\n\tBy SHANNON O. WELLS<\/p>\n
\n\tWhen it comes to the rapid rise of artificial intelligence in society and academia, the more that many Pitt students, staff and faculty learn about how it functions \u2014 and its implications on knowledge, career fields and energy usage \u2014 the more concerns, apprehensions and questions they seem to have.<\/p>\n
\n\tThat was among the takeaways from \u201cWhat Should I Do About AI: A Pitt Community Workshop on AI and Energy\u201d forum on Sept. 2 in the William Penn Union Assembly Room.<\/p>\n
\n\tModerated by David Sanchez, associate professor in the Swanson School of Engineering\u2019s Civil & Environmental Engineering Department and associate director for the Mascaro Center for Sustainable Innovation, event speakers, all from the Dietrich School of Arts & Sciences, included:<\/p>\n
\n\t\t\tMichael Blackhurst, teaching assistant professor and environmental studies program coordinator in the Department of Geology and Environmental Sciences<\/p>\n
\n\t\t\tJustin Kitzes, associate professor in the Department of Biological Sciences<\/p>\n
\n\t\t\tSungwan Hong, assistant professor in the Department of Economics<\/p>\n
\n\tThe interactive event also invited audience members to share their thoughts through question-and-answer and workshop sessions.<\/p>\n
\n\tAfter hearing detailed overviews from the speakers on the definition of AI, how it\u2019s being used and why it requires considerable energy to maintain, audience members were invited to engage with those at their tables and share their topics and ideas with the room. \u00a0<\/p>\n
\n\t\u201cEven though a lot of us came up with environmental concerns, we can see some clear solutions to those,\u201d a spokesperson from one table said. \u201cThis cultural piece kind of comes into it where we\u2019re like, \u2018What is the next generation of our world going to look like if they don\u2019t have to depend on critical thinking and other non-AI skills or pre-AI skills \u2026<\/p>\n
\n\t\u201cIf we create renewable energy solutions that help cut back on the energy concerns with data centers, would that just expand the amount of AI, and I guess strip people from even more of their human qualities?\u201d<\/p>\n
\n\tAnother table of participants questioned how data centers that fuel AI systems could be built sustainably. \u201cCan we lessen impacts? Can we build them in different places that perhaps could reduce the need for (water-based) cooling, or build them in a modular fashion as well?\u201d<\/p>\n
\n\tOther questions addressed AI\u2019s effect on cognitive development and creativity, and the possibility of representatives from different countries coming together in a summit to share concerns and possible solutions \u201cabout how could we handle this.\u201d<\/p>\n
\n\t\u201cIs this something that we could have a Montreal Protocol or a Paris Accords?\u201d the table\u2019s spokesperson pondered. \u201cCould we have a global solution arise, which of course sounds pretty challenging, given the geopolitics today.\u201d<\/p>\n
\n\tDeeper philosophical questions also arose. \u201cAre we even asking for this?\u201d the spokesperson said of AI. \u201cWhy are we doing this? Who is it if we are not asking for this? Who is imposing this upon us?\u201d<\/p>\n
\n\tThe retraction of personal freedoms and personal liberty also was mentioned, along with the potential for the explosion in AI-related investment to become a dangerous economic \u201cbubble.\u201d<\/p>\n
\n\t\u201cThere\u2019s a lot of fear about how AI\u2019s been overhyped, and you can\u2019t even control money in your own pockets, especially if any of you are interested in retirement or 401(k)\u2019s \u2026 with stocks being involved in the Nasdaq and things like that,\u201d one table\u2019s spokesperson said. \u201cAnd AI being used to steal data, not only from things online but also from (surveillance cameras).\u201d<\/p>\n
\n\tWhile the afternoon gathering did not produce hard and fast conclusions or simple solutions, the forum was clearly effective in stimulating thought and discussion based on insights from some of Pitt\u2019s foremost authorities on AI.<\/p>\n
\tPredicting words<\/p>\n
\n\tJustin Kitzes, one of the event\u2019s three featured speakers, reminded the audience that \u201cAI is old.\u201d<\/p>\n
\n\t\u201cThe concept of artificial intelligence is at least 70 years old in academia. It\u2019s not a new thing, and it has a fairly broad definition,\u201d he said. \u201cAI is widely understood as a system, usually a computer system, that can perform a task that otherwise would require human intelligence.\u201d<\/p>\n
\n\tThe evolution of AI in the past decade or so has been spurred by a subset of \u201call possible AI methods that are possibly out there, which is the subset of AI that is machine learning.\u201d<\/p>\n
\n\tThis is AI that can learn to perform a task directly from data, rather than being explicitly programmed.<\/p>\n
\n\t\u201cThink of machine learning as learning to tell images of a cat from images of not a cat by being given images of cats and not cats and figuring it out on its own,\u201d he explained. \u201cThat would be in contrast to making a model where you specify that a cat has two ears, a cute face, an indifferent personality, and anything else that you might decide to tell it.<\/p>\n
\n\t\u201cWithin machine learning, you have (the subset of) deep learning that is based around a particular type of model called a neural network. Within deep learning, there is a particular type of neural network called a transformer, and just one application of transformers are large language models.\u201d<\/p>\n
\n\tThese, he explained, are models behind the popular chatbots that \u201cmany of you are using today: ChatGPT, Gemini, Claude, etc.\u201d<\/p>\n
\n\tKitzes elaborated that AI is used as shorthand for large language models (LLMs), which have what he called a \u201cdictionary of tokens\u201d or word fragments \u2026 maybe 100,000 of them that it can read or say. (The LLM) reads a string of these tokens \u2026 and then generates a probability distribution for what word it thinks should come next in that string.\u201d<\/p>\n
\n\t\u201cThat\u2019s not the only type of AI out there, but it is the type leading to a lot of the impacts and development and discussion going on today.\u201d<\/p>\n
\n\tKitzes emphasized that possibly negative effects of AI should always be considered within other impacts and demands placed on the environment through \u201cour other everyday actions.\u201d<\/p>\n
\n\t\u201cThe cost or impact of an activity, what we would consider acceptable or desirable, always needs to also be contextualized against the potential benefits that that cost may be bringing you,\u201d he noted. \u201cI imagine many of us could be potentially supportive of a large data center that was built dedicated to cancer research, versus a large data center that was dedicated solely to generating images of rabbits on skateboards.\u201d<\/p>\n
\tEfficiency vs. demand<\/p>\n
\n\tIn his presentation, Mike Blackhurst outlined an analysis of the software and hardware LLMs require and the energy required to support them.<\/p>\n
\n\t\u201cI\u2019m sure you\u2019ve all heard stories about how data centers may be influencing the electric power system in terms of price and emissions, so we want to help you understand those connections as well,\u201d he said, noting that data centers, like other energy converting processes, generate waste heat. \u201cThey lose heat as they operate on your data.\u201d<\/p>\n
\n\tMuch like a heating boiler or water heater, Blackhurst said data centers also generate waste heat in converting \u201cchemical energy and natural gas or electricity into heat. \u2026 There\u2019s also waste heat in the sum of the generators that are used to supply electricity.\u201d<\/p>\n
\n\t\u201cIf we burn coal or natural gas or nuclear power, that\u2019s burning something to create electricity, and it\u2019s also generating heat.<\/p>\n
\n\t\u201cOn the other hand, renewable energy doesn\u2019t generate waste heat,\u201d he noted. \u201cSo, if we can use that, we have less waste heat to deal with in the power system.\u201d<\/p>\n
\n\tThe significant need for cooling water at data centers is another source of controversy.<\/p>\n
\n\tOf what he called the \u201ctwo choices for coolant in treating your waste heat,\u201d water requires less energy than air-based cooling systems.<\/p>\n
\n\tBlackhurst cited ways to make the process more efficient and less consumption heavy, including more efficient hardware to convert an \u201cinference\u201d \u2014 asking a model to respond to a prompt as opposed to training it \u2014 into a service.<\/p>\n
\n\t\u201cThe chips (would) use less energy to do that,\u201d he noted. \u201cWe can also have more efficient algorithms. \u2026 It\u2019d be nice if, as we get more efficient, we use less energy.\u201d<\/p>\n
\n\tHowever, some of those efficiency gains are \u201cbasically being cannibalized by AI itself into new services. So, as we make the process more efficient, people are using it more aggressively and finding new ways to use AI.<\/p>\n
\n\t\u201cThis is a very dynamic and uncertain space, but one that I think we should be keeping track of as an institution doing research and teaching about these things,\u201d he said.<\/p>\n
\n\tNoting that the Grand Coulee Dam on the Columbia River in Washington state is the biggest power generator in the U.S., Blackhurst said AI-based \u201chyperscale facilities are starting to get as big as our biggest power generators.<\/p>\n
\n\t\u201cWe have the power draw from about 800,000 homes.\u201d<\/p>\n
\tNew generation<\/p>\n
\n\tIn his presentation, Sungwan Hong highlighted the rising electricity demands from data centers. Based on statistics he shared from 2024, AI and non-AI data centers accounted for 5% of total U.S. electricity use, with AI data centers consuming only a 1% share.<\/p>\n
\n\tBy 2030, however, many projections suggest the share of consumption from data centers will rise to around 12%, he said, with more than half of that coming from AI data centers.<\/p>\n
\n\t\u201cThis 6% of electricity demand is not a small number. It\u2019s the total amount of California\u2019s electricity consumption per year. So, once we have this huge demand shock in electricity, now the question is how would the electricity supply respond?\u201d<\/p>\n
\n\tIf a data center opens tomorrow, the grid operator cannot build a new supporting generator, as they typically take more than a year to build, he noted.<\/p>\n
\n\t\u201cThen we (must) use the current generators more intensively, burn more natural gas or burn more coals, pour on more nuclear, and so on. On the other hand, if we have two years, we already know that in four years we will have a lot of data there \u2026 In the long run, we may start to think about building more generators.\u201d<\/p>\n
\n\tAnother question surrounds the source of fuel, whether we are \u201cexpecting to have more fossil fuel burning in the short run, or going to have more renewables? \u2026 In the long-run period, are we going to have more solar panels or wind turbine fuels in the economy, or (will we) have more fossil power plants that will drive (emissions-related) data?\u201d<\/p>\n
\n\tWith generators able to run on nuclear, renewables, natural gas or \u201cpeakers\u201d \u2014 plants that run only in high-demand situations \u2014 once they are already built, renewables \u201ctend to be the cheapest one to run because there\u2019s no marginal cost\u201d of running them.<\/p>\n
\n\tWhile there are variables in natural gas plant efficiencies that affect the cost, \u201cthey will be more expensive compared to nuclears,\u201d Hong said.<\/p>\n
\n\tResponding to the increasing demand response varies by area.<\/p>\n
\n\t\u201cSome regions respond by more renewables, and some respond by more fossil fuels.\u201d In the Pittsburgh region, \u201cOhio is going to respond with a lot by coal, and Pennsylvania responds more by natural gas. Virginia responds by natural gas, solar and wind.\u201d<\/p>\n
\n\tDespite the regional differences, fossil fuel, Hong explained, is \u201cgoing to work a lot to meet\u201d data-center demands. And where data centers are planned, new fossil-fuel generators will follow, particularly in Pennsylvania, Ohio and West Virginia. \u201cExcept Texas, where the companies are proposing building their own gas turbines.\u201d<\/p>\n
\n\tWith more data centers, therefore, \u201cwe would expect to have more gas turbines rather than renewables.\u201d<\/p>\n
\n\t\u201cHyperscale\u201d users like Microsoft, Amazon and Google have a net-zero consumption commitment in which they \u201ctry to cancel out the environmental impact they make \u2026 by purchasing clean renewable electricity generation from other places to meet the electricity use in their production process,\u201d Hong noted, with one caveat. The region from which the clean electricity was purchased can be very different from where Google data centers are actually running.<\/p>\n
\n\tIn 2024, Google used 30 kilowatts per hour and reported that it matched 100% of its electricity use with clean power contracts.<\/p>\n
\n\t\u201cThey also reported that only 66% was carbon-free on the grids they serve in its data centers. \u2026 They can have a data center in Pennsylvania, which is not 100% carbon-free in their grid, but they are buying electricity from Texas.\u201d<\/p>\n
\tAmeliorating fears<\/p>\n
\n\tDuring the forum\u2019s question-and-answer segment, Blackhurst responded to an audience question about young people being told their jobs will be replaced by artificial intelligence while they\u2019re footing bills for water and electricity use of the new data centers, he said he hoped \u201cyou\u2019re not afraid of it.<\/p>\n
\n\t\u201cAnd I would hope that (the University) would help you prepare to be successful financially, professionally beyond your time here. I think there\u2019s a lot of people here, myself included, who really want to prepare you accordingly.<\/p>\n
\n\t\u201cI think abstaining from it, even if you\u2019re fearful of it, is probably not going to be helpful. \u2026 This is a very safe space for you to dabble,\u201d he noted. \u201cYou have a lot of people with life experience that can help you adjust to that reality here. But I don\u2019t know that I would encourage anybody to be afraid per se, and I would hope we could help ameliorate those fears here in this institution.\u201d<\/p>\n
\n\tNoting that he and other panelists have growing children, Justin Kitzes acknowledged that the influence and \u201cdisruption\u201d from AI on the job market is \u201con the mind of everybody.\u201d<\/p>\n
\n\t\u201cI mentioned the Industrial Revolution. If you look back at other major technological changes that have brought about large social change \u2014 electrification, steam engines, cars, the internet. Some of us remember when the internet was brand new and it was disrupting everything and changing the way that we did everything.\u201d<\/p>\n
\n\tWith the once-strong job-training appeal of digital coding now being usurped by AI, Kitzes admitted the difficulty in predicting future trends.<\/p>\n
\n\t\u201cRemember \u2018Just learn to code\u2019? That probably wasn\u2019t the right answer, it turns out. \u2026 In a time of disruption, one of the best things that you can do for yourself is to be flexible and to put yourself in a position where you can go in many different directions.<\/p>\n
\n\t\u201cBecause quite frankly, we don\u2019t know exactly what\u2019s going to happen, and anyone who says they do is full of it, right?\u201d he added. \u201cProbably the best thing you can do for yourself is consider, \u2018Well, if things went all sorts of different directions, what could I do to give myself a place, no matter what happened\u2019?<\/p>\n
\n\t\u201cRather than trying to find the one option that you think is going to optimize for some future that may or may not come to pass. That\u2019s my personal answer.\u201d<\/p>\n
\n\tShannon O. Wells is a writer for the University Times. Reach him at shannonw@pitt.edu<\/a>.<\/p>\n \n\t\u00a0<\/p>\n \n\tHave a story idea or news to share? Share<\/a> it with the University Times.<\/p>\n