Pengembang simulator instruksi mengemudi berbasis AI ini menghabiskan $1.000 per bulan hanya untuk menjaga demo tetap berjalan

One of the many ways AI will (supposedly) revolutionise the games industry in facilitating highly reactive NPCs via LLM-powered chatbots. Yet so far, only a handful of games like Where Winds Meet have attempted to implement the tech, and the results have been less than spectacular. But there might also be another reason why relatively few games feature LLM-powered NPCs, which is that powering such chatbots can be costly, as recently demonstrated by the developers of an upcoming driving instructor simulator

As reported by FRVR (via Kotaku) Teach My Little Sister How To Drive is a comic simulation game in which you coach your virtual sibling on driving technique from the comfort of the passenger seat. The twist is that behind your sister's dinner-plate eyes is an LLM chatbot (currently powered by an OpenAI model, though the developers intend to use Gemini for the final release) In the role of instructor, players can both converse with the chatbot and issue vocal instructions such as "turn left here", to which the bot will respond accordingly. At least, that's the theory.

Developer Easy Fox released a demo for Teach My Little Sister How To Drive back in February, but in the past month the game has seen a surge in players, which has had an interesting knock-on effect for the demo. "The number of demo players has grown more than twentyfold," Easy Fox explained in a Steam post, adding. "This growth has also increased our operating costs considerably."

Easy Fox goes on to explain that it is "spending over $1,000 per day on AI services for the demo", as players burn through tokens prompting the LLM by proxy. Easy Fox claims that figure "isn't because each player is particularly expensive to support", but rather that "so many people are playing."

As a result of this, Easy Fox has "taken out a bank loan to help keep the demo running". Yet even this may not be sufficient to support the demo long term. "Financial pressure may mean we have to close the demo earlier than expected before the full release," Easy Fox says, though the studio stresses it has "not made that decision at this time".

According to FRVR, that twentyfold increase in players amounts to 450 concurrent. While the overall player numbers will be higher than that, it isn't an enormous figure, and it makes me wonder how financially viable the game will be when it is released. Easy Fox says it will "account for expected AI usage in the game's price". But it also says that players "will not be charged separately for their individual AI token usage after purchasing the game".

That might work in the short term, but surely those token costs are going to pile up over time. It's also worth noting that AI token costs are highly unpredictable, with even large companies like Microsoft and Uber struggling to accurately predict the cost and burn rate of their own internal token use. In its Steam post, Easy Fox says it's looking into options like introducing local AI models for players with "sufficiently capable GPUs". But that's only going to be a fraction of the player-base.

All of this is aside from whether an AI-powered driving instruction game is a good idea. Personally, I think the fact that it revolves specifically around chatting with another character makes this a more sensible application than in Where Winds Meet, where the chatbots don't factor into the game design in any meaningful way. It's much more reminiscent of early chatbot-based social sim Façade.

But the very fact it reminds me of Façade makes me wonder whether the game needs a premium LLM connected to it to achieve its goal. Easy Fox states that current design problems being addressed include "dialogue that feels too artificial and breaks immersion" and "commands being carried out incorrectly, such as turning right when asked to turn left." All of which seems like it could be solved with a good voice actor and a sufficiently rich dialogue system. Easy Fox plans to release the full game in January next year, so we'll find out exactly how the finances stack up then.