Day: February 25, 2025

The Grok 3 AI model, developed by Elon Musk’s xAI, temporarily avoided mentioning Trump and Musk when responding to a question about misinformation. Social media users reported that in Think mode, the model refused to reference these two figures.

According to TechCrunch, Trump and Musk had recently made false claims about Zelensky and the Ukraine war. The controversy escalated after some users managed to prompt Grok 3 into stating that Trump and Musk deserved execution, which xAI quickly corrected.

This incident has raised concerns about xAI’s content moderation policies. When unveiling Grok, Musk presented it as an unfiltered alternative to other AI models. Previous versions had shown hesitation on political topics and were perceived as leaning left. Musk has attributed this bias to the training data and has promised to steer Grok towards neutrality.

OpenAI has blocked access to its services, including ChatGPT, for some users in China and North Korea. The company has also removed several accounts that, according to its claims, were using AI technology for malicious purposes, such as surveillance and influence operations targeting public opinion.

According to Reuters, OpenAI utilized its AI tools to detect these activities but has not disclosed the number of accounts affected or the timeframe of the enforcement. Some users had reportedly used ChatGPT to generate anti-American news articles in Spanish. Additionally, a group of accounts linked to a financial fraud operation in Cambodia leveraged AI to translate content and generate comments on social media platforms like X and Facebook.

This move highlights OpenAI’s increasing efforts to monitor how its models are used and to counter potential misuse.

Solving 84% of IMO Problems with AI

DeepMind researchers at Google have announced that the AI system AlphaGeometry2 has improved its performance in solving International Mathematical Olympiad (IMO) problems, advancing from a silver medal level to gold. The first version of this AI, called AlphaGeometry, was introduced in January 2024. After several enhancements in its second version, its ability to solve complex geometric problems has significantly increased.

In a recent study, AlphaGeometry2 was tested on geometry problems from the past 25 years of the IMO. It successfully solved 84% of these problems, compared to just 54% in the previous model. Furthermore, in the IMO 2024 summer exam, the AI correctly solved 42 out of 50 problems, surpassing the average score of gold medalists.

Kevin Buzzard from Imperial College London believes that computers will soon be able to achieve a perfect score in the IMO.