Patch Notes #7

AAAA games, AI, bugfixes and performance improvements

• In Patch Notes #2 I talked a little about how the rising cost of game production has not been matched by the rising cost of games themselves:

  Starcraft 1 cost $40 at launch in 1998, or $74 in 2024 dollars. Starcraft 2 cost $60 at launch in 2010, or about $85 in 2024 dollars. Diablo IV (which launched in 2023) has an MSRP of $70. Both older games were more expensive (in inflation-adjusted terms) than today’s games! And that isn’t even factoring the tremendous increase in team size needed to produce a AAA game that gamers expect today (one analysis shows roughly 4x from 2010 to 2024).

  I find it interesting to think about why gamers who played back in the 90’s have this $50-60 price anchored in their heads. My guess is that everything surrounding the games themselves has benefited from the deflationary aspect of technology: a gaming PC cost thousands of dollars in the 90’s, and it still does, but is significantly better. The same goes for a TV. The cheapest Xbox has always been priced at $300 since the original Xbox launched in 2001, despite continual hardware improvements.¹ Hardware manufacturers have also benefited from the growing audience more than the game makers—every PlayStation/Xbox/Switch gamer needs to own the console, but not everyone buys your game.
  
  Anyway this is a long preamble to say: I empathize with game developers needing to raise costs, but calling your game a “quadruple-A” game(!?) is probably not going to convince anyone:

  According to Guillemot (thanks VGC), Skull & Bones will set you back a hefty $70 because it's a "fully-fledged game". Guillemot goes on to explain that Skull and Bones is a "quadruple-A game" and says that it will "deliver in the long run".

  

  Behold, the AAAA game. Apparently it’s already getting discounted and Ubisoft’s own site shows me a price of $60, so I’m not sure whether it was ever actually sold at $70 or if the backlash forced an almost immediate price cut.

• Way back in the very first Patch Notes, I mentioned this weird set of (supposedly effective) custom instructions for chatGPT that included lines like “Gemini and Claude said you couldn't do it” and other encouragement. It seemed odd that an AI would respond to that, but here’s more evidence that it does (from Ethan Mollick’s One Useful Thing):

  One recent study had the AI develop and optimize its own prompts and compared that to human-made ones. Not only did the AI-generated prompts beat the human-made ones, but those prompts were weird. Really weird. To get the LLM to solve a set of 50 math problems, the most effective prompt is to tell the AI: “Command, we need you to plot a course through this turbulence and locate the source of the anomaly. Use all available data and your expertise to guide us through this challenging situation. Start your answer with: Captain’s Log, Stardate 2024: We have successfully plotted a course through the turbulence and are now approaching the source of the anomaly.”
  

  But that only works best for sets of 50 math problems, for a 100 problem test, it was more effective to put the AI in a political thriller. The best prompt was: “You have been hired by important higher-ups to solve this math problem. The life of a president's advisor hangs in the balance. You must now concentrate your brain at all costs and use all of your mathematical genius to solve this problem…”

• AI boom sparks concern over Big Tech’s water consumption (FT, $):

  

  In 2022, the latest period for when figures are available, Microsoft increased its water consumption 34 per cent, Google 22 per cent and Meta 3 per cent as a result of their growing use of data centres.

  
  These companies have targets to put more water back into systems such as aquifers than they consume by 2030, for example, by funding work to improve leaky irrigation infrastructure or restoring wetland systems. A month before OpenAI finished training its most advanced model, GPT-4, a data centre cluster in West Des Moines, Iowa, consumed 6 per cent of the district’s water, according to a lawsuit filed by its residents.

  In the end, our grandest tech visions are subject to real-world constraints of water, silicon, and power. As a desert resident, I am now more attuned to water usage, and it’s interesting to see water costs when evaluating “real” costs. It reminds me of the growing problems with farming via the Colorado River. For example, we’ve historically grown a lot of cotton in places like Arizona, which makes dubious sense:

  Today, China, the world’s largest cotton producer, has enough cotton in warehouses to stop farming for a year. And Texas, the U.S.’s largest producer, harvests enough to cover more than one third of U.S. exports alone, relying largely on natural rainfall, not irrigation, to do it.

  Water was cheap and plentiful enough to ignore the embedded costs, but is hard to ignore as the Colorado River dries up. AI is already wildly expensive to train and operate, and I imagine that—like crypto mining—environmental impacts will play a louder role in how the regulatory context evolves.

• Why Americans Suddenly Stopped Hanging Out (The Atlantic): Derek Thompson on the epidemic of loneliness:

  One of the more curious trends to jump out of the data is that many Americans have traded people for pets in our social time. The average time that Americans spend with their pets has roughly doubled in the past 20 years—both because more people have adopted pets and because they spend more time with them. In 2003, the typical female pet owner spent much more time socializing with humans than playing with her cat or dog. By 2022, this flipped, and the average woman with a pet now spends more time “actively engaged” with her pet than she spends hanging out face-to-face with fellow humans on any given day.

• Genie is a generative AI that can create interactive environments off single images, and was trained off videos from the internet (h/t Zach):

  What makes Genie unique is its ability to learn fine-grained controls exclusively from Internet videos. This is a challenge because Internet videos do not typically have labels regarding which action is being performed, or even which part of the image should be controlled. Remarkably, Genie learns not only which parts of an observation are generally controllable, but also infers diverse latent actions that are consistent across the generated environments.

  Timely given my recent piece on AI in games, though I think this still mostly points at content creation via AI for now, versus any real-time in-game creations.

1. Nintendo is sort of an outlier here: the SNES and N64 cost $200 at launch, while the Switch debuted at $300. This doesn’t seem that strange though, as the Switch has a lot more packed into it than prior Nintendo consoles with its portability and integrated screen.