That’s a view from the perspective of utility, yeah. The downvotes here are likely also from a ethics standpoint, since most LLMs currently trained are doing so by using other peoples’ work without permission, all while using large amounts of water for cooling, and energy from our mostly coal-powered grid. This is also not mentioning the physical and emotional labor that many untrained workers are required to do when sifting through the datasets of these LLMs, removing unsavory data for extremely low wages.
A smaller, more specialized LLM could likely perform this same functionality with a much less training, on a more exclusive data set (probably only a couple of terabytes at its largest I’d wager), and would likely be small enough to run on most users’ computers after training. That’d be the more ethical version of this use case.
True, those are awful problems. The whole internet is suffering due to this, I constantly read about fedi instances being literally DDoS’ed by robots.txt ignoring, IP-block circumventing crawlers.
Unfortunately there’s no way to prevent any of this right now with our current set of technologies… the best thing we could do is make it a state- or even UN-level affair, reducing the amount of simultaneous training and focus on cooperation instead of competition while upholding high worker’s rights. However that would also be very anti-capitalistic and supposedly “stifle innovation” (as if that’s important in comparison to idk, our world burning?), so it won’t happen either.
Banning it completely of course is also impossible, humanity is still way too divided and its benefits for “defends” (against ourselves) too high.
In regards to running locally, we currently see a new wave of chip designs that’ll enable this on increasingly reasonable devices. I myself get a new laptop with XDNA2 chip and 32gb RAM today where I want to try running Codestral w/ Linux (the driver arrived natively in 6.14). Technically it should be possible to run any ~30b model on those newer chips (potentially slightly quantized), I’ll definitely try this a little bit and probably write a thread about it in the OpenSuse forums if you’re interested.
That’s a view from the perspective of utility, yeah. The downvotes here are likely also from a ethics standpoint, since most LLMs currently trained are doing so by using other peoples’ work without permission, all while using large amounts of water for cooling, and energy from our mostly coal-powered grid. This is also not mentioning the physical and emotional labor that many untrained workers are required to do when sifting through the datasets of these LLMs, removing unsavory data for extremely low wages.
A smaller, more specialized LLM could likely perform this same functionality with a much less training, on a more exclusive data set (probably only a couple of terabytes at its largest I’d wager), and would likely be small enough to run on most users’ computers after training. That’d be the more ethical version of this use case.
True, those are awful problems. The whole internet is suffering due to this, I constantly read about fedi instances being literally DDoS’ed by robots.txt ignoring, IP-block circumventing crawlers. Unfortunately there’s no way to prevent any of this right now with our current set of technologies… the best thing we could do is make it a state- or even UN-level affair, reducing the amount of simultaneous training and focus on cooperation instead of competition while upholding high worker’s rights. However that would also be very anti-capitalistic and supposedly “stifle innovation” (as if that’s important in comparison to idk, our world burning?), so it won’t happen either. Banning it completely of course is also impossible, humanity is still way too divided and its benefits for “defends” (against ourselves) too high.
In regards to running locally, we currently see a new wave of chip designs that’ll enable this on increasingly reasonable devices. I myself get a new laptop with XDNA2 chip and 32gb RAM today where I want to try running Codestral w/ Linux (the driver arrived natively in 6.14). Technically it should be possible to run any ~30b model on those newer chips (potentially slightly quantized), I’ll definitely try this a little bit and probably write a thread about it in the OpenSuse forums if you’re interested.