AI & THE CRAFT
AI & THE CRAFT ENVIRONMENTAL UNKNOWNS AI isn’t a positive for the planet – that much we know is true. But how much energy does it actually use, and is it more or less than the existing workflows? B y now, it’s no secret that AI isn’t WORDS KATIE KASPERSON
fossil fuel to meet consumer demand. But as Junaid Baig, chief innovation officer at Dimension Studio, points out, “training an AI model is completely different from using an AI model.” Each time a model like ChatGPT answers a query, this is called an ‘inference’. While this has been found to consume more energy than a simple web search, it’s an arguably insignificant difference, Baig claims (although a World Economic Forum article suggests otherwise). “What we hear in the news is mostly about training, rather than inferring models,” he explains. ALWAYS ON In the VFX industry, data centres are referred to as ‘render farms’, which are
‘always on’ and ‘available to any studio’, according to Baig. “In a traditional VFX pipeline, you’re almost always carrying out rendering,” he explains. “Anything I want to present to my client has to go through the render farm. But with GenAI, we don’t have that concept – what we have is, while I’m working locally, I’m inferring on my own computer.” Baig recalls a recent animation project: Mara & Milo . “Traditionally, our pipeline would have taken about three weeks,” he shares. “An artist would model a character, shape that character, add hair and everything. Somebody would then light the character and render it, just to be able to present it.” On Mara & Milo , “we trained a very, very small model on that character. That saved more than two weeks’ worth of work.” Here, using AI was less impactful on the environment than rendering via traditional means – for several reasons. The model was trained and inferred on a local machine rather than in a data centre, preserving water and electricity. It also saved two weeks of time that would have otherwise been spent on carrying out an energy-intensive process. Others, such as Metaphysic (recently acquired by DNEG), are capitalising on GenAI. The tech company uses the tool to digitally de-age actors in real time, for example, which brings VFX a step forward in the production pipeline. In Baig’s eyes, AI doesn’t – and shouldn’t –
exactly a friend to the planet – and that using AI almost always does more environmental harm than good. According to an MIT News article, the problem lies largely in training GenAI models – an energy-intensive process that leads to increased carbon emissions and puts pressure on the electric grid. With OpenAI’s release of ChatGPT-5, conversations about sustainability are coming to the fore. Data centres, for instance, are essential for training LLMs, and require vast amounts of chilled water to keep their computers cool. While they support all kinds of computing tasks, training AI models sucks more energy, more water from local ecosystems and burns more
34
DEFINITIONMAGAZINE.COM
Powered by FlippingBook