Upgrade your account to a Community Supporter account and remove most of the site ads.

Reply to thread

Message: <blockquote data-quote="Ruin Explorer" data-source="post: 9675756" data-attributes="member: 18">It's absolutely being looked into and will continue to be until someone tries it seriously and succeeds/fails (there have been small attempts already but rather unserious), but I think there are two major problems there:1) Current LLMs/chatbots aren't anywhere near good enough to do this yet outside a present-day setting (c.f. all "historical figure" chatbots and so on), they're just totally incapable of staying in-character and in-period. There's no reason to believe they're going to improve here in the next few years, either - though eventually I think people will get good at writing "guardrails" (the trouble with current models is it's all in the prompt, and it's kind of a black box beyond that, you need to access at a deeper level to put proper guardrails). An SLM might well actually a better job, if it was liked trained solely on more appropriate and focused material, it kinda couldn't deviate.2) LLMs are extremely expensive to run, and cannot be run locally - they have to "outsource" to huge datacenters. They also can't react in real time to complex inputs involving memory which you'd need for RPGs (Darth Vader is keeping it simple and snappy, just basically "reaction dialogue") - they can be fast, real fast, but not enough to be natural, and if the datacenters were in a specific country, not worldwide, this is going to be made worse. But the main issue is cost - in a subscription-based MMORPG, this might only bump up the monthly sub (and maybe not even by a lot), but with single-player games or other games that don't continuously charge? You, the publisher or developer, are eating significant ongoing costs, especially if you want fast responses, and especially if you're not the actual owner of the datacenters (meaning only Microsoft of all the gaming companies could get this reasonably cheaply for themselves). Success might even be a curse in this scenario - probably the ideal is people play your game once and forget about it, but if people loved it and kept replaying it and so on, your costs are going to be sky-high (assuming a typical usage-based contract - and I doubt you could get much else in this scenario unless you're Microsoft or Amazon, but Amazon keep shutting down their games division lol). I do think that an SLM, especially if they went less hard on the graphics and instead used modern GPUs (which have stuff that could be used for this) to process all the AI stuff might be able to respond fast enough to seem natural and wouldn't lead to ongoing costs. You'd also face the uphill battle of adoption - most gamers are pretty opposed to AI generated content in games, and with good reason so far. So you'd need a true "killer app" - and a lot of products/devices/approaches just never find that, or are still looking for it (c.f. VR - sims aren't it, Alyx wasn't it). I think there's no way we see an LLM-based game out from anyone but MS or Amazon (more likely MS). SLM I think could from almost anywhere, and I think because it doesn't rely on what is popularly regarded as stealing (which all extant LLMs do), and perhaps more importantly doesn't cost the company a lot of money over time, meaning it's much better for single-player games, and doesn't have to be shut down after a few years.</blockquote>

Verification