AI Pirates

Touched by His Noodly Appendage, parody of Michelangelo’s Creation of Adam, via Wikipedia.org

Reactions to the above image range from religious ecstasy to abject horror. I find it amusing and interesting, so I guess I’m somewhere in the middle.

The Flying Spaghetti Monster was, according to the consensus of scholarly opinion, a satire on creationism, created by Bobby Henderson, a physics graduate at Oregon State University. It was first published as an open letter to the Kansas School Board in 2006, requesting that the theory that a Flying Spaghetti Monster created the universe be taught alongside other versions of Intelligent Design.

Among other things, Henderson explains that the Flying Spaghetti Monster adjusts the results of any scientific test to make it appear that the universe is billions of years old, and also claims that global warming is a direct result of the shrinking number of Pirates since the 1800s.

That’s where the pirates come in, but what about the AI part?

Well, people are starting to try and figure out how to deal with AI in the “real world”. It sometimes seems like we’ve been dealing with it for decades, but it’s actually been less than four years since ChatGPT first came out, in November 2022, and we’ve been spending most of our time since then being shocked by the rate at which the AI space has been changing and growing.

I’ve discussed AI a number of times, including about how it is vulnerable to cognitive bias, whether or not the current AI environment represents a bubble, and whether AI is really “intelligent”, but right now I’m thinking about AI from a more practical perspective.

One of the considerations around how to manage AI models in day-to-day work is token usage. Tokens are a unit of measure used by Large Language Models (LLMs) to reflect the amount of data being processed, and are often used in billing of services. Interesting to note that one of the reasons AI is not generally profitable at present is that most companies are providing free services – most likely as a “loss leader” for future services, or on the assumption that token generation will become more efficient over time.

Recently, an open-source project called “caveman” appeared, claiming to save 75% on LLM tokens by making AI talk like a “caveman”. The claim is that dropping “filler” words, skipping pleasantries and such would dramatically reduce the number of tokens needed to interact with LLMs.

An article reviewing the project noted that the claim was both true and misleading, and the “actual” performance improvement was more modest. Without going into detail, the main questions are around the nature and scope of the instructions passed to the LLM. While the “caveman” idea is somewhat amusing, the reviewer managed to improve on the performance with a more concise set of instructions.

Another consideration I have heard regarding interaction with LLMs is around how to avoid falling into the trap of believing that LLMs are actually “intelligent” - in the “human” sense of the term - because of the degree to which they can mimic human responses. The danger here is around assuming understanding and trusting responses to a greater degree than is warranted, and not falling into what I just saw referred to as “The Claude Delusion”, as Richard Dawkins apparently did recently.

Unfortunately, many working with LLMs are describing them in anthropomorphic terms already. For example, the OpenClaw agentic framework defines a “soul.md” file, which is used to define preferences around the personality traits exhibited by a given bot.

Getting a tool that is designed to mimic human interaction to not do so seems... challenging, but I’ve heard several interesting ideas. One was on Security Now Episode 1084, in which Steve Gibson described the way in which a listener instructed Claude Code to “adopt a voice and interaction style inspired by Agent Smith from ‘The Matrix’”, then provided an example of the resulting interaction.

The goal here is not to most efficiently manage token usage (though that may potentially be a factor), but rather to reduce the degree to which we are liable to anthropomorphize our LLM. Selecting a fictional character makes it far less likely that we’ll “forget” that we’re interacting with an AI, while also providing some degree of entertainment.

Aside from the character, it’s useful to establish guidelines for interaction with an LLM, in order to maximize accuracy, reduce sycophancy, and address whatever other goals are most important for a given interaction. For example, including instructions to be concise may reduce unnecessary reading, while instructions to cite sources or flag uncertainty might help to reduce the incidence of confabulation (ie, “AI hallucinations”).

Ultimately, it depends on your goals, and the degree to which you want to avoid having the LLM “seem” like a real person.

Interestingly, depending on what you are trying to do, you may run up against the “guardrails” established for a given LLM. As an example, using the “abuse” person from Monty Python’s famous argument sketch might be wildly entertaining, but LLMs have frequently been criticized for responses which might be inconsistent with supporting the mental health of certain people. Thus, personal insults will generally be redirected to generic ones which focus on systems, or conclusions, or the human race as a whole.

Still, wouldn’t it be cool to discuss quantum encryption, or AI research, or philosophy with an AI that’s also a pirate?

Cheers!