Artificial intelligence (AI) has been around for some time. We’re familiar with signature verification, natural language processing, unsupervised learning and predictive analytics. In fact, the first chatbot, ELIZA, was created at the Massachusetts Institute of Technology (MIT) as far back as the 1960s.

Much like the development of the internet, mobile and then the cloud, AI is the next tool that organizations can leverage to ensure their survival. The question is no longer should an organization be looking to use AI; it is how should they use AI, and when? Its potential is so great that it will be the differentiator for key market players – and it’s something that we will see realized in the coming decade.

In this article, we share insights on large language models, tackle some of the challenges and risks, and importantly, outline the key opportunities presented by the latest development, generative AI (GenAI).

What has changed in recent years – and why we’re all suddenly talking about AI – is the advent of GenAI. This enables organizations to use AI to create new content in a multimodal format, including text, speech, video and images.

We’ve all heard of ChatGPT: The GPT stands for “generative pre-trained transformer,” as its software provides transformer architecture. Instead of reading information as words, the transformer architecture enables words to be converted into numbers. Those numbers represent tokens, which can be processed as a vector very quickly.

These language models (also referred to as GPT or GenAI) read the books, create a correlation between words, and then use that to sound as if they are truly responding to questions. This means that natural language has become the new programming language.

This is profound because it indicates that we can truly speak to the machines. We can actually start typing human language, and the machines will understand the request and respond to it. As organizations, we need to harness that.

However, it’s a large and confusing landscape because there are many large language models, and they’re multimodal. Which model should be used? What underlying architecture should be used? What are the use cases?

Prompting has become the new programming

The first thing we need to understand is how to ask the question, as that will determine the quality of the response – just like knowing how to program will determine the quality of the output.

The next step is to look at how you should enhance these language models with your own data. If you do, you will need to expand into retrieval augmented generation. This allows you to create knowledge based on your company’s various datasets. The language model will deliver responses based on your company data – which the generic ChatGPT does not have – and that will allow you to harness the capabilities of AI while preserving your intellectual property (IP).

The final consideration would be fine tuning and retraining. When you know what you want your large language model to do, you train it to do precisely that. That is where the true IP will lie. The individuals and organizations that have this expertise will have a distinct advantage over their peers. We have all asked the question, “Will AI replace my job?” We think that the answer is likely no, but someone who knows how to better use AI, might!

Large language models currently have about 60-70 percent accuracy. However, after retraining and fine tuning the models, you can increase the accuracy of the response rate and harness the true potential of AI. From there on, small language models will evolve. Depending on whether your task is fund accounting, custody, reconciliation or to derive specific analytics based on your data, your organization can design language models trained to do very specific things – much like you have subject matter experts for each of the various functions within your organization today.

The risks associated with AI

We can’t talk about AI without considering its risks. If you’re looking to build your own language model, you should consider:

• Explainability – With general machine learning models, you write the model and therefore it’s understood. GenAI is different: It’s not written; It is developed. Think of it like a child attending school each day: Ask a question today and get one answer. Ask the same question in six months, and you’ll get a completely different response based on the learning in those previous six months.
• Significant cyber risks – AI is like a delicate ecosystem. Each piece of data is connected, working together to create a balanced system. Even a single harmful change, like a cyberattack corrupting one piece of data, can upset this balance entirely. Since everything is interconnected, fixing just one part of the problem may not be enough. Therefore, it’s important to be vigilant about the datasets that the language model is being trained on.
• Managing hallucinations – Hallucinations in AI happen when a large language model perceives patterns or objects that don't exist or can't be detected by humans. They deliver responses that are false, misleading and even nonsensical. Organizations must ensure that the technology enables prompts delivering sophisticated, intelligent responses.
• Jailbreaks – Jailbreaking within language models happens when hackers prompt chatbots to circumvent human-built ethical guardrails to produce responses that would otherwise not be permitted.
• Risk of bias in AI – Responses can be learned from biased data, so it's important to train language models on unbiased data. To do that, organizations should ensure they work with ethical AI providers.

At State Street, we’ve simplified the opportunities presented by generative AI into four key themes:

1. Document intelligence – Automate and streamline data extraction from unstructured documents; mitigate risk of errors or incompleteness; increase traceability
2. Data quality controls – Augment and automate data validation; cut false alerts; increase efficiency in resolving true exceptions
3. Client lifecycle/onboarding – Self-service chatbot; predict client needs; educate on product options; automate and streamline onboarding with account intelligence
4. Co-pilot/productivity – Leverage GenAI co-pilots to drive productivity; faster response to internal and external queries; personalize experiences

Our team of over 400 AI practitioners is focused on designing the financial infrastructure of the future, reaching the pinnacle of information value and equipping you with a platform that serves as the foundation for all your organization’s data.

You can learn more about accelerating data-driven investing with artificial intelligence and how we are leveraging AI at State Street in our recent article. Together, we can harness the power of AI to create better outcomes.