What the provider actually sells to you is GPU time / load (oversimplifying). Both tokens and subscriptions are just pretty arbitrary ways to price it, almost unrelated to the actual cost of running the model.
It's not the narrow view, it's the only actual definition of determinism. Words have meaning. Determinism means "same cause = same effect". If the cause is underspecified and requires interpretation that varies depending on the interpreter, the whole concept is meaningless. Intelligence (human or machine) operates in an underspecified domain, it makes sense to talk about undefined behavior, misinterpretation, alignment, anything really, but not determinism.
Random number generators are deterministic. If you give it the same seed, you get the same result. But that does not mean that a human can predict, for a new seed, what the result will be.
Most common things power users seem to want from TUIs on desktop:
- Easy efficient remote access. Pixel streaming and things like GTK Broadway are out. Needs a proper protocol/standard that would send styling-neutral primitives (and that is not X11 forwarding), and the entire toolkit architecture built around that. Something like SPICE/QXL but more semantics-aware? Not sure, might be too low level and chatty.
- Staying in the terminal. This is a matter of window management, though a toolkit can offer good defaults so this doesn't feel janky.
- Keyboard-first. Trivial. "GUI is for filthy mouse casuals" is a ludicrous idea.
- Low bloat. Web and things like GTK Broadway are out.
I haven't seen two similar opinions on actual architecture and stack, however being available in Rust will be a great start.
That is great until you have 12 similar windows open and your os window manager smooshes them into one indistinguishable stack. And if I want a window from app A to stick side by side with a corresponding window from app B you have few options. As opposed to tmux I can organize however I want and combine any arbitrary apps into a workspace. Plus RAM usage difference is night and day.
I like science fiction too, but unfortunately in real life security is only tangentially related to the tech being used, if at all, and "the powers that be" basically have a cyberweapon they want to gatekeep and use, unless open weights models eat their lunch.
Makeup and fingernails with pickaxes and minecarts. Yeah I totally believe it... Even those experimental archaeology influencers with an excavator working between the cuts were more genuine than that.
It would be pretty easy and also pretty unconvincing for anyone who did any manual labor in their life, let alone as backbreaking as mining with a pickaxe.
>I don't think Google or OpenAI use SynthID to include personal data.
Why? I assume by default that it's uniquely identifiable, because it just makes practical sense for OAI and Google (tracking misuse at the very least, but also data), it's trivial to implement and trivial to hide or plausibly deny.
I don't think it's trivial implement. Encoding more bits is harder or more brittle, and not trivial to hide. One could just ask for a completely white image. Any image distortions must then be due to SynthID. If the distortions differ for different accounts, that's strong evidence for varying watermarks.
reply