Charles is a partner at a boutique investment fund. His grey hair and slight stoop reflect years of market knowledge and he blinks from behind round, wire spectacles. Charles is not part of the artificial intelligence revolution.

—–

“AI will draft your reports and save you time.”

“We have a particular way of writing that we’ve honed over many years.”

“The AI will use it. It’s as if you wrote the article and you will proof the final draft.”

—–

“AI is hooked into your data sources to analyze historic patterns and predict trends.”

“I’m amazed no one reads the FT anymore. It’s the source of great insight.”

“You select your data sources whether internal or external, financial or alternative. The AI performs the analysis exactly as you would.”

—–

“Imagine being able to retrieve information instantly by asking a question.”

“I keep paper copies of research for reference piled on my desk.”

“Say there was letter published in the FT last quarter that had a fascinating statistic. You can retrieve it instantly, rather than spending hours hunting for it, or remembering the numbers incorrectly.”

—–

Charles worries about the unknown. His clients are ageing with him and he is not going to risk their retirement savings. Yet he knows his role is to secure those assets for another generation. In a similar way, his firm will pass into younger hands. Peter is bright and enthusiastic and has been mentioning this GPT thing for a while.

An Introduction to GPT

A Generative Pre-Trained Transformer (GPT) is a language model used to produce human-like text responses. It is autoregressive, meaning it is more random than a moving average. Autoregression is common in economics, behavioural analysis and nature and has been around a long time.

GPT uses deep learning. It’s deep because of its multiples layers. The learning is called representation, which in human terms means learning by observation rather than pre-programming. This is much more effective than the engineering of features.

Imagine designing a driverless car by writing down what it will do in every scenario. You form a committee, brainstorm and plan every alternative. Then, when your car is driving, there is a black swan event such as an ice storm in Texas. You didn’t write code to tell the car what to do now.

Driverless cars are built by billions of observations of human driving. The machine learns from us. It sees good habits and bad, but as accidents are rare relative to the number of journeys, the better practises are embedded. This is easier and develops faster in remote areas rather than the chaos of cities.

Sperry Corporation developed the first autopilot for aircraft in 1912. Even today pilots may take control of the plane when necessary. Humans are in control of the technology, just as they are with AI and GPT. Provided you design the system correctly.

An Introduction to BERT

Google has a language model called BERT – Bidirectional Encoder Representations from Transformers. It allows anyone to train a question answering system.

Bidirectional refers to considering context in a left and right direction. A transformer analyzes words in relation to the rest of sentence, rather than one at a time and in order. This allows it to discern the meaning of questions even when there are spelling mistakes and an odd word order. We all get lazy typing search queries.

An encoder is a device that transforms information from one format to another. It senses motion and turns it into an electrical signal. Encoders are used in bottling factories to fill jars without missing the container or overfilling it.

GPT models do not have encoders. This allows them to generate responses quickly, making them ideal for text. But asset management is about more than reading and writing reports. Hence to derive efficiencies from AI you must know the appropriate technology for a task.

Picks and Shovels

Charles has seen many trends comes and go. He remembers Y2K as a hoax, the disaster of making mortgage loans available to anyone and saw the 2010 flash crash as vindication of his methods. Caution has served him well.

His investment strategy is picks and shovels. He has no view on which manufacturer will dominate electric vehicles, but knows they will use technologies purchased from a handful of companies. He understands who supplies those companies.

Charles believes that when AI is necessary the applications he uses today will adopt it. His firm doesn’t use these technologies much because keeping client data secure is essential. The old, manual ways still work.

Charles is concerned about Peter. His interest in AI is growing. Charles is unsure what he would do without his assistant, as it takes a long time to train colleagues and even more to trust them. The business is at risk if people leave because they are not developing modern skills.

On balance, the security concerns about client data outweigh the commercial risk of being left behind. If only there was a way to use AI securely. Charles doesn’t know if there is.

Up Next: AI and Investing for the Security Conscious