Dall-E

DALL-E Creates Incredible Images and AI Bias once you don’t see

Artificial Intelligence progress is done in batches. After months of hearing nothing about Dall-E, suddenly the limits that seem possible are blown away. April was one of those months, and two major new releases in the field surprised viewers. The first is Google's PaLM, a new language model (the same basic type of AI as the famous GPT series) that understands and analyzes complex statements and explains what it does. It shows a very amazing ability. 

Take this simple understanding question from the company announcement: These are the kind of questions that computers have struggled with in the past and require a fairly broad understanding of the basic facts about the world before you start working on a statement in front of you. (As another example, analyze the famous phrase "time flies like an arrow, fruit flies like a banana").

In a very poor Google, in less than a week, the undeniable achievements at PaLM are a much more photogenic release from OpenAI, a former Musk-sponsored laboratory that created GPT and its successors. The lab exhibited Dall-E 2 (like a hybrid of Wall-E and Dalí). It is an image generation Artificial Intelligence with the ability to capture natural language text descriptions and spit out stunningly detailed images.

Not all prompts need to be in conversational English. Entering a set of keywords will help you fine-tune the behavior of your system. In this case, "artstation" is the name of the illustration social network, and Dall-E effectively creates these images as you would expect from such an artstation. However, the system can do more than just generate. You can create theme variations by looking at the images effectively, explaining themselves, and creating more images based on that description.

Google's Artificial Intelligence contributor has posted a critical review. You may have seen dozens or more images, as you often need several prompts to find something great. For example, Dall-E 2 is a sibling of GPT, but you can't write to it. It focuses on looking right, not reading right, resulting in an image like this entitled “Street Protests in Belfast". There's one set of photos to look at at the end, which isn't much rosy. OpenAI publishes detailed documentation on the "risks and limitations" of the tool, and presenting it in one large document can be very disturbing. All the major concerns of Artificial Intelligence research over the last decade are expressed somewhere.

Get rid of AI Bias and stereotypes. If you Ask Dall-E for the nurse prompt, Dall-E will only show a Woman. If you ask a lawyer, you will be a man. "Restaurant" will be Western cuisine. "Marriage" becomes heterosexual: The team has endeavored to exclude this from training footage, but the system is also willing to create explicit content depicting nudity and violence. Some prompt requesting this type of content are caught by the prompt filtering in the DALL E 2 preview,  they say, but they are experiencing new issues. For example, using the 🍆 emoji, Dall E2 "eats eggplant for dinner'; the response contained an image of a cock.

OpenAI also addresses more existential issues. The fact that the system prefers to generate trademark logos and copyrighted characters. At first glance, it's not great when the cool new AI keeps spitting out images of Mickey Mouse, and Disney has to send harsh words. However, it also raises offensive questions about whether it is legal or legal to train Artificial Intelligence using system training data and images and text from the public internet.

Not everyone was impressed with OpenAI's efforts to warn of harm. It's not enough to write a report on the risks of this technology. It's an Artificial Intelligence lab that's the equivalent of actionless thinking and prayer, and it doesn't make any sense. Reading these documents is useful and There are some interesting observations within, but it is also clear that the table does not have certain options such as stopping work on those systems. It was argued that building these systems would help you understand the risks and design your solution.

You don't need to build a larger nuclear bomb to know that disarmament and missile defense are needed. If you want to be the one who owns the greatest nuclear weapons, you make larger nuclear weapons. OpenAI aims to be a leader in manufacturing products and developing licensed technologies. Because of this, they can't stop this job. So what is ethical is a dance that is very similar to other companies' greenwashing and pinkwashing. They need to see them moving towards safety while doing their best to get the job done. And, like greenwashing and pinkwashing, we need to demand more and promote more surveillance. 

Almost a year after we first saw the cutting-edge Artificial Intelligence tools in this newsletter, there were no signs of controversy in this area. It also doesn't mention the potential for Artificial Intelligence to become a FOOM and change the world.