Aaron Stanton

Aaron is an author, founder & investor in AI & XR. His work is often covered by CNN, WSJ, NYT, Forbes, Wired, TechCrunch & more. His previous exit was to Apple.

[Aaron Stanton @AaronStanton]

@skirano I was curious if this could be a hallucination, so I used a guitar tuner to listen to the video. The first time you play the E string, my tuner says it was already correct. When GPT says correct, my tuner says it is now off. Not sure if my speakers would mess that up, maybe?

[Aaron Stanton @AaronStanton]

The original video with lyrics is 5 minutes of time lapse, so I chopped it up into shorter chunks. haha. Continuation of the 32-hours of editing pane 1 of Thoughts by Aaron.

[Aaron Stanton @AaronStanton]

It seem inevitable to me that any Generative AI tool will need this to be a primary focus of their roadmap if they don't want to be left behind once these rules come into play.

https://t.co/95Wi93Qap8

[Aaron Stanton @AaronStanton]

AI tools that "suggest" randomness - instead of upscaling what you original intended - weaken your claim to human authorship. I'm surprised that more people are not talking about the role of control as a key feature of their products.

[Aaron Stanton @AaronStanton]

In order for AI to have a useful role in creative production, it's going to come down to how much control of the output the human author has during creation.

[Aaron Stanton @AaronStanton]

4 hours per pane... I'm about 3x slower.

[Aaron Stanton @AaronStanton]

I timed each step in creating the first 8 panes of my AI-enabled webcomic, including drawing the base references and working to edit the panes. Took me 125 hours of focused work, or about 7. 23 hours per pane. As far as I can research, a traditional artist takes around 2.

[Aaron Stanton @AaronStanton]

It is so tempting to trust LLMs probably more than we should, at this point.

https://t.co/rdmVkS8jhN

[Aaron Stanton @AaronStanton]

A lot of people don't realize that LLMs trained on smaller data sets that don't repeat the correct information over and over actually have very poor accuracy rates. So for common questions and the internet, you'll probably get a pretty nuanced answer.

[Aaron Stanton @AaronStanton]

But for specific topics with less data, 50% accuracy on a data set is considered good on most LLMs, plus 10% to 20% hallucinations.

If I were buying a textbook and it advertised as 50% wrong, sprinkled with 10% straight up hard-to-spot lies... I wouldn't buy that textbook.

[Aaron Stanton @AaronStanton]

And... Inbox Zero is gone. haha.

[Aaron Stanton @AaronStanton]

I finally reached Inbox Zero today after... like... not being that for years. And I'm wondering if I achieved it by being really on top of things, or just finally convincing everyone that they shouldn't want to talk to me. haha.

[Aaron Stanton @AaronStanton]

An AI webcomic I've begun working on... #aiart #ai #aitools #digitalart

[Aaron Stanton @AaronStanton]

So, I' trying an experiment where I train GenerativeAI on my own art, and use that workflow to create an interesting, original, and ethical webcomic within a century of starting. haha. #ai #drawing #process #playingwithai #aiart #digitalart

[Aaron Stanton @AaronStanton]

Thought this posted earlier. :) I'm fascinated by the idea of combining Eleven Labs AI voice transforms - which mirror voice acting really well - and video-to-anime transforms. I think I may try making a 1-man, multi-actor short of some sort. #ai @elevenlabs #aiart