Surprised the community hasn't found better ways to reduce context, 100k is insane. Is it possible to make the agent work on smaller pieces at a time, so less context is needed per piece?
NixOS is like the opposite of Bazzite. On Bazzite you don't configure anything, so no need for a git repo or anything. Just install the OS and start using it. (Bazzite also supports rollbacks btw).
Having to configure and manage every little thing, is like the opposite of the ethos behind Universal Blue distros like Bazzite.
some of my work projects have 80k tokens of just documentation
I feel like there's got to be better approaches than just shoving the entire documentation into the context. Humans don't try to memorize the entire documentation. We just search for relevant pieces and remember those, or have them open in a webpage alongside the code. We probably only have like 10k context maximum.
Do the modern agent harnesses not have automatic compaction?
Cab you help explain it? I thought that the sky was blue because it reflects the ocean and because of rayleigh scattering, but do either of these affect the color of plants?
Brilliant! The data seems to be updated monthly too. I'm surprised at how small some of the websites are, like ServerFault which is only 1.5 GB. StackOverflow is 100 GB but I have space for that.
Well if a fresh install fucks up, then that means the distro itself has an issue. It's very different from how updates fail on traditional systems, which are failures of the update system and not the distro itself.
In other words, any distro can be broken, or simply incompatible with your hardware. But atomic distros get rid of an entire class of failures, where the package manager's update system bugs out and breaks the OS.
I've had bazzite kill it self twice so far on an update.
I'm very curious how this happened since I've never heard of an atomic distros borking on update. I've heard of atomic distros failing to install because of hardware incompatibility, but usually after it installs successfully, it's smooth sailing after that.
People don't function like statistical models. They don't experience the universe in the same way, they don't learn in the same way, and they certainly don't create or interact with artistic media in the same way
Why is this relevant to ethics. Why do you need to learn like a human does, in order for it to be ethical. Why is it only exploitation when a machine does it.
It doesn't have the agency of a human being, nor does have the creativity or individual perspective of one. It has no senses of it's own or lived experiences to draw from, and it is not sentient. Furthermore, it doesn't interact with the world or economy like a human being, it does not create things like a human being, and if has no human rights or even copyrights.
We don't know enough about creativity and sentience and agency to know any of these things. And we definitely don't know whether they are relevant to the ethics of intellectual property. Most people don't even agree on whether intellectual property is ethical in the first place.
I'm starting to sound like a broken record but since nobody else has mentioned it yet: atomic/immutable distros often use something called rpm-ostree, which stages updates for the next boot. Updates are "atomic", so you can't end up with a corrupt partial update, even if you pull the power mid-update (in which case it will just stay on the current version). Everybody gets the same base image, so your OS won't diverge over time and end up like some frankenstein mess that. It's all super cool tech.
Surprised the community hasn't found better ways to reduce context, 100k is insane. Is it possible to make the agent work on smaller pieces at a time, so less context is needed per piece?