The U.S. software company X (formerly Twitter) has rolled out its AI-powered language model Grok for initial users in the United States. In an announcement by Elon Musk’s startup xAI on November 4, 2023, it was stated, “Grok is an AI inspired by The Hitchhiker’s Guide to the Galaxy and is intended to answer almost anything and, even more challenging, provide suggestions on what questions to ask!” German X users can currently only join a waiting list.
According to the announcement, Grok is still in its “very early beta product” stage. They couldn’t achieve more during the two-month training phase. In a sort of warning, the team writes, “Grok was designed to answer questions in a humorous manner and has a rebellious streak. So, please don’t use it if you dislike humor!”
One “unique and fundamental advantage of Grok” is that the chatbot has “real-time knowledge of the world” by accessing X. Furthermore, Grok will also “answer provocative questions that most other AI systems would reject.” However, since X often spreads misinformation, the use of this data by Grok could also become a problem.
ChatGPT 4 is expected to be even better
Grok, which in American slang means “to understand,” is, like Microsoft’s ChatGPT and Google’s Bard, a language model trained on massive amounts of data (LLM – Large Language Model). The developers trained an LLM prototype (Grok-0) with 33 billion parameters. According to their claims, this early model achieved capabilities similar to Facebook’s LLaMA 2 (70B) in standard benchmarks but only consumed half of its training resources.
They also stated, “Over the past two months, we’ve made significant improvements to argumentation and coding abilities, up to Grok-1, a state-of-the-art language model that is much more powerful and achieved 63.2 percent in the HumanEval coding task and 73 percent in the MMLU.”
HumanEval is a format used to test language model programming skills. MMLU (Massive Multitask Language Understanding) allows testing the general knowledge and problem-solving capabilities of language models, according to Bigdata-Insider.de. However, the developers admitted that ChatGPT 4 outperformed Grok in benchmark tests.
Developed with Kubernetes, Rust, and JAX
xAI stated that a “custom training and inference stack based on Kubernetes, Rust, and JAX” was developed for Grok’s training. The training of a language model is described as “like a thundering freight train; if one wagon derails, the entire train gets off track, making it difficult to set it back on course.”
There are countless ways GPUs could fail, including manufacturing defects, loose connections, incorrect configurations, damaged memory chips, occasional random bit flips, and more.
The training involves synchronizing computations on tens of thousands of GPUs for months, resulting in the mentioned errors occurring frequently. One focus of the development was to maximize computing performance per watt.
Developers on X shared their initial experiences with the new chatbot, indicating that it’s possible to engage in multiple conversations simultaneously and switch between them.
