AI Training Wastes 30% of Its Power—Meet Perseus for the Fix
AI training wastes up to 30% of its energy. Perseus optimizes usage without sacrificing performance, shaping a more sustainable AI future.
#6 Your Daily-ish Knowledge Dose
Hi everyone!
I've always found joy in learning or thinking about at least one new thing each day.
It's a small but happy habit that keeps my curiosity alive and my mind engaged. That's why I started a new section to the blog: "Your Daily-ish Knowledge Dose."
In this space, I'll be sharing ideas, and tidbits that spark thought and inspire deeper understanding.
I hope these daily doses of knowledge will enrich your day just as much as they do mine.
And if there's something you're curious about or a topic you'd love to dive into, feel free to share your thoughts—let’s explore together!
P.S. If this post sparked your curiosity, hit the ❤️ button—it helps more curious minds discover this journey!
AI Training Wastes 30% of Its Power—Meet Perseus for the Fix
It’s not the tools or technologies themselves that are inherently good or evil—it’s how we choose to use them that defines their impact.
While AI research is advancing at lightning speed, much of the discourse around it tends to lean toward pessimism. I believe it’s equally important to focus on the positive applications and opportunities it offers.
Artificial intelligence holds extraordinary promise for addressing some of humanity’s most pressing challenges, from optimizing supply chains to accelerating climate research.
Take the climate crisis, for example—a global challenge we’ve been grappling with for decades, often with insufficient progress. In this context, remaining open to all possible assistance, including from technologies like AI, seems not only practical but wise.
But this progress comes with a significant environmental cost: AI’s enormous energy consumption. Fortunately, a new study from the University of Michigan highlights a striking inefficiency—up to 30% of the energy used to train AI is wasted.
This isn’t just an abstract issue. By 2027, data centers could account for 1.2% of global carbon emissions, according to the IMF. Moreover, AI's increasing reliance on these centers could exacerbate water scarcity as cooling demands surge. The solution proposed in the study, however, is as elegant as it is impactful: Perseus, a software tool that reduces energy waste without sacrificing speed or accuracy.
For a deeper dive into the intersection of AI, data centers, and sustainability, check out my article AI Data Center Boom and Renaissance of Sustainability Tech, where I explore these issues in detail.
The Root of the Problem
Training AI models like GPT-3 requires dividing tasks among thousands of processors. These tasks can’t be distributed evenly, much like organizing a library where some shelves hold the entire Encyclopedia Britannica while others hold only a few slim novels. When processors with lighter loads finish early, they waste energy waiting for others to catch up.
Even worse, faulty hardware or network delays can slow down some processors, creating bottlenecks that further squander resources.
The University of Michigan’s researchers identified these inefficiencies and developed Perseus, a tool that slows down less critical processors so all tasks complete simultaneously. This reduces unnecessary energy use while maintaining the same training speed.
The Impact of Perseus
If widely adopted, Perseus could save enough energy to power 1.1 million U.S. homes by 2026.
Beyond its energy-saving potential, it also addresses a critical equity issue in AI.
High energy costs make cutting-edge AI inaccessible to countries with limited power resources, forcing them to rely on smaller, less accurate models or outsource computation to distant locations.
Perseus could help narrow this gap, ensuring that AI technology benefits everyone, not just the most resource-rich nations.
Perseus optimizer in action
My Take: A Sustainability Imperative
The findings of this study are both a wake-up call and a beacon of hope. They reveal how much room there is to improve AI’s sustainability.
While some argue that AI’s environmental costs are outweighed by its potential to combat climate change, that doesn’t justify unnecessary waste.
As the researchers rightly point out, “Why spend something when there’s no point?”
AI has the potential to be a game-changer for sustainability, but only if we approach its development responsibly.
Tools like Perseus remind us that technological progress doesn’t have to come at the expense of our planet.
By reducing waste and optimizing energy use, we can extend the benefits of AI to more communities while mitigating its environmental footprint.
A Bigger Question for AI’s Future
This study also raises a philosophical question: Are we building AI systems to solve problems, to perpetuate unsustainable systems, or to prepare for the rise of dominant cyborg successors as we intoxicate ourselves with power?
We’ve been exploring these questions extensively in the “Thought Seeds Lab” series. If you’re curious, you can dive into these ideas here.
The immense energy demands of AI training reflect a broader trend of prioritizing scale and speed over thoughtful innovation. Perseus challenges this status quo, proving that efficiency and progress can go hand in hand.
In the end, this isn’t just about AI—it’s about how we choose to innovate. Do we want to create tools that extract as much as possible from our planet, or can we reimagine technology as a partner in building a sustainable future?
For me, this is the real promise of Perseus. It’s not just software; it’s a step toward redefining what responsible innovation looks like in the 21st century.
And if we take more steps like this, we may yet harness AI’s transformative potential without compromising the planet we call home.
Thanks :)
P.S. If this post sparked your curiosity, hit the ❤️ button—it helps more curious minds discover this journey!
Enjoying my writing?
I’m keeping this newsletter free for everyone—your support can help keep the journey going strong.
Wanna read more?
#5: Top 5 Positive Sustainability News & Insights - Biweekly
Stay updated with the top 5 positive sustainability news highlights from the last two weeks, accompanied by my insights.
Explore the rise of Green Economic Growth, TNFD's momentum, the EU’s 8.3% emissions drop, Belgium's offshore wind innovation, and Microsoft's bold carbon removal efforts.
#5 TSL: Are We the Bridge to the Next Step in the Cosmos?
Thought Seeds Lab is like a journey—each post has its own story to tell, but when read in sequence, they bloom into something even richer.
COP16 Biodiversity Summit: Key Wins and Unresolved Issues
From October 21 to November 1, nearly 200 countries, NGOs, and companies gathered in Cali, Colombia, for COP16, the first biodiversity summit since the Kunming-Montreal Global Biodiversity Framework (GBF) was adopted in 2022.
In the Mycelium Matrix: Mushroom Hunter’s Insights
Mushrooms 101, insights, a weekend journey into nature's hidden networks...
My mushroom hunt brought me face-to-face with this underground network’s dazzling fruit. I found mushrooms of many forms and hues, each telling a story of its unique role in the ecosystem.
Stay tuned for the next blog post coming in a few days!
Until then, take care!
Very interesting!
Also I think switching to renewable energy makes this less of an issue anyway.
AI energy usage is only an issue if we have limited energy supply or use fossil fuels. As we transition to using more renewable sources of energy or even nuclear, how much energy is used by AI is I think not really an issue.
AI energy use I’d argue is not a problem when we have abundant renewable energy?
So in a way, the AI energy use isn’t even the fundamental problem now, the problem now is how quick we can transition to renewables. Only if we are slow, or for some reason we don’t transition to renewables is AI energy use an issue. But all major countries are transitioning to renewable energy.
I’d say AI energy use is less of a problem than many others, for example AI energy use is tiny compared to the energy use on transportation or domestic energy.
So I think it’s helpful to understand AI energy use not in isolation which can be misleading, but relative to other uses of energy, which are far worse.
And this is only a temporary problem anyway until we transition to abundant renewable/ nuclear energy 🙂👌