Generative AI #1: Why AI Matters
A newsletter on AI, entrepreneurship, creativity, and mindfulness
For years, I have been telling my former students that I plan to launch a newsletter on AI, entrepreneurship, mindfulness, & related topics. The moment is finally here. Season 1 will feature a dozen posts over 5 months that will break down Generative AI and its impact on creative jobs and industries. Thanks for being part of this.
“Generative AI is the NFT of 2023,” said an exec as we were talking about the business implications of AI. While NFTs have some potential and might stage a comeback sometime in the distant future, the NFTs of 2021 were all driven by hype rather than any fundamental value they offered. Is it possible that AI – including its recent avatar, Generative AI, that can generate essays, computer code, visual art and even music – is just another example of the never-ending tech hype cycles?
Let me start by making my position clear. I believe that AI is like steam engine, electricity, or computers … that rare breed of technology that changes life on earth as we know it. There was life before and after electricity; two different ways for humanity to exist. Similarly, we will one day talk about life before AI and after AI. AI will open up new opportunities for innovation and productivity for firms and large-scale economic growth for society (And throw a million challenges at us along the way).
But how do I know this?
General Purpose Technologies
I’ll describe two recent research studies to justify this PoV. The lens that I want to use to elaborate on why I feel that way is a concept known as a general purpose technology. General purpose technology is a term for a technology that has the potential for pervasive use by a large number of industries -- not just a few select ones like high-tech -- and is characterized by what's called “technological dynamism,” which means that they are improving rapidly and have lots of different use cases. Now, if a technology is a general purpose technology, it has a huge impact at the macro and the micro level. At the macro level, general purpose technologies are known to stimulate innovation and economic growth. And at the micro level, it's known to impact competitive dynamics. Recognizing general purpose technologies early and building the right strategy for them can separate winners from losers.
If you look at what makes a technology a general purpose technology, research suggests that there are three main characteristics that are associated with them. The first is that the technology has widespread use across multiple industries. The second is that the the technology is improving rapidly i.e. lots of R&D. And the third is that the research is also spread across many industries. One common way to determine if a technology is general purpose or not is to look at patent applications and approvals related to the technology and its distribution across industries. The problem with using patent data to identify general purpose technologies is that patent filings and approvals are lagging indicators and what we need as managers is a leading indicator. A recent research study by Avi Goldfarb, a professor at university of Toronto, and his colleagues Bledi Taska and Florenta Teodoridis explores a novel alternative. They use hiring data, which is available when companies are making technology investment decisions, as opposed to patent data which can start to come ten or more years after the hiring decisions. The researchers find that the hiring data can be an early indicator of what patent patterns will look like down the road (and therefore whether the technology is likely to be general purpose down the road).
The authors analyzed a database of over 200 million job postings and shortlisted job postings from 2019 that were connected to 21 emerging technologies including machine learning [footnote 1], Blockchain, CRISPR, nanotechnology, and many others.
To assess whether a technology has widespread use across multiple industries, the authors assessed whether the job postings associated with a technology were spread across multiple industries. The authors find that machine learning was ranked fifth among these technologies, behind Telecom, Robotics[footnote 2], cloud computing, and service oriented architectures. Machine learning jobs are spread across many industries, ranging from food services, administrative services, to education, finance, finance, healthcare, information, technology, manufacturing, professional services, retail, and so on. In contrast, they find that job postings for technologies like quantum computing, CRISPR, and Nanotechnology are not that widespread (for example, with quantum computing, most of the jobs are in the “professional scientific and technical services” sector).
To assess whether a technology is improving rapidly, the authors look at the number of research jobs tied to each technology as well as the percentage of all jobs related to a technology that are research jobs? They find that ML had the most research jobs in their database followed by cloud computing. In terms of percentage of job postings that are related to research, ML ranks fourth behind CRISPR, Nanotechnology, and Polymer Science. The final criteria is whether the research is itself spread across multiple industries. To assess this, the authors look at whether the research jobs are spread across multiple industry sectors. ML ranks at the top followed by Cloud Computing and Robotics. In contrast, while CRISPR, Nanotech, and Polymer Science had a high percentage of jobs that are research-related, those jobs are not spread across a wide range of industries.
Combining all of these criteria, the authors conclude that “"a cluster of technologies comprised of machine learning and related data science technologies is relatively likely to be GPT.”
Image Source: A Human’s Guide to Machine Intelligence
If AI is indeed a general purpose technology, the implication is that most industries are likely to change because of AI. This is not unlike how the Internet completely transformed entire industries over a period of twenty years. The other point to note is if one invests in a general purpose technology, the impact is felt much later. The time lag is because R&D can take some time to produce results. And so there're going to be a lot of early failures with investments in ML, some of which we have already seen. For example, in 2015-2016, if you had asked any CTO what they were doing with AI, the first answer would have been “chatbots” because Facebook had just launched a chatbot platform to much fanfare. However, by 2019, most firms had shut down those chatbot explorations because it was too early; chatbots weren’t ready for prime time yet. We will likely see several early failures and costly retreats by firms in response to those early failures.
What About generative AI?
The Goldfarb et al study is based on data from 2019 when most of the AI were “discriminative models,” the kind of AI models that make predictions about whether a credit card transaction is fraudulent or whether an email is spam or not. In short, analytical predictions based on historical data. But tools like ChatGPT and Stable Diffusion are Generative AI; they can generate content and perform creative tasks. Are generative AI likely to be general purpose too? Given the Generative AI wave has just taken off, it’s too soon to look at patent filings or even job postings.
To address the issue, my colleague Dan Rock and his coauthors at OpenAI identified the set of tasks associated with over 1,000 professions and coded whether those tasks will be impacted by large language models (LLMs) like ChatGPT (for example, a sample task for a kindergarten teacher might be to “involve parents and older students in children’s activities”). For each task, they had human annotators assess the level of exposure to tools like ChatGPT. They define exposure as whether a large language model (LLM) like ChatGPT would “reduce the time required for a human to … complete a task by at least 50 percent.” As an alternative to subjective labels by humans, they also had OpenAI’s GPT4 provide similar annotations of whether a system like ChatGPT can reduce the time for task completion (recent work suggests that ChatGPT can be surprisingly good at such annotation tasks).
Based on their analysis of over 1,000 occupations, the authors conclude that “approximately 80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of GPTs, while around 19% of workers may see at least 50% of their tasks impacted.” Further, the level of exposure is not evenly distributed across occupations. Higher wage occupations are generally more exposed to LLMs than lower wage jobs. Jobs that require more preparation (i.e require greater education, job experience, or on-the-job training) are more exposed than jobs that require less preparation. Further, occupations associated with programming and writing skills are more exposed than occupations associated with critical thinking and science. The most exposed jobs, according to the authors, are interpreters and translators, survey researchers, creative writers, PR specialists, proofreaders, court reporters, Blockchain engineers, etc. If one allows for the fact that complementary software can be built in addition to or on top of LLMs , the most exposed jobs include professions like mathematicians, tax preparers, auditors, writers, reporters, administrative assistants. To be clear, their definition of exposure includes both labor-augmenting and labor-displacing effects. So while these professions will be most impacted by LLMs, it is not obvious from their analysis whether the impact will be large-scale labor substitution by AI or not. The authors also find that the impact of generative AI will be pervasive and ultimately conclude that LLMs like ChatGPT exhibit properties of general purpose technologies.
To be clear, there is no clear litmus test for general purpose technologies and neither research study proves that AI is a general purpose technology. One can nitpick and question some of their interpretations. But the weight of evidence from these and other studies is clearly pointing in one direction — AI is a general-purpose technology.
In short, whether you are a creator, manager, or policymaker, or simply a curious person trying to figure out what is the role of humans – you, me, and especially our children – in this new era of generative AI, you better take note of the next general purpose technology in front of us.
And like all general purpose technologies, this one is evolving rapidly. In recent years, the scope of AI was restricted to highly repetitive tasks for which there exist tons of data and which generally seemed like the so-called left-brained tasks (i.e. tasks that involved data analysis or logical reasoning). But what happens now that AI is able to do tasks that we thought were reserved for human ingenuity – writing poems about a Camellia that in its full bloom inspires the observer or composing music like the new Drake song (which I personally prefer over every original Drake song I have heard) or writing short stories or coding software?
When creativity is no longer the exclusive domain of human brains, we have to change our underlying frameworks for how we learn, create, build innovative products and companies, and regulate them. This series of close to a dozen posts on Generative AI will be focused on exactly those themes. In my next post, I will clarify some terminology and fundamentals before diving into the fun stuff of how I believe generative AI will impact us. Please leave your comments as that will help me (and your fellow readers) immensely. I hope this Substack can be a safe space where we educate one another about “AI for Business.”
About me: Since this is my first post and some of you found me on LinkedIn, let me introduce myself. I'm a professor at the Wharton School where my work is focused on technology and digital media, and in particular, on business and social implications of AI. I am the author of A Human’s Guide to Machine Intelligence (what! you haven’t read it yet? you can still wash away your sins here). I was a cofounder of Yodle, a company that used AI to help over 50,000 small businesses advertise online. Recently, I founded Jumpcut Media, with the vision that AI will have a huge impact on storytelling in general and Hollywood in particular. I will draw upon many of these experiences in upcoming posts.
[1] Machine learning is a branch of AI. Almost all of the AI that is having significant impact today is based on machine learning. More on this distinction in my next post.
[2] Robotics itself heavily relies on machine learning. The authors did not merge the two on the grounds that robotics also depends on many other disciplines such as design of electromechanical systems.
Excellent introduction, Kartik. I look forward to reading the next from you.
An aside on "life before and after electricity." I was born two months before the Republic of India. Until the republic was close to 16, the only electrical gadget I used was a torchlight!
Great insights Karthik! Looking forward to the upcoming series.