If it were at least as capable as the 2060 12gb I bought a few years back for under $250, I'd probably try it out just for the novelty. But this looks like another crippled hand-me-down, dregs that the enterprise space doesn't want.
yeah I guess you're right, if you check the operations per second and the memory bandwidth a 4080 is around 10 times faster at each
I think what's going to happen is that Nvidia is going to eventually go with subpar fabs with high defect rates for raw output. Like the recently built ones in the U.S. Then those defective GPUs will just get binned for the consumer and tossed onto pcie cards or garbage like this.
That is the dream. AI is such a cool tool conceptually and even in practice given self-hostable tools like Open Claw and the like, and being able to fully disentangle it from the same issue that most other soy dev projects are faced with (see: Wayland, Mac, Rust etc.), is the cultist community. Once you can run your own powerbox that can help you code without creating a feedback loop to the globohomo giga models, I can easily forsee the advent of "clean" local models that can ape personal coding style and quality much better than current public models. Or, even better, to help with research and text synthesis without feeding said text to public models.
You can make a relatively decent one using SillyTavern and offloading persistent stuff to world books. Use the world books to keep information about the setting in them to keep the context low. The main issue is keeping track of inventory and stat calculations which LLMs have an issue with.
Couldn't have picked a better name.
you already see this with the image generation models, i've only dabbled in it because it really is mostly porn but it's orders of magnitude more advanced and specific and capable than what you get out of the huge monolithic models. if you're clever you can make stuff that people really won't be able to tell is AI. it's actually really funny that people are specifically trying to get the bizarre overcranked HDR look when it's a telltalle sign they're retards who paid for midjourney and are using what is essentially the first thing the AI cranked out
They're all based on the same basic idea of training a neural net on an enormous amount of data. We've only relatively recently gotten to a point with hardware where this is possible at the necessary scale. Doesn't matter if that data is text, images or audio, the same principle seems to apply. If you asked any AI scientist in 2017 when the transformer paper came out whether the results we've seen would be possible by just feeding that architecture the entire content of the Internet, they would be very skeptical.
Yea, small local models are getting much better. I'm using Gemma 4 and Qwen 3.6 for coding. And while I'm pretty sure they're not trained on a ton of C code they know enough to look at headers in the libraries in use and work stuff out. Where in the past it was just a giant model that knew all the stuff. Can I ask them about random facts, probably not, but for "Hey, this program is doing X, please fix it" they can work it out... slowly.
It's the same reason why Vibe Coding is sky-rocketing in usage as well.
