What is Grok, Elon Musk's new chatbot, and how is it different?

Yesterday, Saturday, November 4, Elon Musk's artificial intelligence company xAI unveiled its first product: a chatbot called Grok. It is the newest competitor to ChatGPT, and it arrives with a distinct personality. xAI says Grok is meant to answer almost anything, with wit and a rebellious streak, and warns people who dislike humor not to use it.

Two other things set it apart, according to xAI. Grok can draw on live information from X, the social network formerly called Twitter, which Musk owns. And it is designed to answer "spicy" questions that most other AI systems refuse.

What was announced

xAI describes Grok as a very early beta, built with about two months of training, and expects it to improve quickly with feedback. For now, a small group of US users can sign up on a waitlist to try the prototype.

Musk, posting on X, said that once the early beta ends, Grok will be available to all subscribers of X Premium+, the platform's top subscription tier, which launched the week before at $16 a month.

The name comes from Robert Heinlein's 1961 science fiction novel Stranger in a Strange Land, where "grok" means to understand something deeply and intuitively. xAI also says Grok takes its inspiration from The Hitchhiker's Guide to the Galaxy, Douglas Adams's comic series about a guidebook that answers everything, and that it should even suggest which questions to ask.

Who is behind it

Musk founded xAI in July 2023, AFP reports, recruiting researchers who had worked at Google DeepMind, OpenAI, Tesla and the University of Toronto. Musk himself co-founded OpenAI in 2015 and left in 2018. He has since criticized the company's direction and argues that its models are overly politically correct, a view that helps explain Grok's looser style.

Grok's personality was on display in a screenshot Musk shared. Asked for step-by-step instructions to make an illegal drug, the bot answered with a mock recipe before saying it was joking and would never encourage it.

How Grok works

The engine inside Grok is Grok-1, a large language model xAI says it developed over the past four months. Like ChatGPT's models, it is trained by next-token prediction: it learns to guess the next piece of text, over and over, across a huge amount of writing.

The path there had two stages:

1. Grok-0. Soon after founding the company, xAI trained a prototype with 33 billion parameters, the internal numbers a model learns. xAI says it came close to Meta's LLaMA 2 (70B) on standard tests while using half the training resources.
2. Grok-1. Over the following two months, the team improved reasoning and coding to produce the current model.

xAI built its own training software on Kubernetes, Rust and JAX, three tools for managing large computer clusters, writing reliable code and running machine learning efficiently. It says its runs coordinate tens of thousands of graphics processors for months, so detecting and recovering from hardware failures automatically is central to its work.

Grok's link to X comes from giving the model search tools and real-time information. A model trained on fixed data only knows what was in that data; one that can look things up can discuss today's news. xAI is careful to add that Grok, like any model built this way, can still produce false or contradictory statements.

How it scores

xAI published comparisons on standard tests. Here are a few of its reported figures:

The math exam deserves a note. Benchmark questions are all over the web, so a model may have seen the answers during training. To avoid that, xAI graded its model, Claude 2 and GPT-4 by hand on the 2023 Hungarian national exam, published after Grok's training data was collected. The company says it did no special tuning for it.

xAI's summary is that Grok-1 beats every model in its compute class, meaning models trained with similar resources, including ChatGPT-3.5 (xAI's name for it) and Inflection-1, and is surpassed only by models trained with far more data and computing power, such as GPT-4. As always, these are the company's own numbers, and independent testers have yet to confirm them.

Why it matters

Grok adds a new, well-funded player to a race that already includes OpenAI, Google, Meta, Anthropic and others, and it tests a different philosophy: fewer refusals, more personality and live social media data. Its tie to X gives it a built-in stream of fresh information that other chatbots do not have in the same form.

That same design raises questions. Posts on social media are fast but often wrong, and a model that refuses less may be easier to misuse. xAI lists research priorities that speak to these concerns, including better reasoning, checking answers with external tools, resistance to manipulation, and safeguards against catastrophic misuse. Grok cannot yet handle images or audio, which xAI says it plans to add.

Who can use Grok, and what will it cost?

Right now, only a limited group of users in the United States can try Grok, through a waitlist on xAI's website. xAI calls this a prototype test to gather feedback before a wider release.

Musk says Grok will be included for all X Premium+ subscribers once it leaves early beta. That plan costs $16 a month and also removes ads. No separate price for Grok has been announced, and xAI has not said when the beta will end or when people outside the US will get access.

If you are curious but not ready to subscribe, watch for independent reviews. They will show whether Grok's humor and live data make it more useful in practice, and how often its real-time answers turn out to be right.