D The current and future state of AI/ML is shockingly demoralizing with little hope of redemption
I recently encountered the PaLM (Scaling Language Modeling with Pathways) paper from Google Research and it opened up a can of worms of ideas I’ve felt I’ve intuitively had for a while, but have been unable to express – and I know I can’t be the only one. Sometimes I wonder what the original pioneers of AI – Turing, Neumann, McCarthy, etc. – would think if they could see the state of AI that we’ve gotten ourselves into. 67 authors, 83 pages, 540B parameters in a model, the internals of which no one can say they comprehend with a straight face, 6144 TPUs in a commercial lab that no one has access to, on a rig that no one can afford, trained on a volume of data that a human couldn’t process in a lifetime, 1 page on ethics with the same ideas that have been rehashed over and over elsewhere with no attempt at a solution – bias, racism, malicious use, etc. – for purposes that who asked for?
When I started my career as an AI/ML research engineer 2016, I was most interested in two types of tasks – 1.) those that most humans could do but that would universally be considered tedious and non-scalable. I’m talking image classification, sentiment analysis, even document summarization, etc. 2.) tasks that humans lack the capacity to perform as well as computers for various reasons – forecasting, risk analysis, game playing, and so forth. I still love my career, and I try to only work on projects in these areas, but it’s getting harder and harder.
This is because, somewhere along the way, it became popular and unquestionably acceptable to push AI into domains that were originally uniquely human, those areas that sit at the top of Maslows’s hierarchy of needs in terms of self-actualization – art, music, writing, singing, programming, and so forth. These areas of endeavor have negative logarithmic ability curves – the vast majority of people cannot do them well at all, about 10% can do them decently, and 1% or less can do them extraordinarily. The little discussed problem with AI-generation is that, without extreme deterrence, we will sacrifice human achievement at the top percentile in the name of lowering the bar for a larger volume of people, until the AI ability range is the norm. This is because relative to humans, AI is cheap, fast, and infinite, to the extent that investments in human achievement will be watered down at the societal, educational, and individual level with each passing year. And unlike AI gameplay which superseded humans decades ago, we won’t be able to just disqualify the machines and continue to play as if they didn’t exist.
Almost everywhere I go, even this forum, I encounter almost universal deference given to current SOTA AI generation systems like GPT-3, CODEX, DALL-E, etc., with almost no one extending their implications to its logical conclusion, which is long-term convergence to the mean, to mediocrity, in the fields they claim to address or even enhance. If you’re an artist or writer and you’re using DALL-E or GPT-3 to “enhance” your work, or if you’re a programmer saying, “GitHub Co-Pilot makes me a better programmer?”, then how could you possibly know? You’ve disrupted and bypassed your own creative process, which is thoughts -> (optionally words) -> actions -> feedback -> repeat, and instead seeded your canvas with ideas from a machine, the provenance of which you can’t understand, nor can the machine reliably explain. And the more you do this, the more you make your creative processes dependent on said machine, until you must question whether or not you could work at the same level without it.
When I was a college student, I often dabbled with weed, LSD, and mushrooms, and for a while, I thought the ideas I was having while under the influence were revolutionary and groundbreaking – that is until took it upon myself to actually start writing down those ideas and then reviewing them while sober, when I realized they weren’t that special at all. What I eventually determined is that, under the
