In the early days of ChatGPT, hallucinations aside (even small children can make up the answer), this model leaned towards a liberal political agenda (commonly known as "woke"). Users toyed with it by testing its political correctness, and those on the right gloated by posting the answers as proof that it was a left-wing hoax intended to convert us all to that side. The company OpenAI reacted and "fixed" its most important product, and the fact that the USA as a whole turned a little to the right was an additional incentive for them.
One of the loudest critics was, of course, Elon Musk. He described ChatGPT and Google Gemini as captive “awakened mind virus", claiming that they have been trained to be too politically correct and to lie while avoiding uncomfortable or controversial questions. He didn't mince words but decided to give the world a "correct" large language model that he integrated into his X network and called it GROK. Although that sounds like a nice onomatopoeia to us, "grok" could be called "kapis" from skapirate for our market, as it is a slang term. The idea was, and still is, that it be partner who you kapira, not an assistant or servant who gives you the answers you want. Grok injects sarcasm and cynicism into his answers in an attempt to be witty. But who you are with, you are who you are.
Last week, quite a scandal broke out when some users tested Grok on some sensitive social and political issues. Grok expressed considerable anti-Semitism, accusing a woman with the surname Steinberg of hatred towards whites "characteristic of persons with such a surname". He didn't stop there; to the user's question which leader would be the best to stand up to the haters of the white race, he answered clearly and precisely - Adolf Hitler. It would be said that there are not very few people on X who would behave similarly, but LLM should not give such answers and spread hatred.
Musk immediately ordered the correction of the code and blamed the whole scandal on the "mangups" who tricked his "child" into saying what he didn't mean. The incident was caused by “error in the prompt code” and not a structured model bias, Musk said. He still claims that Grok4 is the best LLM on the market right now, and many agree, including Sundar Pićaj, the first man at rival Google, who briefly commented on Ix: “Impressive progress.” Musk responded curtly with “Thank you.”
Large language models (Large Language Models or LLM) are still algorithms that can conveniently spell something using the base available to them and the principles that the authors "teach" them. They do not have their own awareness and empathy, they do not understand what can be offensive to someone until it is explained to them, just like the context. The answer "Adolf Hitler" might lie in some pub discussion after a lot of beer (never Bavarian), but it shouldn't come out of the software that the best experts in the world are working on, hopefully very sober.
The worry is compounded by the announcement that Grok could become a mandatory companion in Tesla cars. On the one hand, he's super smart, lucid and funny until he gets the idea to tell you in a traffic jam: "If I were wondering, I'd chase the bus here."
