Step Functions in Intelligence. OpenAI, What's Next and a Perspective on Where We're Headed
All Things Venture #093
So originally, I wrote this essay as a reaction to OpenAI’s DevDay, and I wanted to share my thoughts on what that moment in time meant for all of us. But then the literal Succession style made for TV drama of Sam Altman’s subsequent firing and rehiring happened and I (like everyone else in tech) was sucked into that vortex of drama.
Whether it’s the comparisons to Steve Jobs, the resurfaced profiles of Altman’s ambition, or the rumors of an AI breakthrough; it feels clearer now, more than ever, that something consequential is occurring in AI. The default optimist in me says, we’re heading for an exciting new future.
The reluctant skeptic in me asks, is it 1634 and am I just buying tulips?
Nonetheless, I think it’s extremely important to pay attention to this space and I for one, would rather be an active participant proven wrong than a passive onlooker forced into reaction.
So with that in mind, I wanted to share my perspective on what OpenAI’s DevDay and the continued progress of AI as an industry all mean.
First off, OpenAI has clearly cemented themselves as the center of gravity in the AI industry. They have access to the talent, the compute, and the distribution and they are swinging for the fences. Becoming the App Store for an entire technological sea change is heady business. And we have recent precedent to look toward the amount of economic value that can be generated.
For context, Apple recently announced that their app store generated $1.1 Trillion in billings and sales in 2022, and in the process of facilitating those transactions, likely took home somewhere on the magnitude of $30bn - $60bn of revenue for themselves.
This is clearly a tremendous opportunity, and there are credible arguments to make that a large language model in your pocket is orders of magnitude more impactful than a smartphone. The smartphone was a step function in global connectivity.
AI is attempting to be a step function in global intelligence.
The grand narrative of what is occurring in the race to build AGI is what Sam Altman calls, the “slow takeoff” wherein the existence of AGI is gradual, and over a period of time. In a slow takeoff scenario “no one is going to agree on what the moment was when we had the AGI,” but a binary point in time is implied.
In the future some people will say that AGI was achieved in March of 2028, and others will say it was achieved in November of 2029 but in general the implication is that there will be a clear before and after, no different than the internet, the iPhone, or electricity. And in that period of “after” this new technology will spread throughout the world. Access and usage will compound on a multi decade basis, no different than the persistent longitudinal progress that occurred with the creation of the internet.
By no means is it a certainty that we will have this type of up only environment. The market is still in the early innings of figuring out what is enduring innovation and what is temporal. ChatGPT, board coups notwithstanding, feels enduring. Jasper AI feels temporal.
In any event, something is here to stay. And I find the concept of a slow takeoff to be a useful mental model to think about what is going on around us. To simplify my own thinking within the slow takeoff, I break the slow takeoff into eras. I describe these below
In the Iteration Era, the foundation models and the stakeholders around foundational models seek to improve the quality of their models. These individuals and teams focus on latency, accuracy, security, cost, and specificity. This is the era of the “Slow Takeoff” period that we are currently in. This era is dominated by the competition amongst large language models such as OpenAI, Bard, Anthropic, Llama (Facebook), Mistral and the dev ops tooling/infrastructure such as Langchain, Pinecone, Coreweave and others. I call this the Iteration Era because we’re observing an iterative process wherein engineers, researchers, and entrepreneurs iteratively test new ideas that yield positive results that push the industry forward in a positive upward loop.
In the Application Era, a wide swath of entrepreneurs will find repeatable success developing products in unrelated industries by leveraging the improvements made in the Iteration Era. The businesses that are created and survive during the Application Era will be the customers of the large language models and dev ops tooling/infrastructure of today. In effect, this new application layer will continuously and iteratively take the latest improvements deployed during the iteration era and commercialize them as unique products. Commercialization is the name of the game in the Application Era, which we will explore futher later on.
In the Alignment Era, the magnitude of AGI’s impact is clear to a subset of industry leaders and there will be greater precision around the timing through which AGI may be widely available. We enter the Alignment Era only through the success of the Application Era, which is similarly dependent on the success of the Iteration Era. I principally believe the Alignment Era will be defined by broader calls for fiscal responses to AGI. In the fullness of time, this period will also be characterized as the beginning of multi-decade employment declines in certain industries. In addition, I think Larry Summers newly appointed postion as a Board Member of OpenAI is a perfect example of more of what’s to come.
In the AGI Era, the world enters a robust period of multi-decade growth and improving quality of life outcomes on a global scale. Billions of humans around the world interact with multiple different agents on a daily basis. AGI agents are deployed to deliver traditionally high cost, low scalability offerings such as education, healthcare, and financial services across the world. Improved access to the above services creates a more healthy, educated, and wealthier populace on a global basis that leads to greater material wealth and real GDP growth
I personally believe that we’re firmly in the Iteration Era because if you pay close enough attention to the space, you can find individuals and teams looking to improve the quality of the foundational models. From where I sit today, people are generally focused on latency, accuracy, security, cost, and domain specificity. A few examples of teams pursuing different types of progress can be found here, here, and here.
Fundamentally though, I believe that improvements to foundational models will continue to be made and will continue to come. There’s enough recent history showing us so. In the past decade we’ve gone from training AI models to recognize and classify images of cats, to training AI models that can fully replicate the voice, cadence, and style of multi-platinum artists in a near indistinguishable manner.
When I dig a bit into the history of AI research, there’s really nothing that tells me the field is going to materially slow down in the near term. Scaling compute, and access to training data seem like the real limiting factors and the machinery of capitalism seems squarely focused on alleviating those speed bumps.
Further, if the recent history of the AI industry is used as a guide, where researchers continuously find and develop new techniques to train models and make them more efficient (i.e ImageNet, Generative Adversarial Networks, Transformers, RLHF) the path forward seems pretty clear: foundational models will continue to improve.
The billion or trillion dollar question to answer is, by how much and how quickly?
It’s not a question I’m equipped to answer, nor do I believe anyone knows. So I default to my ideological belief that the upward trajectory of innovation is persistent and constant within a capitalist society. Whether it is with the invention of electricity, the internet, or personal computing, history constantly reminds us that betting against innovation in the long run, is a bad trade.
Case in point, here’s a hot take from the NYT economist Paul Krugman that aged like absolute milk:
“By 2005 or so, it will become clear that the Internet’s impact on the economy has been no greater than the fax machine’s.”
Am I unfairly dunking on my guy Paul? Yes, a little bit, but the substance of my message I hope is clear: Don’t be a luddite.
So with that anti-luddite message in mind, and if I operate with a fundamental belief that foundational large language models will only improve in the coming years, and that through their improvement their infrastructure will bring the industry closer to AGI, I naturally think about the intermediate period between now and then.
These intermediate years are the Application Era. It’s during this Application Era when I think we’ll enter into a stage where the old adage of, “the future is already here, - it’s just not evenly distributed” will be most salient.
You could make arguments that we’re in the earliest stages of the Application Era today. Think about what’s already occurring around us. Humane launched their AI pin. Character AI has 4M monthly active users. A fake Drake song went viral and racked up millions of views. Forward launched an AI doctor in a box. There are businesses looking to automate entire swaths of the traditional jobs to do for lawyers and accountants. And it doesn’t stop there. In copywriting, video generation, video editing,security questionnaires, recruiting, outbound sales there are dozens of AI first companies focused on building businesses around these professional services, and to top all of that off OpenAI is on pace to grow to more than $1bn of ARR in a single year.
All of this change around us has primarily occurred in the last twelve months and if we look at the world through the lens of biology, it’s clear that a new organism has entered the ecosystem, and more likely than not it’s here to stay.
Timing is a huge component to the job as an investor, and one of the things I’ve realized about the Slow Takeoff is that there aren’t discreet endpoints between the eras. They overlap. I visualize this roughly below.
No different from how it takes time for private market valuations to adjust to the repricing that occurs within public markets, it takes time for the improvements made in the Iteration Era to be fully absorbed during the Application Era. However, this doesn’t mean that individual actors and teams will operate at a uniform speed.
Information asymmetry exists and it creates opportunities for people to get ahead, so I naturally think that entrepreneurs at the application layer who have the strongest pulse on the improvements coming out of the foundational model layer will be in the best position going forward. Case in point, OpenAI has consistently demonstrated their ability to be ahead of the curve, and even today their team is a testament to the future being unevenly distributed as they continue to make progress on the improvements that will undergird GPT-5 and GPT-6.
What I think will principally define the Application Era is that we’ll have completely new product interactions enter our daily lexicon. I think about it this way, the “like”, the “tweet”, the “swipe”, the “scroll”, the “insta'' these are all now standard actions that billions of people take each day.
They were enabled or accelerated by the mobile computing revolution, and fundamentally they were product adaptations (built by entrepreneurs) that simplified UX. The AI equivalent of these haptic feedback loops are the canary in the coal mine for us being in the Application Era of AI, and when more of them are occurring that lead to behavior change at an individual level, that’s likely a good sign that we’re deeper in the Application Era.
To add some more concrete thinking around the Application Era I think it will also be defined by a few of the following trends:
Broad based margin improvement for enterprise buyers. In order for any of this to work, large corporate entities from Coca Cola to Koch Industries will need to achieve some sort of ROI on AI deployments. This is occurring today, but not at scale.
Decline of Co-Pilot for “X” at seed stage formation. Just like you would rarely see a Salesforce, or Toast, or Microsoft competitor being pitched today, if the Application Era is really taking off, the activation energy of early stage markets will focus on other pursuits. No investor wants to back the nth player in a given market, particularly in scenarios where mind share is consolidating around one or two players.
New product interactions enter the lexicon (talked about this above)
Professional services transitions into a business model with embedded leverage no different from media, technology, or financial services (i.e specifically fund managers); Access to expertise (i.e time), the limiting factor, for professional services providers becomes infinitely scalable
As evidenced by Humane and other wearables, the new lexicon may not be actions that we type or touch but something different will emerge. Chat is clearly the leading interaction and what people are gravitating toward to date, but improvements need to be made. My strongest argument for why we’re in the earliest innings of the Application Era is driven by the fact that not enough teams have really figured out how to nail the UX given current constraints. The current constraint being that the models are not 100% accurate, prone to hallucination, and lack basic reasoning skills, so while chatbots are an intuitive experience they’re not always the best experience. Said another way, the limitations of large language models today limit the ability for valuable user experiences, the limitations around valuable user experiences, limits the repeatable, enduring success of new applications.
Despite the relatively limited success of the Application Era to date, the reason why the Application Era is so important from an investment perspective is because it’s A) the immediate greenfield opportunity, and B) going to have a higher terminal number of enduring companies (at least relative to the foundational models and infrastructure that support them).
Investors reallllllly care about the terminal number of enduring companies because it is functionally the most important output of our work. As an investor I can only meet with so many companies a year, so picking the broad based investment theme that is likely to produce the highest amount of enduring companies is one of the few levers within my control that can dictate success. Picking the right theme still feels more like market beta, but market beta in a venture capital context can still produce incredible returns.
Market beta aside (no one wants to be beta), what I think is so fascinating about this particular point in history is that it generally feels like the future is going to be incredibly different. A thought experiment I ask myself all the time is, if you were an investor in the late nineties, “would you have invested behind the theme of the internet?” In the late nineties the internet was widely available, but it hadn’t necessarily ushered in the “information age” we live in today. In the late nineties there were some odd 150 million global internet users. Today, there are 5.3 billion. The world has shifted from analog to digital, from hyper-local to hyper-global. The Arab Spring, Donald Trump, the #MeToo movement, TikTok, Wall Street Bets all of these cultural, political, and economic events were directly influenced or the result of an internet based society.
Further, all of these events intensely mattered.
The world is flatter, faster, and in the grand scheme of things fairer than our ancestors could have ever imagined, and once again we are standing at some sort of inflection point. I think it behooves all of us to be engaged at this critical point in time. I think it behooves all of us to have an opinion, to share our opinions, to have them disproved, and recalibrated in a manner that allows us to maximize the aggregate value that gets created.
The last thing I’ll leave you with is this. The internet is enabling technology that created a step function improvement in global connectivity. I’m writing this to you all on the Sunday after Thanksgiving at my parent’s place in Texas. I’m sitting next to my grandmother, and my grandfather is behind me as I type. My grandmother and grandfather, who are both in their eighties, have been married for sixty-three years. They were both born in the 1930s.
This morning, like most other mornings this past week, I had breakfast with them, which is a unique and precious joy. To be able to spend simple quality time with them; to hear snippets of their life journey deepens an already singular connection.
In their own ways, my grandparents have both adapted to the step function in global connectivity. This morning we live streamed their local church service on Youtube. After we had breakfast, we walked through our family tree on 23andMe. I was able to see firsthand our family tree, and the lineage we can trace back all the way to the American Revolution (i.e totally separate story but super cool). Our broader extended family stays in touch through a massive WhatsApp and a weekly Zoom call where various aunts, uncles, cousins, and everything in between gather to reminisce, to catch up, and to connect.
I share these anecdotes because they are acute reminders that the world as it is, is not how it will always be. The step function in global connectivity changed the world. It changed how my family and how countless other families across the world operate. It was impactful, it continues to be a daily force in all of our lives, and the world would arguably be much more difficult if we didn’t have the internet or the smartphone.
So knowing the scale, the importance, and the impact of a step function in global connection. Think about it for yourself.
In 80 years, what will the scale, the importance, and the impact of a step function in global intelligence ultimately be?
The internet undeniably changed the world, and the fact of the matter is that there’s a non-zero chance AI does the same. I am deeply interested and excited about what’s to come, and as a default optimist and student of history I don’t intend to miss my generation’s opportunity to invest behind the theme of AI.