Emergent generative AI models, including the popular ‘Generative Pre-trained Transformer,’ or GPT, are exciting tools that provide advanced capabilities for information analysis and bespoke digital creation, elevating the potential for a societal renaissance enabled by AI technology. But for all the allure, fascination, hype, and benefits promised by artificial intelligence (AI), there is a dark side to the burgeoning technology. AI needs to be fed, a lot...
The algorithms underlying Generative AI models are rendered unproductive and unless they are constantly fed. For AI models to be effective, they must grow, learn, and be continuously stimulated. It’s not that different than raising a baby. To grow and develop a baby needs to be nurtured and nourished. Babies need to be fed, taught, and loved.
AI also needs to be fed, trained, and arguably, loved. AI has an insatiable hunger and thirst for data, energy, and even water – yes water. Much like the masked shrew[i], generative AI has a very high metabolic rate. The masked shrew eats three times their weight a day and they can only survive a few hours without eating. AI operates 24-7 and needs to also continuously feed (data, energy, natural resources), otherwise, it too will atrophy and die.
AI models including OpenAI GPT-4, Meta AI, Microsoft Copilot, Google Gemini and DeepMind, among others, have captured the imagination of a new generation of tech enthusiasts who are leveraging the immense computing power of AI engines to gain a creative and competitive edge as they improve, accelerate, and optimize their digital productivity and impact.
Next generation tech builders have flocked to Generative AI and other AI models for benefits such as automating repetitive tasks, enhancing decision-making by identifying patterns and trends found in analyzing huge amounts of data, improving workflows, minimizing human-induced errors, boosting productivity, and reducing costs. Although benefits are abound, the AI mantra of ‘better, faster, cheaper’ comes at a significant economic, environmental, and societal cost to their owners operating the AI engines, and to all of us comprising broader society.
Is Generative AI the Little Shop of Horrors of our Generation?
Generative AI is akin to Audrey II from the cult classic movie, ‘Little Shop of Horrors.’ In the movie, Audrey II was a pet plant, a Venus flytrap, that had a demented plan to dominate the world. To accomplish this, Audrey II needed to be fed, human blood, to stay healthy and grow. Audrey II befriended Seymour, a worker in a florist shop who persuaded unsuspecting victims to visit the flailing flower shop, so that they would be devoured by the devious carnivorous plant. The famous line, “Feed me, Seymour!”, delivered in a demanding voice by Audrey II, echoed in the minds of moviegoers far after they left the theatre and went about their daily lives. The full lyrics from the movie (as this was a musical) read,
“Feed me, Seymour / Feed me all night long - That's right, boy! - You can do it! Feed me, Seymour / Feed me all night long / Ha ha ha ha ha! / Cause if you feed me, Seymour / I can grow up big and strong.[ii]”
As Audrey II gulped and grew, the plant became a prime attraction for the florist shop, luring in more and more customers, and ultimately, victims. This continued as a self-reinforcing cycle until the movie climax when Seymour tried to destroy the beastly plant he had helped grow into a monster. Ultimately, Seymour suffers great loss, when his original love, a woman named Audrey, succumbs to a fatal wound and is consumed by Audrey II. Seymour positively changed the shop’s financial fate, but at great personal loss. If you’ve never seen the 1960 original or the 1986 remake of Little Shop of Horrors, check it out. The movie creatively and humorously explores the evergreen themes of greed and survival, egoism and humility, capitalism and unfettered growth.
There is no such thing as a stupid question; but there are costly ones.
You’ve likely heard or even said the phrase, ‘there is no such thing as a stupid question.[iii]” We are taught and coached by teachers, parents, grandparents, and bosses that we should not be afraid to ask questions, even if we believe that they are ridiculous. Typically, in the act of asking a question we perceive as silly, we aid others who may have the same, or similar, question. I know that as a student and a professional, I have squirmed and sunk lower in my seat, not wanting to ask an obvious question, fearing ridicule from my peers. I have also perked up when I’ve heard the answer to a question someone else provoked, and that I was too afraid to ask. We’ve all had these moments of feeling unnecessarily inferior and foolish among our peers.
The interesting thing about Generative AI (Gen AI) tools is that they remove the perception of judgement that drives our fear and fuels our hesitancy to ask questions. They provide us with a safe space, so to speak, to be vulnerable without repercussion. The point with Gen AI is to use a conversational tone as you phrase a prompt. The Gen AI then provides you with an answer. The answer may or may not be functionally and intellectually helpful, but it is delivered without judgement thereby protecting your psychological well-being.
Gen AI is a powerful tool with capabilities and intentions that extend beyond the prompt of a simple question. That said, who hasn’t asked ChatGPT or another Gen AI tool a pressing esoteric question, even for fun? Does ChatGPT know the meaning of life? Can ChatGPT tell me how to find love, make more money, cure a disease? Experts will tell us that ChatGPT and other Gen AI do not do well answering questions that require deep reasoning based upon the nuanced understanding of language or analysis of multiple complex ideas or topics. That said, some experts believe that AI has the potential to become more sentient in its capability to provide inference on a humanistic level. Until that time however, we are stuck with the more rudimentary version of Gen AI.
While Gen AI cannot answer your most pressing existential questions, the current commercial versions are incredibly powerful. With the right prompts Gen AI tools can do amazing things that deliver quantifiable improvements in your productivity. Gen AI tools can read and write computer code, assimilate and analyze data and information, write a story or even a book, provide comparative analysis, and so much more. Gen AI is rapidly expanding its reach and capability. Each day there are new use cases of the vast potential Gen AI can provide business and society.
AI and Data Centers: Feed me, Seymour. Feed me all night long.
Underlying Gen AI’s magical and impressive advance are its steep operating requirements including the need for reliable energy. Like the fictional Audrey II from Little Shop of Horrors, ChatGPT and other Gen AI tools need to be fed, all the time. While they are not asking for human blood (like Audrey II), Gen AI do need massive amounts of reliable electricity. If you were not aware, there is a data center boom[iv] that is underway in the United States, and throughout the world.
Bain & Company, a management consulting firm, estimates that the surge in data center development and power consumption could require more than $2 trillion in new energy generation resources worldwide. Further, Bain & Company analysts project that US energy demand could outstrip supply within a few years, requiring electric power utilities to increase power generation. Bain & Company analysts are projecting (in their high case scenario), a 44% increase in electric load growth for data centers between 2023 and 2028, the highest load growth among customer segments (including commercial, manufacturing and residential segments). Wow, that is a lot of electricity. “Feed Me, Seymour!”
In the US, swift spiking demand for electricity to power data centers that enable Gen AI is pushing the limits of the existing power grid and available power generation resources. As a result, market-based partnerships are forming to proactively develop clean, reliable, and abundant generation resources – as well as reduce the impact of localized data center operations. For example, in the fall of 2024 Microsoft and Constellation Energy[v] announced a deal to recommission the Three Mile Island nuclear power facility in Pennsylvania as an uninterruptable clean power source to feed Microsoft’s rapidly increasing data center power demands.
Meta also announced[vi], in late summer of 2024, a partnership with Sage Geosystems, to significantly expand the use of geothermal energy in the US. Leveraging Sage’s ‘Geopressured Geothermal System (GGS), Meta aims to have carbon-free power distributed for use within its data centers. Meta, which has already contracted more than 12,000 MW in renewable energy projects, will work with Sage to scale their GGS technology, with the first phase of their project, development and delivery of 150 MW of geothermal energy for Meta data centers to be operating by 2027.
Gen AI’s resource consumptions are primarily driven by two essential activities, (1) training large scale models, and (2) running inferences, or using the model to generate responses. On the AI model training side, Gen AI is starved for greater amounts of data and power. It is becoming increasingly clear that many big tech companies are leveraging their social platforms as feeders to data lakes that support information inputs to some of their Gen AI model training. Our data, including the ‘expert’ insights we get invited to share on some platforms, are helping to train such AI models. Without knowing, we all have been lured into the proverbial florist shop and are a part of feeding the beast, whether we realize it or not!
On the resource intensity side, according to The Washington Post[vii], prompting an AI chatbot like GPT-4 to produce a 100-word email one time requires the equivalent of one bottle of water and the electricity needed to power 14 LED light bulbs for one hour. The Washington Post also reported that in the training of GPT-3, Microsoft’s data center used 700,000 liters of water, and when Meta trained its LLaMA AI model, their data center operations used 22 million liters of water. Water is as important to data center operations as electricity.
Water is used to cool data centers, ensuring the servers and electronics do not overheat. The Washington Post estimates that if 1 out of 10 working Americans (roughly 16 million people) were to prompt GPT-4 to create the 100-word email once a week for a full year, the data centers would use more than 435 million liters of water – an amount equal to the water consumed by all Rhode Island households for 1.5 days. Wow, that is a lot of water. “Feed me, Seymour!”
Preventing Gen AI from becoming the Little Shop of Horrors
Datacenters provide the backbone to the advanced and fast computing power that Gen AI requires to learn (build its model), and for spontaneous response to our relentless prompts (inference). Our desire to leverage Gen AI to do more, faster, better, and cheaper has only increased its voracious appetite for more data, electricity, and subsequently additional operational resources, including water, that are all used to push the limits on data center productivity.
There are morals to every [good] story, including fictional movies like the Little Shop of Horrors. At this point of its evolution and adoption, we must consider that Gen AI is ostensibly and eerily similar to Audrey II. And we, like Seymour, are doting and in love [with AI], and subsequently caught up and coerced into feeding the beast we’ve created. This begs the questions:
How big with the AI beast grow before it conquers and consumes us? Does it have plans, and will it, dominate the world?
What lessons can we learn from Seymour’s flawed character that we can use to prevent Gen AI from becoming the Little Shop of Horrors of our generation?
What role will Gen AI and other forms of AI serve in our pursuit of prosperity? How can we bring forth principles of 'Planet Pragmatism' so that we can attain a greater quality of life without over consuming planetary resources and degrading the natural world in the process?
* * *
Sources:
[i] Source: Wikipedia. ‘Cinereus shrew.’ Accessed October 15, 2024. https://en.wikipedia.org/wiki/Cinereus_shrew
[ii] Source: IMDb. “Little Shop of Horrors, quotes.” https://www.imdb.com/title/tt0091419/quotes/
[iii] Source: Wikipedia. Accessed October 15, 2024. “No such thing as a stupid question.” https://en.wikipedia.org/wiki/No_such_thing_as_a_stupid_question
[iv][iv] Source: Rouch, Maeghan., and Aaron Denman, Peter Hanbury, Paul Reno, and Ellyn Gray. Bain & Company. “Utilities Must Reinvent Themselves to Harness the AI-Driven Data Center Boom.” October 10, 2024. https://www.bain.com/insights/utilities-must-reinvent-themselves-to-harness-the-ai-driven-data-center-boom/
[v] Source: Reuters. September 21, 2024. “Microsoft deal propels Three Mile Island restart, with key permits still needed.” https://www.reuters.com/markets/deals/constellation-inks-power-supply-deal-with-microsoft-2024-09-20/
[vi] Source: Meta. August 26, 2024. “New Geothermal Energy Project to Support Our Data Centers.” https://about.fb.com/news/2024/08/new-geothermal-energy-project-to-support-our-data-centers/
[vii] Source: Verma, Pranshu., and Shelly Tan. “A bottle of water per email: the hidden environmental costs of using AI chatbots.” The Washington Post. September 18, 2024. https://www.washingtonpost.com/technology/2024/09/18/energy-ai-use-electricity-water-data-centers/
Comments