Erin Croxton

Information Management for Government

FA2026 MPA 605-QL

Blog Assignment 1

I did not expect to run down a rabbit hole today, but boy did David Auerbach send me jogging. Auerbach's 2012 article, "The Stupidity of Computers", traces the development, sometimes failures and, dare I say weaponization, of language learning models (LLM) from the inception of memorable search engines like Ask Jeeves to Twitter and OkCupid.

Auerbach's analysis of data management by corporations, governments, and schools initiates important questions and concerns about how the LLMs of 2026 compare to those described up to 2012.

Summary

Auerbach (2012) explains that early LLMs learned in the same manner that one would do cool tricks in an excel spreadsheet– using formulas that create meaning and effectively telling the model exactly what to interpret every single word to mean or do. The problem was that this method, essentially, allowed for none of the nuance that is carried within human interaction and language.

The SHRDLU model, though more advanced than the ELIZA model, could never understand the complexities of sarcasm.

Imagine trying to explain SHRDLU asking if the blue block was the correct one and the responder said "SUre."

The article continues to trace Amazon, Facebook, and Twitter's capitalization on categorization– Amazon's use of already well established categories to market your goods to the public, at a percentage of the cost of course, and Facebook and Twitter's clever idea to make it seem like publicizing your own data was your idea (Auerbach 2012).

Further, Auerbach (2012) says the quiet part out loud. The government, your employer, and big business are watching you!!! And they're doing a bad job, too. The data that is constantly being mined by these entities is being used to keep an eye on you, to determine whether or not you might be a liability to employ, and to market to you– sometimes all at the same time (looking at you LinkedIn). Never mind the fact that as, human beings, it is almost impossible to pigeon-hole an individual's personality based on a few tidbits of data points gathered via internet browsing history.

Implications

As you may be able to tell, I'm curious about all of it, but particularly language learning models's journey from formulaic coding to their ability to infer meaning from language, itself.

Somewhere in the article Auerbach asserts, correctly, that computers are not as intelligent as we give them credit for. It takes the human touch to program the language that LLMs devour. It takes a person to categorize web products (and people) ontologically.

Auerbach leaves the reader, in 2012, with the hope that it might take much longer to reach the point in which the human touch is irrelevant. But, that was 2012, this is 2026... how are the LLMs learning, now?

According to Jurafsky et al. (2026), language learning models– specifically programs adept at determining the probability of word sequences– have become the preeminent artificial intelligence of our time.

The '60s had ELIZA, programmed to be a pseudo-therapist, with her formulaic knowledge, by her formulaic code (Auerbach, 2012); however, LLMs of today no longer require any such plug and play codes. Instead language is broken into "tokens", which are processed in the matter of seconds, managing to also run probability tests that help to determine if the next word to follow in their response is the correct one (Jurafsky et al., 2026). It's actually rather interesting once you think about it because now the human touch is almost imperceptible and a lot of people barley notice.

Now, it looks like we're the silly ones.