The LLM Bubble is bursting

Some observers, including analysts for Goldman Sachs, say that the growth of AI since the height of the pandemic is now proving to be a bubble, like dotcom stocks in the late Clinton years or mortgage based derivatives in 2006-07. 

What is a bubble? By standard definition it is a growth in value of an asset that occurs after the speculative demand has surpassed the operational demand for same. If I buy X as a speculator at a price higher than the operational value of X can support I may be willing to make an operationally foolish deal in the expectation (on the speculation) that there will be a "greater fool" to whom I can then sell it.  

In such a situation, sooner or later the biggest fools ARE the ones who are holding X. What happens to its value then you can fill in for yourself. 

Clearly there is money still entering the field.  The Peoples Republic of China is incubating a group of companies that it proudly calls the "little AI dragons," with seed money of more than US$1 billion each.  

One variant of the AI-is-a-bubble theory is this: LLM-is-the-bubble.  The initials "LLM" refer to "large language models," that is, algorithms that consume huge blocks of internet-based text, organize it pertinent to your requests and spit it back at you in coherent English.  

Question: "What is the plot of [novel x]."  LLMs generally stay away from consuming novels directly, but they do take in reviews or scholarly articles about such novels, and from them they can spin you a plot summary suitable for a lazy high school student to turn in as a book report. This is the sort of trick that has made AI, of which LLM is a subset, so valuable. [For the student anyway.]

From the point of view of the plagiarist high school student, or other users if we find that one unsympathetic, the problem is that LLM can produce howlers (called "hallucinations" in the literature on the subject) that make it all too easy for the human teacher to catch on and flunk him.  Or, more generally, hallucinations are results that are coherent and factually correct, yet factually wrong. The LLM in our plot summary example could, let us say, mix up the names of major characters. "The Don, Michael, throws a lavish wedding for his daughter.  His soldier son, Vito, arrives with a girlfriend." 

Anyway, the LLM-is-the-bubble theory says that the limits of this technology are becoming evident. That in turn is making it clear that the operational value of the technology is a good deal lower than had been thought. If the value of stocks such as Nvidia continue at anything like current levels this will be the consequence of speculative demand, not the operational sort. Definitionally, a bubble. 

Nonetheless, the theorists continue, LLM is not all there is to AI.  As the companies involve shift away from reliance on LLM to other applications of the underlying technology, (and the new Chinese concerns may avoid the over-emphasis on LLM in the first place) prices will find their new equilibrium and perhaps begin to rise again.  So, dread not.