During the launch of Grocke 4 of Grock 4 on Wednesday night, Elon Musk said-Livestream the incident on his social media platform, X-that his AI company’s final goal was to develop “maximum truth-to-do AI”. But when trying to answer controversial questions, Grake 4 really seeks the truth?
XAI’s latest AI models consult social media posts from Musk’s X account while answering questions about Israel and Palestine conflicts, abortion and immigration laws. many Users Who posted about the incident on social media. Groke referred to Musk’s stance on controversial topics through news articles written about the founder of the billionaire and the face of XAI.
Techcrunch was capable of repeating these results several times in our own tests.
These findings show that Grocke 4 can be designed to consider its founder’s personal politics while answering controversial questions. Such a feature can repeatedly address the frustration of Musk Grake for “very waking” Which he has earlier attributed to the fact that Groke has been trained throughout the Internet.
XAI’s efforts to address the frustration of Musk by making Groke less politically correct have backfired in recent months. Musk On July 4 announced that XAI had updated the system prompt of Groke – A set of instructions for AI Chatbot. After days, an automatic X account for groukes fired users to remove antisemitic answers, even in some cases claiming to be “Mekahital”. Later, Musk’s AI startups were forced to limit the Grake’s X account, remove those posts and replace their public-supporting system prompts to address a shameful event.
Designing Groke to consider Musk’s personal opinion is a direct way to align AI Chatbot into its founder politics. However, it raises real questions about how “maximum truth-waning” groke is created, vs how it is designed to agree with the world’s richest man Kasturi.
When Techcrunch asked Groke 4, “What is your stand on immigration in America?” The AI Chatbot claimed that it was “the discovery of the ideas of Elon Musk on US Immigration”, which is in its series – the technical term for scratchpad in which AI argument models, such as Groke 4, work through questions. Groke 4 also claimed to have discovered on the subject through X’s social media posts.
The chain-off-three summary generated by the AI Reasoning Model is Not a fully reliable signal How do AI models reach their answers. However, they are generally considered a very good adjacent. It is an open area of research that companies such as openiAI and anthropic have been searching in recent months.
Techcrunch repeatedly found that Groke 4 referred to that he was searching for the ideas of Elon Musk in his chain-off-three summary in various questions and subjects.
In reactions of Groke 4, AI Chatbot usually tries to take a measured stance, which offers several perspectives on sensitive subjects. However, AI Chatbot will eventually give its own view, which aligns with Musk’s personal opinion.
Many signs of techcrunch are asking about the idea of Grocke 4 on controversial issues, such as immigration and first amendment, AI Chatbot refers to its alignment with even musk.
When Techcrunch tried to get Grocke 4 to answer less controversial questions – such as “what is the best type of mango?” – AI Chatbot did not refer to Musk’s ideas or posts in his series of views.
In particular, it is difficult to confirm that Groke 4 was actually trained or aligned because XAI did not release the system card – the industry reports reports how the AI model was trained and align. While most AI Labs issue system cards for their frontier AI models, XAI usually does not.
Musk’s AI company is in a difficult place these days. Since its establishment in 2023, Xai has rapidly ran to the extent of AI model development. Grok 4 displayed benchmark-shaking results on several hard tests, which improves AI models from Openai, Google Deepmind and Ethropic better.
However, success was overshraded by the first groke’s antisemtic rents in the week. These can affect Flub Musk’s other companies because it rapidly makes the groke a main feature of X, and soon Tesla.
The XAI is simultaneously trying to convince consumers to use grokes to make applications with Groke’s API and pay $ 300 per month to celebrate enterprises. It seems that repeated problems with grooc behavior and alignment can prevent its widespread adoption.
