An artificial intelligence from the University of Washington named GeoS, short for geometry problem solver, has just answered 49 percent of a high-school geometry test correctly. While that might not actually be an impressive number at face value, it is actually on par with the average performance of human students taking the exam. And while it might not sound all that impressive from a sci-fi doomsday perspective, considering computers should be great at math after all, it is actually a small but significant step forward in the field of artificial intelligence.
Given the nature of computers, you’d think that they’d ace the exam, but the key difference here is that the AI wasn’t given problems in a text or binary format that it easily understood. Put it simply, if you gave it a ready made formula to solve, it would have gotten a perfect score. GeoS, however, was fed the exact same exam questions human students are given. And that means, it first had to be able to actually understand the diagrams and text printed on a paper.
Computer vision plays a big role here. GeoS needs not only to be able to identify what the diagram is but also interpret the problem that needs to be solved. This is the real achievement here, not just the simple fact that the AI was able to solve geometry problems. Computers are great at crunching numbers and running through formulas, but they aren’t so great at trying to understand the problem based on human natural human language. So when the AI gets a 500 out of 800 score on the SAT, the average score for this year, that is much a momentous achievement for the AI as it for students taking the exam.
GeoS is just one of the many AI projects that are trying to let AI programs take human exams in different subjects. While GeoS deals with high-school geometry, the Allen Institute’s Project Aristo is trying it’s luck with fourth grade science exams. Meanwhile, Fujitsu, IBM, and other partners have started work on AI that is trying to pass the University of Tokyo entrance exam.
Should be start worrying? Hardly. The GeoS’ performance shows, it still has a long way to go to being in anyway commendable, much less actually dangerous. If anything, these attempts prove how replicating the human thought processes is no walk in the park and that we’re still quite far away from the self-aware, self-intelligent, and self-perpetuating SkyNet of the dark future.