AI like it's 1999 or 1899


In an earlier note, I talked about the natural progression of technology which goes from scarcity to abundance over time and AI is no different. That was in the context of media on the internet and paying per minute with early cell phones. One of the original challenges with AI was context length and token limits. It's becoming less relevant as context sizes continue to increase. However, it's not like you can fit the entire Harry Potter series either. So you still need to cleverly manage your context size especially when using documents and chat history as part of that context. Historically, people have used abbreviations to compress messages either because of cost or medium constraints. That's what this post explores in the context of AI.

Compressing text like T9

Below is a conversation Microsoft Copilot (formerly Bing Chat) using some information from the article From 1890: The First Text Messages:


Summarize this text:

I’m trying something new today. Sometimes in my research I find an interesting old article that I wouldn’t normally post because it’s not from the Sunday Magazine section, or it’s from further than 100 years ago so I’ll never get to it. Instead of letting these go unused, I figure I’ll occasionally post them midweek during what would otherwise be slow weeks. Since this weekend I only have three articles to post, it seems like a good week to try it.

Telegraph operators on opposite sides of the country had some time to get to know each other when they weren’t busy sending other people’s messages. “Metaphorically they shake hands cordially twice a day — when they begin work and when they end it. And when business is dull they hold long conversations, with hundreds of miles — perhaps thousands — separating them, as two friends might do over a dinner table.”

What really caught my eye, though, is that the abbreviations they used seem a lot like the abbreviations used in today’s text messages.


Here is a possible summary of the text:

The author is trying a new feature of posting old articles from the Sunday Magazine section on midweek. The article is about how telegraph operators communicated with each other using abbreviations that are similar to today’s text messages. The author finds this interesting and wonders if there are other similarities between old and new forms of communication.


Now abbreviate it as if using a T9 keyboard for conciseness


That sounds like a fun challenge. Here is a possible abbreviation of the summary:

try new feat post old artcls midwk. artcl abt telgrph oprtrs chat w/ abbrvtns lk txt msgs. find intrstng n wndr othr smilrts btwn old n new comms.

In this conversation, I first had Copilot summarize the original text. It's not a very long piece of text but it works for this example. You could imagine though how a longer piece of text would benefit even more from this summarization. Copilot does a good job of summarizing the text. However, what if I wanted to summarize it even more?

In comes T9 (kind of). T9 is short for "Text on 9 keys" and it was predictive text technology that made it easier to compose SMS text messages using only the 9 keys on the phone keypad. Before T9, it was cumbersome and time consuming to compose full sentences so you might abbreviate them. For example "See you later" might be abbreviated as "c u l8r". You can learn more in the SMS language Wikipedia article. Not only would you abbreviate because it was tedious to compose messages one letter at a time, but also you were limited in the number of characters you could send in a single SMS. Somewhat related, in the early days of microblogging on Twitter, one of the ways you could compose and view tweets was through SMS. There was the cost aspects as well. You were charged per SMS so you wanted to send as much information in a single SMS as possible. Now SMS is only one example where this took place. Telegrams and telegraphs from yesteryear had similar constraints. Whether it was cost, size limits, or general UX challenges, people looked to abbreviations to draft their messages.

Now that we're caught up on this history, that's what you're seeing in this response from Copilot.

try new feat post old artcls midwk. artcl abt telgrph oprtrs chat w/ abbrvtns lk txt msgs. find intrstng n wndr othr smilrts btwn old n new comms.

It is an abbreviated form of the summarized message enabling you to use even less tokens as part of the prompts you submit to AI.

When looking at "word" counts, they are roughly the following:

Text Word Count
Original 58
Summarized 25
T9 18

We can see that the abbreviated version is ~3x smaller than the original message.


Given that models like GPT and others have been trained on the entire internet, it has some knowledge of these types of abbreviations to compose messages. Now, is it scalable and would you want to use this in your production applications, I can't say. Probably not because the abbreviations are lossy. While in general you could read and get the gist of the message using the abbreviations, there are things lost in translation. Now maybe the LLM might do a better job at interpreting the abbreviations. It'd be a fun experiment.

"n d mntm, il kp usng AI lk its 1999. c u l8r!"

Translation: "In the meantime, I'll keep using AI like it's 1999. See you later!" 🙂

Send me a message or webmention