INVESTMENT CENTRE
AI Revolution: Reshaping the Landscape of Investment Management.
Introduction
The past six months have been described as a Cambrian explosion in artificial intelligence (AI), with the initial release of ChatGPT by OpenAI at the end of last year, which captured the world’s imagination by showing the potential of large language models (LLMs).
While AI and machine learning (ML), a subset of AI, have been integral to investment management for half a century, recent advancements have rapidly transformed the landscape. However, the past decade has seen rapid innovation and success across many domains after a long winter in AI when many thought progress was doomed to stagnate.
In this short piece, we briefly explore how the landscape has changed over the past decade, and what we can look forward to in the coming years, which is, of course, impossible to predict. We will put an emphasis on the application of this technology to investment management and related activities, such as financial planning.
A brief history
A significant breakthrough in AI occurred in 2012 when convolution neural networks were used to beat all other contestants – by a large margin – in an image classification competition. The potential for deep learning (DL), a subset of ML, led to a frenzy to explore the limits of this new technology, and how it could be applied to other problems and in other industries.
Another major milestone, among many others that cannot all be covered in this short piece, was the development of AlphaGo by DeepMind which beat Lee Sedol, a Go world champion, in 2016. Some in the field of AI thought this would not be achieved for 100 years. DeepMind went on to develop a more generalised version of this technology called AlphaZero, which, playing against itself, could learn how to become the best player at any two player games, including Go and Chess.
The next major breakthrough came in 2017 with the publication of a paper entitled “Attention is all you need” by Google employees, which introduced ‘transformers’, a new approach to machine learning that revolutionised natural language processing and led to LLMs like ChatGPT.
Other major breakthroughs, which we will not discuss, but may be of interest to some readers include: AlphaFold by DeepMind that is used to predict how proteins fold given their amino acid composition; and stable generative diffusion, which has popularised text-to-image models such as DALL-E, Imagen, and MidJourney, and can be used to produce original images based on user input text.
Where we are now
As mentioned in the introduction, LLMs have captured the imagination of the world. ChatGPT garnered one million users in its first five days of release, and 100 million users in its first two months.
This adoption rate is unprecedented for any technology. Within a few months, OpenAI upped the stakes with GPT4, which they had been testing internally and externally (closed group) for six months. Microsoft had by then injected another $10 billion into OpenAI and started deploying the technology into their products, starting first with Bing, their search engine.
What exactly are LLMs and how can they be used? Essentially, they are models that predict the next word (technically token), given a series of words or tokens – up to thirty-two thousand in the case of GPT4. How can such simple models be so powerful? This has largely happened through emergent behaviour which was not really predicted. As an example, while building the first GPT (generative pre-trained transformer) models, which trained on lots of information from the internet, it was discovered that the model had learned how to code in various languages, something that it was not specifically trained to do.
What else can LLMs do? They can understand text in more than 100 languages and translate between those languages. For example, LLMs can be used to translate a document from Mandarin Chinese to English or Zulu. They can understand the sentiment expressed in text – words, sentences, and longer. They can extract keywords, named entities (people, companies, places, etc.) from documents. They can classify text by topics. They can summarise large pieces of text or extract the main points. They can generate text based on prompts provided, including answering questions, providing ideas on a topic, holding a conversation with a human, etc. New emergent behaviour is continually being discovered. Another emergent behaviour that was discovered – as the model was not trained to do this, and previous natural language models were bad at – was its ability to reason.
Although the technology has many limitations and issues, with hallucinations perhaps being the biggest, it is still phenomenally powerful and useful. It has already passed exams in law, medicine, and business, and many more. It will outcompete most humans on a range of topics and capabilities. And it is getting more powerful every day, with two specific additions that will transform the landscape yet again.
The first is that GPT4 has been built as a multi-modal model, which means that it can ingest text and images. This is yet to be released to the public. While basic image-to-text capabilities have existed for many years, what GPT4 will be able to do is worlds apart from anything we have previously seen. One demo included a basic sketch of a website on a napkin which GPT4 then coded up.
The second is that OpenAI has opened up GPT4 to use plugins. What are plugins and why are they so transformative? Without plugins, GPT4 can only predict the next word in a sequence based on the weights of the parameters of the model and how they are all inter-connected (referred to as the model architecture). It does not have access to the internet to look up facts and it does not have access to tools to perform specific tasks. Plugins change this by allowing software providers to expose their software to be used by GPT4. GPT4 can therefore browse the web to find answers to questions that were not part of its training data. It can perform complex calculations using tools like Wolfram Alpha. It can even find you a great restaurant and go ahead with booking a table using the OpenTable plugin.
Like many other users in many different industries, many are scrambling to understand the potential impact that this technology will have on investment management. While ML has been used for decades in investment management, it has largely (not solely) focused on using numbers, not language. Even where natural language has previously been used, which it has, these language models were nowhere near as capable as what we have today.
Users across the value chain
Let us consider three types of users across the investment value chain to explore some of the possibilities presented by LLMs.
For analysts working on the sell-side or buy-side and analysing companies, LLMs can be used to understand the vast amounts of text that gets produced related to the company being covered. This could include formal documents, such as financial statements, which contain much more information beyond just income statements and balance sheets. It could also include articles from mainstream media, tweets from Twitter, posts from LinkedIn or Reddit, or any other source of information on the web. It could also include information about related companies (including suppliers, clients, and competitors), about the industry or the economy.
Humans just do not have the capacity to consume the same amount of information that machines can. Discretionary fund managers (DFMs), such as ourselves, can use these models to understand fund managers in ways that were just not possible before. We have long used machine learning on returns, holdings and transactions (where available), to understand fund managers in isolation and how they compare to other fund managers. Any information in the form of text – due diligence meetings and answers to questionnaires – had to be ingested by the investment team to be analysed and synthesised into meaning. We can now use these LLMs to analyse the huge amounts of text data, including transcripts from meetings with managers.
Finally, for advisers, these LLMs could be utilised in various applications to assist with the dispensation of financial advice. They may be useful in understanding the risk profile and financial needs of your clients by analysing answers to questions in ways that were not possible before. They may also be useful in communicating with clients. A single message drafted by you could be manipulated by the LLM to sound more formal or informal, more technical or less technical, more comprehensive or more concise, all depending on the specific client that it is intended to go to. You could have your client profile stored in your CRM system, and automatically generate (not send) a custom message to each of your clients based on your generic message.
Please be careful to always include a human in the loop as these models are far from perfect at this stage.
Future stumbling blocks
There is no doubt that these models will become much more powerful with the passage of time. Just how quickly this development happens is difficult to predict and relies somewhat on two major potential stumbling blocks:
- Technology: We may have hit a limit in terms of what the current model will be able to accomplish and development will stagnate until we have the next major breakthrough. Although it appears likely, we are not too concerned for two reasons. The first is that the current technology still has a very long runway. We have more data to train on, more compute to deploy for training, and the ability to fine-tune the current models to achieve even more. We have seen that this is the step changes from GPT3 to ChatGPT (GPT3.5) to GPT4. The second is that the next major breakthrough may just be around the corner. So many resources are being thrown at this technology that new discoveries are likely to follow, sooner rather than later.
- Policy: With great power comes great responsibility and many people (and governments) are concerned that this path could lead to bad outcomes. Going into specifics is beyond the scope of this write-up, but regulation is very likely and it could have an impact on the rate of development. This is not necessarily a bad thing if it helps to prevent a catastrophic outcome, but it may mean that we will need to wait longer for even more transformative technologies like these.
More questions and topics
There are many other important issues that we were unable to cover in this article but below are a few important questions and topics.
First, we need to discuss the ethical implications of using AI in investment management. For example, how can we ensure that AI is used in a fair and transparent way? How can we prevent AI from being used to discriminate against certain groups of people?
Second, we need to discuss the potential impact of AI on the jobs of investment professionals and advisers. As AI becomes more sophisticated, it is possible that some jobs will be automated. This could lead to job losses for some professionals but it could also create new opportunities for others.
Finally, we need to discuss the future of AI in investment management. What are the long-term implications of using AI in this field? How will AI change the way we invest?
Conclusion
There is no doubt that times are changing and the rate of change is accelerating (hyper exponential). Technology has always been a doubled-edged sword, able to benefit humankind in transformative ways, but also capable of being used for destruction, pain and suffering when used by bad actors – or sometimes unintentionally because we simply do not know how to wield the power.
Stay abreast of these technologies as they evolve and consider how they might be applied in your own organisations. As your DFM, we are continuously exploring how these developments can create better outcomes for our advisers and clients. Let us embrace the opportunities presented by this transformative era of AI in investment management together.
Key points
- Rapid progress: AI and machine learning have advanced significantly in the past decade, revolutionising data handling and prediction-making in investment management.
- Large language models: OpenID’s GPT-4 and similar LLMs are transforming the AI landscape with their versatile capabilities, proving to be valuable tools in investment management.
- Future outlook: the increasing role of AI in investment management presents opportunities for increased efficiency, but also poses challenges such as potential technological hurdles, regulatory issues, and ethical considerations.
Knowledge is Power…
Access more insights from the INN8 Invest team’s quarterly report.
For you, the adviser, click below to download a collection of articles to use in your practice.