The key players in the AI market – Anthropic, OpenAI, Google, Meta, and maybe one or two others – are planning to invest trillions of dollars in AI development. Thousands of huge data centers are planned. This is being done without knowing whether current AI approaches will work well enough to make these investments pay for themselves (or if they do work, whether the investments will be undercut by lower-cost competitors).

But regardless of whether the AI software ultimately works as promised or not, the economic downside is huge.

Read more: The Artificial Intelligence boom: one way or the other, it’s bound to end badly

If the trillions invested in AI don’t pay off, it will be disastrous for the economy, because stock-market investors (including those who invest only in index funds) have made huge bets on it. The potential stock-market losses would be massive.

But if it does work, AI can only produce a satisfactory return on those trillions of dollars of investments if it replaces millions of jobs. That, too, would be an economic disaster, and not just for those unemployed. The economy as a whole would be undermined by the loss of purchasing power from those lost paychecks.

In short, the size of the bet on AI has become so large that I can’t see a way that the outcome is positive.

The problem is LLMs. Those of you who have read my previous posts on AI know that I am skeptical about the current generation of AI software. It is based on a technology called “large language models” (LLMs) that is (to oversimplify a bit) fundamentally based on a statistical determination of the probable next word, based on the words that preceded it. LLMs have to be trained on all the available text, including the entire internet, which is one reason the vast data centers are required.

But since LLMs don’t “understand” the text, they are subject to “hallucinations”—coming up with text that sounds like it fits the context, but is completely made up. You can read the details of one of my adventures with hallucinations here.

Some on-line commentators seem confident that the problem with hallucinations can be overcome. I asked Claude for information about this possibility, and Claude’s response was not reassuring: “Zero hallucination is almost certainly not achievable with current LLM architectures.” This response was backed up by references to several recent papers in the field. There are some avenues to further reducing the frequency of hallucinations and to making the responses of AI chatbots less confident-sounding when the result of a query might be erroneous. But it won’t be possible to eliminate hallucinations totally, and therefore no AI response can be totally trusted.  

How about alternatives to LLMs? There are AI approaches that are alternatives to LLMs, and they don’t require vast data centers, nor are they prone to hallucinations. I talked about them a bit in this post. Unfortunately, there is currently very little interest in them outside of academic circles.

That will probably change once we get to the other side of the current AI bubble and AI has a chance to recover from the terrible damage that will have been done to its reputation. If those alternative approaches emerge and start producing promising results,  I may get more optimistic about the future of AI. But as long as commercial AI is based on LLMs, I remain a pessimist.