[zephen]

> If calculators returned even 99.9% correct answers, it would be impossible to reliably build even small buildings with them.

I think past successes have led to a category error in the thinking of a lot of people.

For example, the internet, and many constituent parts of the internet, are built on a base of fallible hardware.

But mitigated hardware errors, whether equipment failures, alpha particles, or other, are uncorrelated.

If you had three uncorrelated calculators that each worked 99.99% of the time, and you used them to check each other, you'd be fine.

But three seemingly uncorrelated LLMs? No fucking way.

[noduerme]

There's another category error compounding this issue: People think that because past revolutions in technology eventually led to higher living standards after periods of disruption, this one will too. I think this one is the exception for the reasons enumerated by the parent's blog post.

[zephen]

Agreed.

In point of fact, most technological revolutions have fairly immediately benefited a significant number of people in addition to those in the top 1% -- either by increasing demand for labor, or reducing the price of goods, or both.

The promise of LLMs is that they benefit people in the top 1% (investors and highly paid specialists) by reducing the demand for labor to produce the same stuff that was already being produced. There is an incidental initial increase in (or perhaps just reallocation of) labor to build out infrastructure, but that is possibly quite short-lived, and simultaneously drives a huge increase in the cost of electricity, buildings, and computer-related goods.

But the benefits of new technologies are never spread evenly.

When the technology of travel made remote destinations more accessible, it created tourist traps. Some well placed individuals and companies do well out of this, but typically, most people living near tourist traps suffer from the crowds and increased prices.

When power plants are built, neighbors suffer noise and pollution, but other people can turn their lights on.

We haven't yet begun to be able to calculate all the negative externalities of LLMs.

I would not be surpised if the best negative externality comparisons were to the work of Thomas Midgley, who gifted the world both leaded gasoline and CFC refrigerants.

[firejake308]

The LLMs are not uncorrelated, though, they're all trained on the same dataset (the Internet) and subject to most of the same biases

[zephen]

Agreed.

This is why I differentiated "uncorrelated" from "seemingly uncorrelated." Sorry if that wasn't clear.