Helena is a major investor in Phaidra, a company using deep-learning artificial intelligence to autonomously operate some of the most energy-intensive and complex industrial processes in the world. In 2022, we invested around ~$11m into Phaidra and began our partnership.
The brainchild of former Google DeepMind alumni, Phaidra marks a new leap in applied artificial intelligence. It uses advanced reinforcement machine learning to analyze elaborate data sets and improve performance in real time.
In 2016, Google used the precursor of Phaidra’s technology to reduce energy consumption in its own data centers by 40%.
Phaidra is now applying this to a wide-range of industrial processes with profound implications for global energy usage, from the production of vaccines to the heating and cooling of cities.
In addition to capital, Helena is providing strategic support to facilitate widespread adoption of a technology we believe represents a new frontier of AI – one that will dramatically augment human capabilities across diverse fields and applications in the coming years.
On March 16, 2016 in Seoul, South Korea, the development of advanced AI reached an apex on the board of a 4,000 year-old game called Go. AlphaGo, a program designed by a London-based machine learning company called DeepMind, had just beaten Lee Sedol, an 18-time world champion, in a five-match tournament, four games to one.
On the surface, this would not appear to be news. Computer programs had been outmatching human players in gaming scenarios for almost two decades when the AlphaGo victory occurred. In fact, the moment had its earliest antecedent in the supercomputer Deep Blue’s 1997 victory over world chess champion Garry Kasparov. Chess, however, is a “closed game,” meaning there is a large but finite set of possible moves in any given situation.
To beat Kasparov, Deep Blue’s designers had supplied it with a huge number of previously played games. For each of Deep Blue’s moves, it searched its database for similar situations, then played the move that gave it the best odds of winning based on its data. Deep Blue’s computational power was impressive (it could evaluate close to 330 million positions a second), but it did not actually understand the game of chess. It couldn’t develop novel tactics, creatively problem-solve, or improvise. In computer science terminology, Deep Blue could only use brute force.
Go, however, is not a closed game. Once considered one of the “four arts of the Chinese scholar,” it rewards strategy and imagination. There are more possible moves than atoms in the universe, so it is impossible for a program to brute-force its way through it. According to DeepMind CEO Demis Hassabis, “if you took all the computing power in the world and ran them for a million years, that wouldn’t be enough compute power to calculate all the possible variations.” What DeepMind did instead was load AlphaGo with the rules of the game and a small data set – just 10,000 games played between amateur players. The team then designed a neural network that allowed AlphaGo to not only recognize, but learn.
Then AlphaGo played. It played thousands of games, both against human opponents and itself. And, as it played, it improved. This is called reinforcement learning. By the time AlphaGo played Sedol, it could use creativity and guile, could improvise and innovate. With deep learning, AlphaGo showed that advanced AIs can surpass humans even in complex, multivariable decision-making.
Since AlphaGo’s success, DeepMind has advanced its deep learning capabilities. The current iteration is MuZero, which can master board games like Go and chess, as well as visually complex Atari games, all without being told the rules. As determining constraints has become part of its learning process, human intervention has become less and less necessary–even counterproductive.
The point of Deep Blue was to win at chess. DeepMind’s goal for AlphaGo was more ambitious. Hassabis described the intention like this: “our mission is to fundamentally understand intelligence, and recreate it artificially. And then once we’ve done that, we feel that we can use that technology to help society solve all sorts of other problems.”
In 2019, two DeepMind employees decided to do just that. Jim Gao, a Team Lead at DeepMind Energy, and Veda Panneershelvam, who was on the AlphaGo team – along with Katie Hoffman, an Innovation Senior Manager at Ingersoll Rand – decided to use their collective expertise to combat climate change through industrial automation.
The industrial sector is the world’s largest energy end-user. According to the U.S. Energy Information Administration, the industrial sector accounted for 33% of total U.S. energy consumption. The IEA puts the number at 30% globally, including 42% of the world’s electricity and 37% of its natural gas consumption. And as the global population grows and further industrialization occurs, the energy consumption necessary to maintain these growth trends will only increase.
This situation is untenable. We are already in a global energy crisis. The world’s energy demand is already putting a great strain on its natural resources, with some groups estimating we have roughly 50 years left of both oil and natural gas. And current renewable energy infrastructure, an absolute necessity to mitigate climate change, does not have the capacity to keep up with the forecasted growth in demand. For context, according to a 2020 Princeton study, to address just the U.S.’s energy needs renewably – until new energy sources are powering the grid – would require a land area equivalent to Wyoming and Colorado combined, just for wind and solar farms. From every angle, the world needs to become more energy efficient.
But industrial processes are remarkably bad at efficiency.
As industry has become more complex, its machinery and networks have become larger and more sophisticated, exceeding the capacity for traditional, human methods of oversight. Take, for example, a typical paper digester in the paper and pulp industry. The amount of data it produces is staggering: inflow volume and admixture, temperature, chemical consumption, steam pressure.
Almost all of these outputs require monitoring in real-time as the digester operates. A single person – or, at best, a small team of people – is in charge of analyzing all of this data, and then making informed, live decisions on how to optimize the system. But human computational power is severely limited. What industrial processes need, then, is a program that can take all that data, analyze it, learn from it, and decide best courses of action. They need an advanced, self-learning AI.
Phaidra designs AI tools for complex industrial systems that are customizable to the industry and plant, that use reinforcement learning to adapt and improve, and that are autonomous, so the changes and modifications can be made in real-time.
Phaidra already has a high-profile success story. Created and led by Phaidra founder Jim Gao, the DeepMind Energy team developed an AI to monitor and optimize Google’s data center cooling.
Data centers are a huge energy consumer; in 2020, they accounted for 1% of global electricity use. The Energy AI began in 2016 as a recommendation system; the team designed a system of deep neural networks to analyze and recommend to its human operators modifications and adjustments. The results were tremendous: 40% reduction in energy used for cooling, and a 15% reduction in overall Power Usage Effectiveness for the center as a whole.
But that was just the beginning. In 2018, they made the system autonomous, which meant it could self-optimize. Within a matter of months, the system was consistently delivering around 30% energy savings.
Google’s data centers were already some of the most efficiently run centers in the world. They had brilliant people monitoring them, with streams of data and metrics. They were, in a sense, the Lee Sedol of data center operations. Which means the improvements Gao’s DeepMind Energy team made were essentially improvements on peak human capability: its AI was 30% better than the best humans after just a few months.
A year later, in 2019, Gao, Panneershelvam, and Hoffman launched Phaidra.
Phaidra builds on DeepMind’s advances in AI and applies them to the industrial sector, including applications in pharmaceuticals, paper and pulp manufacturing, chemical manufacturing, and data center cooling. It provides a closed loop control system that is configurable to each customer’s plants and needs. It uses a learned model of the plant operation to evaluate billions of possible actions, then implements the best one automatically.
The results for the customer: lower energy costs (Phaidra targets 15-30%), a more stable process (up to 70%), and less equipment runtime (up to 50% reduction). Plus, the system will improve the longer it is in use.
Helena was the largest investor in Phaidra’s $25 million series A round of funding and is currently supporting the company’s commercial expansion. At global scale, these results vector toward the exponential, leading to a marked reduction in energy consumption in the world’s most energy intensive sector while liberating resources – and critically, limited renewable resources – for use in other arenas.
Helena’s purpose is to identify solutions to global problems and implement them through projects. Each project is a separate, unique effort.
Sometimes, we believe that the most effective method to implement a project is through for-profit action, including investment and/or the founding and operation of businesses.
These projects are designated as “profit” on their associated project pages on this website. This page is an example of such a project.
To implement efforts through for-profit means, entities including Helena Special Investments, LLC, a privately-owned limited liability company that operates in business activities that have the potential to transformatively address societal problems while targeting attractive returns to investors, are utilized.