#ai

From the creators of DALL-E and GPT. The examples of what Whispir can do are pretty astounding.

A great technical demo (and paper) from Meta. Using only the sensors in the Quest headset (and reinforcement learning) they can recreate the users pose.

A great profile of Greg Rutkowski, an artist who is a more popular prompt than Picasso for AI-generated art (tools like DALL-E allow users to request an image in the style of an adequately prevalent artist). Does this encourage future artists to produce less art in order to avoid AI models copying them? In the future, will these tools offer a way for artists to omit their work (like no-follow in robots.txt for search engines)? While I don’t think these tools will kill art entirely, I’m sure they will harm commercial artists.

Action Transformer by Adept is an AI model that allows you to navigate and use websites and web apps using text commands. I had assumed that these types of features would come to voice assistants (like Alexa or Siri) via voice-first APIs, but this already looks much more capable than any voice assistant. So, maybe this is the technology that will make the capabilities of voice assistants more universal.

Some digital esoterica for the AI generation.

This is an astounding demo of a playable Pokemon game emulation powered by a neural network. This is obviously pretty terrible quality, but it’s still surprising just how good it is already. Imagine this in 10 years. The commentary provides a fantastic read.

A very interesting step-by-step walkthrough of how AI-generated graphics works.

The list of ideas for AI-first products at the bottom of this grant is particularly interesting. For example, how much of the work on UpWork can be automated?

I don’t think animals are ever going to want to talk to us, but it will be very interesting if ML upends the notion that animals don’t do a lot of talking to each other.

Research is locked up in poorly formatted, inconsistently structured PDFs. This makes training models much more difficult than training text or image models. Interesting to see how this is currently being tackled.