Aug 14, 2023·edited Aug 15, 2023Liked by Gary Marcus
This is one of the very best essays I read by you, and I very much appreciate the careful analysis and pointing out the possibilities and the uncertainties.
I fully agree that things are greatly hyped up and in practice generative AI may turn out to result in a set of niche (but important) products rather than a societal overhaul.
The closer parallel may be with the dot-com crash. The internet was a real big deal, but people jumped on the bandwagon prematurely.
I think LLM has been a big step forward in handling very large unstructured data. The insights that lead to it may lead to further progress, if not with the same architecture. And even with LLM, there are applications that have not been there before, even if it falls way short of true intelligence.
And all the more so for drawing out the comment that the cost component derived from current and future litigation around lack of explainability, responsibility, injustice, negligence, copyright (and others yet to emerge), remains far from fully understood.
That's good angle of looking at the AI industry. It is far from being even profitable, not speaking about returning burned money.
By the way has anyone seen any analysis of the economics of each AI value chain layer? I am quite sure that some of them, like FM development, will remain unprofitable.
I find the content licenses, fines and damages rather glaring omissions on the cost side. This is likely a technology with no path to profitability at all.
There were 120+ lawsuits against generative AI companies in July. The “programmers will never go back” from the damage claims of 9 BUSD over breaches of 11 separate software licenses in GitHub CoPilot.
That’s before the publishers and the remaining authors have gone to court, and the FTC probe even got started. Add in the 197,000 pirated e-books in Bibliotik and 200 million copyrighted texts from C4 that went into ChatGPT, all under legally obligatory opt-out and licensing, and you’re looking at algorithm disgorgement and/or fines of up to 4% of turnover.
Same with privacy law. Non-compliance comes with threat of disgorgement and hefty fines, and it’s technically impossible for them to comply with GDPR right-to-be-forgotten.
Generative AI will exceed its potential when applied to existing Web2 foundational models with sound fundamentals. Example: Within tools on a application layer. As a stand alone product in a new category, beware. The big incumbenst rule this domain. There must be business and revenue models first.
Personally I am quite optimistic about the future of OpenAI, given their alliance with Microsoft. This revenue stream is unlikely to end while Microsoft is integrating OpenAI’s in their products and services.
AI is an innovation burst - good analogies would be the canal craze of the early 1800's (e.g., https://en.wikipedia.org/wiki/Indiana_Central_Canal) or the automobile companies of the early 1900's (from hundreds of companies to a handful in 20 years). And as noted below, the dot-com era.
Many startups in all these cases were searching for product/market fit that could prove profitable in the long run. The vast majority failed. The survivors (Erie Canal, Ford, amazon) were wildly successful.
This is the nature of such things: some will make money, most will fail.
I'm very disappointed with this article, I came here semi-expecting some form of rational economic-based analysis about AI and this was entirely one-sided with rose-tinted glasses. It completely ignores the fundamentals of basic economics when factor markets break down, or the sieve action you get with companies who eventually starve out their consumer base before then going belly-up. There's almost no mention of the cost side which inevitably goes up with any monopoly/oligopoly over time for dependencies.
The only thing these products are good for is for replacing humans and eliminating jobs; jobs which won't be replaced in new ventures anytime soon for the bottom and middle population demographics. If you can't earn bare subsistence unrest is not far off; its lessons from basic history coming to life.
This is one of the very best essays I read by you, and I very much appreciate the careful analysis and pointing out the possibilities and the uncertainties.
I fully agree that things are greatly hyped up and in practice generative AI may turn out to result in a set of niche (but important) products rather than a societal overhaul.
The closer parallel may be with the dot-com crash. The internet was a real big deal, but people jumped on the bandwagon prematurely.
I think LLM has been a big step forward in handling very large unstructured data. The insights that lead to it may lead to further progress, if not with the same architecture. And even with LLM, there are applications that have not been there before, even if it falls way short of true intelligence.
Great thought provoking piece.
And all the more so for drawing out the comment that the cost component derived from current and future litigation around lack of explainability, responsibility, injustice, negligence, copyright (and others yet to emerge), remains far from fully understood.
That's good angle of looking at the AI industry. It is far from being even profitable, not speaking about returning burned money.
By the way has anyone seen any analysis of the economics of each AI value chain layer? I am quite sure that some of them, like FM development, will remain unprofitable.
I find the content licenses, fines and damages rather glaring omissions on the cost side. This is likely a technology with no path to profitability at all.
There were 120+ lawsuits against generative AI companies in July. The “programmers will never go back” from the damage claims of 9 BUSD over breaches of 11 separate software licenses in GitHub CoPilot.
That’s before the publishers and the remaining authors have gone to court, and the FTC probe even got started. Add in the 197,000 pirated e-books in Bibliotik and 200 million copyrighted texts from C4 that went into ChatGPT, all under legally obligatory opt-out and licensing, and you’re looking at algorithm disgorgement and/or fines of up to 4% of turnover.
Same with privacy law. Non-compliance comes with threat of disgorgement and hefty fines, and it’s technically impossible for them to comply with GDPR right-to-be-forgotten.
This is a dead-end technology.
I believe there are factual errors. Google’s original page range algorithm was not powered by “AI” in any meaningful sense.
Generative AI will exceed its potential when applied to existing Web2 foundational models with sound fundamentals. Example: Within tools on a application layer. As a stand alone product in a new category, beware. The big incumbenst rule this domain. There must be business and revenue models first.
Thanks for this interesting essay.
Personally I am quite optimistic about the future of OpenAI, given their alliance with Microsoft. This revenue stream is unlikely to end while Microsoft is integrating OpenAI’s in their products and services.
P.s. I wrote an article about this topic before: https://marknuyens.substack.com/p/challenging-openai
thanks so much for this essay. very insightful!
AI is an innovation burst - good analogies would be the canal craze of the early 1800's (e.g., https://en.wikipedia.org/wiki/Indiana_Central_Canal) or the automobile companies of the early 1900's (from hundreds of companies to a handful in 20 years). And as noted below, the dot-com era.
Many startups in all these cases were searching for product/market fit that could prove profitable in the long run. The vast majority failed. The survivors (Erie Canal, Ford, amazon) were wildly successful.
This is the nature of such things: some will make money, most will fail.
I'm very disappointed with this article, I came here semi-expecting some form of rational economic-based analysis about AI and this was entirely one-sided with rose-tinted glasses. It completely ignores the fundamentals of basic economics when factor markets break down, or the sieve action you get with companies who eventually starve out their consumer base before then going belly-up. There's almost no mention of the cost side which inevitably goes up with any monopoly/oligopoly over time for dependencies.
The only thing these products are good for is for replacing humans and eliminating jobs; jobs which won't be replaced in new ventures anytime soon for the bottom and middle population demographics. If you can't earn bare subsistence unrest is not far off; its lessons from basic history coming to life.