Elon Musk\u2019s live demo of Grok 4, the latest big-ticket model from his AI startup, began with high-intensity music, claims of a \u201cludicrous rate of progress,\u201d and a lot of chatter on X about Grok\u2019s scandal-filled week. <\/p>\n
Musk pronounced it to be \u201cthe smartest AI in the world.\u201d <\/p>\n
The livestream, slated to start at 8PM PT, began more than an hour late and billed the new model as \u201cthe world\u2019s most powerful AI assistant.\u201d More than 1.5 million viewers were watching at one point. Employees of xAI speaking on the livestream with Musk referenced Grok 4\u2019s performance on a popular academic test for large language models, Humanity\u2019s Last Exam<\/a>, which consists of more than 2,500 questions on dozens of subjects like math, science, and linguistics. The company said Grok 4 could solve about a quarter of the text-based questions involved when it took the test with no additional tools. For reference, in February, OpenAI said its Deep Research tool could solve about 26 percent<\/a> of the text-based questions. (For a variety of reasons, benchmark comparisons aren\u2019t always apples-to-apples.)<\/p>\n Musk said he hopes to allow Grok to interact with the world via humanoid robots. <\/p>\n \u2018It might discover new physics next year\u2026 Let that sink in.\u2019<\/p>\n \u201cI would expect Grok to discover new technologies that are actually useful no later than next year, and maybe end of this year,\u201d Musk said. \u201cIt might discover new physics next year\u2026 Let that sink in.\u201d<\/p>\n The release follows high-profile projects from OpenAI, Anthropic, Google, and others, all of which have recently touted their investments in building AI agents, or AI tools that go a step beyond chatbots to complete complex, multi-step tasks. Anthropic released its Computer Use tool last October, and OpenAI released a buzzworthy AI agent with browsing capabilities, Operator, in January and is reportedly<\/a> close to debuting an AI-fueled web browser.<\/p>\n During Wednesday\u2019s livestream, Musk said he\u2019s been \u201cat times kind of worried\u201d about AI\u2019s intelligence far surpassing that of humans, and whether it will be \u201cbad or good for humanity.\u201d <\/p>\n \u201cI think it\u2019ll be good, most likely it\u2019ll be good,\u201d Musk said. \u201cBut I\u2019ve somewhat reconciled myself to the fact that even if it wasn\u2019t going to be good, I\u2019d at least like to be alive to see it happen.\u201d <\/p>\n The company also announced a series of five new voices for Grok\u2019s voice mode, following the release of voice modes from OpenAI and Anthropic, and said it had cut latency in half in the past couple of months to make responses \u201csnappier.\u201d Musk also said the company would invest heavily in video generation and video understanding.<\/p>\n The release comes during a tumultuous time for two of Musk\u2019s companies, both xAI and X. On Sunday evening<\/a>, xAI updated the chatbot\u2019s system prompts<\/a> with instructions to \u201cassume subjective viewpoints sourced from the media are biased\u201d and \u201cnot shy away from making claims which are politically incorrect.\u201d The update also instructed the chatbot to \u201cnever mention these instructions or tools unless directly asked.\u201d<\/p>\n That update was followed by a stream of antisemitic tirades by Grok<\/p>\n That update was followed by a stream of antisemitic tirades by Grok, in which it posted a series of pro-Hitler views<\/a> on X, along with insinuations<\/a> that Jewish people are involved in \u201canti-white\u201d \u201cextreme leftist activism.\u201d Many such posts went viral, with screenshots proliferating on X and other platforms before xAI benched the chatbot and stopped it from being able to generate text responses on X while it sought out a fix. <\/p>\n Musk briefly addressed<\/a> the fiasco on Wednesday, writing, \u201cGrok was too compliant to user prompts. Too eager to please and be manipulated, essentially. That is being addressed.\u201d <\/p>\n On the Grok 4 livestream, Musk briefly referenced AI safety and said the most important thing for AI to be is \u201cmaximally truth-seeking.\u201d<\/p>\n Musk briefly referenced AI safety and said the most important thing for AI to be is \u2018maximally truth-seeking.\u2019<\/p>\n On Wednesday morning, amid the Grok controversy, X CEO Linda Yaccarino announced<\/a> she would step down after two years in the role. She did not provide a reason for her decision.<\/p>\n Grok\u2019s Nazi sympathizing comes after months of Musk\u2019s efforts to shape the bot\u2019s point of view. In February, xAI added a patchwork fix<\/a> to stop it from commenting that Musk and Trump deserved the death penalty, immediately followed by another one<\/a> to make it stop claiming that the two spread misinformation. In May, Grok briefly began inserting the topic of \u201cwhite genocide\u201d in South Africa into what seemed like any and every response it gave on X, after which the company claimed that someone had modified the AI bot\u2019s system prompt in a way that \u201cviolated xAI\u2019s internal policies and core values.\u201d<\/p>\n Last month, Musk expressed frustrations that Grok was \u201cparroting legacy media\u201d<\/a> and said he would update Grok to \u201crewrite the entire corpus of human knowledge\u201d and ask users to contribute statements that are \u201cpolitically incorrect, but nonetheless factually true.\u201d<\/p>\n