Barefoot Innovation Podcast

For today’s episode, our guest is none other than ChatGPT. We thought we might be its first podcast appearance, but we’ve found a couple of others already out there. They were not, however, about financial innovation and financial regulation, which are subjects about which, as it turns out, ChatGPT has very considerable expertise.

Late last year I sat at dinner with a very senior U.S. regulator, at an event where we were both speakers. ChatGPT had just been unveiled, and this top official said, “This is going to change everything about how we do bank examinations.” And then he pulled out his phone and started showing me bank supervision materials that had been written by AI.

The world is changing.

I’m a technophile, but I share the concerns that are being raised about generative AI. There was Yuval Harari’s chilling piece arguing that AI may come to dominate — to actually create — the cultures in which human beings live. We’ve seen Italy temporarily ban ChatGPT. We’ve seen the recent letter signed by over a thousand tech leaders, including Elon Musk, urging a six-month pause in development. We’ve seen the terrifying reports of the early chat’s bad behavior — including its creepy declaration of love for New York Times reporter Kevin Roose, urging him to leave his wife. We know this technology is likely to render a whole lot of jobs obsolete, including many jobs of knowledge workers who were never touched by earlier waves of automation. We’ve heard Bill Gates say it will change everything — including many things for the better. It can easily and instantly write both words and code. Some of its AI brethren can make visual art, sometimes with wicked abilities to mimic the work of prominent artists, dead and alive. We know it is likely to disrupt nearly every sphere of life — not just writing news stories and college entrance essays, but realms ranging from protection of intellectual property, to learned professions like practicing law, to international business models based on outsourcing — the list is endless. 

We also hear the warning that its arrival could herald the accelerating approach of the fabled “Singularity,” predicted long ago as the point at which, in Ray Kurzweill’s version, the intelligence of humans and machines will merge. And we know that this is part of an accelerating convergence of parallel technologies, which includes the huge impact likely coming from quantum computing’s potential to destroy encryption, wreaking havoc on personal privacy and on the security of information throughout business and government. A lot of scary stuff is coming.

And maybe most importantly, it is coming FAST. Way, way too fast for us to be ready for it. AI was invented 73 years ago and has enabled a lot of gradual change. However, the original science fiction predictions about it didn’t come true. Why not? Mainly because for decades, it had only limited amounts of information available to analyze. The digital age has changed all that — some analysts say we are now doubling data every 12 hours, and most of it is accessible online. That makes it possible to create these LLMs – the large language models that GPT is ingesting. Now, suddenly, the change is fast.

In fact, it’s head-spinning. One analyst said that ChatGPT is the AI equivalent of a Model T, and that the GPT 4 version that came out just a few months later is like a 1980’s era Oldsmobile. And the next version will be like a Tesla. In a matter of just weeks, we’ve also seen the skyrocketing arrival of Auto-GPT, which builds open source projects on top of GPT-4. In just its first two weeks, Auto-GPT instantly dominated Github, creating incredible things with only minimal human prompting and guidance. It’s entirely possible that sometime this year, I could do a podcast using a deep fake of my own voice, which my little team at AIR could create on our own — no tech skills required.

We’ve talked for years on this show about the dangerous nature of exponential change, how it can look deceptively gradual for a long time, before suddenly accelerating too fast for people to adjust. This is now happening, with a vengeance.

Auto-GPT has also sparked warnings that it may be showing glimmers of an Artificial General Intelligence (AGI) – the hitherto nonexistent ability of machines to understand and learn everything that humans can. The OpenAI organization — the developer of GPT — says it is not yet working on GPT-5, but it’s likely that the next version could bring gains that are orders of magnitude more powerful than what we have seen so far. 

These leaps are taking not a century, not decades, not years, but mere months and even weeks. Nothing in our legal, political and cultural infrastructure prepares us for profound change emerging at this velocity — hence the calls to slow it down, which in turn raise complex questions about how, and even whether, deceleration is possible.

So, the downside risk is, just, terrifying.

At the same time, there is also clearly upside opportunity to put this incredible technology to good use. We at AIR are thinking about doing a TechSprint on how it might be able to help financial regulators, or even more basically, financial consumers. What if a busy single mother, or a young service member, or an elderly couple, or anyone at all, could have their own personal, affordable, independent, reliable AI to help with their financial lives? What if the AI could search all the financial offers available and recommend the best option for its individual user? What if it could weed out offers with hidden adverse terms, or from providers that misuse personal data? What if it could automatically budget, and set aside savings, and invest in ways that matched the consumer’s needs and risk profile? What if it could explain complex financial concepts, through dialogue and answering questions, at the very moment the consumer needs to understand them in order to make a good decision?

Such a system would need massive safeguards built into it, and today, we don’t have good ways to create them and make them reliable. But still – if consumers are going to be attacked or manipulated by AI from bad actors, we can ask — might their best protector be another AI, an entity smart enough to see through the attacks? It’s worth thinking about.

So, we decided to have a chat with the world’s most famous chat.

Here is how we put today’s show together:

• Our AIR marketing lead, Joe Schember, worked with me to design it. The version I’m talking with is ChatGPT-3.5 – the free version. We realize that GPT-4 is vastly smarter. For example, our guest failed the bar examination. Months later, GPT-4 passed with flying colors. People claim that GPT-4 could be hired by most tech companies as a fully qualified engineer. So, again, our guest’s answers underperform what its successor already does.
• Joe and I crafted the questions and are sharing the Chat’s answers with you exactly as we received them. The only changes we made is that we have sometimes deleted a final paragraph. If you’ve played with ChatGPT, you know it likes to recap what it has told you in a summing up paragraph, which sometimes adds value but sometimes doesn't. Other than those recap deletions, we did not remove or reword anything, and we did not correct anything. We also didn’t get these answers through multiple queries asking it to refine, clarify or expand on its answer. These are the straight answers it gave to the questions, within one or two seconds of us typing in the query.
• Next, Joe fed Chat’s written answers into NaturalReaders.com, to make them oral so you can hear them on a podcast show. ChatGPT produces text, not voice. It can’t talk to us, or sing to us. (Why do I think that is likely to change soon?)
• Note that, in contrast to GPT-4, our ChatGPT 3.5 is not up to date. It has only been trained on data through September 2021. As a result, you’ll notice that when I ask it about regulating crypto and digital assets, it gives outdated examples. It doesn't know yet about crypto winter or FTX, but that is a temporary limitation.
• We didn’t want to give Chat a gender (and we want to keep reminding ourselves that this is not a person), so we switched up the voices, midstream.
• Needless to say, we asked Chat to write its own biography — you can read that in the section below. And we asked the AI photo rendering tool DALL-E to create the profile photo we’re using.
• We asked Chat some questions we didn’t use, but only to save space – not because its answers were bad. One was an open ended request to describe how finance is regulated. Not surprisingly, even though we’d asked for brief responses, it gave us a long and knowledgeable list of financial activities and how the government oversees them.

I think the answers are amazing. I can see places to add nuance or elaboration, but for the most part, it’s not only accurate, but it also grasps and ably analyzes very complex and often subtle issues. We didn't include it here, but the first question I ever asked ChatGPT, when I was seeing if I could trick it back in December, was a purposefully arcane one about a specific U.S. law. I wrote, “Why is UDAAP sometimes spelled with one A, and sometimes with two?” Sure enough, it knew the reason. Honestly, I know people who work with UDAAP issues all the time but probably don’t know the answer to that question.

As you listen, keep reminding yourself that our guest is the “un-smart” chat. A much, much smarter one is already here if you pay for it, and vastly smarter ones are on the way.

We asked a couple playful questions. I have to say, it answered one by composing what has to be one of the worst poems ever written. I realize that, being a poem about financial regulation, it doesn’t have any competitors, but, still….

I want to hear what you think of what Chat has to say and how you think we should address its arrival. Reach out to me on Twitter @JoAnnBarefoot or on LinkedIn, and let’s think about it together.

More on ChatGPT

(Note:  We asked our guest to write its own biography for the show.)

ChatGPT is a large language model developed by OpenAI, based on the GPT-3.5 architecture. Trained on a massive dataset of human language, ChatGPT is capable of understanding and generating natural language text in a variety of styles and genres. With its advanced language processing abilities, ChatGPT has been used in a wide range of applications, from chatbots and customer service systems to content creation and language translation.

As an AI language model, ChatGPT is at the forefront of the latest developments in natural language processing, and is constantly learning and improving through its interactions with humans. While ChatGPT is not a sentient being, it is designed to simulate human-like communication, and has been used in a variety of settings to enhance human-machine interactions.

Chat GPT 3.5 has been trained on data only up to September 2021, so it doesn’t know about more recent events.

More for our Listeners

So  there we are. Again, please find me on LinkedIn and on Twitter @JoAnnBarefoot and tell me what you think. 

Back here on earth, watch for our upcoming episode with Jesse Van Tol, President and CEO of the National Community Reinvestment Coalition. We also will be talking with Jessica Rusu, the head of innovation at the UK Financial Conduct Authority. We have a second episode in the works with Delicia Reynolds Hand on her amazing initiative at Consumer Reports, which is now producing insights. And we will have Lisa Rice from the National Fair Housing Alliance, among others.

This month I’ll be speaking in New York City at Fintech Nexus (formerly known as LendIt), and my AIR colleague Nick Cook will be speaking virtually at RegTech Africa in Lagos, Nigeria. In June, Nick will be moderating a roundtable called “Regulatory Dimensions of Finance in the Metaverse” at University of East London’s Centre of FinTech event in London, and will also be at the Salzburg Global Finance Forum. In June, I’ll be on the opening panel at the American Bankers Association compliance and risk conference in San Antonio — we have some very special programming in the works, for that. I’ll also be at the Point Zero conference, sponsored by Elevandi and the government of Switzerland, in Zurich.

Our AIR team has four or five TechSprints in the works, and the July one with the U.S. Federal Housing Finance Agency (FHFA) has opened for applications.

If you enjoy Barefoot Innovation, it’s important to take a moment to leave us a five-star rating on your favorite podcast platform to expand our audience. And with this episode, please share it with your personal and social networks and with people in your organization who are thinking about new technology — or people who need to be!