Grok (neural network)

Main article: LLM (Large Language Models)

2025: Elon Musk launches new LLM into public access - Grok 3

This event took place on February 18, 2025. According to preliminary tests, the thinking Grok 3 Reasoning Beta achieves phenomenal indicators in benchmarks adapted for LLM testing.

The integrated performance measure GIYA (includes MMLU-Pro, GPQA Diamond, Humanity's Last Exam, SciCode, AIME, MATH-500 and others benchmarks) put Grok 3 Reasoning Beta in a solid 1st place among public models with 67 points ahead of o3-mini with 63 points, DeepSeek-R1 with 60, but this is for thinking/thinking models.

Results obtained in aggregation of tests from artificialanalysis.ai.

Ordinary (non-reasoning models) also lead the Grok-3 with 53 points, formally ahead of the best of the public "regular" LLM - Gemini 2.0 PRO - 47, followed by DeepSeek V3 - 46, Qwen 2.5 Max - 45, Claude 3.5 - 44 and extremely outdated ChatGPT-4o - 41.

The density of competition is phenomenal, the gaps are minimal, everything is decided by the flexibility of LLM, the depth of configuration and the ability to solve specific problems.

Ideal LLMs do not exist, each has strengths and weaknesses, so it is better to use a combination of LLMs for different types of problems.

According to formal indicators, Grok-3 this is the best existing LLM in a comparable category, both among the thinking and among the "ordinary," although not the best, since the stronger model is OpenAI o3, which in a limited version is available for $200 per month, but it can hardly be called public.

Grok 3 at this time costs about $30 per month.

Elon Musk started later than everyone else, releasing relatively weak LLMs (in comparison with competitors), but managed to overtake everyone, emphasizing the extraordinary pace of innovation in this industry.

Grok 3 uses the Mix-of-Experts (MoE) architecture, which allows you to strategically activate subsets of parameters for different tasks, making it more efficient in data processing and analysis. It includes 314 billion parameters, which makes it the largest available model, although not the largest, but the quality of the model does not have a linear dependence on the number of parameters, the result is influenced by a lot of factors.

For Grok 3 training, a Colossus supercomputer equipped with 200,000 Nvidia H100 GPUs was used.

Grok 3 introduced new features such as Think and Big Brain modes for complex tasks, as well as the DeepSearch tool for analyzing information from the Internet and from the social network X. Image generation and voice mode capabilities were also added.

Elon Musk stressed that Grok 3 is focused on "finding the truth," even if it goes against political correctness, striving for political neutrality.

The deployment of the functionality should take 2-3 months.

2024: LLM Source Code Discovery Grok-1

In March 2024, a startup xAI Elona Musk developing artificial intelligence technologies announced the discovery of the source code of its large language model (LLM) Grok-1. Developers, companies and enthusiasts around the world can use the platform.

The Grok-1 model is based on the Mix of Experts (MoE) architecture, which significantly increases the speed and quality of request processing. The number of parameters used is 314 billion. The basic model is trained on a large amount of text data, but is not configured to perform any specific task, for example, to conduct dialogs. The Grok-1 pre-training process was completed in October 2023.

𝓲

Фото: xAI

Access to the source code of the model is provided under the terms of the Apache 2.0 license, which gives the right to use the software for any purpose, freely change and distribute changed copies, with the exception of the name. As of mid-March 2024, Grok-1 is one of the largest open source AI models. The xAI report notes that the release includes "base model weights and architecture."

It is noted that due to the large number of parameters, significant hardware resources are required to use the Grok-1 model, including AI accelerators based on GPUs. By publishing the Grok-1 code, Musk hopes to encourage other AI system developers to transfer their products to the open source community. On the other hand, the availability of code for powerful AI models raises concerns from some critics, who talk about the possible unethical and dangerous use of such platforms. In particular, AI systems can be used to generate deepfakes that mislead users.^[1][2]

2023: Neural Network Launch

On November 6, 2023, xAI, which is owned by American entrepreneur Elon Musk, announced and launched its first neural network. It was named Grok.

Grok's artificial intelligence was created inspired by a guide from the book and film "Hitchhiker's Guide to the Galaxy," which tells about exciting adventures in space, the xAI press service said.

xAI has announced and launched its first neural network

Our Grok AI has the ability to answer questions wittily and has a penchant for rebellion, so please do not use it if you do not appreciate humor! - emphasized in the company.

According to the developers, the main advantage of the new AI model in xAI is that it receives real-time world information through platform X (formerly known as Twitter). Grok is supposed to be able to provide answers to provocative questions that other systems cannot answer.

It is reported that the development of the algorithm began with the creation of a prototype of a large language model (LLM) Grok-0 with 33 billion parameters. This test model is comparable in its capabilities to the LLaMA 2 indicators from Meta Platforms (the company is recognized as extremist in Russia, its activities are prohibited in the Russian Federation) in standard tests, but uses only half of its training resources. Over the past two months, the developers have managed to achieve significant improvements in terms of logical analysis and coding capabilities, which led to the creation of a significantly more powerful language model of the Grok-1.

According to Elon Musk, by November 6, 2023, Grok can only be accessed when purchasing a Premium + subscription on platform X. It is not specified when the neural network will be available to everyone^[3]

Notes