On Thu, 20 Aug 2026 at 23:10, Matthias Geiger <[email protected]> wrote:
> On Thu, 20 Aug 2026 21:04, Aigars Mahinovs <[email protected]> wrote: > >That just tells me that you failed to either read or understand the rest > of > >the paragraph that you just qouted. > > > >BTW on the logic of some of this discussion, you using a computer and > >Internet to send this reply has used up some electrical energy and thus > you > >just contributed to the destruction of the environment. And so did > everyone > >else who sent or even read an email. Never mind the cryptographic > >computations needed to sign an email of a proposal or second. > > > >You know, a drop in a flood and so on. > > > >I disagree with usage of such reductionist arguments. > > > > > > This is false equivalency. If I compute by myself, that is vastly > different from a whole technology as LLMs taking 4% (!) of the USA's > total energy consumption. > This quotes *all* data centers together consuming 183 TWh in USA. For context, https://www.energystar.gov/products/recent_program_updates/low-no-cost-tips estimates that consumer electronics in USA homes consumed 143 TWh in 2017. Surely it is more by now. > All that for a slightly better autocomplete? Hardly worth the effort. > It's not a "better autocomplete". It is a pattern matcher. It looks at a text and detects patterns in it - double linked list here, factory pattern there, singleton here, regex for an IP address here, ... And then when generating text it can apply all the patterns it knows in adaptation to the context of the surrouinding text - like implement a sort that works *specifically* with your custom data types and very explicitly avoids accessing properties that cause expensive computations. > Those scrapers work neither ethical nor logical. For > instance, they scrape git repositories multiple times per minute (!), > not just once a week, straight-up ignore robots.txt, and run literal > botnets with residential IPs, making blocks hard. > That is a valid technical concern. Does not mean that LLMs are bad. Some of them are just used badly. I could do the same with Perl scripts, does not make Perl evil. I mean more than it already is. > Machine learning has its limited usage, such as breast cancer detection, > etc. Yeah, no. The usage is *massive* and we've only barely scratched the surface. If there *zero* further LLM tech development going forward, the existing systems already are quite revolutionary in their capabilities especially in coding space. Look at just a random example of what can be done rather simply with a relatively small, local LLM and tell me with a straight face that this is just a "better autocomplete": https://youtu.be/1EzVVj7DFPc -- Best regards, Aigars Mahinovs
