Since its founding, the field of machine learning has used statistical models, including generative models, to model and predict data. Beginning in the late 2000s, the emergence of deep learning drove progress and research in image classification, speech recognition, natural language processing and other tasks. Neural networks in this era were typically trained as discriminative models, due to the difficulty of generative modeling.
In 2014, advancements such as the variational autoencoder and generative adversarial network produced the first practical deep neural networks capable of learning generative, rather than discriminative, models of complex data such as images. These deep generative models were the first able to output not only class labels for images, but to output entire images.
In 2017, the Transformer network enabled advancements in generative models, leading to the first Generative pre-trained transformer (GPT) in 2018. This was followed in 2019 by GPT-2 which demonstrated the ability to generalize unsupervised to many different tasks as a Foundation model.
In 2021, the release of DALL-E, a transformer-based pixel generative model, followed by Midjourney and Stable Diffusion marked the emergence of practical high-quality artificial intelligence art from natural language prompts.
In March 2023, GPT-4 was released. A team from Microsoft Research argued that "it could reasonably be viewed as an early yet still incomplete version of an artificial general intelligence (AGI) system".
Controversy
A picketer at the 2023 Writers Guild of America strike. While not a top priority, one of the WGA's 2023 requests was "regulations around the use of (generative) AI".
In January 2023, Futurism.com broke the story that CNET had been using an undisclosed internal AI tool to write at least 77 of its stories; after the news broke, CNET posted corrections to 41 of the stories.
In April 2023, German tabloid Die Aktuelle published a fake AI-generated interview with reclusive former racing driver Michael Schumacher. The story included two possible disclosures: the cover included the line "deceptively real", and inside the magazine acknowledged at the end of the interview that the interview was AI-generated. The editor-in-chief was fired shortly thereafter amid the controversy.
In July 2023, developments in generative AI contributed to the 2023 Hollywood labor disputes. Fran Drescher, president of the Screen Actors Guild, declared that "artificial intelligence poses an existential threat to creative professions" during the 2023 SAG-AFTRA strike.
Modalities
A generative AI system is constructed by applying unsupervised or self-supervised machine learning to a data set. The capabilities of a generative AI system depend on the modality or type of the data set used.
Generative AI can be either unimodal or multimodal; unimodal systems take only one type of input, whereas multimodal systems can take more than one type of input. For example, one version of OpenAI's GPT-4 accepts both text and image inputs.
Text: Generative AI systems trained on words or word tokens include GPT-3, LaMDA, LLaMA, BLOOM, GPT-4, and others. They are capable of natural language processing, machine translation, and natural language generation and can be used as foundation models for other tasks. Data sets include BookCorpus, Wikipedia, and others.
Code: In addition to natural language text, large language models can be trained on programming language text, allowing them to generate source code for new computer programs. Examples include OpenAI Codex.
Images: Generative AI systems trained on sets of images with text captions include Imagen, DALL-E, Midjourney, Adobe Firefly, Stable Diffusion and others. They are commonly used for text-to-image generation and neural style transfer. Datasets include LAION-5B and others.
Molecules: Generative AI systems can be trained on sequences of amino acids or molecular representations such as SMILES representing DNA or proteins. These systems, such as AlphaFold, are used for protein structure prediction and drug discovery. Datasets include various biological datasets.
Music: Generative AI systems such as MusicLM can be trained on the audio waveforms of recorded music along with text annotations, in order to generate new musical samples based on text descriptions such as a calming violin melody backed by a distorted guitar riff.
Video: Generative AI trained on annotated video can generate temporally-coherent video clips. Examples include Gen1 and Gen2 by RunwayML and Make-A-Video by Meta Platforms.
Robot actions: Generative AI trained on the motions of a robotic system can generate new trajectories for motion planning. For example, UniPi from Google Research uses prompts like "pick up blue bowl" or "wipe plate with yellow sponge" to control movements of a robot arm.
Notable generative AI systems include ChatGPT and its variant Bing Chat, a chatbot built by OpenAI using their GPT-3 and GPT-4 foundational large language models, and Bard, a chatbot built by Google using their LaMDA foundation model. Other generative AI models include artificial intelligence art systems such as Stable Diffusion, Midjourney, and DALL-E.
QUESTION I: How Generative AI concept could be extended outside LLM, say on numeric data?
QUESTION II: What is a Generative Model?
REFERENCE: Wikipedia, An Article