Using AI for boring things and saving the world
If you’ve read this newsletter for a while, you've probably figured out I'm a bit allergic to hype and am always keen to see where we can make use of data and AI in deeply practical ways.
That's why I was intrigued by Phaidra, an AI company focussed on what at first glance might seem to be a 'boring thing' - how do we make industrial plants run more efficiently? Why is that not at all boring? Because more efficiently means by using way less energy and that (not rouge killer AI) is the problem of our generation.
Also, Phaidra was founded by folks who had actually already walked the talk in the two domains that would seem key to success in this endeavour - AI for efficient energy use (reducing Google data centre energy usage by 40%) and industrial manufacturing. So I loved this peek inside their early world
(a) what did they worry about right at the beginning?
(b) how did they de risk?
Turns out the answer to (a) is
If the engineers who ran the plant didn’t trust Phaidra, they might shut it off. It wasn’t enough for Phaidra’s A.I. to work — the human operators needed to know why it worked.
I LOVE that. OK this was a founder team with the big advantage of an actual track record in a fundamentally tricky part of what they were trying to do, but I love that they were focussed on the human - AI interaction that was going to be pivotal for their success. No hubris, no comments on median humans. Instead a clear understanding that it’s not good enough to be right, you have to be right and successfully symbiotic with the humans in the equation to actually make useful change.
And the answer to (b)? How did they de risk and avoid spending pointless and expensive months building what might work? A tight and disciplined design sprint.
And the key questions to answer in that design sprint?
Would people understand and trust what Phaidra was doing?
Could the software demonstrate the business value of Phaidra’s artificial intelligence?
Fast forward to today and take a look at how the Phaidra website sells the product on first touch.
OK, I’ll forgive them the AI Copilot bit - after all, all founders need to speak to the VCs as well. But scroll down and what do you see?
Control systems, safety, transparency. And then right at the bottom if you can read it in this screen capture - security. Yup, they’re selling to the ‘most important customer’.
The most important customer type, (Katie) decided, were energy management engineers who were responsible for the cooling systems in their facilities.
And funnily enough, this most important customer is someone who actually might have a fair crack at understanding the algorithmic smarts that the Phaidra software is built on. So you could bang on about how great your algorithms were and how it was pretty likely that they’d transition to AGI one day real soon.
But instead of that, what might make them come back for a second look? Addressing their safety and reliability concerns and presenting the solution as augmentation not replacement. Smart. And hopefully for us all, compelling and successful in reducing the energy requirements of industrial manufacturing by a fair whack!
While we’re on an energy and climate track, two more of my favourite places where we can use data and AI for practical good right now. Harnessing the wind and grid stability.
Machine learning and wind farms
As a child, I visited Maori pa sites, hydro dams and wind farms. Yes on family holidays. My father is just that kind of geek. In my head therefore, renewable energy sources have been around forever and are an important part of the familiar landscape.
Now, years later, it delights me that it’s possible to use data and machine learning to increase the output and decrease the cost per MWh of the growing global fleet of wind farms, on shore and off shore. From weather forecasting and wind analysis, to efficiency optimisation and drone based turbine inspection, to maintenance optimization and site selection, people are working hard on incorporating AI to push efficiency up and costs down. Predictive maintenance makes enormous amounts of sense - extremely expensive assets where downtime means immediately reduced capacity, you would want to be all over any way to reduce that.
I was also fascinating to see the application of neural nets to optimisation of floating offshore wind turbines. This paper is pretty dense for a lay person but I think what it concludes is that we may be able to both boost the output and reduce wear and tear on turbines by using neural net based active platform stabilization.
AI, the energy transition and grid stability
And finally, that gnarly and deeply worrying problem of electricity distribution on aging grids that were built WAY before we had massive renewable energy sources, rooftop solar and the associated challenges with demand response - short hand for balancing the demand on power grids by encouraging customers to shift electricity demand to times when electricity is more plentiful or other demand is lower.
As we head on past 1.5C average warming, we all know we need to turn off coal fired powered stations ASAP. But we can’t do that unless we become a lot better at multi way flows of generated energy (for instance neighbourhoods or adjacent businesses sharing power that they’ve generated) and of course, energy storage because the wind stops blowing and the sun sets.
AI will play a crucial role maintaining stability for an electric grid that’s becoming exponentially more complex with large numbers of low-capacity, variable generation sources like wind and solar coming online and two-way power flowing into and out of houses.
Jeremy Renshaw, senior program manager at the Electric Power Research Institute
Another interesting startup I came across is WeaveGrid, a software platform that looks to make it easier to integrate the huge influx of electric vehicles into our existing electricity grids. A challenge because all those new EVs need a really significant amount of power and an opportunity because of all the battery storage they represent.
The cloud-based software pulls data from utilities and directly from vehicles and uses machine learning to predict things like what driving patterns mean about when someone is likely to charge, or how quickly the numbers of EVs are growing in a particular area and what that might mean for capacity limits on transmission lines.
Adele Peters, FastCompany
Now tell me that isn’t all much more interesting than selling advertising!