Is it really feasible to run it 100% locally? I know there’s plenty of people with very powerful rigs indeed, but still.
It’s absolutely possible, and done more and more. Smaller models exist, and run on sub 8GB VRAM without any issue. If you have a gaming setup, you can run the bigger models without any issue locally.
You don’t need Astra or Mythos or whatever. As a daily driver for “light” tasks (e.g. summarizing a document) small models perform without a significant difference to frontier cloud models. For bigger tasks (e.g. coding or agentic workflows) you need a bit of beef on your graphics card, but it still runs on regular consumer hardware.
Anything above that is not needed for any normal use-cases.
If yes to the previous: the software doesn’t come from nowhere and ultimately still relies on gas-turbine-powered datacenters and stolen IP and stolen personal data, no?
This has multiple layers:
for running models locally, FOSS solutions exist and are no different to any other piece of software
models themselves are usually trained on beefy data centers and questionable data, but alternatives exist, both for training and curated data sources. Second but - those usually underperform, because LLMs need those massive data sets. Limiting the dataset limits the output quality drastically.
If what I wrote above is true, what exactly are people arguing when they say it’s still possible to use LLMs ethically or true to FOSS philosophy, because … ???
Also multiple answers:
…because first there is a massive amount of hardliners both in the pro- and anti-LLM camp, and differentiated and objective views are rare. Ethical software and “true FOSS” are also very heated topics currently.
…because, while possible, objectively non-ethical use of LLMs does happen more (public vibecoding of products, weaponized use in cyber and conventional warfare, use in surveillance, etc.).
…because there are absurd amounts of money circulating in the AI bubble, it’s too big to fail, and ethical and for-profit usually do not match.
It’s absolutely possible, and done more and more. Smaller models exist, and run on sub 8GB VRAM without any issue. If you have a gaming setup, you can run the bigger models without any issue locally. You don’t need Astra or Mythos or whatever. As a daily driver for “light” tasks (e.g. summarizing a document) small models perform without a significant difference to frontier cloud models. For bigger tasks (e.g. coding or agentic workflows) you need a bit of beef on your graphics card, but it still runs on regular consumer hardware. Anything above that is not needed for any normal use-cases.
This has multiple layers:
Also multiple answers: