Generative AI Hype Peaking
I have been a skeptic of some of the more breathless claims made for Generative AI since at least October 2023, when I participated in a panel organized by Syracuse University. Specifically, I have been skeptical of claims that these technologies will dramatically increase labor productivity in most or all occupations. This is generally what is meant by those who see Generative AI as a 'first step' toward 'artificial general intelligence' and is implicit in claims that AI needs to be regulated to avoid mass labor market disruption.
Time for a New Narrative
I believe we are very close to the peak of investor hype around Generative AI. At time of writing, NVIDIA's stock price is down around 20% for the year to date.
If I could time further falls, I'd be trading on that knowledge instead of writing about it here. That said, I would be surprised if executives are giving "AI strategy" high priority in Q2 of 2026. A new narrative will be required to drive investment in tech.
The Gartner Hype Cycle is a perennially relevant framework for understanding these narrative shifts. If we are entering the "trough of disillusionment", now would be a good time to ask what the "plateau of productivity" might look like.
Real Innovations
LLMs have made possible real process innovations. Software development and customer support services, for example, can be delivered somewhat more efficiently than they were before. Retrieval through LLM document embeddings, though no panacea, represents a genuine advance in search.
In customer support, the early and often routine steps of a call tree can now be handled by chatbots in many instances, with only exceptional cases escalated to human agents. The chatbot is helping some fraction of customers help themselves by walking them through support documentation step by step. Some fraction of end users will find this frustrating, but if they aren't particularly wealthy, the new equilibrium could very well be a degraded customer experience. Those with the means will be able to purchase support services that let them skip the bots and get directly to human assistance.
In my occupation of software development, querying ChatGPT and DeepSeek has largely replaced searching sites like StackOverflow. These chatbots generally save time with prompts like "write a TypeScript type declaration for objects that look like this", "convert this function from Python to Javascript", "provide me with a list of the scientific names of bird species mentioned directly or indirectly in this essay".
LLMs and the Developer Pipeline
Anecdotally, LLMs seem to be substituting for less experienced developers across the industry as a whole. While the job market for more experienced developers (5+ years) has slackened modestly compared to late-2019 due to higher interest rates, the job market for less experienced developers seems positively dire.
We may look back in a decade and lament how self-serving and short-sighted employers stopped hiring less experienced workers, denied them the opportunity to learn by doing, and thereby limited the future supply of experience developers. I imagine the headaches for computer science educators are incredible, as LLMs have been trained on every problem set ever posted online.
Jevons Paradox
The release of DeepSeek just a few weeks ago threw into question how many GPUs would really be needed, in aggregate, to generate all the tokens for which there was effective demand.
Those trying to rescue the investor narrative are appealing to Jevons' Paradox. Jevons observed that rising efficiency in the use of coal coincided with rising coal consumption; AI boosters argue that increasing LLM efficiency can just as well occur alongside rising LLM usage. Brynjolfsson has highlighted software development as an occupation that may see increased demand for workers alongside rising (LLM-driven) labor productivity. But it seems to me that for customer support, for example, demand for low to medium quality chatbot support is fairly inelastic. One hopes the same is true for the "AI Slop" currently inundating social media feeds.
A Lamentable Killer App
The Dead Internet Theory holds that there are virtually no real human beings posting on social media anymore, having all been replaced by bots. This is a wild overstatement. But political influence and scam bots (handling the early intake phase of the scam) may turn out to be the sustainable killer application for this technology.
UPDATE (March 25, 2025): Bryan McMahon, writing in the American Prospect, surveys the the scale of bets on AI so far, and argues that a collapse in investor confidence could have impacts far beyond tech.