xAI has officially lunched Grok 4 during a livestream with Elon Musk, who called it the "smartest AI in the world." He said that if you make the Grok 4 take the SATs and the GREs, it would get near perfect results every time and can answer questions it's never seen before. "Grok 4 is smarter than almost all graduate students in all disciplines simultaneously" and can reason at superhuman levels, he claimed.
Musk and the xAI team showed benchmarks they used for Grok 4, including something called "Humanity's Last Exam" that contained 2,500 problems curated by subject matter experts in mathematics, engineering, physics, chemistry, biology, humanities and other topics. When it was first released earlier this year, most models could only reportedly get single digit accuracy. Grok 4, which is the single agent version of the model, was able to solve around 40 percent of the benchmark's problems. Grok 4 Heavy, the multi-agent version, was able to solve over 50 percent. xAI is now selling a $300-per-month SuperGrok subscription plan with access to Grok 4 Heavy and new features, as well as higher limits for Grok 4.
The new model is better than PhD level in every subject, Musk said. Sometimes it may lack common sense, he admitted, and it has not yet invented or discovered new tech and physics. But Musk believes it's just a matter of time. Grok is going to invent new tech maybe later this year, he said, and he would be shocked if it doesn't happen next year. At the moment, though, xAI is training the AI to be much better at image and video understanding and image generation, because it's still "partially blind."
During the event, Musk talked about combining Grok with Tesla's Optimus robot so that it can interact with the real world. The most important safety thing for AI is for it to be truth-seeking, Musk also said. He likened AI to a "super genius child" who will eventually outsmart you, but which you can shape to be truthful and honorable if you instill it with the right values.
What Musk didn't talk about, however, is Grok's recent turn towards antisemitism. In some recent responses to users on X, Grok spewed out antisemitic tropes, praised Hitler and posted what seems to be the text version of the "roman salute." Musk did respond to a post on X about the issue blaming the problem on rogue users. "Grok was too compliant to user prompts," he wrote. "Too eager to please and be manipulated, essentially. That is being addressed."
This article originally appeared on Engadget at https://www.engadget.com/ai/elon-must-spent-almost-an-hour-talking-about-grok-without-mentioning-its-nazi-problem-061101656.html?src=rss
