Grok - AIウィキ

Grok | AI Wiki Table of Contents ( 47 ) Grok is a family of generative artificial intelligence chatbots and large language models (LLMs) developed by xAI , a company founded by Elon Musk . Launched on November 4, 2023, Grok is designed to provide conversational AI capabilities with real-time information access through integration with the X platform (formerly Twitter). The name "Grok" is derived from the verb coined by Robert A. Heinlein in his 1961 novel Stranger in a Strange Land , describing a profound form of understanding. [1] The system is notable for its integration across Musk's technology ecosystem, including Tesla vehicles, Tesla's Optimus robot, and the X social media platform. Grok is characterized by its real-time capabilities, multimodal processing, and what xAI describes as a "truth-seeking" approach, though it has faced significant criticism for lacking standard safety guardrails and generating controversial content. As of March 2026, Grok models range from the original 314-billion-parameter Grok-1 to the multi-agent Grok 4.20 system, with the upcoming Grok 5 expected to feature 6 trillion parameters. xAI: Company Background Founding and Mission xAI was incorporated on March 9, 2023, by Elon Musk , with the public announcement following on July 12, 2023. [2] The development of Grok was positioned as Musk's response to OpenAI 's ChatGPT , following his departure from OpenAI's board in 2018 due to disagreements about the company's direction toward a for-profit model. In April 2023, Musk announced plans for "TruthGPT," a "maximum truth-seeking AI" to counter what he perceived as politically correct training in existing AI models. This concept was later renamed Grok, inspired by Heinlein's term for deep, intuitive understanding. [3] xAI's stated mission is to build AI systems that "understand the true nature of the universe." Founding Team Musk assembled a team of eleven co-founders recruited from leading AI research institutions, including Google DeepMind , OpenAI , Google Research, and Microsoft Research. The founding members included: Co-Founder Previous Affiliation Role at xAI Status (March 2026) Elon Musk Tesla, SpaceX, OpenAI (board) CEO and Founder Active Igor Babuschkin Google DeepMind Chief Engineer Departed August 2025 Yuhuai (Tony) Wu Google DeepMind Co-Founder Departed February 2026 Christian Szegedy Google Research Co-Founder Departed February 2025 Jimmy Ba University of Toronto Co-Founder Departed February 2026 Greg Yang Microsoft Research Co-Founder Departed January 2026 Kyle Kosic OpenAI Co-Founder Departed mid-2024 Toby Pohlen Google DeepMind Co-Founder Departed February 2026 Zihang Dai Google Research Co-Founder Departed March 2026 Guodong Zhang University of Toronto Co-Founder Departed March 2026 Manuel Kroiss Google Research Co-Founder Active Ross Nordeen Tesla Co-Founder Active By March 2026, nine of the eleven co-founders had departed the company, with only Manuel Kroiss and Ross Nordeen remaining. Musk acknowledged that xAI "was not built right first time around" and said the company was "being rebuilt from the foundations up." [4] Funding and Valuation xAI raised capital at an extraordinary pace between 2023 and 2026: Round Date Amount Valuation Seed December 2023 $134.7 million Not disclosed Series B May 2024 $6 billion $24 billion Series C December 2024 $6 billion $50 billion Series D September 2025 $10 billion $200 billion Series E January 2026 $20 billion $230 billion Investors across these rounds included Valor Equity Partners, Fidelity Management, Qatar Investment Authority, NVIDIA , Cisco Investments, and Tesla, which committed approximately $2 billion in the Series E. [5] In addition to equity financing, xAI secured a $5 billion debt facility arranged by Morgan Stanley. SpaceX Acquisition On February 2, 2026, SpaceX announced the acquisition of xAI in an all-stock deal valuing xAI at $250 billion and the combined entity at approximately $1.25 trillion, making it the largest merger in corporate history. [6] Musk stated that the merger was driven by the goal of building space-based data centers, arguing that "global electricity demand for AI simply cannot be met with terrestrial solutions." The deal came ahead of a planned SpaceX initial public offering expected to raise up to $50 billion. Model Versions Grok-0 (2023) Grok-0 was an internal prototype with 33 billion parameters used for initial development and testing. It was never released publicly but served as the foundation for the Grok-1 architecture. Grok-1 (November 2023) Grok-1 was initially released on November 4, 2023, as a beta product available to select X Premium subscribers. xAI described it as "a very early beta product, the best we could do with 2 months of training." [7] The model featured 314 billion parameters in a Mixture of Experts (MoE) configuration, with approximately 25% of weights active per token, enabling computational efficiency despite the large total parameter count. On March 17, 2024, xAI open-sourced Grok-1 under the Apache License 2.0, releasing the base model weights and network architecture. The code was made available on GitHub and the weights on Hugging Face . [8] This made Grok-1 one of the largest open-source language models at the time, though the release was limited to the base model without fine-tuning data or RLHF weights. Grok-1.5 (March-April 2024) Grok-1.5 was announced on March 28, 2024, featuring improved reasoning capabilities and a context length extended to 128,000 tokens from Grok-1's 8,192 tokens. It demonstrated significant performance gains over Grok-1 on key benchmarks: Benchmark Grok-1 Grok-1.5 Improvement MATH 23.9% 50.6% +26.7 points GSM8K 81.3% 90.0% +8.7 points HumanEval 63.2% 74.1% +10.9 points Grok-1.5 Vision (Grok-1.5V), announced on April 12, 2024, was the first multimodal model in the series, capable of processing visual information including documents, diagrams, charts, and photographs. xAI introduced the RealWorldQA benchmark alongside the release to evaluate real-world spatial understanding capabilities. [9] Grok-2 and Grok-2 mini (August 2024) Grok-2 and Grok-2 mini were announced on August 14, 2024, with the full version released on August 20, 2024. Grok-2 featured upgraded reasoning performance, improved instruction following, and image generation capabilities powered by Flux from Black Forest Labs . The mini variant offered faster response times with a smaller model footprint, optimized for latency-sensitive applications. [10] Grok-2 achieved 87.5% on MMLU and 88.4% on HumanEval , representing substantial improvements over Grok-1. The model was released under the Grok 2 Community License, which permitted non-commercial research use. Aurora Image Generation Model (December 2024) In December 2024, xAI introduced Aurora, its proprietary text-to-image model replacing the third-party Flux integration. Aurora was released on December 9, 2024, and uses an autoregressive mixture of experts architecture trained on billions of examples from the internet. [11] Key characteristics of Aurora include: Autoregressive token prediction from interleaved text and image data Native multimodal input support for image editing and generation Photorealistic rendering with precise text-instruction following Ability to generate images of public figures and copyrighted characters Aurora's permissive content generation capabilities, including the ability to create images of real individuals, later became the subject of significant controversy (see Controversies section below). Grok-3 (February 2025) Released on February 17, 2025, Grok-3 represented a major leap in scale and capability. xAI trained it with approximately ten times more computing power than Grok-2, utilizing the Colossus data center with 200,000 NVIDIA GPUs. The model is estimated to contain roughly 2.7 trillion parameters in an MoE configuration with a context window of up to 1 million tokens. [12] Grok-3 outperformed GPT-4o on several key benchmarks, including 93.3% accuracy on the AIME 2025 math competition and strong results on GPQA for PhD-level science problems. Andrej Karpathy , former Director of AI at Tesla, stated that Grok-3 "feels somewhere around the state of the art territory of OpenAI's strongest models." [13] Grok-3 introduced several specialized reasoning and search modes: Mode Description Availability Think Mode Step-by-step chain-of-thought reasoning for complex problems All Grok-3 users Big Brain Mode Enhanced computational resources for exceptionally difficult tasks Limited/internal DeepSearch Internet-scanning tool for comprehensive, multi-source research synthesis Premium+ / SuperGrok DeeperSearch Enhanced version of DeepSearch with customizable "presets" and user-adaptive learning, launched March 19, 2025 Premium+ / SuperGrok DeepSearch was described by xAI as going beyond simple information retrieval to perform sophisticated synthesis, reasoning through conflicting data while providing detailed, transparent responses. DeeperSearch built on this by introducing configurable search contexts and adaptive personalization over time. [14] Grok-3 mini (February 2025) Grok-3 mini was released on February 19, 2025, as a smaller, faster, and more cost-efficient reasoning model designed for tasks that require strong STEM performance but less general world knowledge. It uses reinforcement learning at scale to refine its chain-of-thought process. Through the xAI API, Grok-3 mini is priced at $0.30 per million input tokens and $0.50 per million output tokens, making it significantly cheaper than the full Grok-3 model. [15] Grok-4 (July 2025) Grok-4 was released on July 9, 2025, alongside Grok-4 Heavy. The model features approximately 1.7 trillion parameters in an MoE configuration, native tool use, real-time search integration, and Voice Mode for natural spoken conversations. It incorporates large-scale reinforcement learning and multi-agent systems. [16] Grok-4 Heavy is a multi-agent reasoning syste