AF_ALG if I remember correctly predates userspace-accessible crypto acceleration and was way more important back when it meant you had actual need for "SSL accelerator" cards in servers, among other things
Yes, I remember that time, it was back when I wasn't allowed to know anything about what servers were doing other than to look it up in the internal leak, which was never maintained
Both the compiler (in absence of inclusion of copyrighted libraries) and the LLM are considered to not add creative work and thus do not change copyright status of the works they transform.
You can consider the training set of the LLM or other AI model to be 3rd party libraries and the level of copyright from them applying to final output to be how much can be directly considered derivative, just as reading copyrighted code and being inspired by it does not pass that copyright to your work unless it's obviously derivative
>> You can consider the training set of the LLM or other AI model to be 3rd party libraries ...
I like this comparison -- training set as '3rd party libraries'. Except, of course, that the authors behind the training set may not have actually granted permission to use, whereas the 3rd party libraries usually have some permission by way of license.
The law only cares about how the work is distributed - if you acquired it legally by purchasing, yes you can train LLM on it, and with exception of moral rights in places like EU the author does not have more to say on it.
It's treated the same as human reading and learning from the work.
You have only the granted artificial monopoly on acts of distribution under US law
Thing is, LLMs level of compression of training set mean that effectively, under the same rules that say you cannot sell that database filled with copyright material, the LLM is fine to sell. Because you have to be able to meaningfully trace each claim to final output (weights). For example, for some older stable diffusion model, it was calculated that each individual work addition or removal resulted in about 1-2 bits of change, meaning the same rules would qualify it as not derivative work.
However, because it is an issue with (at least historical) goals of copyright law, the common pattern that is evolving is that AI is not granted copyright of any work it generates, making it a bit of poison pill for some of the egregious ideas of corporate abuse. Not sure if the weights will be considered copyrightable either.
Copyright works on derivative rules - is the component of the work unmistakenly derived from another copyrighted work.
Under at least EU AI Act, any work done by AI is not granted copyright. But it does not mean copyright does not apply, it means the amount of work credited to AI is set at 0% (simplification). A human working off another's work unless it's perfect copy will have "credit" for changes that are judged creative/transformative, meaning a human plagiarizing something still can claim to have some degree of authorship. An AI won't.
In a sense, the copyright status of final work is a sort of "sum with dilution" were each work involved adds to claims, but AI's output is set at 0 - the prompt or further rework by human is not.
As for employer, details vary but generally "work for hire" rules and contracts do reassignment of material rights (in EU and some other places you can not reassign moral rights which are a different thing).
Gabe and Carmack are probably above Amelio and below Jobs and Gates in impact on the world - but probably above them all in impact when measured on a “desired” axis - people sought out Doom in a way that even the iPhone wasn’t.
I would also say that a lot of people, even people who are professional k8s operators, don't understand enough of the "theory" behind it. The "why and how", to put it shortly.
And the end result is often that you have two tribes that have totally incorrect idea of even what tools they are using themselves and how, and it's like you swapped them an intentionally wrong dictionary like in a Monthy Python sketch.
That part was really surprising to me because for the kind of compute lake he’s talking about building, k8s seems like a pretty good fit for the layer that sits just above it.
We run k8s with several VMs in a couple different cloud providers. I’d love it if I could forget about the VMs entirely.
Is there a simpler thing than k8s that gets you all that? Probably. But if you don’t use k8s, aren’t you doomed to reimplement half of it?
Like these things:
- Service discovery or ingress/routing (“what port was the auth service deployed on again?”)
- Declarative configuration across the board, including for scale-out
- Each service gets its own service account for interacting with external systems
- Blue/green deployments, readiness checks, health checks
- Strong auditing of what was deployed and mutated, when, and by whom
reply