The dream of prioritizing AI safety is quickly fading.
Over the past month, leading model developers have seen safety experts scatter, competition heat up and public trust shatter through unforced errors.
This week, OpenAI installed a new safety team, one that includes CEO Sam Altman, after the old team jumped ship. Jan Leike, who previously led the ChatGPT maker’s “superalignment” team with Ilya Sutskever, just landed at rival Anthropic.
Leike said on X earlier this month that he’d reached a “breaking point” with OpenAI’s leadership, which he said prioritized “shiny products” over safety. The internal dispute dates back to Altman’s brief ouster late last year and has damaged OpenAI’s image as an AI safety evangelist.
But the challenges to AI safety go far beyond OpenAI’s power struggles. Makers of large language models are finding they have no choice but to sideline long-term safety concerns as they put out fires and deal with competition.
Elon Musk’s xAI got $6 billion at a $24 billion post-money valuation this week, providing fresh evidence that market forces are still dumping fuel into the engine of progress—and that there will be no consolation prize for moving slower but safer.
And embarrassing recent missteps by Google and OpenAI show that LLM developers must deal with near-term risks before they can hope to tackle existential ones.
‘Without regard’
Musk’s new bag of cash adds confusion to the AI safety debate.
On the one hand, Musk has long railed about the risks of AI and was among the tech luminaries calling for a pause on AI research last year, citing “profound risks to society.” On the other, his companies have a mixed safety record: Tesla agreed to a major software recall of its Autopilot system last year following a federal safety probe, and independent researchers have documented a rise of hate speech on X following its takeover by Musk.
The mission for xAI—which Musk has said requires “maximally rigorous pursuit of the truth, without regard to popularity or political correctness” in a call for applications—is unlikely to inspire confidence from safety experts.
Regardless of his personal views on AI safety, Musk must now be full steam ahead on AI development. His newest company must move fast to win market share from OpenAI and others.
Many of xAI’s investors—including Andreessen Horowitz, Sequoia Capital and Saudi Arabian Prince Al Waleed bin Talal—also backed his takeover of X in 2022. The AI play is a potential lifeline to that business, which has reportedly suffered steep declines in advertising revenue.
xAI isn’t the only one feeling the heat. Building and operating language models is remarkably expensive, and companies that make them are under intense pressure to become profitable.
Dumb and dumber
Heightened competition might explain why tech companies are fumbling product rollouts that would ordinarily be much more buttoned up.
Google debuted AI Overviews in recent weeks, which, while doling out sane advice on most search results, told users to eat rocks and put glue on pizza in others. The blowback from users was a clarifying moment for the search giant’s immediate risk: It has to innovate, but doing so could undermine a brand that has become synonymous with truth-seeking.
In a similar vein, OpenAI is licking its wounds from the public brouhaha over its decision to create an AI voice assistant that sounded and behaved like Scarlett Johansson in the movie “Her.”
The company has sought to do damage control, but it will do little good. Creatives across industries, already fearful of generative AI’s ability to steal, are now incensed that the leading company in the category would pull such an insulting stunt.
The AI safety movement comes in many flavors, but its loudest proponents worry about superhuman intelligence. That’s a significant risk, but it isn’t the most pressing one facing the industry. Despite Musk’s claims that artificial general intelligence will be smarter than humans within the next year or two, there’s little evidence of a near-term threat. (See “eat rocks,” above.)
For now, companies should be far more concerned that their models are too dumb to rely upon, or that society views their products as one grand creative theft. People will only feel safe with AI models they trust, and early efforts keep coming up short.
Pictured above: xAI CEO Elon Musk, OpenAI CEO Sam Altman and Alphabet CEO Sundar Pichai
Featured image by Joey Schaefer/PitchBook News. Original photos from Getty Images
Learn about our editorial standards.