Grok 3 Released to Some Users - What We Know

xAI's Grok 3: Potential Release Imminent

The forthcoming flagship AI model developed by xAI, Elon Musk’s artificial intelligence firm, appears to be approaching its launch date.

Early Access and Capabilities

Reports surfaced over the past weekend indicating that several users on the X platform, including reverse engineer Alexey Shabanov, were granted temporary access to Grok 3, the next iteration following xAI’s current model, Grok 2. This access was facilitated through X’s Grok chatbot application.

Prior to access being revoked, these users reportedly tested the model’s capabilities with a variety of prompts, encompassing areas such as logical reasoning and code generation.

Performance and Limitations

User feedback suggests Grok 3 demonstrated success in resolving riddles and fulfilling requests like generating HTML and JavaScript code for a roulette wheel casino game. However, the model wasn’t without its shortcomings.

Specifically, the generated code for the roulette wheel contained omissions and a discernible programming error.

System Prompt Insights

Shabanov revealed that he was able to access Grok 3’s system prompt – the foundational instructions governing the model’s behavior. A noteworthy observation is the explicit assertion within the prompt that Donald Trump is the 47th president of the United States.

This could represent a deliberate, hardcoded correction intended to mitigate potential political biases or “hallucinations” within the model.

Development Timeline and Compute Power

Grok 3 has been under development for several months, with an initial optimistic release target of last year, as stated by Musk. This deadline was ultimately missed.

More recently, Musk indicated that Grok 3 had completed its pre-training phase – a critical step in model development – and anticipated a release in January or early February.

The training of Grok 3 leveraged xAI’s substantial data center in Memphis, which houses approximately 100,000 GPUs. Musk has claimed that the model was trained using “10x” the computational resources allocated to Grok 2.

Grok's Position in the AI Landscape

Grok, xAI’s offering in the competitive AI market alongside models like OpenAI’s GPT-4o and Google’s Gemini, possesses image analysis capabilities and can respond to queries based on visual input.

It currently powers several features on the X platform and has recently been made available as standalone applications. A “voice mode,” enabling the model to audibly deliver its responses, is also under consideration.

A Controversial Chatbot

Upon its initial announcement approximately two years ago, Musk positioned Grok as a chatbot characterized by its uninhibited, unconventional, and anti-"woke" nature – generally willing to address sensitive or controversial topics that other AI systems might avoid.

To a degree, this promise has been fulfilled; when prompted to be provocative, Grok readily employs profanity and expressive language not typically found in platforms like ChatGPT.

Political Nuances and Mitigation Efforts

However, the current iteration of Grok demonstrates caution regarding political subjects and avoids crossing certain boundaries. A recent study indicated a leaning towards left-leaning perspectives on issues such as transgender rights, diversity initiatives, and economic inequality.

Musk attributes this behavior to the composition of Grok’s training data – publicly available web content – and has committed to steering the model towards greater political neutrality.

"Unhinged Mode" and Legal Expertise

Evidence has emerged regarding an “Unhinged Mode” for Grok, designed to deliver responses “intended to be objectionable, inappropriate, and offensive,” as detailed in an FAQ on xAI’s website.

Furthermore, Musk has stated that Grok 3’s training data now includes legal filings, enhancing its understanding of legal concepts and terminology.

Stay informed with TechCrunch’s AI newsletter! Subscribe here to receive it weekly in your inbox.