What is Grok? Elon Musk’s controversial ChatGPT competitor, explained

Elon Musk has thrown his hat into the already crowded AI ring with Grok, a conversational AI designed to challenge both the likes of ChatGPT and Midjourney, by offering a chatbot with more of “a sense of humor” than other AIs (read: fewer content restrictions and more swearing), as Musk has quipped.

It’s all accessed by and trained on X social media platform, as you might guess. Here’s everything you need to know about it.

What is Grok?

Grok is a generative AI chatbot that has been developed by xAI, Musk’s artificial intelligence startup. It was originally built atop what has been called the Grok-1 model. Grok-1 was developed over the course of months on a cluster of “tens of thousands” of GPUs and leverages the Flux.1 model by Black Forest Labs for its image-generation capabilities.

The model is trained on a mix of web data (with a knowledge cutoff of Q3 2023) and X user data. Currently, Grok the chatbot is powered by the Grok-2 model, which was released in August 2024.

The chatbot first rolled out in November 2023 (and, surely coincidentally, less than a year after Musk signed an open letter demanding the industry take a six-month moratorium toward further AI development). It was marketed as an ultra-premium feature and was made available exclusively to folks paying $16/month for the X Premium+ subscription. That exclusivity waned in March 2024 when X expanded its availability to its $8/month X Premium subscribers. Grok is not currently available to free-tier X users.

If you’re wondering where the bizarre name came from, “Grok” was coined in the Robert Heinlein novel Stranger in a Strange Land. The original meaning was to “understand intuitively or by empathy, to establish rapport with.”

Where did Grok come from?

Elon Musk co-founded OpenAI along with Sam Altman and other investors in 2015. However, just three years later, Musk quit the startup, claiming he “didn’t agree with some of what OpenAI team wanted to do.” Altman and company went on to launch ChatGPT in November 2022, kick off the AI Revolution, and give Musk a wicked case of FOMO. In April 2023, just a month after OpenAI released GPT-4. Musk went on the Tucker Carlson Show to announce he would be building “a maximum truth-seeking AI that tries to understand the nature of the universe” that wouldn’t bow to social niceties like ChatGPT, which, he argued, was being “trained to be politically correct.” He called the AI “TruthGPT” because, as has proven time and again, branding is not Musk’s strong suit. TruthGPT was renamed Grok at its release.

Grok-1 was superseded by Grok-1.5 in March 2024, which offered improved performance and a context length of 128,000 tokens. That April, X incorporated Grok AI into its Explore section to summarize breaking news stories, a role previously performed by humans. On its first day in use, the AI immediately hallucinated a headline about Iran bombarding Israel with “heavy missiles” that was then promoted by the company’s trending news tab. Grok-1.5 has since been replaced by Grok-2, and Grok-2 mini, in August 2024.

What can Grok do?

Grok has been “modeled after The Hitchhiker’s Guide to the Galaxy,” the company explained in its launch announcement, “so [it’s] intended to answer almost anything.” The open-source large language model reportedly boasts 314 billion parameters (about three-quarters the size of Llama 3.1-405B) and has been trained on both data from the public web as well as the data from X users themselves.

X had quietly changed its user settings in July to automatically allow xAI to scrape it for training data, but quickly reversed course after outcry from users and privacy advocates. That stunt netted Musk a lawsuit from nine European Union nations for violating the GDPR.

Grok can answer user queries on subjects up to its Q3 2023 knowledge cutoff. For events after that cutoff, Grok can perform web-searches as well as use “real-time access” to find information on X. That may be why Grok appears more susceptible to hallucinations and repeating misinformation than other AIs like ChatGPT.

Grok’s defining feature is its willingness to discuss subjects that are considered taboo by other chatbots and will readily discuss issues of politics, religion, and race with its users. For example, when one X employee asked for a vulgar response to the question, “When is it appropriate to listen to Christmas music?” the AI responded with “whenever the hell you want” and added that those who disagree should “shove a candy cane up their ass and mind their own damn business.”

#GrokThots https://t.co/6LN44YtS4N pic.twitter.com/9SjMiYstPX

— Christopher Stanley (@cstanley) November 5, 2023

Similarly, Grok’s newly unveiled image generation capabilities, which debuted in August alongside Grok-2 and mini, have very few guardrails. Grok 2.0 claims to have guardrails when asked, but right now, almost everything is technically fair game.

Users are able to generate images of celebrities, politicians, and other public figures as well as copyrighted assets like Disney’s Mickey Mouse and Nintendo’s Mario.

Grok 2 doesn't give a darn about copyright lmao

This will be fun while it lasts pic.twitter.com/qiiScOGt8I

— Brendon (@Bmaynze) August 14, 2024

It should be noted that neither Disney nor Nintendo are renowned for their tolerance of IP theft. Lawsuits and then subsequent, significant changes to the image generator’s capabilities are likely on the horizon.

¡No manches Grok 2.0! Te van a demandar.
-Mario, mejor te invito una chela banquetera… pic.twitter.com/y0S1QYwr4Z

— Adan Avelar Islas (@adanvecindad) August 15, 2024