*What is generative AI? How do large language models work? SRI Policy Researcher* ***Jamie Sandhu*** *lays the groundwork for understanding LLMs and other generative AI tools as they increasingly permeate our daily interactions.*

Large language models (LLMs) and other generative artificial intelligence (AI) tools are weaving into our daily interactions, yet it seems many of us might be missing or misinterpreting the characteristics that define them and their functional operation. Arguably, the current regulatory landscape of AI is experiencing similar uncertainty as policymakers waver to determine suitable policies to address these new technological advancements. This lack of clarity isn't necessarily a fault but perhaps an opportunity. As we inch towards an AI-driven future, a deeper factual grasp of these technologies might be the key to ensuring a safer future for all.

This piece lays the groundwork for understanding LLMs and other generative AI tools.

LLMs and generative AI

While LLMs represent just one category of generative AI, focusing specifically on text generation, generative AI is named for its capability to generate a more diverse set of outputs, including text, images, audio, computer code, and more. Throughout 2023, a series of examples emerged demonstrating generative AI’s impressive ability to generate content and inform experts, whether it be composing a distinctive music piece, designing graphics, or even detecting and diagnosing diseases through medical images and generating code in various computer languages to support programming.

Whereas generative AI encompasses a wider scope of content generation abilities, LLMs, as a subset of generative AI, are applied to perform language-related tasks specifically. They power software that aids in language related tasks and synthetically generates written text, such as drafting business emails, helping students enhance essays, or summarizing long documents. When you interact with an AI system and receive a language-based response that seems human-like, there's a good chance an LLM is behind it.

Mechanics and methods

How do LLMs work?

As a University of British Columbia researcher recognized when they asked a language model about its process, it replied: “I don’t understand the text I am trained on, but by looking at so many examples, I learn to mimic the style, the context, and the ‘flow’ of human language.”

These mechanics are impressively demonstrated by the Financial Times visual storytelling Team and Madhumita Murgia, who show that the efficiency of text-generating AI stems from identifying patterns in data provided to an LLM system and enabling it to predict language responses and generate content. For context, an LLM might analyze countless online articles and books to refine its understanding of how language is constructed for different topics and genres. Likewise, other kinds of generative AI such as image-making tools, provided with enough data, can, for example, detect recurring patterns in the works of prominent painters and produce images that echo the same style.

Both LLMs and generative AI systems learn to do this by adjusting certain settings, which computer scientists call “parameters.” You could think of parameters as akin to the settings on a camera designed to capture the best photo. Just as a photographer may adjust for focus or lighting, AI systems adjust parameters to create optimal language, visual, audio, and other predictions. The more parameters these models have, the more detailed and specific content they can create. As Helen Toner explains, the “large” in large language models refers to the millions or even billions of parameters used to generate outputs.

Because these tools generate content by analyzing and finding patterns in the massive amounts of data that are given to it (this is called “training”), the data that is used matters a lot.

Generated content is essentially a rearrangement or combination of pre-existing elements, which on the surface may seem to be new and unique, but should be understood as being rooted in—and limited by—the data it was trained on. For them to learn and improve, the foundation of LLMs and generative AI rely on enormous libraries of words, images, audio, and more. Training LLMs, for example, to learn patterns, grammar, and semantics relies on extremely large sources of text data. The sheer size of this data can be illustrated by researchers at the European Data Protection Supervisor, who noted that a recent state-of-the-art LLM was trained with more than 3 billion pages of data from publicly available internet resources.

Often acquired through the method of data “scraping” (also known as “web scraping”) which entails pulling large amounts of information from the internet, these AI systems use this data to train on and improve their performance. They can also improve over time based on the prompts and information we provide as users—this is called reinforcement learning with human feedback (RLHF). Alongside the refinement of performance, LLM and generative AI systems can also use RLHF techniques to continually adapt and evolve based on the prompts and information provided by users. To this end, both technologies hinge on similar principles of predictive analyses accompanied by vast datasets.

The distinction between all kinds of generative AI and LLMs specifically revolves around their applications. LLMs are a subset of generative AI that primarily use language as opposed to other more diverse representations seen through generative AI. But it’s worth noting that these distinctions are becoming increasingly blurred as multimodal AI systems emerge—ones which use both language models and other kinds of representations (pictures, sounds) to function. These reflect a new frontier of AI systems where combining the defining attributes of LLMs and other types of generative AI can elicit new insights and research, as observed by researchers at the University of Toronto.

However, despite the fact that these systems are doing amazing things seemingly “on their own,” we must keep in mind that they are actually leveraging human creativity, intellect, and labour to produce results based on assumptions that, at the end of the day, stem from mathematical concepts and methods.

Understanding the impact

The dependence of these AI systems on our data reveals various challenges and concerns, but it also highlights the complexities in our interactions with and utilization of these technologies. Notably, misplacing complete individual and collective trust can lead to negative impacts, as LLMs, for example, don't always guarantee factual accuracy, despite the coherence of the texts they provide. Indeed, uses of LLMs in the healthcare setting have been shown to provide medical professionals with error-prone or false information.

Similarly, generative AI systems may produce realistic fictional content that might otherwise be inadvertently spread as fact, either falsely by AI systems and users or intentionally by nefarious actors. LLMs specifically and generative AI in general have been at the centre of social debates, with concerns arising from potential biases in their training data, questionable uses of these systems to produce unethical content, and their significant ability to disrupt existing socioeconomic structures.

Many of the challenges and solutions to these problematic scenarios are technical. LLMs and generative AI systems, whether developed by Anthropic, Google, OpenAI, or any other organization, reflect the data they were trained on. This leads to variations to similar prompts, much like different humans might offer different perspectives based on their backgrounds and experiences. This may also mean that the systems disseminate information that contravene societal norms, values, beliefs, and language.

RLHF, as mentioned earlier, relies on human feedback, and there are differences in the evaluation of what constitutes a right or wrong output by different individuals, which in turn shapes how the LLMs and generative AI systems adjust. Humans, for instance, are known to be biased, and this can be transferred into the data itself, which in turn creates a biased AI system. Researchers are increasingly recognizing the importance of assessing the aggregation of this data, along with the impacts of human feedback and its interaction with these systems.

But it’s worth noting that the behaviour of these technologies isn't solely due to the data they’re trained on. Indeed, the objectives and values prioritized during their development, and the specific algorithms and methodologies used by the developers and organizations behind these systems are embedded into their responses and behaviour. This imprinting creates a kind of personality for LLMs and generative AI, a field of research that is still in its nascent stages.

A benefit of this is that developers can responsibly influence the outputs and responses of these technologies through design choices. For example, ethical safeguards can be included during development to ensure that a language model or generative AI system behaves more ethically than it otherwise would. For instance, when ChatGPT is asked for ideas about committing crimes, it responds with a refusal, demonstrating the implementation of ethical constraints. This refusal is a safeguard set by the developers to prevent the AI from providing harmful information. Just as policy decisions help safeguard against negative impacts, developers also play a significant role in ensuring AI is safe for all.

Where do we go from here?

The recent proliferation of LLMs and other generative AI tools finds us in a reality where AI can tell and show us almost anything. But this means clarity and understanding rooted in factual knowledge about these technologies are even more important now, especially when it comes to speaking about the technology itself and formulating responsive policy around it.

While this article only offers a very simple explanation of the technical aspects of LLMs and generative AI, its main aim is to highlight the importance of educating society, decision-makers, and users about this transformative technology and promoting policies and business expectations that recognize the value of AI education. As these technological advancements solidify their normative presence in our daily lives, perhaps AI governance is inadvertently overlooking the greatest societal risk concerning AI—knowledge disparity.

Want to learn more?

About the author

Jamie Amarat Sandhu is a policy researcher at the Schwartz Reisman Institute for Technology and Society. His specialization in the governance of emerging technologies and global affairs has earned him a track record of providing strategic guidance to decision-makers and addressing cross-sector socio-economic challenges arising from advancements in science and technology at both the international and domestic levels. This expertise is supported by an MSc in Politics and Technology from the Technical University of Munich's School of Social Science and Technology in Germany, complemented by a BA in International Relations from the University of British Columbia.