“I don't trust the AI”
The ork horde against human purists — power metal videoclip.
If you're a programmer who loves AI (or SI), I think you'll like this song.
For me, it's pretty funny.
If you hate AI or you're a doomer, maybe you won't find it funny.
Made with MiniMax H3, Qwen-Image2, Suno Hocuspocus and Opus.
(Spanish version in the next comment)
@DYoung_AI Same care as any dependency, and .... oh no, almost nobody does it, because it’s expensive.
Luckily that cost keeps dropping; soon enough compute that our agents can validate the skill before we trust it.
@sethkarten I still find multi-agent useful,keeping things separated helps a lot.
But I also think smarter models will eat a lot of that work themselves, so you won’t need the split as often.
@gsuarezduek Por ejemplo para grok bot.
Ve a esa tediosa web, y haz X.
hace unos meses le decia, me scrapeas las cosas... me las transformas a un script... hoy no, hoy ya le puedes decir que te lo haga en el navegador.
Tambien empiezo a usarlos para mantenimiento.
@aisearchio To me, this is expected. A new era of optimizations and performance boosts is coming.
I'm already doing it with my own tools, which are way easier to optimize, and to build optimized in the first place 😁.
And everything there was no time for before is now unlocked.
@Jason Welcome to the multi-bot club.
Once you give each one a job they start pinging you like coworkers.
I talk more with my grok bots than with humans
@Suhail I’d be quite surprised if there weren’t massive breaches due to the advancement of AI, both right now and for some time past. It’s just a matter of having the resources and impunity. Though if they aren't being detected, maybe it's because they aren't actually happening yet.
@omarsar0 Exactly. I never went deep into agent frameworks back then because it felt like high effort, low reward, and a massive token sink.
My bet was always: wait until models get smarter and tokens get cheaper, and the need for complex setups disappears. Looks like it paid off!
@rauchg This hits. I switched hard to the best model when it matters...
doing it once well beats five half-wrong loops. The under-guide misses still hurt more than the over-guide ones.
At this moment, Claude Sonnet 5.5 is enough, and Opus 5.5 is the key for more complex tasks. Codex, with its 5-hour limit, can't be used to its full potential when you hit a reset, and it's less capable (or dumber, to keep it informal).
Anyway, I love it too. The problem with this is... if you could have 100 smarter employees, why would you choose less intelligent ones even if they were already good enough?
@luccacerf I do the same, but I see I'm less ambitious. I use Godot and try to generate things with Grok Imagine or ChatGPT (2D), but I think I'm going to start integrating Meshy.
It's... awesome.