AI Generated Content (AIGC) Tech Landscape — Apr 2023
Loh Yu Chen and Ian Goh did a quick chart-out of the generative AI tech landscape with their mentor Daryl Lee, during their month-long internship with DSTA. This article was derived from their research.
Generative AI is a subfield of artificial intelligence that involves creating computer algorithms that can generate new and original content. This can include anything from images and music to text and video. Generative AI is revolutionizing the way we think about creativity and is being used in a wide range of applications, from art and entertainment to healthcare and education.
TLDR:
A copy of the tech landscape in XMind mind-map format can be found here. We mantain this mindmap as a live document given the rapid pace of developments in this field. Why did we use a mind-map? It allows a one-glance view of what are the key areas to monitor, where are potential gaps and potential opportunities to explore further.
Two big trends today — Generative Adversarial Networks (GANs) and Generative Pretrained Transformers (GPTs).
GANs are a type of neural network that pits two AI systems against each other to create more realistic and sophisticated content. They are being used to create everything from realistic images and videos to 3D models and even entire worlds.
Generative Pretrained Transformers (GPTs) are AI models that have been trained on massive amounts of data and can generate high-quality text and dialogue. GPTs are being used in a wide range of applications, from chatbots and customer service to creative writing and storytelling.
One approach to keep up-to-date with the latest technology developments is to “follow-the-money”, i.e. what are the VCs looking at. Other ways include tech trends and AI reports such as Future Today Institute and Stanford AI Index Report.
Underlying technology and possible products and services
The most important underlying technology behind Generative AI is deep learning, which involves training neural networks to recognize patterns in data. Another technology is reinforcement learning, which involves training AI systems to learn from their mistakes and improve over time. Natural language processing (NLP) is also a critical technology in Generative AI, as it allows machines to understand and generate human-like language. In our mindmap, we list ideas that cover underlying data, models and metrics notably from Stanford Holistic Evaluation of Language Models (HELM) and LifeArchitect.AI.
There are many exciting developments in the Generative AI space. One example is OpenAI’s GPT-4, which is one of the most advanced language models ever created. GPT-4 can generate high-quality text, answer questions, and even write code. Another example is NVIDIA’s StyleGAN, which can generate hyper-realistic images of people and objects. In our mindmap, we list out generative products such as text, image/video, speech/audio/music, multi-modal include integration with synthetic 3D content generation tools e.g. BlackShark.ai, Unreal DigitalHumans, Reallusion Character Creator.
Recommended readings
We recommend resources include the Generative Deep Learning book by David Foster, which provides a comprehensive overview of the field. Another recommended book is Deep Learning with PyTorch by Eli Stevens, which covers the basics of deep learning and neural networks. We favour PapersWithCode to accelerate learning. For online courses, ClassCentral as well as the Deep Learning Specialization on Coursera by Andrew Ng are excellent places to start.
Open challenges:
- What is the optimal platform for running generative AI models: cloud-based, on-premises with A100 GPUs, on-premises with consumer graphics cards, or running LLM models on portable devices like thumb drives?
- How can we prevent generative AI models from contributing to the formation of information “Filter Bubbles”? Can we enhance their explainability using knowledge graphs? How can we train and fine-tune these models to learn the “right facts” for specific domains?
