OpenAI CEO Sam Altman on Friday defended the useful resource calls for of synthetic intelligence, calling issues about information facilities’ water use “faux” and evaluating the power utilized by AI techniques to that of people.
Altman was talking on the sidelines of the India AI Affect summit in an interview with The Indian Specific when he was requested to handle widespread criticisms of AI, akin to its power and water consumption.
The CEO responded that claims circulating on-line that ChatGPT makes use of gallons of water per question have been “fully unfaithful, completely insane,” and have “no connection to actuality.”
Knowledge facilities historically use massive quantities of water to chill electrical elements and forestall overheating. Whereas information middle cooling applied sciences have promised decreased consumption, some newer information facilities not depend on water in any respect.
Nonetheless, even with bettering effectivity, a report final month from water know-how firm Xylem and World Water Intelligence projected that the water drawn for cooling would greater than triple over the following 25 years as computing demand rises, placing stress on water techniques.
Whereas dismissing fears about water use, Altman stated power consumption stays a good AI concern. “Not per question, however in complete – as a result of the world is utilizing a lot AI … and we have to transfer in direction of nuclear or wind and photo voltaic in a short time,” he stated.
Requested about earlier feedback from Microsoft founder Invoice Gates — who has prompt that the effectivity of the human mind proves that AI can evolve to additionally turn into extra power environment friendly over time —Altman pushed again.
“One of many issues that’s at all times unfair on this comparability is individuals speak about how a lot power it takes to coach an AI mannequin … But it surely additionally takes a variety of power to coach a human,” he stated. “It takes like 20 years of life, and all of the meals you eat earlier than that point, earlier than you get good.”
“The truthful comparability is when you ask ChatGPT a query, how a lot power does it take as soon as a mannequin is educated to reply that query, versus a human, and possibly AI has already caught up on an power effectivity foundation, measured that approach,” he added.
The method Altman is referencing is called inference, which refers to using AI fashions which have already been educated to create new outputs. AI inference is often a lot much less power-intensive than the coaching itself.
Altman’s feedback, notably the AI-to-human comparability, have since sparked some debate on-line amid rising anxiousness about AI’s capability to switch human work.
Sridhar Vembu, co-founder and chief scientist of Indian software program firm Zoho Company, who was current on the summit, criticized the human-AI equivalence. “I don’t need to see a world the place we equate a bit of know-how to a human being,” the billionaire stated in an X submit.
The talk comes as governments and corporations pour billions into new information facilities to assist the computing wants of AI techniques.
In keeping with a Could report by the Worldwide Financial Fund, electrical energy consumption by the world’s information facilities in 2023 had already reached ranges similar to Germany or France, quickly after the launch of OpenAI’s groundbreaking ChatGPT AI mannequin.
In response, some governments have been working to hurry up approval processes to convey new and low-cost power on-line, with some environmentalists warning such strikes may conflict with world net-zero targets.
Some native communities in international locations just like the U.S. have additionally pushed again on growth tasks over fears they are going to pressure electrical energy grids and lift total electrical energy prices.
Final week, the Metropolis Council in San Marcos, Texas, voted down a proposed $1.5 billion information middle challenge after months of public opposition.
Amid such pushback, many tech leaders, together with OpenAI’s Altman, have argued information facilities would require extra power manufacturing from various sources, together with renewable and nuclear power.
