On the 17th (local time), Elon Musk, the chief executive officer (CEO) of Tesla, unveiled the latest artificial intelligence (AI) model 'Grok3' from the AI corporation xAI.
Musk noted that this model has "a performance one dimension beyond Grok2" and emphasized that it has surpassed competing models such as OpenAI's ChatGPT-4o, Anthropic's Claude 3.5 Sonnet, Google's Gemini, and DeepSeek V3.
Grok3 is a super-large AI model developed by xAI, trained at a data center in Memphis, Tennessee, equipped with over 200,000 GPUs (graphics processing units). According to Musk, Grok3 utilized 10 times the computational resources compared to Grok2 and employed an expanded dataset that included vast amounts of data, including court filings.
xAI claimed that Grok3 exceeded the performance of ChatGPT-4o in key AI benchmarks such as AIME, which evaluates performance on math problems, and GPQA, which addresses doctoral-level physics, biology, and chemistry problems. For instance, in the evaluation of the AIME 2024 math competition problems, DeepSeek V3 recorded a 39% accuracy rate, Claude 3.5 Sonnet had a 26% accuracy rate, while Grok3 achieved the highest score with 52%. In the science-related benchmark 'GPQA', Grok3 recorded a 75% accuracy rate, surpassing GPT-4o (50%), Claude 3.5 Sonnet (65%), and DeepSeek V3 (59%).
Grok3 is not a single model but an AI family composed of various versions. In particular, the 'Grok3 Reasoning Model' and the 'Grok3 Mini Reasoning Model' were developed to analyze complex problems and provide answers through thorough fact-checking. This concept is similar to OpenAI's 'o3-mini' and DeepSeek's 'R1', with the characteristic that AI applies the 'Chain of Thought' (CoT) technique by going through a step-by-step reasoning process to derive answers.
The newly added 'DeepSearch' feature is also gaining attention. This AI-based deep search function analyzes data from the Internet and X to provide summaries in response to questions, similar to OpenAI's 'Deep Research.' Musk emphasized that "DeepSearch replaces what users would need to search for over 30 minutes to an hour, providing organized results in just 10 minutes."
Grok3 will initially be available to subscribers of X's 'Premium+' service (29,000 won per month, 300,000 won per year), and additional features will be provided through the newly introduced paid service 'SuperGrok' from xAI. The SuperGrok subscription costs $30 (around 43,000 won) per month and $300 (around 430,000 won) per year, including additional reasoning queries, expanded DeepSearch features, and unlimited image generation.
Musk stated, "We plan to add a voice synthesis feature called 'Voice Mode' to the Grok app within the next week, and in a few weeks, Grok3 and the DeepSearch feature will be applied to corporate APIs as well." He added, "We intend to release the previous model, Grok2, as open-source within the coming months."
Meanwhile, Musk has emphasized that AI models should be able to respond freely without political censorship; however, previous Grok models had maintained a cautious stance on political issues. A study indicated that the earlier Grok models exhibited somewhat progressive tendencies regarding transgender rights, diversity policies, and economic inequality issues. In response, Musk remarked that this was due to "the influence of the training data Grok learned from" and stated that "it will be adjusted to a politically neutral model in the future."
Musk also mentioned that "Grok AI will continuously evolve and will be integrated into SpaceX rocket systems within the next two years."