Generative AI and the Economy
Can Generative AI lead to Economic Growth for Everyone?
This is the fourth post in a series exploring the societal impacts of advances in AI, with an eye towards the future. In this post, we'll discuss AI’s potential to drive business value at both the enterprise and individual level, and how it might not perfectly align with our current capitalist system.
AI - the Next Great Stock Market Hope
Investors were in a precarious position in 2022. With rising interest rates and high price ot earning multiples, predictions of a recession seemed to make it a foregone conclusion, and there wasn’t an obvious growth engine as the crypto frothiness really started to subside. Enter AI!
Investor chatter and hype for AI was already building, but the release of ChatGPT in November 2022 drove it to a new level. AI was referenced 2x as much in quarterly earnings calls in Q1 2023 compared to the five year average, and 4x as much in Q1 2024. Nvidia became the most valuable company in the world, as the GPUs it produces are the model training engine for generative AI models.
All big tech companies scrambled to organize their AI strategies. Microsoft rode to new highs behind its backing of OpenAI, Amazon released Amazon Bedrock and put a hefty investment in OpenAI competitor Anthropic. Facebook took a different approach, and has become the preeminent open-source model provider with their Llama model series, in part to messy the closed source LLM environment. Apple was quiet for a while, but made Apple Intelligence the forefront of their WWDC conference in 2024.
Is AI for Real?
The quick pivot from Crypto to AI led some to draw parallels between the two. There certainly is some truth to that sentiment, but AI’s rise to prominence has much more substance behind it. While many in the crypto/Web3 world struggled to identify any actual use cases for the technology, ChatGPT saw the fastest rise to 100M users of any app in history. And while some of Nvidia’s stock gains are hype driven, they also increased revenue almost 6x over the past five years.
In a 2024 earnings call Google’s CEO Sundar Pichai said “The risk of under-investing is dramatically greater than the risk of over-investing,” as a means to motivate increased spending on AI. Capabilities today are astounding, and if you extrapolate further, it’s easy to see how AI will be a massively disruptive technology in the future.
What’s a bit harder to see, is the business model that aligns with growth-oriented enterprise. Nvidia and similar hardware providers are a glaring exception, as their business model is simple. As long as model training is dependent on GPUs, and as long as Nvidia drives the GPU market, they will continue to make money hand over fist.
Where I think there are more questions are at the application layer, whether you are an LLM foundation model provider, or a application developer on top of open or closed-source LLMs. There are three key challenges to monetizing at the application layer at the enterprise level:
Can you exit from the demo phase to a production application?
As discussed in prior posts, the methodology that powers generative AI inherently does not provide consistency and can also be inaccurate. But if you try something 10 times, it might work really well 9 of them, but that one time that didn’t work could be catastrophic. This lack of consistency is incongruent with traditional software development, and can be an especially large problem in more high-stakes industries, like healthcare.
Machine learning has been deployed in enterprises for decades, so it would be reasonable to assume that we have already solved this problem. Traditional machine learning is responsible for a specific task, so if a prediction is sub-optimal, it might lead to a poor user experience but nothing more. As an example, if a recommendation engine for a streaming service offers movies you’re not interested in, you might just switch streaming providers.
LLM-powered applications in particular introduce more potential issues because of the sheer number of tasks they perform, and generalized nature of the “predictions” they can make. Using the same streaming service example above, if the LLM output is presented back to a user directly, it may lead to issues in data format returned (ex. The response was not a valid JSON object), or it might just produce something completely unrelated to movies as recommendations, or movies that are not in the available catalog.
These are not insurmountable challenges, and there is a tremendous amount of research and effort involved in trying to minimize these issues, but the methodologies behind the current state-of-the-art are inherently probabilistic, so will make mistakes. Combining probabilistic and deterministic solutions is a promising direction, such as using an LLM to understand the semantic meaning behind a question, and calling a Python function to perform numeric calculations, and combining the Python response and LLM-generated text to the user,
Additionally an underappreciated challenge is how hard it is to evaluate LLM-based applications. LLM benchmarks are inherently very general, and are unlikely to be relevant to a specific LLM-based experience. Given the generalized output produced by LLMs, it can be really hard to tell how “good” a response is.
A good example of this is an LLM-based nutrition coach that reviews user meals, and makes recommendations on how to improve them. Even if reasonable looking text that discusses the meal is produced, how can you tell how “good” the coaching was. Did the model take the user’s dietary restrictions and health conditions into account? Did the guidance make reasonable suggestions on how to replace unhealthy meal components with healthy alternatives?
At present the only way to be able to quantitatively measure the quality of coaching is to create your own manual process to review conversations and potentially convert that manual process into a model-based solution, or outsource the evaluation to another LLM.
In summary, while it seems tantalizingly easy to just throw an LLM at a problem and walk away, systems integration, edge cases and evaluation makes a true production deployment much more difficult than initially meets the eye.
Do you have a path to a persistent competitive advantage?
So you’ve worked hard and turned your fun looking demo into a hardened production application, how can you be sure that your application won’t be immediately disrupted?
If you’re a foundation model provider and have moved to the top of the benchmark leaderboards, do you have some methodological approach that you think will give you a persistent advantage? And does your place on the leaderboard actually translate to a better user experience?
Since the ChatGPT release in 2022, we’ve seen a tremendous increase in performance and a push towards multi-modality, and stiff competition in the foundation model market, suggesting that methodological advantages are likely temporary.
Instead, branding, data, and perhaps in the future regulatory or licensing based advantages seem like paths to more a persistent competitive advantages.
Branding is straight forward. To many, LLMs are still just “ChatGPT.” That initial brand advantage in addition to the technical head-start has driven up OpenAI’s valuation to higher levels than Intel as of this writing.
As discussed in a previous post, personalization is not possible for LLM users out of the box, so a user must explicitly provide personal details to a model-based experience to have those details integrated. If a model provider is able to collect personal information from users, and fit responses to their personal requirements, that provides product stickiness as it would take considerable effort for a user to provide those same details to another provider.
Regulation and licensing are still two areas playing out (more on that in the next post), but its logical to assume that at some point these will represent key advantages for companies to gain trust and key branding with consumers and businesses.
For application developers leveraging LLMs, competitive advantages may be even more brittle. In addition to all the risks above, the risk of the foundation model provider being used under the hood completing the task by itself looms large. Using the same nutrition coach described previously, if you are using the ChatGPT as your underlying model, you need to be confident that your process is differentiated enough that users have a need to go to your experience, rather than ChatGPT itself.
We’ve seen this story before with major cloud providers. It has been common in the past decade to see AWS release an in-house service that mirrors a popular open source tool. These cloud providers needed to put in the effort to spin up a team and a service. In the case of LLMs, capabilities may emerge “for free” from model improvements without even dedicated effort in a specific area, so it leaves application developers even more on edge.
For existing companies, is an AI-based approach a better business model than the incumbent approach?
This is obviously only relevant to existing companies, but it’s a very tricky challenge. Google Search is a great example, and a great example of a technological advancement that aligned perfectly with a business model. In Q1 2024 Google earned $46 billion from search advertising alone, over 50% of it’s total revenue.
Search is a very reasonable use case for an LLM-based solution. Competitors like Perplexity AI have emerged combining traditional search with links and LLM-based responses. Google itself has added in “AI Overviews” to it’s search functionality, leveraging its LLM (Gemini) to produce an LLM-based response to a user search in addition to its traditional link-based display. AI Overviews has come with some hiccups, but it seems that Google is committed to at minimum experimenting with integrating LLMs into its search workflow.
Even if LLM-based search becomes a better end-user experience for users, will it drive more revenue for Google? Or will it cannibalize their cash cow, and not produce as much revenue per user or per query despite a better user experience?
I’m sure new monetization techniques will emerge around generative AI, and it already has proven to be a great way for an individual person or small group to compete at levels that would have previously been impossible. At the same time, it’s easy to see the impact of generative AI being at least temporarily deflationary, and perhaps deflationary in a more permanent fashion.
You can extend this same concept to other industries. If adequate healthcare can be provided via LLM, it should reduce costs relative to human counterparts, and you would hope competition in that area would also drive down prices to make healthcare more affordable. The same is true for tax or legal advice, generating business plans, and branding, and the many other applications of Generative AI.
If this is true, what are the implications on the economy as a whole?
Could Generative AI be an Inequality Engine?
If generative AI does have deflationary impacts on revenue, it flies in the face of traditional capitalist motivations of growth. If it starts to meet the promise of its potential, it certainly will create efficiencies by automating what previously what human-focused work. If the combination of these factors occur, it seems natural that organizations will make efforts to be leaner, with existing companies reducing staff, and new companies requiring less staff to begin with. This may occur even if new business models emerge and there isn’t a deflationary impact on revenue.
The positive side of this equation is it could bring costs down for goods and services. If healthcare becomes cheaper and easier to access, that benefits all healthcare consumers, so everyone.
On the negative side, are we leaving people out of this new economy? If companies can operate with less and less people, there’s inherently less available high-paying jobs out there. If there are less high paying jobs available, there’s more of an emphasis on asset appreciation to obtain and maintain wealth than income. We have already seen this play out to some degree in the past 20 or so years, it’s possible that generative AI accelerates this trend forward even faster based on it enabling people to do more with less.
Conflating Technological Process with Capitalist Outcomes
Connecting all these topics together, its not obvious that generative AI technological advancements AND huge benefits to consumers will lead to enterprise value. Or even if it does, it’s possible that the enterprise value is concentrated to a small percentage of the population.
This is not an easy problem to solve. There’s so much potential for generative AI to improve people’s lives, but there is certainly risk that in going to fast we’re not set up for a functional society. This combines the topics covered in prior posts about the cost of education, and potential commoditization of knowledge work, and the individual power that can be gained from leaning into generative AI technologies.
Thinking about how we can bring everyone along for the ride and improve society will be a key challenge in the coming years if generative AI meets its promise. More on this topic in the next blog, when we will dive into existing regulatory efforts, and how they could be improved.