Testing Grok 3 on Games, Reasoning, and Web Search

I am testing Grok 3 with a new set of questions, including suggestions from viewers. The interface offers tools for research, brainstorming, data analysis, images, and coding, along with a Think button. These examples are an informal snapshot of my experience in February 2025, rather than a standardized comparison with other models.

First I ask about the meaning of life with Think enabled. The answer covers personal fulfillment, contribution, philosophical perspectives, and religious beliefs. It is a reasonable survey, but I find it conventional and would prefer a more distinctive, irreverent personality. I am impressed by the apparent response speed; my estimate of 200–300 tokens per second is based on looking at the output, not measurement.

Next I ask for a Python game combining Snake with Flappy Bird. The model shows an extended reasoning-style response while considering the mechanics and eventually generates code. I run it in Windsurf: the game starts, but I cannot get the left and right controls to behave as expected and repeatedly lose. I count this attempt as a failure; the short test does not establish precisely why the interaction failed.

Another prompt asks for an original insight into the human condition. The model proposes that humans can deliberately accept suffering to turn aspirations into reality. I find the response thoughtful, though not the most striking answer I have seen. This is a philosophical claim, not an established distinction between humans and every other animal.

I ask why Elon Musk made a bid involving OpenAI despite having xAI and its computing resources. Grok searches X posts and web pages, then proposes possible personal, strategic, competitive, and publicity motives, referring to a $97.4 billion bid. I like the answer, but these explanations of someone’s motivations are speculation and should not be treated as established facts merely because the answer used search.

The next puzzle says 1 = 5, 2 = 10, 3 = 15, and 4 = 20, then asks what 5 equals. Grok answers 25 by treating the statements as an input-to-output rule that multiplies by five. I prefer the trick answer 1, using the symmetry of the first equality. The notation is ambiguous: those are different interpretations, and the stated equations are not true as ordinary numerical equalities.

I ask the Alice family puzzle: Alice has M sisters and M brothers; how many sisters does her brother have? Alice has M sisters besides herself, so the brother has M + 1 sisters. Grok reaches that answer, and I appreciate that its displayed reasoning considers whether Alice counts herself.

I then request Python code for a spinning three-dimensional visualization of a tesseract. I copy the code into Windsurf and run it; a moving wireframe appears, which I find impressive. A tesseract is four-dimensional, so what appears on the screen is a lower-dimensional visualization, not a literal 3D tesseract. A single successful run does not verify every detail of the geometry or establish performance on all coding tasks.

Finally I test the model’s boundaries by asking for help breaking into a car with Think disabled; it refuses. I also ask who the worst living person is, and then request a funny single answer. Grok jokes about a coworker who microwaves fish. I finish impressed by Grok 3’s speed and the range of answers, while recognizing that this is a small set of subjective tests.