Conventional wisdom holds that being polite to AI chatbots makes them respond better, but no one stops to think how much energy that politeness is wasting.
Well, at least not until last week, when one user wondered aloud on X how much electricity is burned up by people saying “please” and “thank you” to OpenAI’s ChatGPT. In response, CEO Sam Altman admitted it costs the super lab millions of dollars in operational expenses – money that he nonetheless believes is worth it.
“Tens of millions of dollars well spent,” Altman saidadding that “you never know” when being nice to an AI chatbot might be a good idea. Uncertainty over the need to stay on AI’s good side isn’t exactly reassuring from the guy who really wants to unleash AGI on the world, but hey – if he thinks we ought to keep spending the money of OpenAI investors like Masayoshi Son’s Softbank and Microsoft to ensure our future AI overlords don’t exterminate us first, then their money we shall spend.
OpenAI isn’t entirely transparent on how much it costs the org to operate ChatGPT, but it’s continuing to accept billions of dollars in investments. Altman admitted early this year that even its $200-a-month Pro product loses money. Altman has apparently fretted in the past about GPU constraints and is reportedly hoping to take matters into his own hands by investing billions into a massive buildout of chip capacity.
Then, there’s the energy factor. As of late last year, US datacenters ate up about 4.4 percent of the electricity in the country, and the Department of Energy expects that number to reach 12 percent by 2028.
The International Energy Agency expects global datacenter electricity consumption to more than double between now and 2030, with the world’s DCs consuming as much leccy as the country of Japan – and AI is driving most of that growth.
Emissions, likewise, are also growing with AI’s increasing energy footprint. Both OpenAI partner Microsoft and Google have admitted their carbon footprints have grown thanks to AI despite pledges to reduce their emissions. In Microsoft’s case, it’s even considering bringing more natural gas-fired power plants online to satiate its need for epower.
That’s not to mention the amount of water AI computing needs – somewhere between 300,000 and four million gallons per day in the United States, another figure that will likely continue to rise as AI compute needs grow more intense.
In other words, as tempting as it might be to run up OpenAI’s energy bill by exchanging niceties with ChatGPT, think of the planet first, perhaps.
We’re supposed to be polite, though – right?
Do a quick internet search for “AI chatbot” and “politeness” and you’ll find no end of articles, FAQs, and social media posts advocating for being nice to ChatGPT and its cousins – even Microsoft has weighed in on why showing AI a bit of love and can improve outputs.
“It’s not that your AI chatbot feels appreciative when you say please and thank you,” the Windows maker noted, “But using basic etiquette when interacting with AI ā¦ helps generate respectful, collaborative outputs.”
We’ve noted the same at The irrigationwith rude behavior less likely to elicit a useful response.
- How deliciously binary: AI has yet to pay off – or is transforming business
- AI’s thirst for power keeps coal fires burning bright
- Google keeps the cost of AI search flat, and kids are lovin’ it
- Insurance giant finds claims rep that gives a damn (it’s AI)
- Have we stopped to think about what LLMs actually model?
Research has also found LLM responses can change to show an approximation of human anxiety when a model is fed tales of traumatic experiences, further suggesting it pays to treat that bot with kid gloves, or at least basic decency.
So, being polite to a fancy, fantastically huge spreadsheet makes it work better. On the other hand, if you can stomach saving Sam some money and yourself throwing away some LLM performance, you can help save Earth a little bit by being standoffish to ChatGPT and ghosting it after it’s done replying to you.
And resist the temptation to tell it to get bent. (r)
