Experiment Box
(main post begins below this grey box)
For those who are aural learners, Google’s NotebookLM was used to auto-generate a podcast from this post. All I submitted is the URL for this post. No additional “human instruction”. Enjoy.
https://notebooklm.google.com/notebook/36a1045d-b40f-4c0e-ad1d-cef4f179c176/audio
Prologue: Munger
How do we learn and share mental models that help us understand the intricacies of how Climate Change, AI, and the Evolving Grid come together in the current Renewable Energy Transition? Charlie Munger’s notion of Latticework provides a hint.
Let’s begin with a pair of quotes from Charlie Munger:
“What is elementary, worldly wisdom? Well, the first rule is that you can’t really know anything if you just remember isolated facts and try and bang ’em back. If the facts don’t hang together on a latticework of theory, you don’t have them in a usable form.”
“What are the models? Well, the first rule is that you’ve got to have multiple models—because if you just have one or two that you’re using, the nature of human psychology is such that you’ll torture reality so that it fits your models.”
These quotes are both from “Charlie Munger the Complete Investor” by Tren Griffin.
For now, let these Munger quotes percolate at the edge of your awareness as you read this essay; we will return to them at the end.
The Long and the Short
Like a triple helix, climate science, AI, and the energy grid are wrapping around each other in the evolving energy transition. It’s worth stepping back from analysis of individual components and look at the relationships, symmetries, and constraints shaping the emerging structure of a new grid evolving out of the “old” grid. .
This evolving grid arises from the one we have experienced for roughly 40 years and whose historical reliability drives our expectations. Like all complex systems it’s constrained by its history including decisions made at earlier times (often the best decisions possible given reasonable assumptions and knowledge). Like many complex systems it is also through good design and human ingenuity capable of working around current limitations and if needed, replacing them with better mechanisms as they are invented and commercialized
Let’s start with the aspects that layer complexity into considering the trifecta: The Grid (actually many many grids), AI (prediction to pattern finding and pattern generation; simulation and decision sciences) and Climate(long term)/Weather(short term).
Forecasts: We use data and learning algorithms (“AI”) to compute forecasts. Including energy demand/generation/price forecasts. We can also simulate congestion, curtailment, the effects of increasing Available Transfer Capability (ATC) and many other things that affect markets, infrastructure and consumers. We can both forecast and simulate outages due to extreme weather events and simulate the effect of those events given alternate grid designs for the future.
Phenomena: Climate Change related weather extremes are reducing the accuracy of these forecasts leading to “surprise”. However the weather, even extreme weather we experience begin with dynamical processes which the meteorological community has developed expertise in both predicting and simulating at various temporal and spatial scales.
Time Scales: BUT climate change evolves over periods longer than most model training sets. So it may not be explicitly factored into short term forecasts.
Problem to Optimize: Grid Reliability.
How do we reduce the “surprise” of changing weather extremes (rare events) in our short term weather forecasts and downstream forecasts that depend on weather inputs? Particularly with respect to how weather drives and limits renewable power generation.
The question can be asked conversely as “how do we increase the mutual information” between demand/generation forecasts, phenomena that condition those forecasts, and the time scales that separate “then” from “now” so the grid is always up, always reliable with respect to meeting current demand.
Of AI and Energy
AI leverages information to make predictions. Energy is the capacity to do work.
AI takes Energy.
Energy for intense computation.
Energy to run the new AI chips that do those computations; and the data centres that house, coordinate and cool them).
Energy for the plants that manufacture those chips.
(https://e360.yale.edu/features/artificial-intelligence-climate-energy-emissions)
(https://blogs.nvidia.com/blog/liquid-cooling-doe-challenge/)
AI is being used to Optimize energy usage so it is more efficient. From energy forecasts at various time scales, to optimizing for operational efficiency, to assessing and reducing risks. These risks to account for and mitigate include:
available interconnections allowing exports and imports,
having sufficient available transfer capacity to utilize local generation,
adjusting for increasingly distributed and intermittent generation as we switch to renewables,
to brownout (under voltage) or even blackout due to extreme weather events.
(https://www.climatecentral.org/climate-matters/weather-related-power-outages-rising)
AI, in its current incarnation, is becoming a major source of energy demand.
(https://www.bbc.com/news/articles/c51yvz51k2xo.amp)
As well as affecting other scarce resources such as potable water
(https://futurism.com/the-byte/microsoft-arizona-water-ai)
So the sling blade of AI cuts both ways — it finesses our ability to make work efficient via prediction of upcoming conditions and simulation of scenario and design alternatives to evolve grid infrastructure and associated market mechanisms to be more robust to the interplay of renewables and weather. AI also uses (and dissipates) energy in making those predictions. There is a net gain if energy used in the whole training and prediction supply chain (the computation, the chips, the data centres aggregated energy usage) is less than the energy saved by increased efficiencies through AI (we’ll skip over the role of redundancy for now).
All the various and emerging uses AI is being put to are links in a supply chain of energy usage.
Which is rising faster? :
A. Reduction in Energy Demand via efficiencies in energy usage leveraging AI supporting Prediction and Decision?
B. Increased energy demand due to rising adoption of AI across numerous fields of endeavour. We need to consider both the energetic supply chains required to produce the physical hardware, and to run the software of AI enhanced systems.
C. Increased decarbonization of the grid via the rise of renewable generation.
D. Increased intermittency on grid infrastructure due to the rise of renewable generation and limits on incorporating those changes into current infrastructure.
The dual role of AI in being both mediating energy demand efficiency while also being a source of energy demand as well as being a way of resolving increasing renewable generation and intermittency is reminiscent of the Red Queen Hypothesis in evolutionary biology. The Red Queen Hypothesis concerns interdependent (biological) systems with competing (or antagonistic) goals. The Red Queen Hypothesis was a subtle argument on co-evolving systems, that was presented in analogy to the well known Red Queen from Alice in Wonderland. Consider insects and plants. As a plant species across generations develops a new mechanism to avoid predation, the insect species across generation develops a new mechanism to get around the plants defences. And so on and so forth as insect and plant species race faster and faster around each others’ armaments in evolutionary times. In our case, the rise of renewables and mechanisms to reduce intermittency due to renewables must co-evolve and AI may be able to play a mediating role in that race.
"In our country," said Alice, still panting a little, "you'd generally get to somewhere else—if you run very fast for a long time, as we've been doing."
"A slow sort of country!" said the Queen. "Now, here, you see, it takes all the running you can do, to keep in the same place. If you want to get somewhere else, you must run at least twice as fast as that!"
(https://www.cwauthors.com/article/Alice-Wonderland)
(https://fs.blog/the-red-queen-effect/)
(https://royalsocietypublishing.org/doi/full/10.1098/rspb.2014.1382)
(https://www.ugr.es/~jmgreyes/Host-parasite%20coevolution%20Red%20Queen.pdf)
So to ask “which is rising faster” decarbonization of the grid due to the rise of renewables, and intermittency of the grid as a consequence points us towards competing mechanisms. Market (and other) mechanisms funding decarbonization of the grid and infrastructural mechanisms assuring the emerging grid is reliable. AI is likely to have a strong role in both.
How do we know the emerging system will be stable?
How do we mitigate if it is not?
Surprises and the Limits of Prediction
Given that recent observed climate change related increases in weather extremes surprise us, and are challenging to forecast.
Given that with the rise of renewables in the Renewable Energy Transition, energy generation is increasingly dispersed and sensitive to weather surprises. The Renewable Energy Transition is often simply called The Energy Transition. But that conflates it with past energy transitions.
(https://vaclavsmil.com/wp-content/uploads/2010/02/2016-ERSS-Debating-Energy-Transitions-1.pdf)
Given that AI methods can contribute to improving our issues of forecasting and controlling aspects of the energy grid(s) needed for the grid(s) to function (ranging from with intermittency , brownouts, congestion, planning at multiple time scales, distributed infrastructure design, etc).
Given that the underlying technology supporting AI itself has high energy demands (at least in current incarnation of AI).
Given these four trends, we can infer that we can leverage AI to introduce predictability and control to our human response and adaption to climate change (the “energy transition”) if
(a) the energy devoted to predictive control that can stabilize systems is less than
(b) the energy needed to construct and run the AI systems that would be able to do this work of anticipating and stabilizing systems so energy needs are met.
This sounds curiously like a variant of Maxwell’s Demon, a thought experiment in thermodynamics, the field of science that deals with energy, heat, entropy and order.
Maxwell’s demon is a thought experiment posed by James Clerk Maxwell as a way to reason about whether it is possible (at least in theory) to violate the second law of thermodynamics that limits how much actual work we can obtain from a given amount of energy due to a tendency towards disorder.
(https://we-are-berkeley-lab.lbl.gov/spooky-science/maxwells-demon#)
Maxwell’s demon neatly ties information, prediction, and energy together. While formulated long before AI, it anticipates the constraints on AI which are
( a) the information needed to make a prediction and
(b) the energy that must be utilized to make the prediction. Predictions are not “free”. As well as the energy used to make a prediction we must also consider the additional energy used to act on the prediction.
In short, Maxwell’s Demon must use information (data) to make a prediction (AI algorithm used and it’s computational complexity bounding running time on a physical system using energy) and then act (sorting molecules being the actions in Maxwell’s demons case; sorted being the outcome).
If we think of the grid as a complex system, AI seems to be playing a role analogous to Maxwell’s Demon: to anticipate and help bring order.
Of course, both Maxwell’s Demon and it’s banishment are thought experiments.
The increasing variability of weather, the complexity of our transition to renewable based generation, the utility of AI in helping us solve prediction and control problems along the way are real.
They hinge on AI acting in Maxwellian Demon fashion to enhance prediction and control of the grid while not creating a level of energy demand that would negate those benefits.
How would we track this?
What data would we need?
What of the needed data is currently easy to obtain?
How would we account for balancing AI’s increasing energetic demands and its obvious benefits which are much wider than the energy transition?
As a starting point, consider error, as characterized by the well known residual calculation,
residual = actual_Y_value - predicted_Y_value
Based on the notion of a residual consider the following inequality:
Energy_Saved_Via_ReducedPredictionError
Is Greater Than or Equal to
(Energy_Used_Prediction + Energy_Used_Actions).
In principle this is a calculation we can make on actual systems. Most forecasters report their prediction errors. What is missing is converting a reduction in error into increased energy availability . What is also missing is tracking and reporting on the energy usage going into making those predictions.
To put it succinctly, if a linear regression reduces error in predicting demand from a wild ass guess; will a generative AI applied to transformer technology sufficiently further reduce prediction error to justify it’s much higher compute costs and associated energy expenditure (as well as greater inherent complexity)?
Latticework:
Climate <==> AI <==> Evolving Grid
“As unique and as unprecedented the unfolding global energy transition may be, it shares the key feature with its predecessors: it will be a gradual, multidecadal, inter-generation process with different and diverging national pathways and rates of progress.” - Vaclav Smil in “Energy Transitions: Fundamentals in Six Points” (https://www.funcas.es/wp-content/uploads/Migracion/Articulos/FUNCAS_PE/009art03.pdf)
To recap the argument so far:
- Climate Science & Energy Forecasting: Climate Change related weather extremes are reducing the accuracy of these forecasts leading to “surprise”. The “surprise” level over the last several years (2021 to 2024 roughly) has moved in our consciousness from “explainable” one offs to nervous expectations of the future; akin to the tale of the boiling frog with humanity collectively in the starring role of frog. Ribbit! Gulp.
(https://www.ncbi.nlm.nih.gov/pmc/articles/PMC534568/)
- Energy Demand & AI : AI, in its current incarnation, is becoming a major source of energy demand. All the various and emerging uses AI is being put to are links in a supply chain of energy usage.
- AI Growth & Energy Constraints: The increasing variability of weather, the complexity of our transition to renewable based generation, the utility of AI in helping us solve prediction and control problems along the way are real. … They hinge on AI acting in Maxwellian Demon fashion to enhance prediction and control of the grid while not creating a level of energy demand that would negate those benefits.
We have been experiencing (and arguing about) climate change for decades. The current renewable energy transition is a response to climate change via decarbonization of the energy supply chain. However between the years 2000 and 2005 there was actually a slight drop in global use of renewable energy sources; the increase starts about 2010 and picks up pace.
There has been a steady uptick in VC and Government funding since 2015
(https://www.nber.org/system/files/working_papers/w29919/w29919.pdf)
Then COVID 19 came and shut the world down by early 2020, Remote became the new normal for much of the working world and the renewable energy transition funding jump coincides with the current COVID and Post-COVID (or COVID tolerant) period we are currently in.
COVID was seen as disrupting the energy transition; and post COVID, (current) period is widely seen as a return to a trajectory or renewable based generation with an even steeper rise.
(https://energy.sais.jhu.edu/articles/how-covid-19-disrupted-renewable-energy-transition/)
Climate Change, Renewable Energy Transition, then COVID in 2019 . A perfect storm that could not be predicted? The view from deep time (evolutionary biologists) notes that climate change leads to host parasite switching; which can in some cases be predicted. COVID is just the most obvious recent example of such climate change related switches.
(https://orb.binghamton.edu/cgi/viewcontent.cgi?article=1093&context=alpenglowjournal)
(https://pubmed.ncbi.nlm.nih.gov/36176231/)
(https://royalsocietypublishing.org/doi/10.1098/rstb.2020.0360)
(https://www.pnas.org/doi/10.1073/pnas.1705067115)
(a few more references for those interested in the fascinating literature on climate change and host parasite switching coevolution)
Which takes us back to AI — whose usage rose dramatically during the pandemic. COVID related to Climate Change adds another thread from evolutionary biology onto the latticework of interdependencies binding AI, Energy, and Climate Change.
My latticework of mental models is likely a little different than yours.
The arguments here have drawn on
evolutionary biology
tall tales of Demons and Angels in Thermodynamics and Statistical Physics.
Alice in Wonderland (tied to more evolutionary biology)
multi disciplinary knowledge of AI and statistical sciences with a long focus on the analysis of prediction errors.
In short, bits and pieces of everything I have worked on in my adult life. A lattice work of models that form an (ideally) covering set of models I can deploy as I approach new problems. My latticework is a personal synthesis of concepts.
As is Charlie Mungers.
As is yours.
For some of us, there is a lot of mathematical machinery beneath the surface of these models. But in seeking to understand the world — we are more often deploying our conceptual understanding, which often becomes intuitive as the concepts become personal knowledge. The math comes later. If a bear is racing towards you. Don’t compute the best escape route — Run!
Climate change. AI. Renewables. Evolving grids. These are the bears.
How do we learn from each other? By listening, comparing, contrasting. Adding to our latticework . I’ll start with four questions (below) that can be openers to sharing differing perspectives and syntheses. I would love to hear your top questions connecting Climate Change, the Evolving Grid and AI. Please feel free to throw them in the comments.
Opening Questions:
Is the Energy Grid Transitioning into a Latticework as we diversify energy sources? That is a set of overlapping but not necessarily fully covering set models.
How do we account for changes in Financial Risk to energy systems due to changing climate? We resolve a lot of risk via financial tools operating in financial markets. Can we refine finance risk concepts like Alpha and Beta to incorporate climate variables — Climate Alpha and Climate Beta?
How do we account for changes in Operational Risk to Physical Infrastructure and Information Infrastructure?
How do we resolve market mechanisms for pricing stability with the physical infrastructure required for voltage stability?
How do we account for both AI’s role in reducing risk via computation, and the risks from AI via the energy demand, and requisite supply chains (from cloud costs to chip manufacturing) that have to be maintained (and are disruptable) in a complex political and biological landscape where a single host switching incident can stall the world and bootstrap AI and catch most of us by surprise?
If you wish to further explore latticework’s, here are some starting points online
(https://fs.blog/munger-worldly-wisdom/
https://fs.blog/mental-models/
https://fs.blog/munger-worldly-wisdom/)
I would also highly recommend “Latticework. The New Investing. ” by R.G. Hagstrom which was my first introduction to the Latticework concept as it is used in investing.
As you evolve your personal latticework
Climate <==> AI <==> Evolving Grid
Please share here in Comments and in Notes. I want to learn your Latticework(s).
Subscribe for free to receive new posts and support my work.