Could Artificial Intelligence have Invented Jazz Music?

References:

Stuff that’s good to know.

  • 🧠 LLM’s and Transformers

LLM stands for Large Language Model. It’s a very specific type of AI, a Neural Network using a Transformer design (like a brain). An LLM predicts the next word in a sentence – billions of adjustable dials allow the model to map meaning between words.

  • 🤖 AGI and ASI

AGI is Artificial General Intelligence (human level general intelligence, ASI is Artificial Super Intelligence (beyond human level intelligent). Remember, binary definitions are misleading. Better to imagine a spectrum – narrow at one end, broad or more general at the other. Humans exist toward the general end because we can apply our intelligence across multiple domains. Current AI is toward the narrow end – good at summarizing existing scientific theories; not so good at formulating new ideas or designing novel experiments to test them.

  • ⛔️ LLM’s and Transformers will Lead to AGI (debate)

There are currently two camps. Camp one – believes that LLM’s on the current trajectory will lead to AGI simply through scaling up, increasing the compute and training. Camp two – believes there are significant limitations in the current design of LLM’s which won’t allow us to scale to AGI and ASI on the current path; we’ll need a significant shift. There are very smart people on either side of the debate, which leads me to believe we’ll probably land somewhere in the middle – ie, current LLM’s with some added functionality will lead us to AGI and ASI.

  • 🌎 Flexibly manipulable abstraction

This is just a fancy way of saying that human-like intelligence is flexible and comes from a higher level of abstraction – the unique interactions we have with the world around us.

Could Artificial Intelligence have Invented Jazz Music?

Both Ben and Joscha agree some form of General and Super human intelligence is coming soon.

Joshcha is reluctant to take a definitive stance against Large Language Models (LLMs) in the current form, as a potential path to Artificial General Intelligence (AGI).

Ben takes a relatively harder position – that Language Models in their current form are unlikely to take us to AGI without some drastic modification.

Ben puts forth then following thought experiment.

If an LLM was trained on all training data up to the 1900’s, could it have invented Jazz music?

The answer is, ‘probably’ – no.

And it’s worth considering why.

Jazz wasn’t just the evolution of two genres of pre-existing music.

It was the fusion of African Rhythms, call-and-response patterns, Blues from the American South, Ragtime, European Classical and harmony, Spiritual hymns from African-American Churches, Brass Band, and much more.

And more importantly, it was more than just music.

It was cultural innovation born out of hardship, slavery, oppression, darkness, resilience and ingenuity. The very best and the very worst of the human-ness.

Ben is basically arguing that there is a fundamental difference between how we human monkeys think about ourselves and the world, and how current AI systems process information.

We have a form of flexible self-awareness of ourselves, and others around us developed through our interactions with the world.

So while current LLM’s will be able to abstract away many of the current things humans do in the world – it won’t be able to reproduce truly novel and creative cultural breakthroughs, like Jazz music.

Yet.