Generative AI (4th installment of a series of articles about AI)
Generative AI (4th installment of a series of articles about AI)
The latest developments in AI, particularly Deep Learning, have/are/will penetrate many fields. The latest is what is called Generative AI or AI Generative Content. With generative AI, users can input text and generate poetry, music, designs, paintings, and others that match the text input provided by the user. An example of this is DALL-E from OpenAI, a deep learning model that accepts text input and produces image output according to the user's request. For example, a user can ask DALL-E to paint a person in the style of Rembrandt's paintings through text, and DALL-E will produce output in the form of several human paintings that resemble or closely resemble Rembrandt's painting style. Another example is MusicLM from Google where users can input text about the music they are asked to create and the output is music according to the user's request. For example, a user can write "Slow tempo, bass-and-drums-led reggae song. Sustained electric guitar. High-pitched bongos with ringing tones. Vocals are relaxed with a laid-back feel, very expressive", and MusicLM can generate music audio and vocal sounds according to the characteristics requested by the user. The development of AI is also known as a Large Multimodal Model, which is essentially a AI model that can generate certain modalities (text, images, video, audio) as output from input in the form of other certain modalities. GPT-4, which was just released by OpenAI on March 23, 2023, and is a further development of Chat-GPT, is also a Large Multimodal Model because it can process input in the form of images and output text according to the image provided by the user. For example, a user can provide a photo containing a strange image and then we ask GPT-4 what is funny/strange about this photo, and GPT-4 can show the strange thing about the photo using text. Many of AI's Generative technologies are based on the development of Deep Learning based on Transformers and Generative Adversarial Networks trained with multi-modal data sets.
The latest development of AI is certainly very useful and helps us in creating/producing paintings, designs, music and others. Never before imagined that now we can input text and AI will produce creative works according to what is requested in this text. Most of the technologies/applications based on Generative AI are still not very good in performance or operate in limited domains/topics, but it is certain that they will get better in the next few years, along with improvements to the AI model/algorithm, including the possibility of involving human feedback in the training process, and the increasing number of datasets that are trained to be learned by the AI model.
On the other hand, the development of Generative AI has raised debates and concerns in several circles, for example regarding copyright and others. Some AI tools that generate images/paintings have even begun to be sued in court by several famous artists/painters because their AI- models were trained using images spread on the Internet, including paintings from these famous painters without permission. DeepFake which can produce photos/videos/audio that look as if they are real but are actually fake (created by AI) is another worrying example in the hands of irresponsible people. The ethics of AI are indeed very necessary at this time for all stakeholders. The development of AI has raised concerns and worries for many parties. As a note related to this, the developers of GPT-4 have actually tried to minimize the negative impact of prompts/queries given by users. For example, if there is a question that is dangerous or corners a certain group or is very sensitive, GPT-4 will refuse to provide an answer and state the reason. As a side note, we at Korika (Artificial Intelligence Industry Research and Innovation Collaboration) are collaborating with Unesco to promote the AI Ethics in Indonesia, taking into account that AI technology, despite its extraordinary benefits for civilization and humanity, can also be dangerous, unsafe, and biased towards minority groups. These negative implications of AI technology need to be minimized and, if possible, eliminated through the principles outlined in the AI Ethics.
Bandung 26 Maret, 2023
Bambang Riyanto T.