LLMs write by picking which next word best correlates to a response from the prompt, so they tend to follow grammar extremely well but make logical mistakes.
Humans write by forming their thoughts into word sounds, then transcribing them, but a single pronunciation can have multiple spellings, depending on the context, but the context is somewhat abstracted by that point, so humans regularly write homophones or malapropisms of the appropriate word.
If bun and node.js run I'm sure the agents will, you might have to fight tool calls since the system utils differ from GNU core utils a tiny bit here and there, but you could toss the agent and whatever tools into a jail and have a nice package, use zfs snapshots between prompts so you can disect it later.
Some may call me a devil's advocate (I love FreeBSD as my first alternate OS, it's more sensible as a whole, and less chaotic then linux) it provided the network stack for Microsoft and a base for darwin (OSX) Maybe it's just the low brow art, maybe it's the root of evil and the label is on the tin, how cute!
With these big players who are regularly found supporting people with evil intentions: Don't attribute to incompitence what could be ascribed to malice, nay you must trust the gods of the clouds to keep your secrets for you, all for the low low price of $x.99 a month a seat, you may only cancel your service with an arcaine dance and the sacrifice of your first born!
The technology is irrelevant, at one point in my career I worked for a ~200 person co. that was aquried by another massive (20,000+ seat) company who was even less useful than Oracle; trying to figure out the rough mechanics of why such a useless tower of people is built and maintained, I realized the magic: big juicy government contracts!
reply