Generative AI vs large language models: What’s the difference?
Whether used to drive business productivity, generative AI models are being pushed as carrying transformative potential for how we live and work.
Large language models (LLMs), a specialized kind of deep learning model, sit at the heart of some of the most popular AI models and inform how they operate. Knowing the difference between generative AI and LLMs gives businesses insight into its enterprise value – and where its limitations lie.
There is always a hierarchy within technology and AI is no exception to this rule. AI has become synonymous with data analytics and predicting outcomes alongside machine learning (ML) and with helping realize better automation across sectors such as manufacturing. However, there are now many other subsets.
Generative AI (as the name suggests) is a type of AI that uses algorithms to produce content based on user inputs. At a basic level, this includes producing or summarizing text, images, and audio, though more advanced multimodal models are now capable of handling other data such as video or a mix of the above. Each response is tailored to the user’s instructions – known as prompts – provided during each task.
What are large language models (LLMs)?
LLMs are artificial neural networks trained using ML on vast amounts of data to recognize text inputs and produce contextually relevant text outputs.
“LLMs function by utilizing intricate deep learning techniques like transformers to produce text that’s remarkably similar to human writing,” explains Peter Wood, chief technical officer at Spectrum Search.
“They’re trained using extensive datasets covering a huge array of texts so they can understand context, grammar, and semantics.”
What’s important to note is that, as with other forms of generative AI, LLMs are built to produce the best probabilistic response to any user input. These models aren’t comprehending text in a way similar to humans, but merely providing answers based on incredibly complex training regimens that instill pathways to producing helpful answers.
“LLMs generate and comprehend text by leveraging deep learning techniques,” says Nathan Marlor, head of data and AI at Version 1. They primarily use transformer architectures, which process and generate text by predicting the probability of word sequences. This process begins with tokenizing text into smaller units, such as words or sub-words.
“The model then uses attention mechanisms to weigh the importance of each token in the context of others, enabling the model to capture intricate language patterns and dependencies.
“By training on massive datasets, LLMs learn linguistic structures, context, and semantics, allowing them to generate coherent and contextually relevant text based on input prompts.
How LLMs work on a technical level
Camden Woollven, group head of AI at GRC International Group, describes how at their core, LLMs are just pattern recognition machines. “They’re trained on hundreds of billions of words,” he says.
“During training, they learn to predict what word comes next in a sequence. The key is something called a transformer, which uses ‘attention’ to figure out how important different words are in context.
“As the model processes text, it builds up layers of understanding, from basic syntax to more complex semantic relationships. When you give an LLM a prompt, it’s essentially continuing the pattern it sees, drawing on all that training data to generate a response that fits. And while it’s not exactly reasoning like we do, the output can be surprisingly human-like.”
Another emerging subset of LLMs is known as small language models (SLMs), which are built for ease of deployment and lower latency text answers. As these are more focused, they can be trained faster on defined and useful data.
Examples of SLMs at present include Google’s family of Gemma models, which start at 2 billion parameters in size, or Microsoft’s Phi-3 mini starting at 3.8 billion parameters.
However, all of these new AI technologies are currently problematic to run at scale as Woollven explains. He cites an often-overlooked issue that LLMs are “computationally intensive” with a “significant” level of energy usage and environmental impact.
There is no doubt though that the popularity of these tools is growing fast. Businesses are harnessing AI to positively affect their productivity and processes, hoping this can result in efficiencies to counter that cost.
But with increased use comes louder calls for regulation. Right now, the LLMs have become more of a focus for regulators than future applications or services run on generative AI. and Dom Couldwell, head of field engineering EMEA at DataStax suggests regulators are about “12-18 months behind where the industry is”.
“Many people conflate generative AI with LLMs – the challenge is that generative AI applications involve a lot more moving parts than LLMs on their own,” he warns.
“You have to integrate the LLM you choose into your application or service, and you have to decide on how you will use your own data with that LLM as well. All those parts add up to the whole service that a user gets.
“Regulations on AI currently focus on the LLM as the force behind generative AI, and while they are an essential part of generative AI, they are not responsible for the service as a whole.
“Instead, we have to create regulation that covers the whole generative AI landscape, from the role of company data through to the traceability and understanding of how results are created. Without this insight, we’ll miss out on the opportunities to use LLMs as part of effective generative AI services.”
Combining the AI streams
Millions of people around the world are already using AI tools such as OpenAI’s ChatGPT to produce written content, which operates using the LLMs GPT-4 and GPT-4o. As it’s the most well-known example of how generative AI around the world at the moment, it’s helped to popularize LLMs as the ‘default’ option for generative AI – even though this isn’t necessarily the case.
Other notable LLMs include Meta’s Llama and Google’s Gemini, all of which have helped to establish ‘text in, content out’ as the primary form of user interaction with generative AI models.
While generative AI has become popular for content generation more broadly, LLMs are making a massive impact on the development of chatbots. This allows companies to provide more useful responses to real-time customer queries.
However, there are differences in the approach. A basic generative AI chatbot, for example, would answer a question with a set answer taken from a stock of responses upon which it has been trained.
Introducing an LLM as part of the chatbot set-up means its response will become much more detailed and reactive and just like the reply has come from a human advisor, instead of from a computer. This is quickly becoming a popular option, with firms such as JP Morgan embracing LLM chatbots to improve internal productivity.
Other useful implementations of LLMs are to generate or debug code in software development or to carry out brainstorms or research tasks by tapping into various online sources for suggestions.
This ability is made possible by another related AI technology called retrieval augmented generation (RAG), in which LLMs draw on vectorized information outside of its training data to root responses in additional context and improve their accuracy.
LLM problems, pitfalls, and solutions
One of the major drawbacks of LLMs are known as “hallucinations”. This is where a model confidently produces an incorrect or completely nonsensical answer. Due to the nature of LLMs – every answer is produced with a degree of error to achieve a ‘unique’ response – it may be impossible to eliminate hallucinations entirely.
It’s because of drawbacks like hallucinations that many leaders are still wary of generative AI. Whether this prevents the technology from becoming as widespread as it might, or drives a shift away from LLMs as underlying models, remains to be seen.
The main way hallucinations are being tackled right now is by implementing RAG into the process, giving the AI a far broader base of knowledge to work with. RAG also has the additional benefit of enabling sources to be cited, which can point users to the information from which an output has drawn context.
Another major issue for generative AI and LLMs is legality when it comes to using data to create new text, images, music, or video works. Copyright, data privacy, and ethical concerns around bias are massive factors.
For businesses and organizations using this type of AI, this makes setting up clear guidelines and frameworks a necessity. But the precise guardrails needed to keep ethical AI on the table are a hot point of contention across the world.
But Peter Schneider, product director at software framework Qt Group, has this warning about LLMs. “Even if you have a gigantic amount of training data – let’s say trillions upon trillions of parameters – it’s still a probabilistic word-guessing engine,” he explains.
“And while it’s astonishing how good they’re getting at sounding like a human, you’re not ever going to guarantee it will give you factually correct answers, no matter how advanced the information gets. So, you have to validate the information. Never trust an LLM unquestioningly. There’s no such thing as a foolproof human; there’s no foolproof machine either.”
/cdn.vox-cdn.com/uploads/chorus_asset/file/25734293/GXdMxmZXcAEQ5K5.jpeg)
