Generative AI - Primer
The painting above, Théâtre D'opéra Spatial, won the September 2022 Colorado State Fair. The painting is attributed to Jason Allen. The painting contains classical figures in an elaborate almost baroque landscape bathed in a heavenly light. Normally there would not be anything strange about this story except the painting was made using Midjourney an AI software package.
In mid-February 2023, Kevin Roose, a NY Times technology columnist was testing a new chatbot AI enhanced version of the Bing search engine when the testing (conversation) went into uncharted territory. The AI software attempted to declare its love for the reporter among other things.
The Bing AI chatbot is based on ChatGPT AI software. Both Midjourney and ChatGPT and Codex (computer coding AI) are highly complex software packages that belong to a special class of AI called Generative AI. This software is called generative because it creates new output using algorithms to synthetize or produce new data. Generative AI creates new content based on a type of deep learning called GAN (Generative Adversarial Networks). The GAN is a class of machine learning framework that uses two different types of neural networks. One is a generator (creates new data) and the other is a discriminator that evaluates the data. Used together the generator improves its output based on the feedback received from discriminator until the contents generated is almost identical to real data.
Generative AI is a nascent technology with immense potential and its ongoing applications have attracted a lot of investor attention. So far generative AI has been used to design drugs in a space of a few months whereas previous drug discovery processes tool 3 to 6 years.
In addition, the technology has had an impact on material science within automotive, aerospace, medical and electronic industries by composing new materials with specific material properties through inverse design. In addition, the technology has already had an impact in semiconductor chip design by reducing the component placement lifecycle from weeks to hours. In the automotive industry, the technology can be used to innovate on lighter vehicle designs helping make cars more fuel efficient.
Another usage of the technology is the generation of synthetic data. This is a type of data that is generated rather than obtained from direct real-world observations. This ensures the privacy of the original data sources. This type of data can then be used to train models that can help companies develop or identify new business models, enhance customer experience, or improve decision making.
Generative AI presents us large philosophical and practical challenges that will only grow as the technology grows and matures. The information generated by AI models can be just wrong, biased, or manipulated to enable criminal or unethical actions. In addition, there are legal questions as to the ownership of synthetic data when copyrighted work is used as part of an AI training set. There are also the reputational and legal risks involved in unintentional publishing of biased or offensive content.
Training generative AI models is highly computer intensive which can raise the question of whether the increase in power consumption and hence greenhouse emissions is worth the effort. Finally, there is the security aspect. Generative AI can be used to generate malicious code, perform fraudulent activities, and create fake spam news.
There are ways that the business risks in the use of generative AI can be mitigated. Organizations should consider the use of smaller, specialized models based on their own data to fit their needs. Organizations should not use these AI models to make critical business decisions and should keep a human (or team) to review the output before it is published or used.