|A Tale Of Two AIs - Ai-Da And DALL-E|
|Written by David Conrad|
|Sunday, 10 April 2022|
Our knowledge of the earliest humans comes from their art. Now a humanoid robot is able to create works of art using the same methods as human beings and AI can generate realistic images by understanding a description in natural language. Is this the best of AI and the worst of AI?
This time last year I was prompted to address the question What Is Status Of Art In Our Digital World? prompted in part by the news that the robot artist Ai-Da, the humanoid robot created by a team of programmers, roboticists, art experts and psychologists, was preparing for "her" first major exhibition at London's Design Museum. Now Ai-Da is about to make her debut in a solo exhibition at the 2022 Venice Biennale titled Leaping into the Metaverse which, according to a report in the UK Guardian newspaper:
will explore the interface between human experience and AI technology, from Alan Turing to the metaverse, and will draw on Dante’s concepts of purgatory and hell to explore the future of humanity in a world where AI technology continues to encroach on everyday human life.
Ai-Da's latest skill is to paint in the same way as artists have traditionally painted. With a paintbrush clamped in a bionic hand and camera eyes regarding the subject, the robot uses AI algorithms to "decide" what brush strokes to make. This, according to her creator, Aidan Meller is “mind-blowing” and “groundbreaking” stuff. The motivation for the project is however ethical rather than artistic and he told the Guardian the the question he wants to raise is not “can robots make art?”, but rather “now that robots can make art, do we humans really want them to?”
In my opinion Ai-Da is more an experiment in how easy it is to con people into beliving that this is a great step forward in AI. Ai-Da has a cute name, recalling Ada Lovelace, but isn't much more than a doll with some motors. The software that backs it seems to be fairly standard and the pitch is all about convincing the onlooker that this is something like an intelligent and creative android. The fact that it has an exhibit at the Venice Biennale is more a statement of how easy the art world is to fool than any real achievement.
I guess if you have no idea what the technology is all about you are going to be vulnerable to reading more into what you see than is actually there...
Ai-Da Robot with creator Aidan Meller
By contrast to Ai-Da, OpenAI's work really is ground breaking and it doesn't need to be dressed up with a motorized tailors dummy. It stands on its own two feet (not that it has or needs feet) and it creates amazing illustrations. This is the work that prompts us to ask the question "what is art" and "what is creativity".
OpenAI's approach to producing art is via text. DALL-E 2 is the second generation of an AI whose name is a portmanteau of the artist Salvador Dalí and WALL-E, the robot star of the 2008 Disney/Pixar fim.
DALL-E made its debut in 2021 and the peer reviewed paper introducing it, DALL·E: Zero-Shot Text-to-Image Generation from OpenAI was the top choice by Louis Bouchard of AI Review for the best AI research paper for 2021. The I Programmer report by Nikos Vagallis includes Bouchard's video about how OpenAI applies its GPT-3 transformer architecture and CLIP, a neural network which learns visual concepts from natural language supervision. DALL-E was trained using 250 million text-image pairs taken from the internet, mostly from Wikipedia and learned to generate new images based on text inputs.
This Open AI video introduces DALL-E 2 and its new capabilities such as in-painting - which is a facility for editing images by providing descriptions.
You can join a waitlist for access to DALL-E 2 and meanwhile there's a lot of fun to be had with demos provided on the Explore Section of its website.
Here's an astronaut, playing basketball with cats in space, as a childrens book illustration:
and teddy bears, shopping for groceries, in the style of ukiyo-e:
For the technical details, the OpenAI researcher's paper Hierarchical Text-Conditional Image Generation with CLIP Latents is available as a pdf.
or email your comment to: firstname.lastname@example.org
|Last Updated ( Sunday, 10 April 2022 )|