Happy Customers are Always Right
Wealth creation is centered around customer value. Wall Street likes to tout big ideas. Those are often nothing more than entertaining narratives. Stories don't make money. Happy customers do.
It's "Big Ideas" season. Have you received an email recently with the phrase "Big Investment Ideas for 2023" in the subject line? You must have gotten one unless you live in a cave. Whether it’s investment firms, venture capital shops, or just finance blogs, for some reason people think that the arbitrary choice of ending the year in December and starting a new one in January corresponds with the birth of new investment ideas. Obviously, this has nothing to do with ideas and everything to do with marketing. The investment business is very peculiar in this matter. While most of their customers care about returns, investment managers pretend that it’s about ideas. It’s not. It’s about returns. But if you can’t generate high and/or steady returns, you have to refer to something else. That’s where the "Big Idea" convolution steps in. In this essay, we argue that "Big Ideas" might have some merit as inspirational goal posts. But they are dangerously misleading investors since they project the illusion of control about making money in financial markets. If you want ideas, read novels. If you want returns, follow entrepreneurs. Focus on those entrepreneurs that create customer value by obsessively iterating around engineering problems. Ideas don’t pay the bills. Happy customers do.
Investment professionals like to think of themselves as "idea generators." It's strange how a profession that, on the inside, could not be more dull and boring prefers to present itself as a creative idea generator. Merriam-Webster has lots of definitions for the word "idea." Let’s stick with "a formulated thought or opinion." Creativity is essential when formulating thoughts or opinions. In this essay, we argue that ideas are only worth their salt if they are born through a creative process. Most so called ideas marketed by the investment community are not. They’re just prose disguised as creative thought. And adding "Big" to "Idea" really makes it sound important. Let’s analyze this a bit more. What does it take for a human to come up with an idea? Where do ideas come from?
David Deutsch is my go-to philosopher for questions like this. In this book, "The Beginning of Infinity," Deutsch argues that ideas are the fundamental building blocks of reality. They shape and define our understanding of the world. Ideas, such as scientific theories and mathematical concepts, provide frameworks for understanding and explaining the workings of the universe. In this sense, ideas serve as a kind of "software" that runs on the "hardware" of the physical world. This means that ideas play a central role in shaping our perception of reality and that the progress of science and technology is driven by the creation, dissemination, and refinement of ideas over time. Deutsch treats ideas literally, like software that evolves through human ingenuity. While we don’t exactly know how it evolves, we have a good sense for the process. It’s very much like evolution by natural selection. Ideas evolve in a similar fashion. Karl Popper formulated this idea (note the recursive nature of this!) in his book "Conjectures and Refutations". New ideas are proposed as tentative solutions to problems or explanations of phenomena (conjectures) and then subjected to critical testing and evaluation (refutation). If an idea withstands testing and experimentation, it becomes more robust and reliable and may continue to be developed and refined. If it fails to withstand testing, it is discarded or revised. Popper argued that this process of conjecture and refutation is essential for scientific and intellectual progress, as it allows for new and better ideas to emerge through a process of critical evaluation.
Deutsch is the contemporary continuum of Popper. He uses modern language to illustrate Popper’s ideas. And like it or not, his status as a theoretical physicist with accolades in quantum computation brings panache to his arguments and thus contributes to their adoption. My interpretation of Deutsch and Popper is that ideas are not just similar to natural selection, they are natural selection. When organisms evolve by genomic mutation, they are, as Popper would say, "proposing tentative solutions to problems." In other words, they’re formulating a new idea of how to deal with a specific problem posed by the environment. In that sense, genomic mutation is as much an idea as inventing gravity or quantum computation. Ideas solve problems and thus contribute to the process of evolution. The main driver of new ideas is creativity. Deutsch views creativity as a fundamental aspect of human nature and argues that it is the ultimate driver of evolution in science, technology, business, and the arts. In fact, creativity is at the core of human agency. And according to Deutsch, creativity is not just a matter of chance or luck but is a deliberate and intentional process that can be cultivated and enhanced through effort and practice. He also emphasizes the importance of critical thinking and evaluation in the creative process, as these skills allow individuals to validate their ideas and refine or reject them based on their merits. Crucially, creativity is a process that happens despite our lack of knowledge about how it actually happens. Deutsch ascribes high value to creativity since it is the source of new ideas. And new ideas are the software that runs our world, with an emphasis on "our," since ideas are a creation of the human brain. They help us make sense of the world and hence help us function in the world. We shape our environment with knowledge and knowledge fundamentally relies on ideas.
I deliberately dove deep into the weeds of philosophy to illustrate the point that ideas are a pretty serious thing. We should be very careful when talking about them. They are the things that separate us from cavemen—freezing in cold weather, succumbing to totalitarian hooligans, and pretty much everything else humans had to endure to create the civilization we live in today. Civilization is what happens when you introduce creativity to the evolutionary process. The power of ideas is to give us structure and an understanding of the world. With that, we can make predictions. Useful ideas are those that effectively solve a problem, address a need, or improve upon existing solutions. They have practical applications and bring tangible benefits to people's lives or to society as a whole. Gravity is an idea that helps us build stuff and is thus very useful. Another recently evolved idea is the representation of complex dynamics through neural networks. Ideas are useful when they pass the test of time and the marketplace.
How do investment ideas fit into all this? Where do investment ideas come from, and how do they evolve? What are useful investment ideas?
Let’s look at a few examples of "big idea" talk. Most of these ideas are from Arch Invest’s "Big Idea" deck for 2023.
Technoloical Convergence
Five innovation platforms are converging to create unprecedented growth. Artificial intelligence is the most important catalyst; its velocity cascades through all other technologies. One example of this convergence is AI and robots. Tesla is using advances in computer vision and tokenized language models to fine-tune its autopilot technology. Another area of convergence is language modeling and genome sequencing. Pacific Biosciences is using tokenized language models developed by Google to better execute long reads in genome sequencing. The purpose of these models is to increase accuracy while lowering cost.
The idea here is that several technologies (AI, robotics, gene sequencing, energy storage, and public blockchains) will converge and spur massive innovation and growth across all platforms.
This is a great example of an inspiring narrative that misleads investors into the illusion that money will be made here. The convergence of technology is a good thing. But it doesn’t help you invest in profitable endeavors. In fact, the convergence of technology is a tautological statement. Technologies always converge. Breakthroughs in railway technology in the middle of the 19th century were only possible because of breakthroughs in steel manufacturing, and steel manufacturing was advanced by processes in mining, metallurgy, and chemical engineering. Every technology is embedded in a web of adjacent technologies. Arch Invest argues that today’’s technologies converge more rapidly and cross-fertilize themselves more intensely. That might be true. But what is an investor to do with that information? It’s like saying the jungle is a fertile habitat for many life forms that will propagate each other through symbiosis. Yes, that is definitely true. But which life form will triumph over the others? Which will die prematurely, and which will evolve into even more lucrative niches? Those are the questions investors need to answer. Useful ideas help solve practical problems. The best way to gauge whether problems are practical in the business context is to follow happy customers. The more happy customers you have, the more likely you are to solve a practical problem.
One important factor in Arch’s research is the focus on cost. They argue that technological convergence is a driver of lower costs. We will revisit this argument.
Artificial intelligence increases the productivity of knowledge workers
Here is another powerful idea. AI helps software engineers get work done and thus increases their productivity. This is reminiscent of the time when Word and Excel entered the world of business. Productivity went up dramatically. The trend toward higher productivity in knowledge work has been going on for decades. Has it?
I disagree. AI is not a boost for knowledge creation. AI accelerates white collar assembly-line work and boosts the productivity of paper pushers, bean counters, and other menial office jobs. In aggregate, that might have an impact. But for companies and investors, this type of collective productivity increase is not going to move the needle. AI is not helping companies become more competitive and generate higher returns, in the same way that Microsoft Excel is not helping investors become more competitive. Everybody has access to the same tools, and everybody will take advantage of them. Differentiation will happen in the same old-fashioned way: by entrepreneurs using their creativity, ingenuity and tenacity to drive solutions for engineering problems, lower costs, and increase customer value. Creativity and grit eat technological progress for lunch.
Digital Wallets Will Disintermediate Traditional Banking
The rapid adoption of digital wallets is driving network effects. Think of digital wallets as an app that enables financial transactions without an intermediary. You buy your coffee at Verve and send Satoshis directly from your digital wallet to theirs. No bank or credit card processor is involved. All you need is online access and a blockchain to validate the transaction. Arch calls this type of transaction "a closed loop transaction." By enabling this type of transaction, digital wallets could disintermediate traditional banks. Alipay and Cash App are two examples. Wallets based on the bitcoin lightning network.
Vitalik Buterin, the co-founder of Ethereum, has emphasized the importance of secure digital wallets for holding cryptocurrencies. He has stated that the proper management of private keys, which are the means of accessing a digital wallet, is crucial for ensuring the safety of assets stored in it. He has also mentioned the need for user-friendly and user-controlled wallets, as opposed to centralized wallets that are controlled by a third party, in order to empower individuals and give them more control over their funds.
This idea has legs for investors. It is indeed lowering the cost and increasing customer value. The problem is technology. Is blockchain technology advanced enough to offer closed-loop wallet-based transactions?
Are digital wallets similar to online trading platforms or online travel agencies? Or are they more like Google?
Public Blockchains, smart contracts
It’s hard to formulate the value proposition of public blockchains. Let me try. Public blockchains are decentralized validation networks that solve for identity. On today’s internet, you can transact with people, but you need third party validators to make sure validation is taken care of. For example, if you buy a coffee at Verve, you need somebody to assure them that your money is good and that you paid them. Typically, this is a credit card processor that talks to your bank and communicates with Verve to make sure the payment is okay. All this stuff can happen through a decentralized public blockchain such as Bitcoin, Ethereum, Algorand or Solana. The reason there are several public blockchains is that some of them are trying to solve different types of problems. Public blockchains are optimizing for speed, security, and transaction cost.
The project of the internet of identity is exiting, but unlike the early internet, this new type of web, also called Web3, is struggling to come up with use cases or what the industry calls "killer apps." The killer app of Web1 (the original internet) was email. The killer app for Web2 was sharing pictures. What is it for Web3? Gaming and NFTs are touted by industry insiders as killer apps. But they are not. Gaming is very popular with young people, but it’s still just a niche. It’s like saying the NFL is the killer app for TV. Yes, the NFL is popular. But only sports fans watch games live. There are lots of sports fanatics, but their distribution in the population is limited to something like 20% or 30%. That’s a lot, but it’s not like email, the iPhone, Facebook, or cars, where everybody from a five-year-old to your grandma needs to have one. Gaming is like sports, hunting, ballet, or reading novels. It attracts a certain niche segment of the population. It’s not a killer app. This is one of the misleading arguments many technophiles use to tout Web3 as an investment thesis. In fact, if you have to formulate a killer app a priori, it’s not a killer app. Killer apps happen organically and establish themselves in the popular narrative after the fact. Web3 is not going to make you money. Happy customers will. Where are they? Who are the happy customers in Web3?
We are excited about the idea of Web3, but we don’t see it as an investment idea. It’s more like a new political concept. It reminds me of the early Renaissance, when European cities like Florence, Venice, or Antwerp introduced new forms of contract law and accounting to do business. It was a huge boon for business, and money was made. However, the actual contract law and accounting practices were only a catalyst, not the money maker per se. Web3 is an evolution in contract law. You could theoretically introduce algorithmic contract law. Let’s think about this. Why was contract law introduced in the early Renaissance? What drove that change? It was a shift in authority. Commercial hubs such as Florence or Antwerp detached from the Ancient Regime of feudal lords and monarchs. Modern commerce required more flexible accounting and legal practices. You couldn’t always run to your Lord to settle disputes. When shipping goods from far places around Europe you had to rely on a more adaptive system for account settlement.
The key question here is: did contract law evolve because of new developments in commerce, or did commerce evolve because new contract law enabled it? I believe it was both. It was a symbiotic relationship. This has implications for today. Will Web3 spark the development of new types of commerce? And will new types of commerce inspire rapid development of Web3? It’s a good bet to assume that both will happen. Just as Web1 and Web2 sparked new types of commerce, so will Web3. In order to figure out how to make money with that, ask yourself, "Where are the happy customers?"
Bitcoin as decentralized money (this is a good idea)
This is a killer app. We see bitcoin (the currency) as a viable contender to become a global reserve currency. Now, what exactly does that mean? It means that humans will use bitcoin to store value like digital gold. The reason they do that is because bitcoin promises not to lose value due to inflationary policies. Risks to bitcoin are not inherent in the protocol but are mundane, such as seizure of assets, seizure of mining operations, etc. It would be difficult to uphold the Bitcoin protocol if all governments in the world decided to imprison anyone who mines bitcoin. But let’s assume that this scenario is unrealistic. In that case, you can expect bitcoin to become a global reserve currency. This is a good investment idea. Why? because there are happy customers here. Ask people in Lebanon or Venezuela how they feel about bitcoin. They cheered because it helped them move money out of the country.
Now, what exactly is inflation? Inflation is when your currency loses value relative to the goods and services you consume. In other words, it's what your money can buy. If you want to travel, live in a nice house, and spend money on your kids’ tuition, then measure the cost of that stuff relative to your money. If you have to spend more money for the same stuff, you have inflation. Note that it doesn't matter what you use as money. Whether it's dollars, rupees, cigarettes, or bitcoin, if the stuff you desire gets more expensive, you have inflation. We argue that no matter what money you use, inflation can happen because it's not a money-related phenomenon.
The ultimate measure of inflation is not money but time. Don’t ask how much it costs me, but "how long do I have to work to afford this stuff?" This is key since it not only looks at labor but also at investment returns to gauge your affordability. For example, if a house costs $3 million and you have $1 million in capital, how long will it take you to afford this house? Let’s do the math with bitcoin. Let’s assume bitcoin appreciates 6% per year and the house appreciates 3% per year. How long will it take you until you’re able to buy the house? It will take you something like 40 years to be able to buy the house. Now assume that instead of investing in bitcoin, you invest in a productive asset such as Tesla or Microsoft. We expect Tesla to return 25% per annum in the coming 10 years (this is not investment advice, just a guess). If we’re right, it would take you six years to be able to afford the house.
This is to illustrate that bitcoin is a good idea but not necessarily a solution to inflation. It mitigates inflation but doesn’t solve it. In fact, bitcoin doesn’t solve any of the societal problems we have, such as high government debt, irresponsible behavior in Congress, or even the risk of inflation, as I just outlined in my example. Bitcoin is better than Fiat but not a solution to inflation. Bitcoin competes with many other productive investment solutions. It is a good alternative to Fiat money but not a great investment. Hopefully, this illustrates how "Big Idea" talk can mislead investors into thinking that something is a solution to their money problems. Mostly it’s not. Wealth is not generated with ideas, it’s generated with happy customers.
By the way, just as a tangent, the best way to lower inflation is not to invent digital money but to create political institutions that are adapting to demographic changes such as immigration, housing, labor, etc. Those are the real reasons for inflation. While we disagree with the creation of the Fed and believe the Fed does much more harm than good, it’s not the main cause of inflation. Inflation is caused by people making decisions. For example, you could lower the cost of housing by having flexible zoning laws and efficient regulation for construction. You could lower the cost of education by introducing competition and breaking the stronghold of academia and the academic industrial complex, which has hijacked our education system. The same applies to public health.
In their big ideas deck, Arch argues that Bitcoin’s main advantages are decentralization, auditability, and transparency. I disagree. The only advantage bitcoin has over anything else is that it can be used as money. That’s where we agree with bitcoiners. It has potential for money. But where is the investment opportunity here? That’s not clear to us. We believe it’s a great tool to enable new, interesting forms of commerce. But the notion that bitcoin by itself will become a valuable asset is overreaching. Any money is only worth what it can buy. Without things to buy, bitcoin has absolutely no value. You might argue that’s a tautology, like saying that without things to buy, working for a salary makes no sense. But when it comes to money, it’s paramount to include the things you can buy with it.
Let’s think about this. Imagine a world with only bitcoin and no other currencies. Is that going to be a world with or without inflation? To answer this question, just look at your own life. I live in Los Altos, CA. My son is in high school, and I like to travel. My main expenses are housing, education, health care, and travel. Will bitcoin make these things more or less expensive? Using our previous definition of inflation as a measure of how much time it takes you to afford X, bitcoin is not going to make a difference. The time it takes me to afford my house or my son's education is not a function of the money I use. It depends on my own ingenuity and the institutional framework I live in. Khan Academy is one example. Thanks to Sal Kahn, I can educate my child pretty much for free and provide him with a world-class education up to the senior high school level. This is a huge boost to my purchasing power and reduces the time it takes me to afford an education for my son. And yes, Mr. Friedman, inflation is not a monetary phenomenon! And that's why bitcoin is not a hedge against inflation. Bitcoin is a hedge against the dollar.
This is a very important point. Inflation is not a monetary phenomenon. In fact, it has nothing to do with money at all. Inflation is all about how much time it takes you to afford something. And that function doesn’t have any money in it, whether it’s bitcoin, dollars, or pesos. It’s funny; I never really understood why people are so in awe of Milton Friedman. Now I understand. He misled generations of economists into believing that inflation is a monetary phenomenon. It’s not. Inflation is the time it takes you to buy something. And that’s driven by human action, policy, laws, institutions, zoning laws, competitive dynamics, and all the other stuff people actually face in their daily lives. Monetary policy is like the wind at sea. It can be helpful or harmful, but it’s not what drives a ship. The latter is determined by creativity, ingenuity, and sometimes just raw muscle power.
To summarize this section, Arch argues that blockchain technology will unleash new ways to interact on the internet. Web3 is the internet of identity, which allows users to exchange data in a more personal and controllable way. Smart contracts allow for the algorithmic execution of transactions with conditions attached. Many aspects of contract law will be executed without human supervision. Bitcoin is a source of money. It will help proliferate a new area of digital currency. But we argue that bitcoin is not a hedge against inflation.
Generative AI (This idea came to us from Microsoft)
At a recent AI event, Microsoft CEO Satya Nadella proclaimed that search and software development are entering a new area of innovation. He refers to the concept of Co-Pilot, a technology Microsoft is already successfully applying to Github. GitHub Copilot assists developers with code suggestions in real-time as they write code on the platform. It uses machine learning models to suggest contextually relevant code snippets that automatically complete common tasks for the user, thereby reducing the time and effort required to write the code. Andrej Karpathy spoke about this months before the Microsoft event. I am not sure whether he is promoting Github. It doesn’t matter. Co-pilot technology, according to Karpathy, is like autopilot at Tesla. It began with lane keeping and will quickly expand to more functionality. Nadella emphasized the copilot idea in search and browsers. The key concept here is that instead of just receiving search results as links, you interact with the system to get better answers.
We believe that Microsoft is well positioned to profit from this idea. Why? First, let’s look at co-pilot technology in software development. Karpathy cheered about using it at Github and how he’d be paying for it if it wasn’t free. That’s a huge statement. He’s a happy customer. Second, let’s assume there is copilot technology in search and browsers. It will fundamentally change our search experience. Today, search is based on compiling a list of links and summarizing them. Microsoft wants to use your search as a prompt to deliver a generated result. In other words, what Nadella is talking about is a search result that is tailored to your query. You spend more time figuring out what you want and less time how to find it.
Now, let’s think about this. What is co-pilot technology? It’s a language model that delivers answers when given a prompt. In Github, it looks at code and suggests better alternatives. How does it do that? By using transformers. What are transformers? This is a pre-trained language model that can generate results when given a prompt. What is a pre-trained language model? It’s a model that finds relationships between text tokens. What are text tokes. Text tokes are letters, words, sentences, etc. It’s whatever the model designers decide to use as a unit of inquiry. For example, if you use letters as tokens, you have 52 tokens, which are the 26 letters of the alphabet in small and capital form plus a number of signs such as periods, question marks, etc. Then you look for common relationships between those letters in texts. How do you look for relationships? Here is what Karpathy calls the crux of GPT models. You create two vectors, one is the key, and the other is the query. For example, you create a key-vector for the word America where all vowels are expressed in one hot encoding like [1,0,1,0,1,0,1]. Then you run this vector as a query on text, where all the vowels are also represented in one-hot encoding. You get the dot product of those two vectors, where the vowels with similar positions will express a high number and everything else will be zero. The key idea is that you create vectors and run dot products on them to express similarity in form and position. At its core, the GPT model is a large number of vectors running dot products on each other. That’s the math behind the model. The model is trained by masking parts of a text and then finding out how well the model predicts that masked part of the text when given the unmasked part. Weights are formed based on the accuracy of the prediction.
What GPT needs are large amounts of compute to do the matrix algebra and a large amount of memory to store the weights.
Once GPT is trained, you give it a prompt, and it predicts letters, words, and sentences. That’s how it works. It predicts the next sequence of text. If you tell it, "I want to program a function that sorts numbers in Python," it will give you a suggestion. It might ask you questions such as what your enduser expects, the size of your dataset, what your hardware constraints are, etc. to better assist you.
This concept of co-pilot technology can be applied to search and even just browsers. That’s where Microsoft wants to make a difference.
We believe this type of language based modeling can be used for non-language applications, too. Tesla is already using it for lane prediction, as Arch noted in their research report.
The best way to analyze this from an investor's perspective is to focus on the customer. Who are the happy customers? Today, we see Microsoft as having happy customers. Github is a great tool to iterate. Github is to Microsoft what driving is to Tesla. It’s fertile ground for collecting data, learning, iterating, and increasing the value of the product.
Github-Copilot is the Autopilot of software development.
Precision Therapies
As the central dogma of biology states: DNA ->RNA->Protein. It means the DNA encodes the expression of RNA, which again encodes the sequence of proteins. Proteins are the workhorses of biology. They do stuff. DNA and RNA are codes; proteins are the things that actually have a physical effect on living organisms. Arch argues that targeting DNA, RNA, and/or proteins allows for more personalized medicine. What exactly does that mean? For example, you could isolate one specific DNA mutation as the main cause of a disease. Then you could edit that mutation and fix the disease. Some of the Crispr companies, like Intellia or Beam, are doing just that. RNA targeting is one step above the DNA source code. Let’s assume you isolate the cause of a disease as a misspelling in the RNA that causes the wrong assembly of proteins. The way to fix that is to edit the RNA. Now, you could skip that whole process and just fix the protein itself. That’s the last step in the trilogy of DNA -> RNA -> Protein.
We recently attended an AI seminar where researchers spoke about the use of AlphaFold technology to accelerate discovery in precision medicine. AlphaFold is an AI-based solution to the question of how proteins fold. In other words, you look at a sequence of amino acids and predict the folding structure of the protein. Let me elaborate. DNA is the source code of life. Here, the necessary information is stored so RNA can fold. RNA molecules then create an amino acid sequence, which folds into a protein. The key with proteins is that it’s not only the sequence of amino acids but also the 3D structure of the protein that determines its function. Two proteins with the same amino acid sequence but different 3D folding structures will execute different functions. Hence, we believe the introduction of AlphaFold into the work flow of discovery will have huge implications. Now you have a closed loop of DNA->Protein -> DNA. You can go back and forth. You can iterate. Any wealth creation ultimately depends on the speed of iteration, the cost of error making, and error correction. With the introduction of genomic sequencing, Crispr and AlphaFold, this process is getting a massive boost in speed.
This brings us to the main problem with biotech. Where is the happy customer? Who is the customer? Who pays? Those questions are not clear. And that’s why iterating around happy customers is difficult, and that’s the main reason why Moore’s Law is not yet happening in biotech. Even if you introduce Moore’s Law into the workflow of biotech through digitization of biology and AI applications such as AlphaFold, you still have to deal with the cumbersome FDA approval process. You can accelerate drug discovery, but you can’t really accelerate the regulatory process. That’s why biotech is still a Zero/One type field, where you invest lots of money in R&D hoping to hit the jackpot. It’s faster and less costly than before, but still way too slow.
Still, we are in a better place today than we were 10 years ago. Before, we used to literally hit shit at the wall, hoping something stuck. Drug discovery, until recently, was a glorified version of alchemy. Today, we have more deterministic ways to discover drugs, and engineering practices are entering the field at speed. What do we mean by "engineering practices"? Take the Golden Gate Bridge. This is a marvel of engineering. It was built by engineers who predicted very well how the structure would eventually fulfill its function. Medicine doesn’t work like that. If we built the GGB like a drug, we'd build hundreds of them and then choose the one that doesn't collapse, if any. With the introduction of sequencing technology, Crispr gene editing, RNA editing, protein editing, and AlphaFold, we are closer to the engineering world. However, the key problem is the speed and cost of iteration. That’s where biotech still falls short.
For investors, this is a serious problem. At Orange Capital Partners, we aim to create wealth by investing in fundamentally new concepts and engineering practices. Biotech falls into that bucket. But it doesn’t allow for rapid iteration. That’s a problem. One way to mitigate this problem is to invest in platform companies that apply technology to solve specific problems across a platform. For example, Intellia is innovating in Crispr gene editing and delivery in vivo. We believe that this approach allows for faster iteration, at least in the preclinical stages. It’s not perfect, but it gives us more confidence that customer value can be massively increased.
The problem with biotech is the FDA. It’s a regulatory body that controls the iterative process around drug discovery. Imagine if software development was regulated by a federal regulatory body. We’d still be in the age of mainframes or maybe transitioning to client-server computing. Regulation might be necessary, but it definitely comes at a high cost. The same applies to self-driving technology in cars. If automakers could iterate faster, cause more accidents, but learn quickly and improve, we’d already be looking at self-driving robots. Those robots would be much safer than human drivers. But the transition is slow because federal regulators are the bottleneck of iteration.
We’re not arguing against regulation here. We’re simply stating that the problem with biotech as a wealth creating sector is the FDA. As a society, we need to find better ways to iterate, make errors, and correct them while still maintaining enough safety. That’s a real challenge. Something economists should be thinking about.
When it comes to Arch’s Big Ideas deck, this is another example of feel-good talk with serious roadblocks. Where are the happy customers? Stories don’t make money, happy customers do.
Molecular Cancer Testing
Tumor cells emit certain DNA fragments that healthy cells don’t emit. You could theoretically find these types of DNA fragments, let’s call them circulating tumor DNA, or ctDNA. Imagine a blood test that captures all ctDNA and gives the patient a good picture of cancer risk. Early detection of cancer is the best cure, since operative or other non-invasive treatments show the best results when treated early. What is going on here? We are looking for methods to detect cancer that are better than our own body’s methods. Sounds promising? Have we ever found diagnostics that are better than our own nervous system? I don’t think so. We might not know exactly why, but we humans are pretty good at sensing when something is wrong. So why all this excitement about early-stage cancer detection through molecular diagnostics? There are two ways this can work. Before we go there, let’s state a fundamental assumption, which is that humans are very good at detecting cancer. The human body has evolved over thousands of years to detect illnesses. We have a very powerful digital computing infrastructure called "the nervous system," which works at amazing speed and at an astonishingly efficient power budget.
When we get sick, we know it. We feel it. We lose our appetite, we get tired, irritated, etc. Our body sends lots of signals to the frontal cortex, our CEO function in the nervous system. The problem with detecting illness is not the detection but our innate cultural bias to cover up such illnesses. In other words, we know when we’re sick but we cover up. This has two reasons. First, it’s an evolutionary habit since in the old times human would just leave sick people behind so not to waste resources on them. In modern times our reluctance to admit illness is because we’re afraid to be held hostage by the public health industrial complex. The latter is optimized not for health care but for making money off sick people. Nobody wants to be held hostage. If we really thought the modern public health industrial complex was there to help us increase our health span, we'd be happily testing for illnesses. Unfortunately, that's not our reality. Testing for an illness like cancer means subjecting yourself to a test to determine whether you will be held hostage by the public health industrial complex. It feels like a prison sentence. No wonder people try to avoid it as much as they can.
Having established that our own nervous system is actually pretty good at disease detection, the question that comes up is whether we can do better with external digital technology such as computers, AI, etc. Can we be faster, more efficient, and more effective than our own nervous system?
We can be even faster at computing the risk of cancer than our own nervous system. By adding external compute through applying Moore’s Law to diagnostics, we can actually out-compute our own detection system. This is possible. It took humans thousands of years to master horses and thus outrun our own ability to move forward. It took us another few millennia to figure out how to apply more force to things through wheels, machines, etc. There is no law of physics that prohibits us from out computing our own nervous system. Let’s think about that. Can our brain come up with something that is better at computing than our own nervous system? Can something create something that is better at doing exactly that thing that something is already good at? Absolutely yes. The human brain is by no means the limit of computational and analytical power in the universe. It’s just one pretty advanced form of thinking. But not the ultimate limit.
With the out of the way we are confident that eventually we will build diagnostics that are better at predicting cancer risks. But who is the happy customer? Unfortunately, like in biomedicine, diagnostics is caught at the crossroads of the public health industrial complex. Rapid iteration is difficult because of regulatory and institutional barriers. The problem for entrepreneurs is to break the stronghold of the public health industrial complex.
Here is one example of how absurd the incentive structure is. Hospitals can bill health insurance providers for expensive cancer treatments. They can’t bill them for forgone treatment that was prevented due to early cancer detection. GDP will decline if we reduce expensive cancer treatments and instead introduce effective early detection mechanisms. Like in biomedicine, entrepreneurs in the diagnosis space have to find a way to iterate faster, make errors, and correct them to truly build wealth. We don’t see a path to such a scenario yet.
For example, we bought Twist Bioscience for exactly that reason. We thought that the company could evolve with increased interest in molecular diagnostics and next generation sequencing. Twist is a synthetic DNA company with adjacent businesses in molecular diagnostics. After holding the shares for three years, we sold them because the company has not been able to deliver profitability. In fact, we believe the company is forced to underprice its services to show revenue growth. It’s not a good situation. We use Twist as an indicator that the molecular diagnostics space is not ready yet. Customers, whoever they may be, are not paying true value. There is a problem with capturing the true value of diagnostics. Take the example of a cancer patient who costs the public health system two million dollars. Now take a diagnostic that, through early detection, could reduce that cost to 100,000 dollars. That’s a 20-fold decrease. Who gets that money?
Molecular diagnostics is a great idea. But we don’t see a path to wealth creation due to the stronghold of the public industrial complex. How are entrepreneurs supposed to circumvent the beast? That’s a question we need answered before we get excited about the future and digitization of biotech.
Electric Vehicles
Cost declines in battery technology make electric vehicles more compelling. We agree. But that’s not the whole story. We argue that EVs are not a wealth-creating disruption. However, Tesla is. It’s the company and its relentless culture of iteration that will drive EV adoption, not the EV factor itself. Let’s parse that argument.
Arch is arguing that Wright’s Law and other factors will drive down EV prices. We disagree. EVs are not fundamentally something like semiconductors, where Moore’s Law drove down prices. Moore’s Law is something that all industry practitioners have experienced. It was Gordon Moore who coined the term. There is an innate feature of semiconductor technology that makes it follow the trajectory of exponential growth. Crucially, it’s not one company driving the process but the field itself. It reminds me of martial arts, where athletes are improving rapidly. Crucially, it’s the field as a whole that is moving forward, not one specific company or individual. EVs are different. Here it is Tesla moving the field forward, not the field moving Tesla. You can see that in the many initiatives Tesla is involved in, like manufacturing technology, vertical integration into the supply chain all the way up to mining, battery research and development, charging technology, etc. Tesla is innovating across the value chain. Vertical integration is key here. It’s similar to Amazon and could computing which was driven by innovations at AWS. Or Apple and the IOS for mobile. Mobile computing was not innately driven by some industry law but by the vertical integration, innovation, and relentless iteration Apple applied to the space.
As a consequence, we agree with Arch about the opportunity for EVs, but for different reasons. EVs will only succeed if Tesla succeeds. Tesla is the driving force. There is no innate law or process that benefits EVs. It’s the entrepreneurial ingenuity of the Tesla team, which is driving down costs and improving the customer value of their EVs, that is driving the rapid adoption of EVs.
Tesla is driving the adoption of Teslas, not EVs. EVs don’t follow any specific law like Moore’s Law or Wright’s Law. It’s Tesla's vertical integration and its relentless push to iterate that are driving down costs and increasing customer value at Tesla. To achieve this type of customer value growth, Tesla is innovating across the value chain.
Autonomous Ride Hail
Robots driving people from point to point. Let’s dissect that. What is a robot? Merriam-Webster defines it as "a machine that resembles a living creature in being capable of moving independently (as by walking or rolling on wheels) and performing complex actions (such as grasping and moving objects"). For our purpose here, we only need the "moving around part." It definitely doesn’t have to resemble a living creature. Or does it? Here is an interesting quote by the then COO of Baidu, Qi Lu, from Wired magazine in 2017: "If you want to truly build digital intelligence to be able to acquire knowledge, make decisions, and adapt to the environment, you need to build autonomous systems. In autonomous systems, the car is the first major commercial application that is going to land." This quote encapsulates everything important about AI. It’s about digital intelligence and a system that acquires knowledge, makes decisions, and adapts to the environment. That’s the key: adapting to the environment. A robot has to acquire knowledge, make decisions, and adapt. There is no law of physics disallowing that process. In other words, it will happen. The question is, "When, and will it happen in this multiverse?" That’s trickier to answer. When species evolved billions of years ago, they evolved exactly like that. They acquired knowledge, made decisions, and adapted to the environment. A mollusk floats through water, adapts, learns, and adapts. Later species acquired eyes and the ability to see. That fostered the Cambrian Revolution because now species could see better, better find prey, or better avoid being preyed upon. Specialization kicks in. Competition kicks in. Competition fosters more development and specialization, and the flywheel goes on and on. Robots are just another form of species that acts in the environment. The task of building them is to evolve a system that learns, makes decisions, and adapts. Its all about learning, decision-making, and adapting.
There are a number of players out there driving around and collecting data. What does collecting data mean? Some players like Waymo use many types of sensors, such as lidar, radar, vision, etc. Tesla only uses vision. The latter is key to scaling, since cost is an important factor in scaling. Tesla is relying on vision only because it is the most cost-effective and scalable solution. Vision is how species evolved. Well, that’s not quite true. Species employ a variety of other sensors, including skin-based pressure and temperature detection, as well as a variety of other signals. So it’s not just vision. But vision is a powerful sensor. Why? Because it gives you a good sense for where you stand in relation to your environment. You know where you are and where everything else around you is located. With that, you can make decisions and adapt. Using more sensor data doesn’t change this equation. In other words, you still just want to know where you are and then adapt. Knowing where you are is a function of sensor data. The question is, how much of that do you need? So far, we don’t have a definitive answer. The powerful concept behind Tesla’s approach is to use vision because it is cost effective. Cost effective solutions allow for more scaling and, thus, faster iteration around real customers. Scaling AI solutions for cars depends on cost. You have to find a way to deploy sensors and AI computers that don't cost as much per car and also don't consume too much power. Tesla is optimizing for low-cost solutions. Another powerful concept behind "vision only" is that if you constrain yourself, you find better solutions. For example, instead of relying on lidar for depth, you can use camera vision and computational tools to determine depth.
Tesla is using vision to determine what the car sees. Vision serves the purpose of answering the question: "Where am I in relation to my environment?" One of the concepts Tesla is using is occupancy networks. Ashok Elluswamy describes the problem as follows:
Predict geometric occupancy
Use multi-camera & video context
Predict dynamic occupancy flow
Persistent through occlusions
Resolution where it matters
Efficient memory and compute
Runs in ~10 ms
The first four are obvious. But the latter three are typical Tesla. "Resolution where it matters", "efficient memory and compute" and "runs in 10ms". Those requirements are the basis for scalability, cost consciousness, and customer value creation. You can’t build a robot at scale without considering these constraints.
To assess the current robot systems and their likelihood of success, we use our familiar metric, which is "happy customers are always right." We can’t assess the relative quality of Waymo versus Tesla versus Baidu, etc. It’s too early. Nobody has a complete end-to-end system. But what we can do is use our principles of wealth creation to test the current systems against them. In order to build something relevant, you have to iterate around happy customers. In order to build wealth, you have to iterate around happy customers. Happy customers are paying customers. Happy customers are parting with resources to consume your product. That’s the ultimate yardstick. Use it.
Will robotaxis happen? Will it be similar to an internet search? Search is a great AI-based tool to find information. It’s algorithmic, which means it crawls the web and indexes it based on algorithms. It gets better with usage. Driving robots is the same. They get better while being driven because you can learn from mistakes. Today we are using humans as annotators. For example, when you do a left turn, the computer looks at what you do and compares it with what it has predicted. Tesla calls this "shadow mode." That’s just one example. The key is to learn from experience, make decisions, and adapt.
How do language models come into play here? What are language models? They are tokenized models with letters as the smallest unit of analysis. Tokenized means you use tokens as the basis of analysis and then find relationships between tokens. Tokens can be letters of the alphabet, words, sentences, paragraphs, etc. It’s up to the algorithm designer to define what the tokens are. GPT stands for "Generalized Pertained Transformer." What is that? It’s an algorithm that finds relationships between tokens. The idea of GPT is that once you train the model on billions of words of text, you can then give it a prompt and it will predict the next word. Like in this sentence, my brain is predicting the next word. Next-word prediction can be generalized. How about you tokenize voxels? Now predict the next move based on the voxels you see and the planned route. In other words, based on what you see, you could use GPT technology to predict which lane you should take. We believe that GPT technology can be transferred to the robot space. It will happen.
Summary of robots
Robots will happen because the laws of physics don’t disallow them. In order to build robots, we need a system that learns, makes decisions, and adapts to the environment. The key to success is to develop scalable, low-cost robots that can be iterated around happy customers. Tesla is most focused on this path. Others take different paths, where cost is not the primary consideration. We disagree with this type of approach. China is a wild card. GPT technology will transfer to the robot space.
Now, what about the point to point thing? Will we use robotaxis in the way we currently envision it? Is "cost per mile" the right metric? Will people consume miles or experiences, or what? Will transport be transport or maybe morph into something else? Youtube was meant to be a dating site, then became a site for cat videos, and today is the world's largest entertainment and education network. What will robotaxis be? What will happy customers care about?
It’s fair to assume that typical transportation concepts will apply to robotaxis. These are the cost per mile, travel time, and comfort. Any type of transport optimizes for cost, time, and comfort. Robotaxis will be the same.
According to Arch, the robotaxi providers will gain the most value. They assume there will be a service called robotaxi, independent of the manufacturer. Those service providers will optimize for cost, time, and comfort. Today, it’s not clear who these providers will be. Maybe Tesla can vertically integrate into the cost, time, and comfort space. We believe they will. Time is a particularly important vector to optimize for. Actually, we believe that time and comfort are much more important than cost per mile. Even today, you see express lanes on highways to speed up your trip. The same applies to toll roads. They are a signal that people are happy to pay for time saved. Time saved can also be stuff you do while you travel, like work, play, sleep, etc. Tesla is working on this with powerful onboard computers and Starlink, which connects them. Eventually, we agree with Arch that the most customer value accrues to the providers of transportation. But it’s not just about cost per mile. It’s about time and comfort. Importantly, this does not only apply to robotaxis. Even at today's level of technology, co-pilot technology can massively increase comfort and save people both time and money.
Transport -> Cost per mile, time, comfort
Robobutler, not robotaxi
Actually, the taxi is not so important. It’s the butler. The robot takes you from A to B, picks you up after a walk, parks and charges itself, picks up groceries, etc. It’s a robobutler, not a robotaxi. Cost per mile is not the key. It’s time and comfort. Saving time is probably the most important thing. Imagine if you could go to the gym and have the car pick up your kid from training. The time you’re saving is worth a lot. Time is the largest cost factor.
Value of Transport = (time + comfort) / cost
Think about the history of transportation. In ships, we transition from using sails to coal to diesel. On the road, we went from foot to horse to rail to car on the road. We went from nothing to flight. There is, of course, a cost component here, but time and comfort are by far the most important factors. What if I told you that you could travel by carriage for 5 cents per mile or by Tesla for 50 cents per mile? Which one would you choose? Ask Ernest Shackleton whether he prefers to walk to the South Pole at 5 cents per mile or go there with a snowmobile for 5 dollars per mile.
Cost matters, but time and comfort are more important decision drivers.
Conclusion
Autonomous ride-hailing is when robots drive you from A to B. We argue that it's less about ride hail and more about butler. We believe the concept of a robobutler is more compelling than a robotaxi. Both have the potential to offer customer value. However, the butler seems more compelling to us. We further argue that cost, time, and comfort are the key drivers of customer value when it comes to transportation. Hence, the successful provider of robobutlers will optimize for cost per mile, time and comfort. The most important factor is time, followed by comfort, and finally cost.
Orbital Aerospace
Reusable rockets reduce launch costs.
Lower launch costs allow for more flexible satellite technology. Small, medium, and large, which opens the door for different use cases. Lower cost increases the market because there are many more use cases.
Satellite launches enable Starlink. Low-cost, high-bandwidth satellites enable low-cost terrestrial broadband.
Orbital Aerospace is like electric vehicles (EVs). There is nothing innate in the industry driving costs down; it’s Space X, ingenuity, iteration, and a relentless focus on lowering costs.
Conclusion
Big Idea talk in the investment community is just that: talk. We argue that such narratives are misleading. They do serve the purpose of setting directions and goal posts. But they are not useful for investors who seek to create wealth. Stories don’t make money. Happy customers do. We discuss and critically analyze a number of so-called ideas, mostly based on the Big Ideas deck with Arch Invest. We appreciate their work and find it inspiring. However, it’s important to recognize such presentations for what they are. They're neither ideas nor creatively evolved thoughts and opinions. They're just compiled narratives that float around Wall Street and serve as a summary of what people talk about in the financial markets. Crucially, these are not investment ideas. We identify two important investment ideas. First, there is no such thing as the adoption of EVs. There is only the adoption of Teslas. It’s Tesla's relentless drive to lower costs and increase customer value through vertical integration that is enabling the rapid adoption of their electric vehicles. Second, Microsoft is the beneficiary of GPT-based co-pilot technology. They will apply it to software development, which allows them to iterate around happy customers.