MISC 23: Interacting with AI 3

Its been a year or so and I started interacting with the so-called ‘large language model’ AIs.

Search the archive for AI and see the previous two articles which relate to impact of it on Music.

Its evolving fast and some of the most powerful vested interests are throwing the book at getting it to the next level so here is my current assessment of the pros & cons of their abilities, for all those out there who have not tried it but who are concerned for their unsuspecting dependents at home and work.

To summarise:

• Pros: they perform what looks like magic when asked to generate code and summaries of large texts or the wisdom of the masses. If you are not familiar then they are free to try. Just go to OpenAI and ask ChatGpt4 to collate and summarise all the data on your favourite topic

• Cons:

  • they make mistakes and make up facts and sources, so people who are not experts cannot tell the difference, because they are predicting the next word based on consulting the wisdom of the masses (all the published material on the net) plus training by clever humans and are built from clever coding that simulates how we humans do it. The wisdom of the masses is not trusted by experts but it might as well be facts to the non-experts.

  • more frustratingly, if you are an expert and begrudgingly accept the errors (based on the fact that it is still a great assistant for data processing), when you point out the errors the bots can appear to be convinced and can even pass tests to check they have understood the mistake, like a helpful and astute human assistant in person…but if you log in at a later date and start a new conversation, the bot will repeat the error because it starts again from the original central dataset it is trained on, like having a slacker as an assistant who nods along but is not interested.

    • see SCALES 32 for an example of this. Curiously, after I had published the article, the next time I asked it the bot got it because it found my article afresh and could simply reconstitute it to satisfy me that it understood, but the mistake is to think that it has understood because by changing just one parameter of that puzzle, it was back to making obvious errors and not grasping the problem.

  • they can be used for serious skullduggery

With this in mind, I found this anecdote and speculation by Rory Sutherland in his chat with Rick Rubin on the podcast Tetragrammaton really helpful way of thinking about them.

In the latter stages of the podcast chat he described how his assistant had used AI to do a CV of Rory and asked him to check it over. He was both fascinated and worried to discover that the CV was completely wrong but in the broadest sense, ‘plausible’ - as if the AI had correct identified his ‘character class’ (an advertising executive) and then produced a generic and believable version of him to satisfy the request. Rory described this as the work of a ‘plausibility engine’.

I decided to ask the bot what it though of this description:

• Would it be correct to describe you as a Plausibility Engine to help a newcomer understand how you work?

  ChatGPT said: Describing me as a "Plausibility Engine" is actually a pretty good starting point—especially for a newcomer. It captures an essential part of how I generate responses. Here's what that means and where it helps (and falls short):

  • Statistical prediction: I generate text by predicting what is most likely to come next in a sentence, based on a huge amount of training data. That means I prioritize what's plausible and contextually appropriate.

  • Fluent, coherent responses: I aim for natural, reasonable-sounding language—what a knowledgeable human might say.

  • Useful metaphor: For someone just getting started, "plausibility engine" sets expectations: I’m not retrieving facts from a database—I’m producing what sounds right based on patterns I’ve seen.

  • Plausibility ≠ Truth: I don’t inherently “know” what’s true—I can produce convincing but incorrect information (“hallucinations”), especially when the question is vague, niche, or ambiguous.

  • Lack of reasoning or beliefs: I don’t “believe” things or “think” like a person—I don’t weigh evidence in the traditional sense. My reasoning is learned pattern completion, not conscious logic.

  • No awareness or understanding: I don’t “understand” things the way humans do. I simulate understanding via language.

  • Calling me a Plausibility Engine is a useful and mostly accurate metaphor for beginners. Just add the caveat: "It sounds convincing, but it's not always right—verify important info."

Rory went on to suggest that the an ‘inquisitive engine’ would be more useful - one which consulted the data and made speculative suggestions, like an imaginative human. Imagine that.

Now how about sensory detectors coupled with some next-level ‘emotional circuitry’ so the bots can actually design things that affect humans. Back to SCALES 32, I realised during this interaction that the bot could eventually be guided to solving the problem of making a scale that only resolved over two octaves but could not understand the word ‘resolve’ in the way we feel it.

Another way to think of the current models is to treat them like really good mimics or stage magicians, which appear like magic to the unwitting but are just really difficult skills to acquire by humans. This said, I think its fair to take our so hats off to the designers because it was not long ago that a pocket calculator had the same reaction.