Grok 3 Goes Live: Elon Musk’s AI Model Shows Promise and Flaws

xAI’s next flagship AI model, Grok 3, may be nearing release, as several users reported gaining access to it over the weekend.

Grok 3 is the successor to Grok 2 and is accessible through X’s Grok chatbot app.

Early Access and Capabilities

Users, including reverse engineer Alexey Shabanov, managed to test Grok 3 before their access was revoked. The model responded to a range of queries, including logical reasoning and coding-related questions.

Notably, Grok 3 answered riddles and generated code for tasks like a roulette wheel casino in HTML and JavaScript. However, it made some coding errors and missed a few details in the roulette wheel example.

Political and System Prompt Insights

Shabanov also revealed that Grok 3’s system prompt explicitly identifies Donald Trump as the 47th president of the U.S., possibly to fix political hallucinations in the model. This could be a sign of xAI addressing the model’s previous political biases.

Training and Expectations

Grok 3 has been in development for several months, with Elon Musk previously stating that it would be released last year.

Musk later confirmed that Grok 3 completed pre-training earlier this month, and the model is expected to arrive in January or February 2025.

xAI has been using its large data center in Memphis, containing 100,000 GPUs, to train the model, with Musk claiming Grok 3 was trained with 10x more compute than Grok 2.

Features and Controversies

Grok 3 can analyze images and power various features on X. It also recently gained standalone apps and may soon include a voice mode to read responses aloud.

Musk initially pitched Grok as an edgy, unfiltered model, willing to answer controversial questions other AI systems avoid.

However, Grok’s current responses on political topics, like transgender rights and diversity programs, lean to the political left, though Musk has pledged to shift it closer to political neutrality.

In addition, evidence of an Unhinged Mode has surfaced, promising responses that are “objectionable, inappropriate, and offensive,” which Musk has defended as part of the model’s training.

Legal and Training Data

Grok 3’s training also incorporates court filings, aiming to improve its understanding of legal subjects.