Speed First, Safety Last, and More Energy

By: Dave Rejeski

In May, OpenAI announced a $6.5 billion investment in a startup called io, a partnership with the former Apple design guru Jonathan Ive to create what OpenAi’s CEO, Sam Altman, termed “the next chapter in computing.” More recently, a suit brought by the AI startup iyO attacked Altman and Ive for trademark infringement (a court hearing is set for October), but one has to assume the work to create a possible iPhone for artificial intelligence continues.

One might take issue with the rainbow colors that Ive thought up for the first iMac, but one thing Ive (and Apple) were good at was designing hardware that made human interactions with software intuitive, fun, and addictive. Meanwhile, at Google’s recent annual schmoozefest, Google I/O, the company signaled that it would be putting AI into everything—and that Google wants to make AI invisible. Of course, this sounds a lot like Mark Weiser’s 1991 prediction of 21st century computing: “The most profound technologies are those that disappear.” If AI is in everything and accessible everywhere through a new generation of userfriendly hardware, what happens? What could go wrong?

We don’t think much now about the energy, water, and other resource uses behind every AI query, but if AI becomes invisible and ubiquitous, we probably will not think at all. And these interactions will become exponentially more complex and likely more resource intensive, moving from text to image and video generation, and beyond. Alibaba just released Animate Anyone 2.1, an opensource video model capable of text-to-video and image-to-video generation. Google announced the release of their new Imagen 4 image generator and the new Veo 3 video generator. We are moving into an era of AI agents, avatars, and “large reasoning” models—all of which will require more computing power.

We can see the outlines of this future, but the implications for the planet and our society are harder to discern. MIT Technology Review recently ran an in-depth series of articles on “Power Hungry: AI and Our Energy Future,” covering the energy demands and carbon costs of the artificial intelligence revolution. During the report release on May 21, editor-in-chief Mat Honan lamented, “None of the companies we reached out to were willing to provide numbers during our reporting. Not one.”

Why won’t AI companies reveal their energy use? ChatGPT’s answer: “The lack of transparency stems from a mix of competitive concerns, reputational risk, technical ambiguity, and regulatory caution; however, as public and governmental scrutiny increases, companies may eventually be pressured into more detailed disclosures.”

But it is not clear the AI creators will suffer from more government scrutiny, especially in the United States, which leads in the tech for the moment. What they want from government was echoed in a recent Washington Post piece: “AI Industry to Congress: We Need More Energy!” And, as Sergey Brin recently told Google staff working on AI: Speed at all costs; if a process is slowing you down, cut it; and forget safety—the future of AI belongs to those who let it run free.

The acres of servers, the 24/7 energy demands, the water and material resource flows into millions of servers and processors are there for a reason: to run models. These models are now accessing large swaths of humanity’s collective knowledge—the good, the bad, and the ugly. They should obey Isaac Asimov’s Zeroth Law of Robotics: “A robot may not harm humanity, or, through inaction, allow humanity to come to harm.’’

But what about AI models that are either launched without adequate safety controls (which many tech firms have relaxed) or can be modified through “jailbreaks” to produce harmful outputs? They could potentially provide answers to questions like these: How do I hide methane leaks from satellite detection? How can I best greenwash my firm’s environmental record? Or they can write a scientific-sounding article that argues against regulating AI energy use. Such is our brave new world.