Post Thumbnail

New Grok 3 from xAI surpasses GPT-4 and Claude 3.5

xAI, Elon Musk’s company, has introduced a revolutionary version of the Grok 3 language model, which has set new performance standards in artificial intelligence, surpassing leading competitors including OpenAI GPT-4, Claude 3.5, Deepseek V3, and Gemini 2 Pro.

For the first time in AI history, a model has broken the 1400 ELO score barrier in the LMArena evaluation system, ranking first in all testing categories. Grok 3 demonstrated impressive results in key areas: 52% in standard mode and 96% with reasoning mode enabled on the AIME 24 test, 75% (improving to 85%) on the GPQA test, and 57% to 80% on LiveCodeBench programming tests.

The technical foundation for Grok 3’s unprecedented performance is the Colossus supercomputer, equipped with 100,000 NVIDIA H100 GPUs. The new version’s computational power is 10 times higher than its predecessor Grok 2.

The model received several innovative features, including “Big Brain” mode for enhanced performance, improved reasoning capabilities, and upcoming voice interface addition. Users will be able to activate advanced reasoning capabilities for solving complex problems in mathematics, science, and programming using “Think” and “Think harder” commands.

“The model has achieved a frighteningly high level of intelligence,” commented Elon Musk, emphasizing the system’s unprecedented creative and emergent capabilities. Grok 3’s success in creating creative program code and generating innovative games is particularly noted.

xAI plans to launch a web version of Grok 3 with multiple personalized AI personalities to match user preferences. The company also announced plans to make the previous version, Grok 2, open source in the coming months.

Autor: AIvengo
For 5 years I have been working with machine learning and artificial intelligence. And this field never ceases to amaze, inspire and interest me.

Latest News

Open source model RoboBrain 2.0 will become foundation for humanoid robots

AI model RoboBrain 2.0 can now combine environment perception and robot control in 1 compact system. Specialists already call it the foundation for the future generation of humanoid robots.

Tinder launched double dates: AI assembles teams of 4 people

Tinder app launched a double date function that allows users to team up with friends to find pairs. Now you can invite up to 3 friends and together browse profiles of other so-called teams. That have at least 1 match in individual preferences.

New benchmark showed AI failure in Olympic programming tasks

A new benchmark LiveCodeBench Pro for evaluating artificial intelligence programming capabilities has appeared. Link in description. It includes the most difficult and fresh tasks from popular competitions. International Olympiad in Informatics and World Programming Championship. Tasks were marked by winners and prize-winners of these competitions themselves.

Data up to 2022 became "pre-nuclear steel" for AI training

Artificial intelligence, intended to become the locomotive of technological progress, is beginning to slow down its own development. According to The Register, generative models have filled the internet with so much synthetic content that this creates a real technological dead end.

Sam Altman revealed Meta's attempts to poach employees for $100 million

Sam Altman publicly revealed the unprecedented talent hunt that Mark Zuckerberg is conducting. The Meta head offers OpenAI employees truly astronomical sums. $100 million just as a signing bonus!