Generative AI: what should we know about it and why is it important?
Generative Artificial Intelligence (GenAI) is a general term for models like ChatGPT that can create new content: texts, sounds, programming code, images, or videos. Recent groundbreaking discoveries in this field have the potential to significantly impact our approach to automating content creation.
AI Talks #01: Kaśka Śledziewska and Michał Paliński. A Review of Large Language Models.
Machine learning, one example of which is generative AI, has been changing industry, services, science and our daily lives for years. Recommender algorithms help discover new movies or songs, are used in medical analysis, production automation, personalization of marketing content, or detection of anomalies in financial data. You have probably used AI without even knowing it – voice assistants such as Siri operate on the basis of this technology, as do chatbots that help find answers to questions about online services. It is very possible that AI has also been used on you, using your data or supporting decision-making (e.g. in credit risk analysis).
Until recently, machine learning was mainly used in predictive models that analyzed patterns in data, supporting tasks such as classification or clustering. An example of a problem that such algorithms can solve is analyzing a large number of photos to learn to recognize classes of objects (e.g. whales in aerial photos) and detect them in subsequent photos (which supports the study and monitoring of these animals). The breakthrough was the emergence of generative AI, which not only recognizes patterns but can also create content on its own.
The first machine learning models that processed text were trained to classify data based on human labels. For example, a model could learn how to label social media posts as positive or negative.The type of learning is called supervised because a human oversees the process, guiding the model on how it should interpret data and make decisions.
Newer models utilize a method known as self-supervised learning. In this approach, text models are trained on massive datasets, such as internet forums, books, or scientific articles, allowing them to independently generate predictions. For example, models like GPT can anticipate a sensible sentence ending in nearly any context based on just a few words.
Initial studies are emerging that shed light on the impact of generative AI on various occupational groups. Elondou et al. (2023), a team associated with OpenAI, predict that one-fifth of workers may experience the influence of GPT models on half of their professional tasks. Automation is no longer limited to merely routine tasks, those performed according to repetitive, easy-to-program procedures. Thanks to AI models, the scope of automatable tasks now includes those requiring creativity or analytical thinking. As a result, the topic of automation begins to affect new groups of workers, including researchers and academic teachers. However, a group of researchers from ILO emphasizes that GPT has a greater potential to support rather than replace work (Gmyrek et al. 2023). The increase in productivity using generative AI has already been studied experimentally, for example, in the context of writing tasks (Noy, Zhang 2023), programming (Peng et al. 2023), or customer service (Brynjolfsson et al. 2023). Generative AI has the potential to level the effects of work between less and more skilled workers—both Noy and Zhang (2023) and Brynjolfsson et al. (2023) noted greater benefits from using generative AI when performing tasks among individuals with less professional experience.
References:
Brynjolfsson, E., Li, D., & Raymond, L. (2023). Generative AI at work. DOI
Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023, March 17). GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. arXiv.org. DOI
Gmyrek, P., Berg, J., & Bescond, D. (2023). Generative AI and Jobs: A Global analysis of potential effects on job quantity and quality. SSRN Electronic Journal. DOI
Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. DOI
Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv (Cornell University). DOI