Your Weekly Disruption: AI Has Finally Woken Up to Its Power Problem

I'm James Hinton, and this is The Disruption Continuum: an independent, non-sponsored, subscriber-supported blog about AI, technology, and business.

Are you new here? Get free emails to your inbox 4 times a week. Would you rather listen to content? You can find my podcast here.

Todays Read: 7 Minutes

What I cover

AI Macro View: My opinion on the macro trends shaping the Age of AI and why it matters to you.
AI News Roundup: A summary of the major news events that happened during the week.

How AI is Used to Create This Content

In the age of AI, generating authentic content is more important than ever. Being clear about where, how, and why AI is used ensures readers like you have content that represents my genuine human opinion.

I use AI to:

Provide editing suggestions for text (i.e. Grammarly, not including GenAI components)
Test any tools mentioned in the news that I have time for
At times, generate images (because I'm definitely not a good graphic designer, and also, you can do some pretty hilarious things with it)

I do not use AI to:

Generate entire newsletters, paragraphs, or long-form content. This doesn't mean you wont see the odd '-' but that's because I want it there. Not because it was put there by a random GenAI tool. Also, if that's your way of seeing if something is generated by GenAI, then you really need to level up your understanding of AI #rant
Provide opinions on things I am writing
Auto-summarize news, blogs, and videos (unless it is to test a new tool and see how accurate it is)

This week's opinion piece: AI has finally woken up to its power problem

Early in my electrical engineering degree, I learned the importance of testing and modeling. I quickly learned that any time you can effectively model and instrument a real-world outcome using software, you make your life easier.

Not only can you prototype good, bad, and frankly ugly ideas faster, but you can also save yourself tons of time as your project gets more and more complex.

However, I also learned that creating these modeling engines can be a pain in the *ss - especially at the start.

You have to create all sorts of different inputs and outputs.

Each input and output has to model a specific part of your environment. It also needs to record what happens before, during, and after each event so that when something inevitably goes wrong, you know what happened.

Then, each input and output has to interact with all the other inputs and outputs seamlessly. It doesn't take long before you start feeling like you're spending more time modeling the outcome than you are actually creating it in the first place.

Even though you eventually learn that the time spent now doing this modeling will save you billions of hours later, it's still an inordinately tedious, annoying, and frankly soul-destroying process. It's probably why so many software developers loathe it (while also secretly knowing how necessary it is).

On top of that, you also have a bunch of physical constraints you often have to work with. It's pretty pointless creating a wind tunnel simulator to test your latest speed-racer design where the wind only 'blows' in one direction.

That's not real life where you can be driving along the highway, minding your own business and then BAM - a 100 km/h (sorry, 60 mph for my American friends) gust of wind comes out of nowhere.

The problem is that by the time you've modeled and instrumented a truly accurate model, you end up with incredibly complex mathematical calculations.

Sometimes, these calculations can be so complex that they take computers the size of my house days and days to calculate, and cost so much that the average entrepreneur working out of their basement can't access them anyway.

In fact, one of my professors once told me that the amount of processing power required to instrument testing software effectively can be over five times more than what it takes to run the actual product!

AI Makes These Challenges Seem Insignificant

The thing is, as frustrating as I personally found building effective modeling systems for whatever project I was working on, they're nothing compared to the problems the AI ecosystem is facing.

It's like the comparison between an ant shaking its hand at the giant human foot about to crush them like a, well, ant.

Large Language Model (LLM) AI systems (basically ChatGPT, Claude, Gemini, etc) are incredibly powerful. Whatever your opinion is on Artificial Neural Networks, it's undeniable that they can generate some pretty amazing things.

You can generate whole blog articles. They might read a bit bland, and sound a bit canned, but they can also give you direct answers to your question, saving you from crawling through ten billion blue links, 50% of which are useless SEO stuffed blog articles.

The problem is that LLM's consume a crazy amount of energy. So much so that entire electrical grids are being put under strain as they struggle to keep up with the sheer volume of requests people are making.

Think about that for a minute. ChatGPT was launched in November 2023, and now we're looking at entire electrical grids being put under strain.

Entire data centers are being delayed because electrical grids can't handle the strain.

The problem gets worse. Imagine you're an AI company looking to cash in on this amazing new technology.

So you put your best engineers to work on the problem, and they come up with an amazing new frontier model (a fancy name for a new AI model trained from scratch).

You proudly release it to the world, and because people are pretty smart and a little bit mischievous, they create a whole new subculture just waiting for 'your' AI to do something dumb.

Proving that Murphy's Law is still relevant, within days of your brand new release your fancy new AI system that cost you billions of dollars recommends that a user use glue to stick their pizza cheese to their pizza base.

Oops.

Why Can't We Just Instrument AI?

Now, it's easy to pile on that company, point fingers, and bemoan the enshittification of the internet, blah blah blah.

The thing is, all the other LLM providers out there were probably sitting in their boardrooms thinking,

"Thank goodness that wasn't us."

Of course, we all know that SOMETHING HAS TO BE DONE.

So, some high-up person with a fancy title like vice-president, has a bright idea. They ask what they think is an obvious question:

"Why can't we just change how it's programmed so that it only gives accurate answers? Let's just look at the logs and fix it!"

You might be thinking that too.

It sounds pretty obvious.

Forget about the fact that Artificial Neural Networks don't really tell you how they think because they use techniques with fancy names like 'Deep Reinforcement Learning' and 'Generative Adversarial Learning.'

Forget about the fact that AI, just like any computer system, is only as good as the information that it is fed into it.

Forget about the fact that there's a whole world of people, all with their own ideas about what's right and wrong. Instead, just do it.

Now, not only do you need the power to run your LLM, you also need the power to instrument it properly.

Which, of course, requires more 🔌

The smart people who build LLM's get to work.

They figure that while they don't want to be the arbiters of right and wrong (or maybe they just don't want the lawsuits), they can at least instrument it.

At least that way, you can see the exact moment when it goes off the rails and starts suggesting crazy things.

And if you get upset? Well, the logs are there, and we can just slap a beta tag or legal disclaimer on it.

So, models like OpenAI's o3 model are born.

And earlier models are retrained and recalibrated, and hopefully, things are good.

So far, so good. AI instrumented ✅

The Problem is Power

The thing is, now you've got an even bigger problem.

While you've been beavering away on your fancy new models, the rest of the world has woken up to the power of AI.

They're thinking to themselves: "These advances in AI are pretty great. What if I go down to my local pub (or coffee shop) and have a yarn with my mates, and together we can come up with a new idea?"

Sure enough, they do.

A couple of drinks or coffees in, out comes an iPad, and some plans are drawn up.

Then, because we've got all these amazingly smart people building all these incredible prototype tools, you might even be able to bust out your laptop and build it right there. I mean, check out Claude's AI-powered apps, released this week.

You can literally tell it in words what you want, and it will build a prototype of the whole thing for you. I'm even thinking about creating an episode on my YouTube channel about how I saved $50,000 and six weeks' work using it in one afternoon.

No Joke.

All of a sudden, all over the world, there are people going to coffee shops and pubs and restaurants and coming up with these incredible ideas and testing them right then and there.

Want to write an article, but you're not a copywriter?

Bust open ChatGPT and you have the bones of the article done in about 30 minutes, and probably the whole thing done in an hour. Amazing.

The problem is that some data center is processing every single one of these events. Every single time you or I bust open ChatGPT, we're consuming more and more power. Our electricity grids are already struggling with demand, yet new use cases, powered by ever-more-capable AI models, continue to be released.

This week alone, we had AI-powered education software released by Gemini (Google), which claims to radically improve learning experiences for children.

We had another Alphabet company, DeepMind, release a product called Weather Lab that claims to be very effective at cyclone prediction (it's still in Beta, so don't replace your normal warning process), and some interesting research from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) that uses GenAI to optimize robot design.

All these new products, with all these people using it, and meanwhile our power consumption is skyrocketing.

AI might be amazing, but it's pretty useless if the outcome means no 🔌

A Power Solution

With all this in mind, you can imagine my delight when I read about an AI company that's actually trying to do something about AI's Power Problem.

Instead of moaning and pointing fingers at whatever government agency they feel is responsible, they've actually put their heads down and brains to work on solving AI's power problem with...AI. Their tagline is pretty cool too -> Powering the AI revolution. Lol.

Anyways, the company is called Emerald AI, and what they do is employ a whole bunch of really clever techniques at the data center level to scale up and down AI workloads.

For instance, one of the things they figured out is that there are a whole bunch of background tasks that are useful for AI, but don't need to be immediately useful. It's kind of like taking those overnight database syncs that used to be the bane of every IT person's existence and applying it to AI.

If that sync could happen dynamically and seamlessly across an entire data center, and by overnight, you mean any time the grid looks like it’s under strain.

Their value proposition is basically this: When there's no power in the grid, we can scale down your workloads until there's power left, and then when the power comes back, you can scale back up again. Without you lifting a finger.

To prove their point, and presumably get their next venture capitalist funding injection, they went ahead and did a test. When they did this, they found that AI data centers could scale down their power usage by 25% during peak periods.

Put into perspective, they estimate this would free up 100 gigawatts (a flipping massive amount) of electricity consumption, which would equate to around $2 trillion in data center investment.

Of course, all of this information came from an Nvidia Blog post, so take it with a grain of salt. However, it's well worth a read, and definitely a positive sign.

Why This Matters to You

Here's the thing.

Very few people in the world need to invest any time trying to figure out the relative impact of 100GW of electricity being saved by some new startup.

It's a number that is so far out of our day-to-day experience that its borderline farcical to worry about it.

I get that.

However, I would argue that in many ways, this is similar to the change that happened when cars first got released.

Today's modern infrastructure of smooth, sealed roads and petrol stations would have been impossible to imagine for people raised in the era of horses and carriages.

For our ancestors poking around on horses, carriages, and feet, it's not like people didn't have roads back then. However, the sheer scale of our network of roads would have been inconceivable for all but the most outrageous visionaries.

I tend to think of AI in the same way.

Now, instead of building and sealing roads, we have to think about ways to efficiently and effectively power this amazing revolution that's come upon us.

If we want to build a future where any person on the planet can have an idea and quickly build it, then we MUST solve AI's power problem.

And that's why this article was my article of the week.

Top Five Interesting AI Things in the News

Your A.I. Term Glossary & Cheat Sheet

Still learning what all of these terms mean?

I get it.

It’s a lot if you’re not always actively keeping up with it.

Use this as your way to stay in-the-know about everything I talk about in this newsletter.

AI Terms Table

TERM	DEFINITION
AI (Artificial Intelligence)	A field of computer science focused on building systems that can perform tasks typically requiring human intelligence, such as understanding language, recognizing images, or making decisions.
LLM (Large Language Model)	A type of AI model trained on massive text datasets to understand and generate human-like language. Examples include ChatGPT, Claude, and Gemini.
Artificial Neural Network	A computing system inspired by the human brain's structure, consisting of interconnected layers of nodes ("neurons") that learn to recognize patterns and solve problems.
Deep Reinforcement Learning	A training method where AI learns to make decisions by interacting with an environment and receiving rewards or penalties, gradually improving its behavior over time.
Generative Adversarial Learning (or Generative Adversarial Networks, GANs)	A technique involving two neural networks, a generator, that creates content and a discriminator that evaluates it - that compete with each other to produce more realistic outputs.
Frontier Model	A term used for the most advanced or cutting-edge AI model developed by a company, often trained from scratch with new techniques or data
GenAI (Generative AI)	AI systems that can create new content, such as text, images, audio, or code, rather than just analyzing existing data.
On-Device AI	AI models designed to run directly on local devices (like smartphones or laptops), rather than requiring cloud or data center processing, which helps with privacy and reduces power needs.
AI-Powered Apps	Applications that integrate AI capabilities to perform tasks automatically or more intelligently, such as writing assistance, design, or data analysis.
Instrumenting AI	The process of monitoring and tracking how an AI system makes decisions or outputs results, often through logs and other data, to ensure accountability and improve reliability.
Data Center	A large group of networked computer servers used to store, process, and distribute large amounts of data, including running AI models.
AI Factory	A concept referring to AI-driven systems or facilities that can continuously process data, generate outputs, and adapt at scale, similar to a physical factory producing goods.
Emerald AI	A company that develops systems to reduce the power consumption of AI workloads in data centers by dynamically scaling tasks.
MIT CSAIL (Computer Science and Artificial Intelligence Laboratory)	A leading research lab at MIT focused on computer science and AI research and innovation.
Nvidia Blog	A content platform from Nvidia (a major AI and GPU company) that often shares updates and research related to AI and computing.