stat_pg is a StatsD-inspired native PostgreSQL metrics ingestion plugin that supports a billion metrics per second on modern hardware, and millions per second on a Raspberry Pi.
It's in alpha. It's free, open-source, and MIT licensed.
Link in first comment.
Introducing Hill Climb
Inspired by @theo's hill climbing workflow, I've built this skill
It builds the right check for your task & makes your agents prove every step
You give it a goal it writes a test, locks it, and climbs in small steps. A bad step is undone
I noticed that my agents output quality got 100x better while increasing the success rate on complex & seemingly impossible tasks
@harleyfoote_@oldstackjournal Definitely, there's lots of work to do to convert nondeterministic things to deterministic scripts that the agents rely on. That strategy works well in practice so long as you can think of the script to do it.
Part of the reason I'm bullish on AI is to help people create solutions to their own issues.
Here's a few itches I've scratched in the last 90 days.
* stat_pg
* mechalinters / mecharuby
* https://t.co/vA3aoGm5L6 / xnk player
* kamal-watch
* stripped down, simplified CI system
... and more
If you think by putting "invariant" into a prompt that it's an ironclad guarantee, that's silly. That's no different than telling a coworker a set of steps/rules to follow - nobody is perfect, including the agents. You can't beat the consistency of a deterministic check.
There are tools to deterministically parse and evaluate code. Some things are less of a pass/fail, and more of a judgment call, but for the pass/fail items (the return value of a function, for example) adversarial reviews aren't enough.
@SAIHMemory@oldstackjournal It's not about token usage. It's about the quality (or lack thereof) of the guarantee you get from sending an agent to do a deterministic job.
@SAIHMemory@oldstackjournal I mean if all we're doing is throwing agents against agents, what are we doing other than spending credits faster? I get the approach of adversarial reviews but I just wind up feeling like I get two wrong opinions instead of one. Maybe I'm just doing it wrong.
@carlocapocasa@oldstackjournal Agree that all that tests can/have ever done is "nail the implementation in place" - TDD is a methodical system by which you write down your assumptions so you can be notified if any of those assumptions are violated. They don't prove that your assumptions were correct.
To clarify a bit, Ali is probably just trying to say that vibe coding on its own is not enough. And I would agree with that.
Sometimes, unintentionally, we discourage people before they've even tried. When that happens I think it's a shame.
I'm 100% for people trying.
Been coding for 30+ years, and have some of the highest standards of quality of amongst professionals that I know. Even so, I can't endorse this mindset.
I get it, but the next generation is coming, and they're going to be AI native. And that's okay.
Vibe coding is a generational scam that sells people the fantasy of bringing their own idea to life without knowing how to code. The reality is that production software requires real software engineering knowledge and practices