Our CEO Joseph Twigg was joined by Iria Del Rio, our lead NLP engineer to talk about the explosive rise of ChatGPT and other large language models, what got us here and what this means for the financial advice and services space.
Watch the full webinar replay here.
We’re witnessing history
The release of ChatGPT has brought a technological tectonic shift. It has brought AI and terms like large language models (LLMs) and generative AI into the consciousness of pretty much everyone on the planet. ChatGPT is the fastest growing application in history, reaching 100 million users in two months. There are almost 70 million news articles talking about ChatGPT and other large language models.
We’re at the start of a new era. LLMs are going to fundamentally change the way people work over the next four or five years, with every industry impacted. There will be large impacts in wealth management and financial advice.
Large language models – tell us more
ChatGPT is a change in the paradigm of Natural Language Processing and an example of a LLM. These are models that learn the underlying distribution or combination of words in a particular language – be it the spoken word or even programming languages. These models are great at predicting the proceeding most probable words given a certain context. They can do that for really long paragraphs and long pieces of text. They are trained on data using text that is available on the internet.
LLMs are black box technology, reflecting the fact that it’s not always clear why they produce the type of response they produce. However, this is not a reason to panic – there is a lot of research about how to understand the reason for these responses and techniques such as fact checking the models. We can also use traceability techniques to make sure the model is producing an accurate and correct summary, for example.
A brief history in AI time
In 2018, AI took a big shift with the appearance of an architecture called the Transformer. It was a type of probabilistic model designed to produce and generate content. This model was quite powerful, but nothing compared to GPT2 and then GPT3. We then moved to instruct GPT, which not only changed the size but the way we trained the models. They were trained on instructions, not how to give a general answer, and this was augmented with a new technique called Reinforce. This meant learning from human feedback and being more aligned with the type of output a human would expect. This helps to control outcomes and ensure they are not biased and toxic.
Reinforce is what the ChatGPT model is based on and that is why when you ask a question that can be considered toxic or inadequate, the model politely refuses to give you an answer. With this technique, the model is also more aligned with the type of output that a human would expect. ChatGPT is almost ‘chatty’ keeping information that has been discussed previously and setting the right tone for human interaction.
The next iteration is GPT4, but we don’t have a lot of details about this. Open AI hasn’t revealed how it was trained, its size or how much information it can handle. We do know it can handle images as well as text, so it’s another big jump and potentially trillions of parameters.
Where to next?
The phenomenon of emergence – something we can see in nature and in physics in different scenarios and the ability to show completely different behaviours and capacities. This happens with LLMs when they reach a certain size, they can suddenly start to perform tasks that they were not able to before. This is what we’ve seen with the GPT models moving from millions to billions, and potentially trillions of parameters the model learns during its training process. The model is starting to show capabilities that we haven’t seen before.
So what for AI and financial advice?
We can use LLMs with a combination of tools, and this is really beneficial in financial services. However, because LLMs have been trained on data from a certain period of time, it’s really vital to keep the information current and accessible on the internet. Domain expertise is particularly important in a highly regulated area like financial services and the point- in-time knowledge base is fundamental. However, risk management is also essential and even with the appropriate guard rails in place, it is imperative that the research and development to manage these risks is really emphasised and undertaken.
Don’t panic, AI is not taking over financial services…yet
AI is (probably) not going to take your job anytime soon. AI-derived advice from the black box is a regulatory minefield and we also don’t know if it is what consumers want. It is important for us to acknowledge the effective utilisation of AI by humans, as that will bring about a productivity revolution. It is coming and will fundamentally change the economics and the approach of financial advice. So, if you’re an advisor, you will be able to see 2-3 times more clients whilst delivering an equivalent or better service. That’s something that really can’t be ignored.
Humans are most definitely needed
This productivity revolution all revolves around what we call a ‘human plus’ activity. In terms of AI adoption to date, it has been quite a niche. Larger companies have fraud, cyber or AI type analytics in place. For example, AI looking at customer experience or automating compliance monitoring. But where we are going now is more general, and AI assistants will be every day parts of our life, reducing admin burden. Humans will become the validators and reviewers of the advice that is being driven by machine outputs. This will ultimately allow the main focus for advisors and managers to be on relationship and customer support.
Aveni Assist – how can we help you?
We’ve developed Aveni Assist on the model underpinning ChatGPT called GPT3.5 Turbo, and combined it with a range of our own internal models that have been developed specifically for financial advice and wealth management. We’ve created gold standard prompt engineering designed to drive excellent customers and created important information retrieval mechanisms and traceability algorithms. This enables you to trace back from any piece of data that’s machine generated to the source of that data. It provides the control but also the comfort and trust that the output is accurate. This can be deployed face-to -face if you record meetings and over phones, but it is optimised in the video conferencing setting. The meeting is captured, outputs processed in our natural language pipeline and within a minute or two the report is available. It provides an admin workflow; instant compliance and vulnerability checks; generates emails; updates CRM systems and generates highly detailed suitability reports; and all easily searchable with briefings notes incorporated.
This will take a process that is typically three hours down to 15 to 30 minutes. We’re very excited to bring this to market and believe it will revolutionise the administrative and risk management burden. Be the first to try Aveni Assist: sign up to our waitlist here!