On November 5, Musk's xAI announced its first artificial intelligence model, "Grok," on X.

Musk has repeatedly mentioned in public that he was a big fan of the science fiction novel "Hitchhiker's Guide to the Galaxy" since childhood. Under the influence of the boss, Grok also incorporates elements of this science fiction novel. When introducing Grok, the official stated that it is an artificial intelligence that mimics the Hitchhiker's Guide to the Galaxy. Grok can answer almost any question and even suggest what questions to ask.

Musk's influence goes beyond that. As a boss with a distinct personality, their artificial intelligence is not as conventional. XAI specifically reminds that Grok is designed to answer questions and has a tendency towards wit and rebellion, and recommends that those who dislike humor not use it.

Like AI tools such as ChatGPT, Grok is also open for external use to continuously improve the model. However, currently, only American users are provided with the opportunity to try the Grok prototype, and the quota is limited.

xAI: Grok-1's computational reasoning ability has exceeded ChatGPT-3.5

xAI is an artificial intelligence company that was only established in July this year. At the beginning of its establishment, xAI received attention from the outside world due to its strong talent lineup. The company's team is led by Musk and consists of multiple engineers and experts. Many of the personnel inside are former employees of large technology companies such as Microsoft and Google. In addition, Dan Hendrycks, the current director of the Artificial Intelligence Security Center in the United States, serves as an advisor to the company.

Supported by so many top talents and backed by billionaire Musk, xAI launched an artificial intelligence model within less than four months of its establishment, which is extremely efficient.Although the development time is relatively short, according to official test results, Grok's performance is very superior.

The current engine that powers Grok is Grok-1, a Large Language Model (LLM) developed by xAI over the past four months. Grok-1 has undergone multiple iterations during this period. At the beginning, the company trained a prototype LLM with 33 billion parameters, which is Grok-0. XAI stated that this early model approached Meta's LLaMA 2 functionality on a standard language model benchmark, but only used half of its training resources.

In the past two months, xAI has made significant progress in reasoning and coding capabilities, ultimately giving birth to the more powerful Grok-1.

In terms of reasoning, xAI conducted a series of evaluations on Grok-1 using standard machine learning benchmarks aimed at measuring its mathematical and reasoning abilities. In these benchmark tests, Grok-1 demonstrated excellent performance, surpassing all other models in its computational category, including ChatGPT-3.5 and Inflection-1. Currently, only models trained with a large amount of training data and computing resources (such as GPT-4) can surpass Grok-1.

Since these benchmark tests can be found online, xAI cannot rule out that the model was unintentionally trained on these benchmarks. Therefore, in the 2023 Hungarian National High School Finals, xAI conducted a new evaluation of its model, as well as Claude-2 and GPT-4. The results showed that Grok passed the exam with a score of 59% and achieved a B level, Claude-2 also achieved a similar score with a B level of 55%, and GPT-4 achieved a B level with a score of 68%.

In addition, in terms of coding, Grok-1 achieved 63.2% on the encoding task of HumanEval (a dataset provided by OpenAI to evaluate AI's ability to solve programming problems) and 73% on the current mainstream LLM evaluation dataset - MMLU (an English evaluation dataset containing 57 multiple-choice question answering tasks covering mathematics, history, law, etc.).

Backing against "X"？

When introducing Grok, xAI specifically mentioned its unique advantage of being able to access real-time information on the X platform.

Now, as a global social media platform, X has a huge amount of information appearing and flowing on a daily basis, which is a huge database for Grok. And for artificial intelligence, if it wants to answer more like a human, it needs to learn from a large number of human conversations, which can also be trained from X.

Grok also inherited the bold qualities of Musk's speech. XAI indicates that Grok can also answer sharp questions that most other artificial intelligence systems refuse to answer. XAI believes that designing artificial intelligence tools that are useful for people from various backgrounds and political perspectives is crucial. In an interview in April this year, Musk expressed concern that existing artificial intelligence companies would prioritize "politically correct" systems.

However, the information on X is not entirely high-quality, but also contains a large amount of junk information. How to distinguish the authenticity of information and how to avoid false information in training models is also a challenge for Grok and the Xai team behind it.

Since Musk's acquisition of Twitter for $44 billion in November last year, it has been controversial. The most criticized aspect is that due to Musk's relaxation of content censorship rules and the restoration of many banned users, there has been an increase in hate/negative comments on the platform. According to a study by Newsguard, which tracks online false information, within a week of Musk's acquisition of Twitter, the most popular but least trusted accounts saw a nearly 60% increase in engagement. Moreover, among the thousands of accounts that were unblocked after Musk took over, over one-third of them even spread hatred or false information after being unblocked.

But Musk denied it. In an interview with the media in April this year, he claimed that the company was working to delete robot accounts, and after he took over, false information on Twitter decreased. He said, "My experience is that false information has decreased, not increased."

Regardless of whether false information has increased or decreased, it is undeniable that X is indeed filled with a lot of false information. In this case, X is like a double-edged sword for Grok, and it remains to be seen whether the benefits outweigh the drawbacks or the drawbacks outweigh the benefits.

At the recent Artificial Intelligence Summit held in the UK, Musk stated that artificial intelligence will be the most disruptive force in history, but his conclusion is that it will ultimately "become a force for good."

When xAI was founded, it was said that the company's goal was to understand the true essence of the universe. At present, this original intention has not changed. When introducing Grok, the xAI official mentioned that the development of Grok was aimed at creating artificial intelligence tools to help humans seek understanding and knowledge. The company hopes to provide users with artificial intelligence tools while complying with the law. XAI hopes that Grok can become a powerful research assistant for anyone, helping them quickly access relevant information, process data, and propose new ideas.

xAI: "Our ultimate goal is to have our artificial intelligence tools assist in the pursuit of understanding."