The left side of figure 5-1 indicates that when there is a performance gap — when product functionality and reliability are not yet good enough to address the needs of customers in a given tier of the market — companies must compete by making the best possible products. In the race to do this, firms that build their products around proprietary, interdependent architectures enjoy an important competitive advantage against competitors whose product architectures are modular, because the standardization inherent in modularity takes too many degrees of design freedom away from engineers, and they cannot not optimize performance.

The issue I have with this analysis of vertical integration — and this is exactly what I was taught at business school — is that the only considered costs are financial. But there are other, more difficult to quantify costs. Modularization incurs costs in the design and experience of using products that cannot be overcome, yet cannot be measured. Business buyers — and the analysts who study them — simply ignore them, but consumers don’t. Some consumers inherently know and value quality, look-and-feel, and attention to detail, and are willing to pay a premium that far exceeds the financial costs of being vertically integrated.

Google trains and runs its Gemini family of models on its own TPU processors, which are only available on Google’s cloud infrastructure. Developers can access Gemini through Vertex AI, Google’s fully-managed AI development platform; and, to the extent Vertex AI is similar to Google’s internal development environment, that is the platform on which Google is building its own consumer-facing AI apps. It’s all Google, from top-to-bottom, and there is evidence that this integration is paying off: Gemini 1.5’s industry leading 2 million token context window almost certainly required joint innovation between Google’s infrastructure team and its model-building team.

On the other extreme is AWS, which doesn’t have any of its own models; instead its focus has been on its Bedrock managed development platform, which lets you use any model. Amazon’s other focus has been on developing its own chips, although the vast majority of its AI business runs on Nvidia GPUs.

Microsoft is in the middle, thanks to its close ties to OpenAI and its models. The company added Azure Models-as-a-Service last year, but its primary focus for both external customers and its own internal apps has been building on top of OpenAI’s GPT family of models; Microsoft has also launched its own chip for inference, but the vast majority of its workloads run on Nvidia.

Google is certainly building products for the consumer market, but those products are not devices; they are Internet services. And, as you might have noticed, the historical discussion didn’t really mention the Internet. Both Google and Meta, the two biggest winners of the Internet epoch, built their services on commodity hardware. Granted, those services scaled thanks to the deep infrastructure work undertaken by both companies, but even there Google’s more customized approach has been at least rivaled by Meta’s more open approach. What is notable is that both companies are integrating their models and their apps, as is OpenAI with ChatGPT.

It may be the case that selling hardware, which has to be perfect every year to justify a significant outlay of money by consumers, provides a much better incentive structure for maintaining excellence and execution than does being an Aggregator that users access for free.

Google’s collection of moonshots — from Waymo to Google Fiber to Nest to Project Wing to Verily to Project Loon (and the list goes on) — have mostly been science projects that have, for the most part, served to divert profits from Google Search away from shareholders. Waymo is probably the most interesting, but even if it succeeds, it is ultimately a car service rather far afield from Google’s mission statement “to organize the world’s information and make it universally accessible and useful.”

The only thing that drives meaningful shifts in platform marketshare are paradigm shifts, and while I doubt the v1 version of Pixie [Google’s rumored Pixel-only AI assistant] would be good enough to drive switching from iPhone users, there is at least a path to where it does exactly that.

the fact that Google is being mocked mercilessly for messed-up AI answers gets at why consumer-facing AI may be disruptive for the company: the reason why incumbents find it hard to respond to disruptive technologies is because they are, at least at the beginning, not good enough for the incumbent’s core offering. Time will tell if this gives more fuel to a shift in smartphone strategies, or makes the company more reticent.

while I was very impressed with Google’s enterprise pitch, which benefits from its integration with Google’s infrastructure without all of the overhead of potentially disrupting the company’s existing products, it’s going to be a heavy lift to overcome data gravity, i.e. the fact that many enterprise customers will simply find it easier to use AI services on the same clouds where they already store their data (Google does, of course, also support non-Gemini models and Nvidia GPUs for enterprise customers). To the extent Google wins in enterprise it may be by capturing the next generation of startups that are AI first and, by definition, data light; a new company has the freedom to base its decision on infrastructure and integration.

Amazon is certainly hoping that argument is correct: the company is operating as if everything in the AI value chain is modular and ultimately a commodity, which insinuates that it believes that data gravity will matter most. What is difficult to separate is to what extent this is the correct interpretation of the strategic landscape versus a convenient interpretation of the facts that happens to perfectly align with Amazon’s strengths and weaknesses, including infrastructure that is heavily optimized for commodity workloads.

Meta’s open source approach to Llama: the company is focused on products, which do benefit from integration, but there are also benefits that come from widespread usage, particularly in terms of optimization and complementary software. Open source accrues those benefits without imposing any incentives that detract from Meta’s product efforts (and don’t forget that Meta is receiving some portion of revenue from hyperscalers serving Llama models).

The iPhone maker, like Amazon, appears to be betting that AI will be a feature or an app; like Amazon, it’s not clear to what extent this is strategic foresight versus motivated reasoning.

achieving something approaching AGI, whatever that means, will require maximizing every efficiency and optimization, which rewards the integrated approach.

the most value will be derived from building platforms that treat models like processors, delivering performance improvements to developers who never need to know what is going on under the hood.

Copywriting, tax preparation, customer service, and many other tasks are or will soon be heavily automated. I can see the beginnings in areas like software development and contract law. Generally, tasks that involve reading, analyzing, and synthesizing information, and then generating content based on it, seem ripe for replacement by language models.

Anyone who makes a living through  delicate and varied movements guided by situation specific know-how can expect to work for much longer than five more years. Thus, electricians, gardeners, plumbers, jewelry makers, hair stylists, as well as those who repair ironwork or make stained glass might find their handiwork contributing to our society for many more years to come

Finally, I expect there to be jobs where humans are preferred to AIs even if the AIs can do the job equally well, or perhaps even if they can do it better. This will apply to jobs where something is gained from the very fact that a human is doing it—likely because it involves the consumer feeling like they have a relationship with the human worker as a human. Jobs that might fall into this category include counselors, doulas, caretakers for the elderly, babysitters, preschool teachers, priests and religious leaders, even sex workers—much has been made of AI girlfriends, but I still expect that a large percentage of buyers of in-person sexual services will have a strong preference for humans. Some have called these jobs “nostalgic jobs.”

It does seem that, overall, unemployment makes people sadder, sicker, and more anxious. But it isn’t clear if this is an inherent fact of unemployment, or a contingent one. It is difficult to isolate the pure psychological effects of being unemployed, because at present these are confounded with the financial effects—if you lose your job, you have less money—which produce stress that would not exist in the context of, say, universal basic income. It is also confounded with the “shame” aspect of being fired or laid off—of not working when you really feel you should be working—as opposed to the context where essentially all workers have been displaced.

One study that gets around the “shame” confounder of unemployment is “A Forced Vacation? The Stress of Being Temporarily Laid Off During a Pandemic” by Scott Schieman, Quan Mai, and Ryu Won Kang. This study looked at Canadian workers who were temporarily laid off several months into the COVID-19 pandemic. They first assumed that such a disruption would increase psychological distress, but instead found that the self-reported wellbeing was more in line with the “forced vacation hypothesis,” suggesting that temporarily laid-off workers might initially experience lower distress due to the unique circumstances of the pandemic.

By May 2020, the distress gap observed in April had vanished, indicating that being temporarily laid off was not associated with higher distress during these months. The interviews revealed that many workers viewed being left without work as a “forced vacation,” appreciating the break from work-related stress and valuing the time for self-care and family. The widespread nature of layoffs normalized the experience, reducing personal blame and fostering a sense of shared experience. Financial strain was mitigated by government support, personal savings, and reduced spending, which buffered against potential distress.

The study suggests that the context and available support systems can significantly alter the psychological outcomes of unemployment—which seems promising for AGI-induced unemployment.

From the studies on plant closures and pandemic layoffs, it seems that shame plays a role in making people unhappy after unemployment, which implies that they might be happier in full automation-induced unemployment, since it would be near-universal and not signify any personal failing.

A final piece that reveals a societal-psychological aspect to how much work is deemed necessary is that the amount has changed over time! The number of hours that people have worked has declined over the past 150 years. Work hours tend to decline as a country gets richer. It seems odd to assume that the current accepted amount of work of roughly 40 hours a week is the optimal amount. The 8-hour work day, weekends, time off—hard-fought and won by the labor movement!—seem to have been triumphs for human health and well-being. Why should we assume that stopping here is right? Why should we assume that less work was better in the past, but less work now would be worse?

Removing the shame that accompanies unemployment by removing the sense that one ought to be working seems one way to make people happier during unemployment. Another is what they do with their free time. Regardless of how one enters unemployment, one still confronts empty and often unstructured time.

One paper, titled “Having Too Little or Too Much Time Is Linked to Lower Subjective Well-Being” by Marissa A. Sharif, Cassie Mogilner, and Hal E. Hershfield tried to explore whether it was possible to have “too much” leisure time.

The paper concluded that it is possible to have too little discretionary time, but also possible to have too much, and that moderate amounts of discretionary time seemed best for subjective well-being. More time could be better, or at least not meaningfully worse, provided it was spent on “social” or “productive” leisure activities. This suggests that how people fare psychologically with their post-AGI unemployment will depend heavily on how they use their time, not how much of it there is

Automation-induced unemployment could feel like retiring depending on how total it is. If essentially no one is working, and no one feels like they should be working, it might be more akin to retirement, in that it would lack the shameful element of feeling set apart from one’s peers.

Women provide another view on whether formal work is good for happiness. Women are, for the most part, relatively recent entrants to the formal labor market. In the U.S., 18% of women were in the formal labor force in 1890. In 2016, 57% were. Has labor force participation made them happier? By some accounts: no. A paper that looked at subjective well-being for U.S. women from the General Social Survey between the 1970s and 2000s—a time when labor force participation was climbing—found both relative and absolute declines in female happiness.

I think women’s work and AI is a relatively optimistic story. Women have been able to automate unpleasant tasks via technological advances, while the more meaningful aspects of their work seem less likely to be automated away.  When not participating in the formal labor market, women overwhelmingly fill their time with childcare and housework. The time needed to do housework has declined over time due to tools like washing machines, dryers, and dishwashers. These tools might serve as early analogous examples of the future effects of AI: reducing unwanted and burdensome work to free up time for other tasks deemed more necessary or enjoyable.

it seems less likely that AIs will so thoroughly automate childcare and child-rearing because this “work” is so much more about the relationship between the parties involved. Like therapy, childcare and teaching seems likely to be one of the forms of work where a preference for a human worker will persist the longest.

In the early modern era, landed gentry and similar were essentially unemployed. Perhaps they did some minor administration of their tenants, some dabbled in politics or were dragged into military projects, but compared to most formal workers they seem to have worked relatively few hours. They filled the remainder of their time with intricate social rituals like balls and parties, hobbies like hunting, studying literature, and philosophy, producing and consuming art, writing letters, and spending time with friends and family. We don’t have much real well-being survey data from this group, but, hedonically, they seem to have been fine. Perhaps they suffered from some ennui, but if we were informed that the great mass of humanity was going to enter their position, I don’t think people would be particularly worried.

I sometimes wonder if there is some implicit classism in people’s worries about unemployment: the rich will know how to use their time well, but the poor will need to be kept busy.

Although a trained therapist might be able to counsel my friends or family through their troubles better, I still do it, because there is value in me being the one to do so. We can think of this as the relational reason for doing something others can do better. I write because sometimes I enjoy it, and sometimes I think it betters me. I know others do so better, but I don’t care—at least not all the time. The reasons for this are part hedonic and part virtue or morality.  A renowned AI researcher once told me that he is practicing for post-AGI by taking up activities that he is not particularly good at: jiu-jitsu, surfing, and so on, and savoring the doing even without excellence. This is how we can prepare for our future where we will have to do things from joy rather than need, where we will no longer be the best at them, but will still have to choose how to fill our days.