The Race in Generative AI: Understanding the players, the game, and how it will affect financial services

Is the current race in Generative AI an inflection point in the future of humanity or a suicide attempt? Depends on who you ask. What is becoming increasingly clear is that Generative AI can have multiple use cases in the financial industry, such as knowledge mining, summarization, compliance, and chatbots.

And before we really get into it we have to go back to the Second World War.

Alan Turing, the man who built a machine to break the German cipher machine Enigma, quickly followed this work with writings about human-like machine intelligence. He proposed the famous Turing test, which tests the ability of a machine to exhibit human-like intelligence. To succeed a machine would have to converse with a human evaluator and provide answers that are indistinguishable from actual humans. The reason why this is important to the current conversation around AI is because Turing’s hypothesis is the first popular instance we see of what we now call chatbots.

The kind of chatbots currently capturing the public’s imagination are powered by Generative AI. Use of AI in chatbots is by no means novel – in fact techniques within the umbrella term AI, like Natural Language Processing and Machine Learning, are often used to power chatbots today. The power of today’s AI celebrities like ChatGPT and Bard lies within how these programs understand prompts and generate responses.

Chatbots of yore relied on a rule-based or script-based mechanism, which means given a response, the program follows through a set of instructions and regurgitates pre-written responses. ChatGPT and Bard however can construct their own responses given the context and needs of the question. Similar to chatbots, DALL-E, an image generator developed by OpenAI, also relies on a conversational structure. For DALL-E the user inputs a text prompt -- “Sitting in front of a computer, a journalist is pulling her hair out, trying to understand AI, in the style of a comic book” -- and lo and behold, you get this:

Yes, DALL-E made that image. No, no part of this article was written by ChatGPT. But offerings by OpenAI (ChatGPT and DALL-E) and Google (Bard) are not the only ones making waves in the AI landscape.

Here’s a who’s who of Generative AI.

The Generative AI Landscape

Open AI - Chat GPT and DALL-E

ChatGPT was optimized using “Reinforcement-Learned Human Feedback” (RLHF), which was performed by workers in call centers who were asked to rate a huge amount of GPT responses. This kind of feedback allowed the system to learn what makes a good or bad response. 

On the other hand, DALL-E is a text to image generation software. It takes in a text prompt and throws out a collection of pictures it made itself. The data for the models comes from already existing scraped data sets. 

The payments for access to Open AI’s models go through Stripe. In fact the company also uses ChatGPT in its own operations to offer GPT-powered Stripe Docs. This feature allows developers to pose questions in “human-speak”, not code, and the software answers by summarizing the relevant documentation it has and extracting information from it. 

Google - Bard

Much like Open AI’s RLHF based ChatGPT,  Bard is optimized using the same system and is a text generator. The underlying training model however is based on the company’s own proprietary software called LambDa. 

In fact Google announced this month that it will allow developers to use the company’s generative AI model to build out listings on Play store. This announcement comes at the heels of Google lowering walls around Bard and removing the waitlist around the tool, launching it in 180 countries. 

Microsoft - Bing and Azure

Microsoft is a major investor in OpenAI and this partnership has allowed the company to utilize OpenAI’s resources in its own products, like the search engine Bing -- giving direct competition to the search leader Google. In March this year Microsoft announced the Bing Image Creator using DALL-E. This feature will be integrated in its “new” Bing, which moves away from the search engine model to one that is conversational.

Microsoft’s involvement with Open AI means that its tools can be embedded in the company’s cloud computing service Azure. Senior Director of Microsoft’s Global Partner Solutions Thomas Mathew adds that it is important not to turn to these services for use cases that might impact people’s opportunities or tasks which require updated information and accurate outputs.

OpenAI’s services still have some wide applications for the financial industry. 

These use cases include enhanced chatbots and analysis of contact center conversations that can be done by leveraging Azure cognitive services in tandem with OpenAI’s models. Similarly, the services can also be used to analyze legal and financial documents to extract information and summarize it. Moreover, these services may also help with translation of banking documents and customer interactions. 

Amazon - Bedrock

As is characteristic to Amazon, the company is more about providing building blocks for products rather than user-facing solutions. In April this year, the company announced its first generative Large Language Model (like ChatGPT and Bard) but one that can be used by businesses for their own use cases, called Bedrock. Amazon’s customers can fine tune the model according to their own needs and use cases and the company's data is not fed into training Bedrock itself. 

And they already have partners on board. Amazon said “Bedrock will be a massive step forward in democratizing FMs, and our partners like Accenture, Deloitte, Infosys, and Slalom are building practices to help enterprises go faster with Generative AI”. 

While we are still waiting on banks to announce integration with these Generative AI models, it is helpful to remember that banks like Truist already use Amazon Lex, a conversational AI tool for their chatbots. It is likely that as these products mature, announcements regarding the use of the enhanced models will follow.

Alibaba - Tongyi Qianwen

The AI race is unfolding across the globe, and internet platform company Alibaba unveiled its own Large Language Model called  “Tongyi Qianwen” which can be roughly translated to “truth from a thousand questions”. The company plans on embedding the solution across its entire ecosystem. Its model can respond to questions in both Chinese and English. The company’s enterprise communication platform DingTalk and intelligent voice assistant Tmall Genie are going to be launching the model first. One of the early enrollments into Tongyi Qianwen’s program is CICC Wealth Management, though it is unclear exactly how the model is currently being utilized at the company. 

AutoGPT

AutoGPT relies on OpenAI’s latest AI models to “autonomously” perform tasks. The program uses the AI models to understand the prompt and then interacts with apps and services online like web browsers and word processors to realize the goal set by the user. 

Unlike its competitors above, AutoGPT is freely available on GitHub. And there is speculation that due to its adeptness at handling task automation and management, the program might be used to autonomously trade crypto. 

Tearsheet Take: Fools rush in where angels fear to tread

While talk of Generative AI is powering the zeitgeist, it is unlikely that we will see FIs deploying these models in every part of their business. While the progress made by these technologies are novel, they are also new. Too new. Financial institutions are generally risk averse and have very low tolerance for error. On the other hand, all of the listed software above is prone to inaccuracy.

This doesn't mean that widespread use of these models wont materialize in the financial industry. In fact as we saw above there are already some implementations, albeit limited, that are  underway. For example the most obvious one is the use of AI in chatbots deployed by banks like ERICA by BofA or Truist Assist by Truist. FIs already use some form process automation similar to AutoGPT in back office tasks. Use of Generative AI is powering investing too, as investing service Public announcing Alpha yesterday, a program that helps retail investors cut down on time invested in market research through a conversational interface.

Excitement about ground breaking AI is not new. In fact AI has become a nebulous word. It loses meaning everytime it is used. “As soon as we solve a problem, instead of looking at the solution as AI, we come to view it as just another computer system,” says Martha Pollack, professor at the AI Laboratory of the University of Michigan in an interview with Wired News. 

It is possible that when the dust settles on these technologies, our general understanding of what AI is capable of will have changed. It is also possible that this new understanding will slot into everything we already know about computational improvement: it grows and gets better, it is still a tool, a software and the AI apocalypse is still far ahead and avoidable. 

It is also possible that we are the fools.