> AI is opaque... The amount of papers I’d have to read and mathematics that I’d have to ingest to really understand why a certain prompt X results in a certain

> AI is opaque... The amount of papers I’d have to read and mathematics that I’d have to ingest to really understand why a certain prompt X results in a certain output Y feels overwhelming. Modern deep learning is almost wholly built off of partial derivatives, the chain rule, and matrix multiplication. It looks complicated from the outside because there are so many people publishing variations on this same basic formula, but it is the same basic formula over and over and over again. I may be a "new fart" myself, but when I set out to become an AI researcher I was honestly kind of surprised how transparent the mechanisms underlying AI actually are! I had expected to need to spend years to learn what I needed to understand it, but in reality it took me about 6 months of dedicated study to get to a point where I felt more-or-less comfortable with how it all worked. Granted, learning how Transformers and attention worked specifically took me a little longer (not sure why, was just some kind of mental block) > AI is Not Approachable I would argue that AI is more approachable than it ever has been. A used RTX 3090 with 24GB of VRAM can get you shockingly far for ~800 bucks. Will you be training GPT-5 at home? No. But there is still so much fascinating fundamental research waiting to be done in this field. Personally, I really enjoy using my home PC to implement and train funky novel architecture designs and see if they work on common toy problems. My hope is that one day I'll stumble onto one that becomes a Transformer-killer and then I could get a compute grant or something to try scaling it, but that's likely a pipe dream. Still, I find the amount of cool stuff you can do with consumer hardware today astounding. > AI is Not Open At the risk of sounding like a broken record, I think AI is more open than it ever has been. Yes, the state of the art is closed-source right now, but Open Source AI has never been as active and exciting as it is right now. Mistral LLMs, CLIP, Whisper, TheBloke and his crazy quantizations, DINOv2, LLaMA-2, Stable Diffusion, and so much more have been released open in just the past year or two. Local LLMs are slowly (but surely) catching up to GPT-4, even if they are lagging behind by about 6-8 months. The best open LLMs these days are (anecdotally) above the level of ChatGPT-3.5 when it launched. That's exciting! Of course all the coolest, most sexy AI research is happening behind closed doors, but to see that and conclude that AI is "not open" feels shortsighted. In conclusion: What a time to be alive! I think it's perfectly valid to feel nervous about the future, because things, they are a-changing! And change is scary! But I really do believe that there is more to be excited about than there is to be worried about, and that everything will Work Out™ in the end.