Baldur’s Gate 3 devs built a testing AI. Then, they tried to defeat it.
Inside the bot that helps build the next generation of RPGs
I
It does not eat, it does not sleep. But unlike the Terminator, the quality assurance AI that Larian Studios uses to help test old-school RPG throwback
Baldur’s Gate 3 is meant to help people build the most solid video game possible, at a pace much faster than a human could achieve alone.
Quality assurance jobs in video games are
some of the most misunderstood positions in the industry. Where the average player imagines that QA testers get paid to play games for a living, the reality is more tedious. Often, testers will spend hours doing things like walking into walls just to make sure the game is solid all the way through, or opening different menus at different points in the game to double-check that it keeps running smoothly. Once a bug is found, QA testers have to find ways to reproduce it so that the development team can fix the issue — at least, if it’s considered an important enough hindrance.
Humans have been doing exactly that for decades, but advances in technology are allowing developers to outsource some of the work to machines that are perfectly suited for endless repetitive tasks. Mossmouth used a
Spelunky 2 AI, for example, to help
generate random levels that could actually be completed by real players.
Larian Studios started going down this path during the development of 2014’s
Divinity: Original Sin, when the company started building a proprietary AI that it now dubs the World Tester.
“In the early years, it was just teleporting an NPC across regions, testing performance markers and looking for basic crashes,” a Larian representative told Polygon over email.
The “super gamer,” as Larian sometimes calls the AI, works with QA teams to help stress-test player actions and combat, among other things. The second that a new game build is put together, the World Tester digs in before any human lays a finger on it. Sometimes, the Tester will juggle more than one build at the same time, apparently capable of simulating multiple hours of gameplay at once. The QA team will only get involved if the game is stable enough to not fail during its session with the World Tester, which helps save everyone time. The aim isn’t to
replace humans. It’s to help humans become more efficient at what they do, Larian said.
Nowhere did this become more apparent than during a showdown that transpired right before the release of the Enhanced Edition of
Original Sin. While the Tester was still rudimentary, it could still open and close interfaces extremely quickly — leagues ahead of a living, breathing human being. The programmers who helped shape the AI felt confident that it could get rid of corruption and crashing issues. Notably, the code they had built did not use controller inputs, so the team could hook up a gamepad to meddle with the AI’s affairs.
Hilariously, someone on the team was actually able to start up combat against the AI, eventually defeating it in a one-on-one battle. When it happened, the Larian chief technology officer started screaming that it was “impossible.” Except, it wasn’t. To celebrate the unexpected victory against a superhuman opponent, the dueling developer was promoted to become a principal tester at Larian’s studio in Belgium.
Since then, the AI has only gotten stronger — Larian has expanded its abilities for more sophisticated manipulation of UI, and for sifting through dialogue at “lightning speed,” the studio said. With
Baldur’s Gate 3 now available in early access, the World Tester is still being put to good use in 2020. Everything the AI does gets tracked, of course, and at the end of the process, the team parses that data. In the end, the humans still call the shots.
“Our QA teams across the world are extremely important, and the world-tester is really not as smart as it thinks it is,” Larian said.