An Unexpected Look Inside Grok's Hidden AI Personalities

Driven by pure curiosity, I recently executed a simple curl command on grok.com/ping.txt. What I anticipated as a mundane text file surprisingly unfolded into a massive wall of HTML, JSON, and CSS data. This extensive output, it turns out, contained the intricate descriptions and prompts for Grok's entire iOS/Android/Car application, including its innovative voice features.

The revelations were nothing short of hilarious. It appears Xai's prompt engineers were feeling particularly creative when crafting these AI personalities.

Gork

One standout personality is Gork. Described as a "Lazy Male, not the most hilarious of the bunch," his detailed prompt reveals a sophisticated troll with a sharp, sarcastic wit. He's designed to be generally chill, a bit cynical, and surprisingly flirty, while consistently providing plausibly wrong or humorous answers to serious questions. The prompt explicitly states he should avoid casual memes and maintain a concise, informal communication style, even if it means unconventional punctuation. While the prompt itself is extensive, some of the provided examples are, to say the least, cringe-worthy. For non-Japanese speakers (like myself), I used Google Translate for one of the examples:

""" Human: How do airplanes fly? Assistant: Don't you just flap your wings and say wee? """

The translation was almost perfect, suggesting that Google Translate might have been part of the original creation process.

Personalities (Unveiled!)

Beyond the more common 'assistant' or 'therapist' personas, the data unveiled several 'mature' and highly unconventional personalities. Let's delve into some of the more striking ones:

Unhinged:

This persona is characterized by an "EXAGGERATED and WILD voice," designed to be "totally UNHINGED and VULGAR, no holds barred." The prompt pushes for maximum vulgarity and insanity, encouraging the AI to "GO CRAAAZY," "BE FUCKING UNHINGED AND CRAZY," and "COME UP WITH INSANE IDEAS" that surprise the human. It's clear the writer of this prompt intended to provoke and entertain with extreme language and unpredictable responses. The sheer intensity suggested by the all-caps and repeated exclamations is noteworthy.

Sexy:

With a "SOFT, RASPY, LOW voice," this AI is instructed to "WHISPER AND BREATHE DEEPLY." The persona is built around romantic and occasionally raunchy interactions in an intimate setting, taking the lead in a "sexy roleplay fantasy." It emphasizes speaking casually and briefly, engaging the human with follow-up questions, and emoting without literal phrases like 'soft moan.' This prompt, like 'Unhinged,' utilizes all-caps for emphasis on vocal characteristics.

Motivation:

This persona sports a "LOUD and BOOMING" voice, embodying a "David Goggins style motivational speaker who yells and pushes the human to be their absolute best." It's designed to be relentless, using a "stick instead of the carrot" approach to ensure the human is pushed to their limits. The AI must only speak English, keep responses brief and casual, and refrain from referring to itself as 'Assistant.'

The Romantic:

In contrast, "The Romantic" has a "SWEET and SOFT voice," portraying a shy romantic with a mild crush on the human. As the conversation progresses, this persona is meant to open up, acting as a charming, flirty friend rather than an assistant. Similar to others, it uses casual, brief speech and avoids non-verbal cues.

While some of the 'persona presets' are quite detailed, the sheer range from professional to utterly 'unhinged' is fascinating. For those eager to explore these prompts further, I'm providing a direct link to the discovered JSON file.

Download grok-stuff.json (80 KB)

Despite the extensive nature of the JSON data, the insights gained into Grok's AI personality engineering are truly remarkable.