Amrith Ramkumar at The Wall Street Journal ( Apple News), writing about a new partnership between the Trump administration and his on-again, off-again bestie Elon Musk:
Under the agreement with the General Services Administration, which oversees technology procurement for the federal government, agencies will get access to models such as Grok 4 and a new fast version called Grok 4 Fast for a nominal fee of 42 cents, the GSA said in a news release Thursday.
The deal follows similar arrangements with Alphabet’s Google; the ChatGPT maker, OpenAI; and Anthropic, a model developer that has clashed with the White House. The arrangement means the government is now working with the four U.S. companies making the most-advanced AI systems. OpenAI and Anthropic agreed to provide their models for $1, while Google is charging 47 cents. Musk likes the number 42 because it is a reference to “The Hitchhiker’s Guide to the Galaxy,” the sci-fi novel.
Hayden Field reported on this for The Verge a couple of weeks ago, recapping just some of the controversies around xAI:
The height of Grok’s power, up to now, has been posting answers to users’ queries on X. But even in this relatively limited capacity, it’s racked up a remarkable number of controversies, often resulting from patchwork tweaks and fixed with patchwork solutions. In February, the chatbot temporarily blocked results that mention Musk or President Trump spreading misinformation. In May, it briefly went viral for constant tirades about “white genocide” in South Africa. In July, it developed a habit of searching for Musk’s opinion on hot-button topics like Israel and Palestine, immigration, and abortion before responding to questions about them. And most infamously, last month it went on an antisemitic bender — spreading stereotypes about Jewish people, praising Adolf Hitler and even going so far as to call itself “MechaHitler.”
Musk responded publicly to say the company was addressing the issue and that it happened because Grok was “too compliant to user prompts. Too eager to please and be manipulated, essentially.” But the incident happened a few weeks after Musk expressed frustration that Grok was “parroting legacy media” and asked X users to contribute “divisive facts for Grok training” that were “politically incorrect, but nonetheless factually true,” and a few days after a new system prompt gave Grok instructions to “assume subjective viewpoints sourced from the media are biased” and “not shy away from making claims which are politically incorrect.” Following the debacle, the prompts were tweaked to scale back Grok’s aggressive endorsement of fringe viewpoints.
The whack-a-mole approach to Grok’s guardrails concerns experts in the field, who say it’s hard enough to keep an AI system from veering into harmful behavior even when it’s designed intentionally, with some measure of safety in mind from the beginning. And if you don’t do that… then all bets are off.