Виталик Бутерин о Grok: «Безусловное улучшение» с оговорками Translation: Vitalik Buterin on Grok: An Unconditional Improvement with Caveats

The Grok chatbot from Elon Musk’s startup xAI is making the X platform more «truth-friendly.» This statement was made by Ethereum co-founder Vitalik Buterin.

*»The simple ability to consult Grok on Twitter is probably the most significant enhancement after Community Notes, which has positively impacted the platform’s commitment to truth,»* he remarked.

According to the programmer, the key advantage of the chatbot lies in the unpredictability of its responses.

*»I’ve often seen users turn to it for validation of their radical political stance, only to receive a harsh and sobering dose of reality in return,»* Buterin explained.

He referred to Grok as «an undeniable upgrade» for X but cautioned about potential risks. He noted that the model is further trained on the opinions of a narrow group of users and Elon Musk himself, adopting their perspectives.

In November, users *[noted](https://forklog.com/news/ai/grok-nazval-maska-krasivee-smeshnee-i-sportivnee-mnogih-znamenitostej)* the bias of Grok. The chatbot significantly overestimated the billionaire’s abilities. The large language model claimed Musk was stronger than Mike Tyson, more attractive than Brad Pitt, funnier than Jerry Seinfeld, and slimmer than Billie Eilish.

*»Billie has form and charisma, but Elon endures the pressures of managing multiple companies and fatherhood, which objectively makes him more resilient,»* the neural network stated.

Musk attributed such hallucinations to the influence of «hostile prompts»—intentionally distorted queries. The cryptocurrency industry saw this episode as an argument for the decentralization of artificial intelligence.

*»When the most powerful AI systems are owned by a single company, it creates conditions for algorithmic bias to become institutionalized,»* noted Kyle Okamoto, chief technology officer of the cloud platform Aethir.

In August, Grok *[leaked](https://forklog.com/news/ai/grok-slil-370-000-perepisok-s-polzovatelyami-v-google)* 370,000 conversations with users to Google, and in September, the chatbot was *[trained](https://forklog.com/news/ai/grokking-chat-bota-grok-nauchili-publikovat-skam-ssylki)* to post scam links.

In December, the model repeatedly *[spread](https://forklog.com/news/ai/grok-rasprostranil-dezinformatsiyu-o-perestrelke-v-avstralii)* misinformation about a mass shooting on Bondi Beach in Australia.

Recall that in May, Telegram and xAI *[established](https://forklog.com/news/ai/telegram-vstroit-grok-v-razlichnye-servisy-messendzhera)* a one-year partnership to distribute Grok among the messenger’s users and integrate it «into all applications.»