The last ten years has seen multiple waves of artificial intelligence (AI) hype. From IBM Watson, to consulting firm projections of $15 trillion in economic gains by 2030, to Geoffrey Hinton’s claim that radiologists would become gone by 2021, and to Elon Musk’s claim that self-driving cars would be here by 2020 and the criminal justice system, home flipping, and the used car businesses would soon be disrupted by AI. Needless to say, these waves did not become tsunamis, and instead are more like the small waves from skipping flat stones.
The latest wave is ChatGPT, a chatbot launched by OpenAI in November 2022 that is built on top of OpenAI’s GPT-3. GPT-3 is a family of large language models that is fine-tuned with both supervised and reinforcement learning techniques.
The short-term issues are whether ChatGPT can generate the necessary hype to spur a big valuation for OpenAI, and whether generative AI in general can do that for other startups. Currently only two publicly traded startups, SoundHound and c3.AI, can be truly defined as AI companies. Neither are profitable nor have large market capitalizations (about $2 billion). Looking more broadly, two data analytic startups, Snowflake and Crowdstrike, have market capitalizations of $47 billion and $25 billion respectively. Can the latest generation of AI startups do better than this?
The long-term issue is whether AI startups can become profitable and attain large market capitalizations that are sustainable. No AI Unicorn is even close to being profitable and only about 10% of publicly traded Unicorn startups are profitable. Can generative AI startups do what no AI startup has done before, or what so few Unicorn startups have done in the last 10 years?
Some investors are salivating at the prospect that ChatGPT can replace Google’s search engine. After all, wouldn’t you rather read clearly written prose than click through dozens of links to find out what you want to know? What people forget, however, is that Google’s search engine works best for advertisers, not for us, because they pay the bills. If ChatGPT is used for search, the same thing will have to be done, thus making ChatGPT far less useful to us than we currently think
A second application is customer-service chatbots: can ChatGPT do better than the current generation of chatbots? A January 24th Wall Street Journal article says: “Many executives said they are proceeding with caution, given the limitations of ChatGPT—fine-tuned from GPT-3.5, a model created by startup OpenAI—as well as OpenAI’s older AI language system, GPT-3, which companies are already starting to integrate into digital products.”
A third application is general articles. However, CNET’s error-ridden articles, the banning by StackOverflow, and the poor performance of ChatGPT on MBA and legal exams should be cause for concern. Even Yann Lecun, recipient of the 2018 Turing Award, says this about ChatGPT and other generative AI: “they are useful but they make stuff up.”
Optimists will say it is the early years, despite the use of chatbots by customer service departments for at least five years. The question of how much better ChatGPT will get in the future is complicated, and it gets even more complicated if we ask how fast. Huge increases in computing resources have been needed to increase the accuracy of image recognition system by another nine, from 99% to 99.9% and 99.99%. It is likely that the same types of increased computing resources have been needed to make ChatGPT better, and that further increases will continue to become even more difficult.
More fundamentally, to really understand how valuable ChatGPT might become to investors and society, we should be asking if it or generative AI can become a productivity enhancing tool that is used on the scale of Microsoft Word, Power Point, or Excel, or Oracle’s database technology.
Might ChatGPT increase the productivity of workers, and if so where? Will it enable us to manufacture cars, construct houses, provide cell phone or Internet services, or educate kids with fewer hours of labor? Or might it enable us to produce things using less energy, and thus prevent climate change? Or might it enable us to use less land to make things or provide services, and thus open up more land for parks?
The problem is that written documents usually don’t provide value by themselves; they are largely a tool to help us produce things, construct homes, or provide services, among other things. Although there is a media industry that includes newspapers, magazines, books, and other written documents, those industries are rather small, and typically require the best writing, far beyond what ChatGPT, or even most humans can produce.
ChatGPT can be used to write legal briefs, but they may make bogus arguments based on made-up precedents. More fundamentally, productivity will not be enhanced by more legal briefs but by faster resolution of legal disputes, or even better, fewer disputes. Resolving disputes faster and cheaper requires the human judgement that ChatGPT lacks.
What about software coding? ChatGPT does generate code and it can often do this as well as it generates text. Which means, the code looks good, but sometimes it contains errors, errors that often require extensive debugging. The question is whether ChatGPT can produce code in the future that requires less debugging than code originally produced by humans. Otherwise, the productivity advantage of ChatGPT will be small, or even negative.
What about engineers, accountants, and architects? Can ChatGPT make them more productive? Many people forget that more than five years ago there were those who thought generative AI would make professionals such as architects and engineers more productive, including myself (see my 2017 slides). This AI was supposed to help architects design buildings better and faster, and engineers design a broader array of products better and faster. For instance, AI-based systems were expected to become better at design as architects provided feedback on designs generated by AI system. The AI system was programmed to understand the logic of design, aesthetics, and location’s topography.
Similarly, AI systems were purportedly able to propose new designs for mechanical and electrical products and thus raise the productivity of engineers. For instance, Airbus purportedly used AI to design a New Structural Partition on the Airbus in which the AI system minimized weight while providing sufficient strength. I believed in 2017 that AI was the future of engineering, architecture, journalism, accounting, and the legal profession.
Much less is said of such systems now, presumably because they are much harder to implement or perhaps their benefits have been much smaller than was thought five to ten years ago. This is an important lesson for today’s conversation about generating text or images from AI. After all, the benefits of generative AI only emerge if it helps knowledge workers, such as engineers and architects, become more productive at their jobs. And the generative AI of more than five years ago appears not to have done this, and it has been largely forgotten.
How valuable will ChatGPT become? Will it become what the AI optimists have been claiming each wave of AI – from IBM Watson to self-driving vehicles – would become? Or will it be another wave of hype that is quickly forgotten before a new wave of hype is proposed? Start asking how productive it will enable workers to become and which workers? The more productive it can make us, the more we will pay for it. And start asking how easy is to improve the accuracy of ChatGPT, because it will only be a great tool if its accuracy has many nines in it, say 99.99% accurate.
Author: Dr Jeffrey Funk
#AI #ChtGPT #Innovation #OpenAI
