Id rather NOT use something that will share my personal details with questionable institutions, but deepseek won't remember anything I typed in my current session (and for the libs out there, no, I don't care if China has my personal details; it's not China that can and might use my personal details against me).
Id hate to craft a whole setting in shadowrun only to have to do it again next time. I've been working in Qatar for over a decade and wanted to craft a Qatar-based game (for which by the way, deepseek has gone above and beyond in crafting a shadowrun Qatar; they know about the local traditional attire and suggested shadowrun items themed around them without prompting, and locales they rewrote to reflect a shadowrun aesthetic, and it's calling the underground hacktivist resistance 'silk ghosts' (it does a silk road thing throughout)). I only suggested robots occupy most jobs and it crafted a Qatar where robots are 70% of the workforce while the locals are either execs or destitute, and the (in real life nationally owned) energy and oil company are now privately owned by 'Sustainitech'.
After this impressive world building all I can think is I really don't want craft all this from scratch every time.
I'm cool with being wrong about this, but unfortunately haven't found a solid source recently to indicate either way (maybe another day I'll have time to find one, but not today)
Unfortunately my comparison was obv apples to oranges, but id find it at least a little bit surprising if a continuously running and rendering process like a game would use the same amount of energy in the same amount of time as a user interacting with a chatbot that only uses a GPU on request. Yes, if a user was continuously feeding input in and generating output from a llm that seems trivially true that it would use more energy, but thats not how it would be used in this case.
Also, training is a fair point for end-to-end analysis but I don't even know how we would start that for total energy cost for developing a game, hence leaving that out of discussion. (Edit: I assume, because of all the attention on llms in recent years, that training an llm costs significantly more than developing a game (how ever we would fence off both measurements), but that is even more vibes based analysis than my original assumption earlier in the thread)
Yeah, the time spent between responses does reduce the energy consumption, but games don't use 100% of the GPU all the time either - they'll raise or lower the power they're pulling in response to how intensive the graphics currently are, and can be bottlenecked by the CPU. BG3 is kinda of a bad example game because despite the polish it's a CRPG, so most of the processing is directed towards NPC models and behaviours, which is handled by the CPU - the GPU is very unlikely to spin all the way up to max power like it will be the entire time it's processing an LM query.