast abilities to good use.

If you own/work for/know a company that fits this description, do us a solid and reach out. gainsec.com

I'd love a full minimal demo of this exact class in use without JavaScript - my understanding is that this aria value must necessarily be controlled with JavaScript. <details> provides an automatic "open" attribute, but not aria-expanded?