A Few Things: The Rising Cost Of Compute, David Foster Wallace, Relax For Same Result, Seth Godin & Tim Ferriss On Meaning, The Precipice, Coming Wave Of AI and How To Invest, Amplifying Humanity....
June 1, 2023
I am sharing this weekly email with you because I count you in the group of people I learn from and enjoy being around.
You can check out last week’s edition here: The Mind of Napoleon, Yann LeCun on AI, The Beginning of Infinity, Hurtling Through Space, Jeffries AI Report, MSFT CTO on Work, How Scalable Is Your Job?
Believe it or not, that “♡ Like” button is a big deal – it serves as a proxy to new visitors of this publication’s value. If you enjoy this article, don’t be shy :)
Quotes I Am Thinking About:
“An era can be said to end when its basic illusions are exhausted.”
- Arthur Miller
“You miss 100 percent of the shots you never take.”
- Wayne Gretsky
“The fact that an opinion has been widely held is no evidence whatever that it is not utterly absurd.”
- Bertrand Russell
“The essence of strategy is what you choose not to do.”
- Michael Porter
“The world would be happier if men had the same capacity to be silent that they have to speak.”
- Baruch Spinoza
A. A Few Things Worth Checking Out:
1. 13D - WILTW had a great piece on: AI’s soaring carbon footprint will drive huge investments into more power-efficient chips, AI model compression, and accelerated clean energy production.
Key bits (emphasis mine):
Machine learning is on track to consume all the world’s energy, if left unchecked. A single prompt in ChatGPT consumes hundreds of times more energy than a single Google search (0.3/kWh vs. 0.0003/kWh).
In January 2023 alone, ChatGPT’s 13 million unique daily users consumed electricity equal to 175,000 people on average, while GPT3’s training led to 550 tons of carbon dioxide emissions, notes the University of Berkeley.
The factors driving this exponential increase in AI’s carbon footprint include:
The size of LLMs is growing exponentially. In 2019, GPT-2 had 1.5 billion
parameters, which increased 116x to 175 billion in GPT-3. Now, GPT-4 has
about one trillion parameters.
Ian Bratt, senior technology director at ARM, summarizes: “The compute
demand of neural networks is insatiable. The larger the network, the better the
results, and the more problems you can solve. Energy usage is proportional to
the size of the network.”
However, power consumption can increase even faster than model size.
To train AI models, engineers iterate over the same dataset multiple times.
Godwin Maben from Synopsys explains:
“If you look at the amount of energy taken to train a model two years back, they were in the range of 27-kilowatt hours for some transformer models. If you look at the transformers today, it is more than half a million kilowatt hours. The number of parameters went from maybe 50 million to 200 million. The number of parameters went up four times, but the amount of energy went up over 18,000X.”
As a result, developing more power-efficient chips and AI models will become a priority, as will an acceleration in clean energy investment:
More power-efficient chips will gain share;
Compressing AI models is emerging as an essential tool to improve energy efficiency. The two current model compressions techniques are: Pruning (reducing number of neural network parameters within a model), Quantisation (compressing original network by reducing number of bits required);
The drive for AI energy efficiency will accelerate investment in neuro-morphic computing. Today’s microprocessors perform operations following instructions programmed in 1s and 0s, processing information in a linear fashion. In contrast, neuro-morphic processors mimic biological synapses of the human brain.
Another factor to consider when you are considering investing into AI.
2. David Foster Wallace, on Consumerism: Few people understand modern American culture like David Foster Wallace did. In particular, he saw how young people in the upper-middle class lived sad and empty internal lives, even though their external ones were defined by tremendous comfort.
One of his more interesting observations is the absence of silence in modern life. We listen to music when we’re on our computers, move through our homes with TV in the background, and insist on playing pop music at our restaurants. What’s going on there?
This interview is interesting beyond Wallace's ideas. It's also a revealing window into his psyche: his fears, his insecurities, and the trepidation he feels about telling the truth.
3. Relax for the Same Result: Our hustle culture encourages us to work more and try harder, but it's worth asking when this advice falls flat.
I think about this Derek Sivers piece often and ever since reading it a few years ago, it changed how I approach life: Relax For The Same Result.
It begs the question: "When is effort superfluous, and when is it what makes all the difference?"
4. The US government is spending billions of dollars to build out state-of-the-art domestic semiconductor manufacturing capacity. But spending money is no guarantee of success.
In fact, there are already worries that the CHIPS Act passed by the Biden administration isn't succeeding, due to various roadblocks, speed bumps and unforced errors. What are the odds that it will pay off? And what should we be watching for as evidence of its efficacy?
The Bloomberg Odd Lots podcast spoke to Dan Wang, technology analyst at Gavekal Dragonomics and Adam Ozimek, chief economist at the Economic Innovation Group about: This Is How We Will Know The CHIPS Act Is Working.
5. Seth Godin spoke with Tim Ferriss on The Pursuit of Meaning, The Life-Changing Power of Choosing Your Attitude, Overcoming Rejection, Life Lessons from Zig Ziglar, and Committing to Making Positive Change.
Seth is a true creative right brained thinker and always makes me think differently and at a deeper level.
Here are some key quotes I enjoyed:
“Some people believe that the purpose of business is to enable culture, to enable humanity. And some people believe that the purpose of humanity and culture is to enable business. And I think those people have too much influence right now, and they’re wrong.”
“Milton Friedman just made up this nonsense about the only purpose of a corporation is to maximize its profit. It lets people off the hook and they become tools of a system that grinds stuff out.”
“Boomers have driven our culture since the day I was born. When we were draft age, that was when the draft really mattered. And when we were listening to rock and roll, that’s when music really mattered. And when people needed to make money for their family, that’s when Wall Street really mattered. And now boomers are dying. And so we are living in a culture where there’s an overhang of all these people with loud voices talking about the end of the world because it’s the end of their world, but it’s not the end of the world.”
B. The Precipice by Toby Ord
This book has helped me think about humans, our potential and the risks ahead in a whole new light.
Toby Ord is an Oxford philosopher who has written a beautiful and deeply researched book. Toby has advised the WHO, the World Bank, the World Economic Forum, US National Intelligence Council and UK Prime Minister’s Office.
The premise of Toby’s book - The Precipice: Existential Risk and the Future Humanity, is that if all goes well, human history is just beginning.
We are at the beginning of amazing things and as such Toby (and I assume all of us) wants to maximise human flourishing.
But this vast future is under threat.
The book has 3 parts: The Stakes, The Risks and The Path Forward.
The book explores the threats we face and what we can do about them.
A critical quote on page 31 for context on the significance of this moment:
During the twentieth century, my best guess is that we faced around a one in a hundred risk of human extinction or the unrecoverable collapse of human civilisation. Given everything I know, I put the existential risk this century at around one in six. If we do not act together, if we continue to let our growth in power outstrip our wisdom, we should expect this risk to be even higher next century, and each successive century.
Key quote on why if the risks are so high, why don’t they receive more attention:
Economic theory tells us that existential risk will be undervalued by markets, nations and even entire generations. While markets do a great job of supplying many kinds of goods and services, there are some kinds that they systematically undersupply. Protection from existential risk is a public good: protection would benefit us all and my protection doesn’t come at the expense of yours. So we’d expect existential risk to be neglected by the market. But worse, protection from existential risk is a global public good - one where the pool of beneficiaries spans the globe. This means that even nation states will neglect it.
The same effect that causes this undersupply of protection causes an oversupply of risk. Since only 1% of the damages of existential catastrophe are borne by the people of UK, their government is incentivised to neglect the downsides of risk-inducing policies by this same factor of 100.
This means management of existential risk is best done at the global level. But the absence of effective global institutions for doing so makes it extremely difficult, slowing the world’s reaction time and increasing the chance that hold-out countries derail the entire process.
The books spends three chapters discussing
Natural Risks such as: Asteroids, Volcanos and Supernovas;
Anthropogenic Risks such as: Nuclear Weapons, Climate Change, Environmental Damage;
Future Risks such as: Pandemics, Unaligned AI, New Tech (for example Nanotech).
He arrives at this risk landscape. You can agree or disagree, but seems like a good framework to start with:
So how do we decide what to work on and when. Our ultimate aim is to spend the resources allocated to existential risk in such a way to reduce total risk by the greatest amount.
Here he comes with a simple equation:
Cost Effectiveness = Importance x Tractability x Neglectedness
Importance = value of solving a problem
Tractability = how it’s to solve a problem
Neglected = Extent of resources spent on it
Toby’s Grand Strategy for Humanity has three stages:
Reaching Existential Security - reduce existential risk by as much as possible
The Long Reflection - develop mature theories that allow us to compare the grand accomplishments our descendants might achieve with aeons and galaxies as their canvas.
Achieving Our Potential - he sketches out a canvas based on duration, scale and quality.
What could our potential be?
He starts with a beautiful H.G. Wells quote:
“It is possible to believe that all the past is but the beginning of a beginning, and that all that is and has been is but the twilight of the dawn. It is possible to believe that all the human mind has ever accomplished is but the dream before the awakening.”
Now Toby looks at the size of the canvas ahead of us.
Human history so far has only been 200,000 years of Homo Sapiens and 10,000 years of civilisation. We have had civilisation for a hundred lifetimes on end and humanity for thousands. But the universe we inhabit is thousands of times older than humanity itself. There were billions of years before us; there will be billions to come. In our universe, time is not a scarce commodity.
In just five centuries, we have gone from the dimmest understanding of our Solar System- unable to grasp any coherent picture of our Sun, Moon, Earth, and the wandering points of light called ‘planets’- to breath taking high-resolution images of all our planets and their moons.
Our Solar System’s greatest contribution to our potential lies with our Sun, and the vast bounty of clean energy it offers. The sunlight hitting Earth’s surface each day carries 5,000 times more energy than modern civilisation requires. It gives in two hours what we use in a year.
But the most important thing we have learned from the sky may be that our universe is much vaster than we had ever contemplated. Each time we thought we had charted the limits of creation, it transcended our maps.
What beauties are we blind to? Our descendants would be in a much better position to find out.At the very least, they would likely be able to develop and enhance existing human capacities - empathy, intelligence, memory, concentration, imagination. Such enhancements could make possible entirely new forms of human culture and cognition: new games, dances, stories.
C. The Technology Section:
Key bits (emphasis mine):
An LLM's goal is to input a sequence of data into it, access its vast training data and relationships it has learned between the data, and then output a correct response given what it knows and the sequence inputted.
Each model has a few key attributes: the model's size (parameters), the size of the training data (tokens), the cost to compute the training, and the performance after training. I will briefly discuss three categories: size, data, and compute.
The important takeaway is that GPT and other LLMs will improve but likely improve from increasing data, size, and computing. Given that, something must store that data (memory) and something must train the models (compute), this is very good for the underlying infrastructure. In this case, semiconductors.
AI models like LLMs and Stable Diffusion take a lot of computing power. These models are huge. One way to think about it is that each parameter or learned trait takes up data manipulated and used massively. GPT-4’s 1 trillion parameters are 4 terabytes of data being loaded into memory, passed information, and computed. That’s a non-trivial amount of computing for inference, and the cost of training can spiral into the 100s of millions of dollars, given the number of times the model parameters are tweaked.
I cannot stress this enough. AI-enabled workloads like training and inferences are more hardware intensive and costly than before. Let’s compare AI workloads to “legacy” workloads, like traditional SaaS or social media operations.
Each query for ChatGPT costs as much to run a CPU in a data center for an hour, which can be consumed in seconds.
And a reminder that the GPUs powering OpenAI are specialized for the task. The amount of computing it would take to host a website for an hour is being used in a single query, and the ChatGPT model and LLMs are often multiple queries over and over. In a future where AI is added into many more aspects of our lives, you could likely run up a bill that would cost orders of magnitude higher than most websites or SaaS applications in just an hour on ChatGPT. This is a step function higher in demand, and as the world becomes more AI-intensive, hardware demands will soar.
I mentioned above the amount of data and size of the models being trained and inferenced, and I want to spend a second talking about one of the key constraints of AI training and inference, which is interconnect.
Let’s take training, each time the model is trained, the parameters have to be held in memory near the GPU or accelerator, and there isn’t a single rack in the world capable of holding half a petabyte in it’s memory, so the problem has to be split into smaller units and parallelized across other computers. This would be like cutting the larger problem into larger slices like a pizza, and having each slice cooked separately then put back together into a larger pizza. There isn’t an oven big enough to cook 1 whole pizza.
Nvidia has probably the best and most parallel process today. Not only are all the GPUs connected in a single rack, but each of the racks is connected to the other with NVLink, making each GPU connected in the rack and the racks connected at a very fast speed. This interconnect speed is one of the biggest bottlenecks, and currently today, the DRAM bandwidth in particular, is among the biggest problems for how to scale AI training in the future.
In reality, training models like GPT-4 take more than a single 500k server blade, but rather thousands of blades working together to train the model in parallel. This is the concept behind the SuperPod, which aims to scale many DGX blades into a single larger GPU cluster. The problem of the extremely large model is then split into smaller pieces and worked on in parallel as if the entire data center was one giant computer. This is the concept behind the phrase “Datacenter as the New Unit of Compute” that Jensen Huang likes to say so often.
This didn’t intend to become an Nvidia primer. The reality is that it is because of how dominant Nvidia is in this entire ecosystem. I will talk about what I see as their concentric levels of competitive advantage because Nvidia has something very special here. When you ask Jensen his thoughts on competition, I think he is unwilling to compete where he thinks it is commoditized or Nvidia cannot dominate. Nvidia is the clear leader in each of the three big places they service but offers the whole solution as a product.
2. My friends at Atrum Global in London shared their 46-page AI investment framework. Their conclusion for what they will be looking to invest in:
Specialised models that leverage proprietary data to target niche use cases and domains.
AI-driven applications exhibiting a user feedback loop that drives product differentiation.
AI-augmented applications where AI supplements an already attractive underlying business.
3. Just finished reading Impromptu: Amplifying Our Humanity Through AI by Reid Hoffman with OpenAI's GPT-4 and it got me thinking How AI will reshape what finance professionals can do.
The short book explores how AI, and especially Large Language Models (LLMs) like GPT-4, can elevate humanity across key areas like education, business, and creativity.
What makes it really cool is that the book is not just a book, it’s a conversation. Hoffman doesn’t just write about GPT-4; he interacts and writes with GPT-4, letting you see it's capabilities and weakness.
You can go deeper with Reid through his awesome new podcast series: Possible.
4. A16Z shared their AI Canon which is a great place to go deeper into all things AI.
5. LionTree, the investment and merchant bank shared their Artificial Intelligence Perspective. Here are 3 pages worth sharing:
D. Charts That Made Me Think:
Believe it or not, that “♡ Like” button is a big deal – it serves as a proxy to new visitors of this publication’s value. If you enjoyed this, don’t be shy.