• Which companies will achieve measurable value from the implementation of late-generation AI (as opposed to deterministic ML)?

• Which capabilities are best delivered through a heavily productized application layer vs. interacting with a foundation model itself?

• Will startups around AI today take a similar form to the SaaS companies of the last 10-15 years, or will they choose a different path?

• How do individuals want to interact with AI, and what new form factors, interfaces, and communication modes will emerge? 

Today’s newsletter covers several trends, stories, and technologies that address the above questions. In addition to the commentary around each specific link, a couple of other meta-thoughts come to mind. 

For the past 2 years, text has been the lingua franca of AI interaction. While the ability to converse with generative technologies through natural language is alluring and familiar, I wonder if people and businesses will eventually tire of it. After all, there are plenty of statements, instructions, queries, and actions that might better be transmitted through less verbose means of communication. Is the concept of a chat prompt akin to the command line of the AI computing age? If so, what form factors will take the place of the graphical user interface in this analogy?

There’s a lot of talk in the startup investment world about the end of SaaS and the rise of “service as software”. Many observers contend that AI technologies will democratize software development, leading to more custom solutions being built within organizations rather than bought from vendors. However, it’s possible to look at this a different way. The companies that will utilize AI to build internal software are those that already have a competency in technical product development, but such organizations are in the minority. Building software is hard, but so is buying it. The number of people and organizations that could benefit from a software solution is several orders of magnitude larger than the number who could competently procure, finance, integrate, and deploy that solution. New types of workflows enabled by AI have the ability to extend automation capabilities to these previously unreachable organizations, massively growing the addressable market for productive technological innovations. The ability to innovate not only on the technology itself but on the form factor and commercial model for delivery may in fact be the best argument for the future of services powered by software.

As always, please share your thoughts, ideas, comments, and any interesting content. If you like this newsletter, please consider sharing it with your friends and colleagues. Happy reading!

-David

P.S. If you’ll be at Money2020 and would like to say hi, please drop me a line!

Recent News & Commentary

Which companies will achieve measurable value from the implementation of late-generation AI?

The AI Hype: $600B question or $4.6T+ opportunity? - Foundation Capital

In this post, Ashu Garg at Foundation capital discusses the balance between hype and opportunity in AI. He begins by noting that much of the initial unqualified enthusiasm around AI has cooled off, with frequent content from major publications and research organizations promoting the idea of a bubble destined to pop. This negative instinct displayed by many is supported by the disappointing performance of public software stocks over the past several years. This piece does a good job of faithfully representing the bear cases - and yes, some people are just lighting money on fire - but ultimately lands in a place of cautious optimism. Ashu and the Foundation team believe that AI is catalyzing a fundamental shift in the role of software. The combination of massive cost efficiencies along with the ability to fundamentally transform workflows leads him to believe that we may actually be underestimating the potential of AI. 

Which capabilities are best delivered through a heavily productized application layer vs. interacting with a foundation model itself?

Former Brex COO who now heads unicorn fintech Figure says GPT is already upending the mortgage industry - Techcrunch

Even in the absence of definitive statistics, it’s probably safe to say that the majority of applicants for a mortgage or HELOC have likely complained about the process being too manual and overly time consuming. Much of this delay is due to the need for human underwriters and loan processors to manually compare information across multiple data sources. Figure, a fintech company specializing in HELOCs, has announced their deployment of an LLM-based (GPT-4 in this case) tool designed to automate such tasks, reportedly taking time and cost out of the process. The techcrunch piece also briefly mentions some of the systems Figure has put in place to run such a system, including a strict privacy agreement with OpenAI along with various model testing and evaluation frameworks. This seems to be a pretty straightforward and obvious use case for AI tools, and it will be interesting to see if Figure releases any customer-centric stats to demonstrate the impact.

Intuit Pioneers Done-for-You Future for Consumers and Businesses with Agentic AI - Intuit

Intuit is betting heavily on its ability to use generative AI and its massive repository of financial data to build agentic automation across its customers’ financial lives. This ambitious and wide-ranging press release promises upcoming capabilities, including cash flow automation for small businesses as well as an advanced, conversation-driven analytical engine. The release also references GenOS, apparently the name of their data architecture that makes such things possible, as well as other internal tools it has built to accelerate development. It’s always hard to tell from a press release exactly what will be widely available, and when, but Intuit does seem incredibly well positioned to build this version of ‘self driving money’ for its customers, and it will be interesting to watch for adoption.

Will startups around AI today take a similar form to the SaaS companies of the last 10-15 years, or will they choose a different path?

Every White-Collar Role Will Have An AI Copilot. Then An AI Agent. - a16z

It seems straightforwardly clear that intelligence is helpful in most ‘information age’ jobs. In this piece on the a16z blog, Angela Strange and James daCosta explain how every white collar job will have an AI ‘copilot’, likely followed by an ‘AI agent’. I doubt many informed people would argue with that premise, but many assume that the incumbents powering centralized systems of record (e.g. Intuit, referenced above) hold an insurmountable advantage, making it difficult for new startups to emerge. The authors here list 3 approaches that upstart competitors might use to overcome the advantages of incumbency. These include gathering data upstream of the system of record, automating workflows that happen outside that system, and unifying data across systems. Once new technologies are able to solve problems in these areas, they gain a foothold that earns them the right to expand their reach across the enterprise, even ultimately creating newer, more comprehensive AI-native systems of record. This logical framing makes a lot of sense and is already being pursued by startups in various verticals. The one thing I’d add is that it’s not a given that the application of AI copilots and agents need to map directly to the present organization of human labor. Perhaps it would be more appropriate, eventually, to say that every business problem will be addressable with a combination of AI agents and humans with copilots, but likely in workflows that don’t neatly resemble those of today.

The Agent Development Life Cycle - Sierra

Over the past couple decades, most software companies have adopted some version of what has become a fairly standard ‘software development lifecycle’ (SDLC). Of course, every technology organization does some things differently. But, there is a well-paved path, and you’ll notice a lot of things in common across the product and engineering teams of many companies. There’s value here, as teams can build upon plenty of hard-learned lessons in software building. However, AI-native applications, particularly those that make use of ‘agents’, may have different opportunities, risks, costs, and constraints than those of traditional deterministic software. In this piece, the founders of Sierra, a platform for building branded, intelligent agents for customer service, argue that the development of agentic, AI-intensive systems requires a new type of software development lifecycle. They lay out the key components of the process on which they’ve presumably converged, including Development, Testing, Release, QA, and Alignment, discussing the AI-driven differences present in each of these steps. This post is full of valuable advice and makes useful reading for practitioners in the space. That said, the computing model, cost structure, and capability set of AI technologies is so different from that of deterministic software, and so early, that we’re likely quite far from settling on the optimal lifecycle. My hope is that a new “ADLC” doesn’t become quite as standardized as things have become in the traditional software world. Perhaps AI might catalyze a new heterogeneity of approaches, leading to a creative explosion of technologies and addressible use cases.

How do individuals want to interact with AI and what new form factors, interfaces, and communication modes will emerge?

The rise of AI Agents in Financial Services - Fintech Brain Food

In this long and well-thought-out essay, Simon Taylor lays out the state of AI in financial services, including how to properly separate hype from substance, why the concept of ‘agents’ is likely to be very meaningful, and why fintech is a ripe market for the application of AI. Parts of the discourse around AI in financial services are misleading or unhelpful, he points out, including concerns around “responsible AI”, “hallucination” and the accusation that it’s just a “trick”. Even when the critics are well-meaning, these are all concerns that can be managed, and those who are working to make a real impact are building with them in mind, unlike the “aging rock stars” he aptly references. Simon is particularly bullish on the potential of agents as a winning form factor for this particular vertical. He constructs a pyramid of ascending value with agents at the top, also arguing that most practitioners should skip the “AI as copilot” phase, contending that a copilot is an awkward interface and sits outside true workflows. Finally, he recommends that people in fintech experiment aggressively with these new technologies. In Simon’s words: “Let go and dance with the AI”. Well said! 

Path to high-quality LLM-based Dasher support automation - Doordash

In a complex system like Doordash, with thousands of delivery people delivering millions of orders, there is a lot that can go wrong! I was intrigued to learn that Doordash has a large ‘Dasher’ support team, which has long been a combination of automated and human interaction. However, the existing automations were highly deterministic in nature, making it impossible to solve new problems that didn’t look similar to existing ones already present in their documentation. This somewhat technical blog post outlines Doordash’s development of a RAG-based system to use generative AI technologies to automate the process of Dasher support, along with the systems put in place to ensure accuracy, timeliness, and cost. The piece includes flowcharts of the component systems, including their ‘response guardrail’ and ‘LLM judge’. The architecture is impressive, as is the thoughtfulness and detail apparently present in the system. It’s an example of the value of truly understanding a problem and iteratively building a system to solve it, in a measurable way, with previously unavailable technologies, rather than simply “tossing” the problem to a generic LLM. While this application is in a different industry, there are a lot of valuable lessons for financial services firms. The architecture of a system that provides a high quality user experience while saving time and money, grounded in factual content and with LLM and human failsafes, is something we’ll likely see across domains.

Better.com launches Betsy™, the First Voice-Based AI Loan Assistant for the US Mortgage Industry - Better

The mortgage application process in the United States tends to involve a lot of back and forth interaction with the applicant, often by phone. These phone calls involve requests for documentation, questions about terms, and coordination of the multiple steps involved in bringing a loan from application to closing. Better Mortgage has now announced its voice-based AI loan assistant, Betsy, along with some impressive demos. It’s smart of Better to begin with the ‘non-licensed’ areas of the mortgage process, where a well-calibrated AI assistant can probably outperform some meaningful percentage of human customer service agents while also staying away from interactions that require human oversight. Apparently, the voice bot is integrated into Better’s proprietary loan origination system to ensure accurate data and help the interactions fit smoothly into a customer’s workflow. I’d enjoy reading an explanation of the technical architecture, similar to the Doordash piece covered above. It will certainly be exciting to see how this works for Better and its clients, and I’m sure we’ll see a lot more moves like this one. 

AI agents have brains, but where are their wallets? - Amias Gerety & Prateek Joshi

As we build more workflows around the capabilities of AI agents, the actions we ask them to take on our behalf will expand, and that will surely include making purchases or otherwise moving money. Amias Gerety at QED and Prateek Joshi at Moxie argue that the existing payments system is designed for humans and may therefore have difficulty in adapting to a world with a meaningful share of payments being made by AI agents. For example, there are plenty of risk and fraud prevention mechanisms in payment systems that aim to differentiate between humans and bots. Now these systems would have to assume the existence of good bots, somehow authenticating them and distinguishing them from malicious bots. This piece lays out several points of development around how payment systems might need to evolve to support an agent-heavy future. The question of whether existing payment systems will adapt to suit the needs of agents or whether it is possible to create new, agent-native dominant companies in payments and associated fraud prevention feels similar to the debate around humanoid vs. non-humanoid robots. Those betting on humanoid robots assert that since the built environment is centered on the human form, robots taking on that form will fit in and be more generally productive. Those on the other side contend that without the constraint of conforming to a human form, specialized robots can excel at specific use cases. In the case of payments - a massive market and potentially huge target for agent-driven transformation - there will surely be plenty of ambitious new entrants looking to make their mark.

Today’s newsletter covers a variety of topics sure to be of interest to builders, operators, and investors in artificial intelligence and financial services. These include a window into large-scale adoption by the largest institutions, technical innovations to build higher-performing knowledge systems, optimistic and pessimistic takes on the state of AI, and news around analytics, risk, and fraud use cases.

It can be difficult to distinguish between inputs and outcomes when reading accounts of AI adoption among the largest financial services institutions. It’s one thing to hire a ton of people, spend a lot of money, and issue press releases. It’s another to report on outcomes - products and services that create demonstrable value for firms and their customers. A report from Constellation Research provides much more detail than usual into J.P. Morgan’s data and AI strategy. nCino also released a “Banking Advisor” product that can be used to drive automation and efficiency among its many bank clients. Of course, building in AI is not as simple as throwing a bunch of data at an LLM and asking it some questions. Real-world applications in a highly-sensitive domain require well-tuned and thoughtfully-deployed models and high-performing end-to-end data pipelines. Predibase released an index of fine-tuned models and benchmarks to demonstrate the viability of task-specific fine-tuning, and Contextual.ai launched its concept of “RAG 2.0”. When I first learned about RAG a while back, my reaction was: “This is cool, but it also seems like kind of a hack. I wonder what will replace it.” I’m glad to see it as an active area of research and development.

At this stage in the AI hype cycle, there are a lot of takes. Even avoiding the unhinged (“AI will kill us all”) or the simplistic (“AI will solve every problem immediately”), it’s quite easy to find a range of opinions and predictions at varying levels of optimism and pessimism. Chris Mims wrote in the WSJ that AI is already losing steam. Ben Evans wrote a solid essay on the differences between general-purpose technologies and specifically useful products, wondering whether AI can actually be both. Retool released a survey and report on how companies (presumably including their many clients) are using AI today. Matt Harris of Bain shared a presentation on the fintech “hero’s journey”, and how things can only go up from where we are right now as an industry. 

So many of the practical applications of AI in financial services are, from a user perspective, behind the scenes, and there’s a lot happening here. Taktile and Ocrolus (where I work) announced a partnership to bring high-quality data to lenders via a modern decisioning platform. Sentilink shared thoughts on the reality of GenAI-based fraud risks. Coris introduced an “agent” to help automate risk workflows. I also recorded a webinar with 2 industry colleagues that ended up being a really fun and wide-ranging conversation on the potential for AI in risk management.

As always, please share your thoughts, ideas, comments, and any interesting content. If you like this newsletter, please consider sharing it with your friends and colleagues. Happy reading!

Recent News and Commentary

Windows into adoption in banking

JPMorgan Chase: Digital transformation, AI and data strategy sets up generative AI - Constellation Research

While size and incumbency may have been barriers to adoption in earlier waves of technology (e.g. mobile, open source, cloud), there’s a solid argument that the largest, most well-capitalized organizations are best positioned to benefit from the potential of modern AI, given the importance of large datasets, substantial capital investments, and talent. Of course, realizing the potential of AI in a massive organization isn’t easy; it requires a thoughtful data strategy and real dedication. This piece from Constellation Research is nearly a year old, but it provides some valuable insight into the data strategy of J.P. Morgan Chase, which does seem to be at the forefront of megabanks in the productive utilization of new technology. The report mentions how JPM is adopting a multi–public-cloud approach, even while modernizing its own data centers. The bank’s head of technology strategy stated that given their scale, having access to massive public cloud compute capacity is essential: "If you look at our size and scale, the only way to deploy at scale is to do it through platforms.” The report also provides some insight into the bank’s data mesh architecture, which in theory allows for the development of many data products across the bank while preserving control and lineage. Of course, what would be even more compelling is information on outcomes rather than inputs (i.e. not how much money they are spending or people they have working on it, but the magnitude of outcomes they have been able to achieve for their customers). Nevertheless, it is clear that JPM is making a massive bet on AI technologies, and it’s valuable to get a window into how this is being pursued.

nCino Deploying Banking Advisor, a Generative AI Solution to Drive New Efficiencies in Financial Institutions - nCino

Cloud-based banking software platform nCino announced the release of Banking Advisor, a co-pilot and set of automation tools that allows its bank clients to more easily make sense of client data, streamline workflows, and spend more time on uniquely human tasks such as managing client relationships. The release and accompanying demo video suggest a fairly broad set of capabilities, and it would be interesting to get more information on what technologies are used under the hood. However it works, these new tools appear to be quite useful for organizations engaged in relationship-based banking at scale, and nCino, with its large dataset and long client list, is well-positioned to develop and commercialize such an offering.

Technical innovations in knowledge systems

The Fine-tuning Index - Predibase

As builders and companies race to take advantage of the latest AI use cases, an important factor is the choice between commercial and open source models. While GPT-4o is nearly always an out-of-the-box leader on many benchmarks, the development of high-performance open source models has led many companies to consider whether they could use these as a lower-cost, higher-control option for more specific use cases. As with anything in engineering, there are a set of tradeoffs, the answers to which depend on the goal and definition of success. Predibase, a fine-tuning and inference platform for LLMs, has published a leaderboard of fine-tuned small and/or open-source models, comparing them to GPT-4 and GPT-3.5 Turbo on various benchmarks. Some of the open-source models show impressive results, particularly the Llama3 family, open-sourced by Meta. Of course, fine-tuning a model requires a large-enough repository of high-quality data, making such approaches viable for narrow, specialized tasks, but less likely for cases requiring reasoning or some form of generalized “thinking”. In financial services applications, there are certainly solid arguments to both approaches, depending on the criticality and generality of the use case. The most sophisticated and technically capable firms often build on high-quality, commercial foundation models to prove a use case, but many may shift to a hybrid approach over the long term.   

Introducing RAG 2.0 - Contextual.ai

Retrieval-augmented generation (i.e. “RAG”) is a widely used technique for extending the capabilities and accuracy of an LLM-based system using a specific corpus of knowledge. This is particularly attractive in industry-specific or company-specific use cases, where there is value in allowing an LLM to access a large proprietary knowledge base in a highly-orchestrated and purposeful way. In a relatively short period of time, RAG has become so broadly adopted that several companies have emerged to focus on the facilitation of building and deploying RAG systems. One such company, Contextual.ai, has introduced a system they refer to as “RAG 2.0”. To draw contrast, Contextual in this blog post describes typical RAG systems as “frozen”, meaning that each component (embedding model, retriever, LLM, etc.) is static and that the full system is stitched together in a brittle way. RAG 2.0, in contrast, integrates its components end-to-end specifically for the purpose. Based on the benchmarks they show in the post, this approach appears effective. This type of evolution has the potential for real impact in financial services, where industry players are eager to take advantage of AI but are also highly focused on building systems that are relevant to the task, grounded in real-world (and often non-public) data, and have a lower risk of hallucination. My take is that we’re incredibly early in the journey toward fully-integrated, high-accuracy knowledge platforms, and it’s good to see so many people working on novel approaches.

The State of AI - Takes ranging from pessimistic to optimistic

The AI Revolution Is Already Losing Steam - WSJ

In forecasting the future, it is far easier to predict the what than the when. Technologies that appear novel and truly capture peoples’ imagination tend to spur wild visions of rapid change. Then inevitably, certain things happen slower than expected, and many changes take a form far different from when was originally imagined. Gartner famously calls this the hype cycle. Bill Gates famously said: “Most people overestimate what they can do in one year and underestimate what they can do in ten years”. Carl Sagan famously said “It was easy to predict mass car ownership but hard to predict Walmart”. It’s just incredibly early. In this WSJ article, Chris Mims points out the ways that the AI revolution is “losing steam”, referencing slower-than-expected progress in many areas and the unpleasant surprises many businesses have had regarding the cost, speed, and value of read-world AI deployments. It’s a worthwhile read, though I have a different perspective. It’s not really that AI is losing steam, and in fact many of the technical advances that are not as visible to the lay public are happening at an astonishing pace. It’s just that big things do take time - particularly when they involve questions of broad importance to society - and the impacts are often non-obvious.

Building AI products - Ben Evans

It’s one thing to have a set of new and impressive technologies. It’s another thing to build these into useful products that gain serious adoption. Perhaps even harder is when the technology in question is a new probabilistic computing paradigm that often “gets things wrong” (i.e. ‘hallucinates’). These are the topics that Ben Evans discusses in this characteristically well-thought-out essay. One particularly compelling point is his comparison of AI to the invention of the electric motor, which enabled a ton of products that otherwise would not be possible. Of course, as he says, “Electric motors are a general-purpose technology, but you don’t buy a box of electric motors from Home Depot - you buy a drill, a washing machine and a blender.” He wonders if some of the excitement about AI is its potential to actually break that pattern, to be a general-purpose technology that is also, itself, specifically useful. A good read for those building products and companies on AI building blocks.

The State of AI H1 2024 - Retool

Retool - the internal business tool development platform - conducted a new version of its periodic “state of AI” survey and released a comprehensive report on the findings. The sample is fairly comprehensive, comprising around 750 people across multiple industries, roles, and levels of seniority. The report - which, by the way, is presented with an admirably unique and captivating visual design - reveals sentiment across many different areas of AI adoption. For example, there are data on levels of adoption, specific technologies, ROI, areas of usefulness, developer tools, and predictions of the future. It’s a valuable read but also comes with a handy  table of contents for anyone wishing to jump to particular areas.

Only Up From Here: 2024’s State of Fintech and the Hero’s Journey - Matt Harris @ BCV

Matt Harris and his colleagues at Bain Capital Ventures are fintech veterans who enjoy a comprehensive view of the financial services and technology ecosystems. It’s always valuable to read or hear what they have to say. In this presentation, given at a joint Fintech CEO Summit with NYCA Partners, Matt explained his thesis for why we are at a “bottom” in the fintech markets and painted an optimistic case for the future. Using the conceit of a “hero’s journey”, this presentation reviews some of the lessons from our collective path to the bottom, including the end of “middleware BaaS” and regulatory arbitrage. A core belief looking forward is that the traditional model of banking (i.e. net interest margin) is facing an existential crisis over the next several years, and that this will provide opportunity for resourceful and creative founders. It’s a thoughtful and forward-thinking analysis, and it’s worth the time to watch the video or read the transcript.

Analytics, risk, and fraud

Taktile and Ocrolus partner to unlock real-time underwriting for small business lenders

Lenders rely on high quality data, insightful analytics, and repeatable automation to make good credit decisions at scale. Yet, while there exist numerous data sources for understanding the financial health of people and business, many lenders lack the infrastructure to rapidly integrate, test, simulate, and deploy workflows using the entirety of available information. Ocrolus (where I work) is an AI-based document automation platform that provides highly-accurate data to lenders in small business, consumer, and mortgage lending (among other use cases), and Taktile is a modern and highly capable decision engine for lenders. I’m excited about this partnership given the clear compatibility of both companies’ technologies and the potential to streamline the integration and useful application of data in the underwriting processes of literally hundreds of lenders.

SentiLink’s Thoughts on GenAI Fraud

Anyone who’s explored the capabilities of state-of-the-art AI to generate realistic images, audio, video, and text has probably considered the risk of a greater incidence of fraud made possible by the technology. Of course, the reality is quite a bit more nuanced. This blog post from Naftali Harris at Sentilink, a leader in synthetic fraud prevention, considers the specific areas where genAI poses a greater or lesser risk and provides valuable insight on how to think about it. Interestingly, he does not believe genAI will result in an explosion of synthetic identity fraud, as the bottlenecks to the creation of credit-worthy synthetic identities aren’t easily alleviated by this technology. He does, however, mention several common fraud controls that could be severely threatened, including those involving video, voice, and behavior monitoring. While it’s reasonable to conclude that the solution to bad guys with AI is for the good guys to have better AI  - and this is true in many cases - he also points out some reasons why this is not entirely straightforward, and the longer-term is unclear. Anyone in fintech should read this post, as if you’re in the business of money, you’re in the business of fraud prevention whether you know it or not! 

Introducing Risk AI, your agent for risk management workflows - Coris.ai

In any risk decision process, there are always cases that are an easy yes or an easy no. Unfortunately, plenty of cases end up in a gray zone of uncertainty, where a quantitative prediction model or deterministic ruleset does not perform reliably. Decisions made in this zone often massively determine the fate of a lender, payment facilitator, or other provider. The solution in such situations has traditionally involved heaps of manual review, often with inconsistent decision quality, spotty results, and poor customer experience. Many hope that agents built on generative AI technology have the potential to enhance decision-making of this sort. Coris, a small-business-focused data intelligence platform, released RiskAI, its version of such an agent. This blog post explains the applications where such a tool is useful and includes details on its features, including the ability to slack with an AI agent that has access to merchant profile data and can be configured to assist in decision automation for particular use cases, specifically endeavoring to reduce false-positive errors.

What’s next for AI in risk management? Webinar recap and on-demand playback

All financial services businesses are fundamentally about risk management, and the advancement of AI technology provides optimism that we might be able to (metaphorically-speaking) catch more tuna while ensnaring fewer dolphins, all the while providing better customer experiences at a lower cost and passing on the savings to end users. I recently recorded a webinar with my friends Vinodh Poyyapakkam of Coris.ai and Ryan Hildebrand of Bankwell. It was a wide-ranging and intellectually-stimulating discussion of AI applications we’ve presently observed, the greatest potential problems that AI could solve, and what we may see over the next several years. You can watch the webinar in full or read a summary in this blog post from Coris.

While spring has sprung in the northeast United States, I’m leaving the beautiful flowering trees of Connecticut behind for a week and am writing today from Delhi, India, where I’m looking forward to spending time with my colleagues on this side of the world. Traveling provides valuable perspective on all things, including the differences in the need for, nature of, and access to financial services across the globe. 

Today’s newsletter covers a variety of topics, including the challenges and opportunities in finding AI use cases, several compelling applications of AI in financial services today, the opportunity for generative AI to facilitate the expansion of credit in areas lacking a mature and well-structured credit data system, and AI-related security considerations. 

As always, please share your thoughts, ideas, comments, and any interesting content. If you like this newsletter, please consider sharing it with your friends and colleagues. Happy reading!

Recent News & Commentary

AI is Like Water - NfX

The thing about any new technology is that - no matter how amazing - once widely available, it ceases to be a differentiator. A long time ago, factories that used electricity were revolutionary, and then the use of electricity became widespread and simply table-stakes. More recently, this happened in the consumer internet and with mobile apps. At first, having a website or an app was a big deal, and it quickly became an expectation. This shift will happen even faster with AI. As AI capabilities become widely available, mere integration of the technology itself does not make a valuable business. Rather, the success of a solution comes down to things like product design, customer experience, and brand. This excellent thought piece by Morgan Beller at NFX compares AI to bottled water - a product whose core technology is clearly a commodity but where novel brands continue to innovate and win. She asks: “If you took AI out of your pitch deck, is it still a good business?” Applying her thesis to applications in financial services, we are now in the era of 'adding AI' to existing use cases. While the technology certainly can provide benefits - particularly in speed, accuracy, or efficiency - it likely doesn't change the essence of the core financial product offered to customers. Once everyone 'adds AI' to products, it ceases to add distinction and simply becomes the norm. What products or experiences could not exist without AI? Build those! 

Looking for AI use-cases - Ben Evans

“This is amazing, but I don’t have that use-case.” This is the central question asked by the always thought-provoking Ben Evans in his latest essay on generative AI. He writes of the birth of the personal computer (PC), how it was celebrated for its general potential but truly reached adoption when Visicalc and other spreadsheets made this potential obvious and specific. Likewise, he considers if, while chatGPT, Claude, etc. feel super impressive, LLMs haven’t had their Visicalc moment. What Ben does exceptionally well is to put technological trends and developments into a broader historical context. His general assertion here is that just like with past innovations, products that take advantage of them need to be researched and built, one-by-one. 

Gauging Generative AI x Financial Services - a16z

While financial services is a highly-regulated industry, this has not caused it to delay adoption of generative AI. This post by Angela Strange and the a16z team mentions several real-world examples of companies solving practical, tangible use cases today. Interestingly, while employees at many institutions avoid mentioning AI in the earshot of a compliance person, this piece mentions multiple examples of how the technology can assist with compliance. For example, software available today can help write SARs (Sardine), interpret long regulations (Norm), or perform automated screening of suspicious customer activity (Greenlite). AI is also being used to streamline complex workflows (Vesta, Casca) and has the potential to automate various customer actions. Solid post and original tweetstorm as well. 

Opportunities for Generative AI in Financial Services - Visa & This Week in Fintech

Visa and This Week in Fintech collaborated to write a thoughtful and comprehensive white paper on the current state and future possibilities for generative AI in financial services. The paper covers a wide range of applications, including those in fintech, banking, customer support, investing, commerce, and payments. What I particularly appreciate are the numerous real-world examples, including specific companies and links. It also discusses topics of potential uncertainty, including regulation, international adoption, and implications for the underbanked. Notable for its breadth as well as depth, this report is definnitely worth a read.

JPMorgan Says Its AI Cash Flow Software Cut Human Work By Almost 90% - Entrepreneur.com

JP Morgan has apparently given corporate clients access to a new AI-based cash flow forecasting tool. Boasting 2,500 current users and claiming a 90% savings in terms of manual work, the tool appears to be a success in streamlining the work of cash flow management in a complex enterprise. Bank of America and RBC have also launched similar capabilities. While the article here references the potential that banks could eventually charge for such a product, it’s short on details. For example, how much data for a particular corporation does the tool need to make accurate forecasts? How much human review is required to get from machine output to acceptable accuracy? It’s also not clear whether banks are best-suited to deliver such a capability or whether it will eventually be offered by accounting and ERP systems, which would logically have more real-time information on the operations and financials of a business.

Towards Generative Credit - Matthew Flannery, Branch

In the United States, we have relatively well-structured and consistently available credit data that does not exist everywhere. In other parts of the world, including India and Africa, making credit decisions can be substantially more difficult and requires the acquisition and analysis of a more diverse set of data points than would be practical (or compliant for that matter) to use in the U.S. In this Medium post, Matt Flannery of Branch discusses the role of generative AI in improving credit assessment for microfinance in geographies with less mature credit systems. LLMs can already be quite useful in labeling unstructured data, making it viable for use in statistical models to predict risk. This post also contemplates future developments, including automatic feature engineering, predictive synthetic data, and ‘AI as credit officer’. We certainly don’t have a good compliance framework for many of these concepts in the U.S. However, in markets without a well-established and reliable source of credit history, there could be great benefit in using non-traditional approaches enabled by AI to expand credit access.  

Managing Artificial Intelligence-Specific Cybersecurity Risks in the Financial Sector - U.S. Dept of Treasury

The U.S. Treasury released a report on cybersecurity risks around AI in financial services. It’s a long read (52 dense pages), but it’s worthwhile to note a few interesting points. The report mentions a “fraud data divide”, where larger institutions have the data to train models, and smaller ones may not. It also mentions a skills gap, where many institutions lack the personnel and sophistication to tackle a rapidly-expanding set of AI-related security challenges. Most interesting perhaps is the concept of a “nutrition label”, stating what data was used to train models offered by vendors to financial institutions. This is relevant particularly for institutions relying primarily on vendor-contributed technology and wary of the privacy and liability implications of models trained on data from unknown sources.

What’s next for AI in Risk Management

On May 16th, I’ll be joining my friends Vinodh Poyyapakkam, CEO of coris.ai, and Ryan Hildebrand, CIO of Bankwell for a webinar to discuss what’s next for AI in risk management. You can expect a wide-ranging and enthusiastic discussion of how AI is being used by various partciipants in financial services and the many possibilities that exist for transformation. If you’re interested, please register. I’m looking forward to an exciting conversation!

Email header image: on the beach in Villers-sur-Mer, Normandy, France last month

Reasons for enthusiasm include the hope that advances in the application of AI will provide a boost to what has been a tough fintech market over the past 18 months. Not too long ago, there were plenty of fintech companies talking about AI or claiming to use it, but in superficial, pie-in-the-sky ways with few real proof points. Today, while we’re still extremely early, there has been enough practical, real-world usage to facilitate better-informed conversations and more realistic assessment of opportunities and challenges.

There are, of course, many open questions! Among these: 

• Will AI be dominated by a few giant foundation models that acquire such incredible general intelligence that they can be applied to nearly every problem but require immense computing resources to train and run, or will there be numerous small models, tailored to narrower areas of intelligence and specific use cases? Or both? Recent advances help bolster the small-model case, particularly Mistral’s impressive benchmarks with its 7B and 8x7B models and Predibase’s incredibly cool "LoRA Land" library of fine-tuned small models that outperform GPT-4 on particular tasks and can be run on a single GPU.

• Will proprietary AI models continue to be more prevalent in the ecosystem, or will open source AI eventually take over? How will the open/closed dynamic compare to that of previous tech innovations, including PCs, smartphones, and the internet? This is a heated debate unlikely to be resolved any time soon.

• Will the best models be trained and served on expensive hardware from a small number of public cloud providers, or will more tasks (particularly inference) be pushed to the edge/client? Does it make sense for large financial services companies to run their own AI hardware?

• Will AI be a boon to incumbents who already have the capital and data to build the best AI-powered products and the distribution to sell them, or will it be rocket fuel for new disruptors who build companies in an AI-first way, from scratch?

• If a company uses AI in its product, is it best served putting that AI usage out front in a public way, or is it better to focus on customer value regardless of the underlying technology. In a recent 30 Minutes to Presidents Club Podcast, Nick Casale of Handoffs argued that the term “AI” should never be in your sales pitch. 

The preceding list applies to the entire AI field but is particularly interesting to consider in the realm of financial technology, given its intersection with policy, privacy, and so many aspects of our economic lives, personal and professional.

Today’s newsletter covers a variety of developments and perspectives on fraud, something that can be both enabled and protected against by AI. It also covers the theme of idea recycling in fintech and asks whether new technologies will make it possible for previously unsuccessful ideas to finally thrive. In addition, we consider the ability for AI solutions to streamline small business credit application workflows and increase access to capital.

If you’re at Fintech Meetup and would like to chat, please let me know. Looking forward to the next issue soon, which will cover highlights and learnings from the conference and much much more!

As always, please share your thoughts, ideas, comments, and any interesting content. If you like this newsletter, please consider sharing it with your friends and colleagues. Happy reading!

Recent News & Commentary

Machine learning or Generative AI: What's better for Fraud Prevention? - Sardine

Given how clearly useful generative AI tools are in committing fraud, it follows that there would be great interest into using these technologies to prevent fraud. In this well-thought-out blog post on the Sardine website, CEO Soups Ranjan explains the differences between ‘traditional ML’ and generative AI, why ML solutions are often the superior approach, and where generative solutions do have potential in the fraud prevention world. Given the need for accuracy, the importance of domain-specific feature engineering, and the deployment and cost advantages, ML is currently better than AI at fraud detection. Soups highlights gradient-boosted trees for prediction on structured data and clustering algorithms for anomaly detection and fraud ring identification. Sometimes, XGBoost is all you need. In the view of this post, the best use for generative AI in fraud today is as a co-pilot for compliance operations. For example, generative tools can perform case reviews, generate SAR filings, or evaluate the usage and impact of a particular rule set. He also imagines more ambitious uses for genAI in fraud detection that would require a plentiful supply of labeled training data that does not yet exist in sufficient quantity.  

Signals: Why is there so much idea recycling in fintech? - This Week in Fintech

Fintech aficionado Nik Milanović wrote an excellent piece on why we keep seeing the same ideas over and over again. His examples (PFMs, autopilot for your money, etc.) ring true, and the central theme is that while many of these products are quite good and seemingly useful, they haven’t seen success, because many builders haven’t studied the market history well enough to understand why previous attempts didn’t work. As plenty of consumer-facing applications have struggled to find enduring success time after time, many of the real breakout winners have been B2B fintech infrastructure companies (e.g. Stripe, Plaid). However, even the potential of great infrastructure is fundamentally limited by the underlying health of end-user-facing products. Selling shovels to gold miners only works as long as some are striking gold! To avoid the endless recycling of ideas that don’t work, it’s critical for founders and builders to understand the previous attempts, what went wrong before, and why “now is different” or “we are different”. Nik’s post is so good, and you should read it. What will be really interesting is if AI capabilities help answer the “why now” question by delivering products that couldn’t previously exist that solve real problems in a way that was not previously viable.

FCC Makes AI-Generated Voices in Robocalls Illegal - FCC

It’s trivial to imagine how useful AI-based voice generation would be to those wishing to perpetrate robocall scams. In this recent ruling, the Federal Communications Commission affirmed that AI-generated voices would be considered “artificial” under the TCPA (Telephone Consumer Protection Act). Of course, running a fraudulent phone scam to manipulate and steal money from vulnerable people is already a crime, but the escalation in the quality of AI-generated voices and the ease of using them for impersonation led the FCC to adopt a more targeted approach. Automated dialing systems are not only used by scammers. In fact, plenty of financial services firms utilize automated calling in their customer service or sales operations, activity that is governed by TCPA rules. These rules specifically prevent the use of an artificial or prerecorded voice in an automated call without a consumer’s consent. With this latest ruling, AI-generated voices fall under this category even if any difference from a human voice is imperceptible. There may be situations, however, in which an AI-generated voice is tolerable by a consumer, even preferable. In such cases, a consumer could opt in. What comes to mind are cases where someone is truly reluctant to speak to a person, for instance if they feel embarrassment around financial hardship. If such an individual could have a conversation with a well-informed and helpful non-human AI agent, would he be willing, or would he anthropomorphize the bot thoroughly enough that discomfort would be preserved? My guess is that the technology, thinking, customer experience, and regulation in this area will continue to evolve. 

11 Frauds and Scams That Will Be SuperCharged With AI - Frank on Fraud

If you’re in the fraud prevention business and you haven’t read Frank McKenna’s prolific writings at Frank on Fraud, you should really check it out. Side note, if you work in financial services, you are in the business of money, which bad guys would like to steal, and you therefore have a fraud problem, whether or not you have the data to identify and measure it yet. In this post, Frank created an infographic/poster of 11 fraud techniques that are likely to be further enabled by advances in AI. A few standouts: extortion scams where fraudsters generate explicit images of a target and threaten to release them if demands are not met; ransom scams where fraudsters use AI to clone the voice of a purportedly kidnapped family member and demand ransom; AI-constructed fake retail sites that appear real and are used to harvest personal information and credit card details. The full list of scams here is thought-provoking and merits preparation and planning if you’re potentially vulnerable. Not a bad thing to print and put on the wall for your fraud team!

Bankwell Bank pilots generative AI in small-business lending - American Banker

So many small businesses rely on financing to bridge the gap between expenses today and revenue tomorrow. Yet applying for credit, particularly at a bank, can be an onerous, time consuming, and confusing process. In fact, in a recent survey, 67% of SMBs opting for non-bank financing cited banks’ difficult application processes and lengthy decision timeline as reasons. Bankwell, a $3 billion asset bank in CT, is exploring the potential for generative AI to streamline the operational process around SMB credit application and deliver a better customer experience. The pilot uses Cascading.ai’s AI-driven loan origination system to automate key components of the approval workflow but also to engage with applicants 24/7, reducing communication gaps and apparently increasing completion and funding rates. I’m interested to see the progress here, as there aren’t many great systems available in small business lending. If Cascading and Bankwell can demonstrate the abilities and impact of an AI-first LOS, it can go a long way in streamlining access to capital for small business owners. 

Deepfake face swap attacks on ID verification systems up 704% in 2023 - CyberSecurity Magazine

In unsurprising news, deepfakes make it a lot easier to commit fraud, and a proliferation of tools - many of which were designed for perfectly legitimate purposes - are being used for unsavory ends. Whereas live video was previously an effective way to prove true identity, fraudsters can now use ‘face swap’ apps to create realistic, full-motion emulations of an individual, often bypassing both human and algorithmic detectors. As with any fraud problem exacerbated by AI, the creative use of this technology is critical to its solution. Building identity verification systems that are resilient to the use of generative impersonation capabilities is a very hard problem and a multi-billion dollar opportunity.

When businesses first started to adopt electricity in their operations, beginning in the 1880s, those who did so were sometimes called ‘electric companies’, as the use of this new technology represented a meaningful distinction. Eventually though, electricity became widespread, and the only businesses that still called themselves ‘electric companies’ were those actually in the business of producing power. The same thing happened in the late 1990s with the internet’. Many businesses who had a website when it was not the norm started to be called ‘internet companies’, even if the technology was not their core product. Today, it’s just as odd to call Ikea an ‘internet company’ just because it does a lot of business online as it is to call your local coffee shop an ‘electric company’ just because it uses electric bulbs rather than whale oil lamps for lighting.

In 2023, we saw many companies rush to call themselves ‘AI companies’ if they were using AI in their products (and sometimes if they weren’t). This positioning seems to add some distinguishing value for many constituents (e.g customers, investors, employees). At some point, however, the integration of AI capabilities into every product and service will become so widespread that it ceases to be independently distinguishing. It’s hard to say whether this will happen in 2024 or 2028, but I do think we’re likely to move along that axis substantially this year, and it will happen faster than with prior technologies1. 

Today’s newsletter covers a variety of topics, from investor research and thought leadership to highly technical explainer videos to chronicles of real-world applications for AI in financial services. 

As always, please share your thoughts, ideas, comments, and any interesting content. If you like this newsletter, please consider sharing it with your friends and colleagues. Happy reading!

Latest News & Commentary

AI: The Coming Revolution - Coatue

The team at Coatue put together a long, well-researched, and opinionated deck on the AI revolution and its likely impacts over the next couple years. It’s extremely wide-ranging, covering everything from adoption velocity, pace of investment, and where value accrues in the stack, to developer culture, regulation, hardware wars, open source, the future of work, and AI’s impact on the power grid. There are a ton of interesting graphs and numbers, all with sources footnoted. Overall, the report takes some positions, but it asks even more open-ended questions. If you’re in the space, you owe it a read. 

Vertical AI: Why a Vertical Approach is Key to Building Enduring AI Applications - Greylock

While vertical-specific SaaS companies have created large and impactful businesses in several industries, there are plenty of industries with far shallower SaaS penetration. In this thoughtful and detailed post by Christine Kim at Greylock, she outlines the very large opportunity in using AI technologies to bring next-generation vertical software to these industries. One of the key factors here is the ability of LLMs to work with and add structure to unstructured or messy data. This, in theory, would allow companies to build products for markets that don’t have highly-consistent systems of record. She describes 6 pillars to her vertical AI investing framework and then also gives examples of its potential application in professional services, financial services, and healthcare. These of course are not small industries! In fact, if we understand the market to be ‘industries with insufficient data structure/consistency/quality’, my view is that this describes….every real-world market! The opportunity here is very large, particularly for founders with deep industry expertise and the technical acumen to bring creative product solutions to life.

Intro to Large Language Models - Andrej Karpathy on Youtube

In this 1 hour video, Andrej Karpathy, an elite AI developer and OpenAI employee, describes how LLMs work in a sophisticated yet extremely accessible way. What’s great is the way that he is able to speak about an incredibly complex subject and make it understandable to a broad audience without dumbing it down. He explains how LLMs are built from the bottom up, how they are tuned to be useful, and some of their more recent new capabilities. He also shares some frameworks for reasoning about how these models ‘think’. For instance, they generally have a ‘system 1’ (instinctive thinking) but not a ‘system 2’ (slower, logical thinking), although this is developing (FYI my copy of Thinking Fast and Slow is on the bookshelf behind me as I write this). In addition, Karpathy explains how an LLM is more than a ‘chatbot’ and can actually be thought of as a new type of computer or new type of operating system, an analogy that seems to work quite well (e.g. an LLM’s context window is like a computer’s RAM). And, just as security matters in traditional computing, there are many rapidly developing security issues in the LLM world, many of which are demonstrated here. LLMs will be such a foundational building block of technology over the next few decades that they are important to understand, even if you’re not a hands-on practitioner. Just as broadly understanding relational databases has been massively useful over the past few decades (and will continue to be), understanding LLMs will be incredibly useful over the next few decades.

Building a financial large-language model in-house - NTropy @ Fintech DevCon

Ilia Zintchenko of NTropy gave a presentation on the very cool technical progress they have made building specialized LLMs for transaction classification. In the presentation, he demonstrates how and why, all other things being equal, high levels of accuracy would require a longer prompt, which of course involves higher cost and higher latency. Then, he discusses a few novel techniques they have used to achieve high accuracy at a reasonable cost and low latency. These involve fine-tuning, using models of different parameter size, entity recognition, and caching. He also describes some caveats for the use of such approaches. It’s interesting to watch Ilia discuss these challenges and solutions, as they will likely apply to many use cases, particularly in financial services. Companies will want to develop AI capabilities that deliver high accuracy on specialized tasks in a particular domain. Accomplishing this at a high level of quality while controlling costs and maintaining a performant system is a real technical challenge, and it’s great to think about all the ways it can potentially be overcome.

Coris AI launches Merchant Real Industry using GPT-4

The ability to classify the particular industry of a business is highly relevant in financial services, particularly in the domains of lending and merchant onboarding.  Often, a small business lender will apply different decisioning frameworks to different types of businesses, and they may avoid certain industries entirely. For example, it could make sense to apply a particular cash flow analysis and underwriting framework to service businesses and a very different one to retail businesses with inventory. The most common taxonomy for business classification is NAICS. Despite the importance of this attribute, there is no perfect source of truth, and lenders often struggle to find a solution to obtain it with accuracy and coverage. While various data vendors attempt to fill this gap, the availability of high quality LLMs has given many people hope of a better solution. In this blog post from last April, Coris announced the availability of a GPT-based solution to determine the industry code of a business with high accuracy. It works very similarly to how a human might approach the task: find the merchant’s website, read the site, analyze the site’s content alongside the NAICS taxonomy guide to determine which category is the best fit. Classification of a business into a defined taxonomy of categories based on publicly available text seems like a great use case for LLM-based approaches. This matches what I’ve heard from several clients and friends in the space, who have adopted this method and achieved high accuracy at lower cost. Of course, an MCC or NAICS code is a fairly blunt categorization, so this is the tip of the iceberg. It will be interesting to consider other ways that language models could be used to turn large amounts of unstructured text into finite, structured data that could be used to better understand the financial dynamics of a business.

How Ally Financial Is Delivering Business Value From Generative AI - Forbes

Over the past decade, the century-old General Motors Acceptance Corporation (aka GMAC) has transformed itself into Ally, the nation’s largest all-digital bank, offering a wide variety of consumer financial products, in addition to auto loans. Ally has received accolades for its high-interest savings accounts as well as emphasis on digital customer service. It’s no surprise that Ally would make efforts towards the use of generative AI to enhance customer service and operational efficiency. This piece from Forbes discusses several of the bank’s early experiments, led by CIO Sathish Muthukrishnan. Ally has built an in-house platform to serve as a ‘secure bridge’ between external LLMs and Ally’s customer data, enabling employees to test use cases while retaining compliance with data security controls. Marketing is among the first production use cases for the technology, where generative AI is being used in campaign creative development, search engine optimization, and content drafting. In addition, Ally has run a pilot program for generative AI in customer service, using the technology for live transcription and summarization of phone calls, which reportedly reduced the time needed to handle customer inquiries. Ally’s Muthukrishnan notes that they have developed an AI Playbook and AI Governance Group, also integrating ‘Human in the Loop’ (aka HITL) processes to ensure accuracy. 

Noetica Announces $7.8M in Funding to Bring AI Driven Data and Insights to the Corporate Debt Markets - Businesswire

The corporate bond market is nearly $11 trillion in the U.S. alone. Despite the size and importance of this industry, most corporate debt transactions are based on incredibly complex, lengthy, and arcane agreements. This makes deals difficult to understand and hard to benchmark, creating barriers to information access, risk to bond investors, and ultimately, a less efficient market. Noetica develops its own AI to understand and benchmark corporate debt transactions, apparently without the use of external LLMs. I’m curious whether this means self-hosting a fork of an open-source LLM or actually developing its own models from scratch. Either way, training and tuning an LLM-based learning engine to understand the intricacies of complex financial and legal structures seems like a productive application of the technology and something that could save a lot of time, money, and risk. Of course, in any field with massive precedent and entrenched ways of doing business, accomplishing innovation requires not only technological but also cultural adaptation. It will be interesting to see how Noetica is able to grow using this round of funding and what it will ultimately take to drive meaningful change in credit markets.

The more we learn, the more questions arise, and the more potential paths emerge. Marc Andreessen’s Techno-Optimist Manifesto heralds the power of technology to improve any aspect of humanity. The Biden White House’s executive order on AI provides a sprawling perspective on government involvement in the field. AI OG Andrew Ng points out the risk of regulatory capture, where many of the government proposals to restrict AI are downstream of incumbent company efforts to stoke fear.

While LLMs are not ready to be used directly in making underwriting decisions, they can be quite valuable in creating some of the inputs required for more precise risk assessment, particularly in applying structure and labels to unstructured data. This was the subject of a very well-done panel I witnessed at Money 2020 last month. Meanwhile, there is robust quantitative evidence for the value of reinforcement learning algorithms in the optimization of credit limits, further demonstrating that ‘non-generative’ AI is highly applicable in financial services.

While many banks’ enthusiasm for generative AI has endured, even showing product demonstrations to regulators and rolling out customer-facing capabilities, other large institutions have struggled, particularly due to issues of talent, data infrastructure, and cost. Finally, I’ve included links to a couple recent videos: an interview where I discussed what the best lenders are doing to adopt AI and automation in lending, as well as a long, technical deep dive on the ability for LLMs to provide analytical assistance.

As always, please share your thoughts, ideas, comments, and any interesting content. If you like this newsletter, please consider sharing it with your friends and colleagues. Happy reading!

Latest News & Commentary

New models and developer products announced at DevDay - OpenAI

OpenAI hosted its first ever developer day this past week, and it was packed with big announcements. First, the company announced a new version of GPT-4 (GPT-4-Turbo), which has a 128k context window (i.e. you can put way more information in a prompt, 16x the previous version) and is also several times cheaper (one-third the price for input tokens and one-half the price for output tokens). It’s also trained through April 2023, adding roughly 2 years to the model’s ‘knowledge’ about the world. In addition, OpenAI introduced the ability to develop and use custom-built versions of chatGPT, known somewhat confusingly as “GPTs”. Developers, and even people who don’t code, can build custom agents, providing specific instructions, augmented knowledge, and the ability to call 3rd party APIs. GPTs can be shared privately or publicly, and an upcoming “GPT Store” will let developers monetize their creations. There was also a slew of developer-focused announcements, including but not limited to: improved function calling, a version of gpt-4 with vision support, a conversation threading API, a new text-to-speech API, and the ability to fine-tune gpt-4 and even work with OpenAI to build custom models.

The magnitude and cadence of new releases here is pretty impressive, and the impact on the industry is potentially significant. First, it demonstrated that many early-stage “AI products” developed by startups as thin wrappers over OpenAI might just become features of ChatGPT, thereby eroding any competitive moat. Next, while the GPT Store has not yet launched, there’s clearly potential for it to become the primary ‘app store’ for conversational AI agents. Previously, developers who wanted to create applications on top of LLMs with a particular knowledge base and 3rd-party API integrations had to build more complex toolchains, often using frameworks such as LangChain or LlamaIndex. Now that it’s possible to do this in a no-code environment within ChatGPT itself, it will be fascinating to see what apps become available given a theoretically much wider developer base. For financial services, the impacts are less clear, other than to say that whatever you were building may have just gotten easier, cheaper, and faster to deploy. In addition to public-facing AI apps, I’d be eager to witness the proliferation of “GPTs” within financial institutions as shareable productivity-boosters. 

The Techno-Optimist Manifesto - Marc Andreessen, a16z
In this rousing manifesto, Marc Andreessen lays out a passionate and detailed case for techno-optimism, the belief that the proper development and use of technology is critical for and capable of solving most any problem, ultimately increasing the well-being of humanity. He begins with the assertion that we are being lied to, manipulated in the discourse by those who preach pessimism, resentment, and doom. Rather, he contends, technology has been the driving factor in the improvement of the human condition and a ‘lever on the world’. You may have already read Marc’s manifesto, but if you haven’t, you really should read it in full. It’s thought-provoking, bold, and for those of us who build and invest in technology out of a love for humanity, a refreshing antidote to the doomerism so pervasive in popular culture and politics.

Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence - The White House

I can’t promise a twenty-thousand word executive order will be as exciting as the Techno-Optimist Manifesto covered above, but it’s obviously valuable to understand all sides of the AI regulation debate, including the impact on financial services, mentioned as a “critical field” in terms of AI impact. Biden’s order contains some good ideas, some bad ideas, and some misguided ideas, all too numerous and detailed to discuss in a single issue of this newsletter.  The EO is much less a last word than an opening volley, and we’ll clearly see a ton of AI-related policy activity going forward, especially heading into the presidential election season of 2024.

Google Brain cofounder says Big Tech companies are inflating fears about the risks of AI wiping out humanity because they want to dominate the market - Business Insider
Amidst plenty of big news in AI policy and regulation, AI legend Andrew Ng weighed in with his view, namely that the biggest tech vendors are creating a climate of unjustified fear to achieve regulatory capture. According to Ng, "There are definitely large tech companies that would rather not have to try to compete with open source, so they're creating fear of AI leading to human extinction. It's been a weapon for lobbyists to argue for legislation that would be very damaging to the open-source community." There’s definitely a big-tech vs. open-source war brewing around AI. Many of the fear-based arguments look similar to those used by large incumbents trying to restrict more widespread development of previous technical advances, including encryption and the internet itself. As we’ve seen in so many other sectors, over-regulation can lead to many unintended consequences, including the entrenchment of the few large incumbents who can bear the regulatory burden. In fact, given how many times we’ve seen this exact phenomenon, in industries as diverse as banking, energy, pharma, housing, and agriculture, it’s almost odd how not-attuned people are to the risk of regulatory capture. Ng published some more thoughts on this on his company’s blog.

Money20/20 Vegas Panel Recap: ChatGPT can write, but can AI underwrite? - Taktile
A couple weeks ago, I was in Las Vegas for Money2020, and even with an incredibly packed schedule, this panel was one of the few I made sure to attend. One of the panelists, Maik Taro Wehmeyer of Taktile, a modern decision engine for financial services, wrote an excellent recap of what I also found to be a compelling yet grounded conversation. Unlike other recent would-be revolutions in fintech (e.g. P2P lending, crypto), the panelists were quite rational and measured in their estimations of the capabilities of LLMs in underwriting. This very much aligns with my point of view as someone who has developed many risk models and strategies over the years using multiple generations of machine learning and has also experimented quite a bit with LLMs. While the non-deterministic and relatively unexplainable nature of LLMs make them too risky for actual underwriting decisions (especially in highly-regulated markets like the U.S. and E.U.), they do have the power and potential to contribute to better risk prediction. One of the most compelling use cases is the ability for LLMs to add structure to unstructured data, creating labels that can then be used as inputs into better-understood and more observable machine learning architectures. This allows lenders to incorporate datasets that intuitively have some explanatory power but would otherwise be impractical to include in credit decisioning without manual intervention and all its inherent problems (cost, time, bias, etc.). I’m optimistic for this approach to using AI in the service of more accurate and faster credit decisioning using more diverse data sources and glad to see others ignoring ‘hype’ and approaching these issues in well-thought-out ways.

Optimizing Credit Limit Adjustments Under Adversarial Goals Using Reinforcement Learning - Alfonso-Sánchez et. al.
I’ve often observed that while a ton of time and effort is spent by lenders, investors, data companies, and researchers trying to analyze and predict borrowers’ propensity to default, comparatively little attention is paid to the other strategic levers that affect the profitability of a credit portfolio, such as the optimization and management of credit limits. In my experience, there is great benefit to applying thoughtful analytics across the entire customer lifecycle of credit, which is why I was glad to come across this paper from researchers at Western University in Canada and the Universidad Nacional de Colombia.

“Although credit limit setting is an essential problem for traditional banking industries and Fintech companies, since identifying the adequate credit limit will define the profit and therefore the sustainability of the credit card portfolio, this question has not been widely studied. This contrasts with, for example, the default prediction problem, in which a large number of banking analytic research papers are published.”

The authors set out to explore the effectiveness of reinforcement learning (a family of machine learning techniques, a.k.a. “RL”) for developing an optimal policy for adjusting limits on credit cards, thereby balancing two conflicting objectives: maximizing portfolio revenue and minimizing expected losses. To do so, they used data from the credit card product of a Latin American financial “super app”. There are some interesting technical/mathematical details in the paper, including data structuring and experimental design work to account for the fact that the ‘rewards’ from a credit limit adjustment are not immediate, unlike those in many common RL applications. The researchers found that the “Double-Q” reinforcement learning algorithm they developed outperformed other strategies in determining optimal credit limit adjustments. An additional finding was that ‘alternative data’ collected from the super app did not add value above typical financial data in developing a secondary ML model used to predict card balances post credit limit adjustment. Though highly technical, the methodology and results are quite interesting and also a great example of the value of what one might call ‘non-generative AI’ in lending risk decisions.

Gretel's Tabular LLM

Gretel, a developer-focused synthetic data generation platform, released a “Tabular LLM” in early preview. The tool allows users to generate, augment, and edit tabular data with natural language prompts. While I haven’t personally experienced this product, this is the sort of format-specific model that would have many intuitive and valuable use cases in financial services. To name just a few fairly common problems potentially made easier: generating realistic synthetic data that has the same information value as a real world dataset but preserves privacy; filling in missing values in a sparse dataset, generating a dataset of realistic customer profiles with diverse financial profiles across thousands of scenarios for the purposes of testing an application or simulating performance; constructing ‘stressed’ portfolio scenarios while incorporating random noise in order to stress test models and risk strategies. Of course, the risk here is that the generated datasets would be overly dependent on the model’s training data and therefore not truly useful for a seriously specific analytical use case. If so, the tool would be great for product testing and user demo creation but fall short of its true potential. Nevertheless, this is a fantastic idea with great promise, and I’d be excited to give it a try. 

A year after ChatGPT’s launch, how do banks stack up? - BankingDive
The launch of ChatGPT and its subsequent iterations have catalyzed a massive wave of enthusiasm across multiple sectors of business, including financial services. It’s hard to find a single large bank that hasn’t dedicated substantial talent, budget, and mindshare to exploring and developing uses for AI tools, whether to streamline internal processes or perform client-facing interactions. This piece in Banking Dive provides an update on the efforts of several large institutions, including Goldman Sachs and JP Morgan, where enthusiasm has apparently not waned. JP Morgan, for example, has been demonstrating its generative AI projects to regulators, seeking their involvement and feedback, particularly around the controls they would put in place. This is as expected, given generative AI’s potential to upend the current paradigm of model risk management generally used in the industry and by supervisory bodies. While Goldman Sachs claims not to be using generative AI in any client-facing projects, other banks are specifically targeting client interaction as an application of the technology. Morgan Stanley, for instance, is apparently testing an AI bot meant to provide assistance to financial advisors. It will be interesting to see how these banks continue to work both internally and with regulators and to observe how soon such applications will be deployed in production. As J.P. Morgan’s Chief Information Officer remarked: “It’s not a tomorrow thing”.

Companies struggle to deploy AI due to high costs and confusion - Axios

It’s been amazing to see so many companies large and small dive into experimentation with AI technologies, perhaps in an effort to avoid the often too-late adoption of previous technologies, such as mobile computing and the cloud. This is particularly true in financial services, where even large institutions generally not seen as early adopters have announced extensive efforts to incorporate AI into their products and businesses. This piece from Axios reveals that despite the potential, many such companies are struggling with the real-world deployment of AI. The reasons for this include cost, data infrastructure, and talent. AI models can be incredibly expensive to train, tune, and run, particularly when a use case is overly broad. Unlike general open source software or public cloud technologies, even the experimentation is expensive. AI also requires the right sort of data organization, a capable and flexible technical architecture, and the talent to build and deploy what are often very different types of computing workloads. Interestingly, Deloitte and Nvidia have a partnership to provide consulting services for just these problems. Now that companies are shifting from the ‘this is so exciting’ to the ‘this is harder than we thought’ phase, it will be interesting to see which efforts cross the chasm into production and which end up abandoned by companies that find it too difficult to perform the sufficient technical or cultural transformation.

AI and automation in lending: What are the best lenders doing? - Finovate Interview
Back in September, while I was at the Finovate conference in NYC, I spoke to research analyst David Penn about how the best lenders are using AI today. We touched on how some of the top lenders in business and consumer lending are using Ocrolus (where I work), explained how lenders are using cash flow analytics to manage risk, and tackled some of the most common misconceptions around the use of AI in financial services. It was an enjoyable conversation, and I hope you enjoy watching it!

Can an AI be your analytics intern? - FintechAIReview

In case you missed it, I recently recorded a technical deep-dive on LLM-based data analysis tools, including ChatGPT Advanced Data Analysis and LIDA, an open-source project from Microsoft Research. One of the most useful ways to think about AI is that it gives you a bunch of interns. Just like an intern, it’s not perfect, it doesn’t know everything, and you probably have to check the work, but it might help give you more leverage and amplify your work. In the video, I use publicly available securitized auto loan data to explore the capabilities of AI tools for analysis. You can see how the tools plan analysis, write code, generate graphs, and explain their conclusions. I even show some of the code underneath LIDA to reveal how such a tool ‘prompt-engineers’ the LLM to produce high-quality results. Interestingly, interacting with an LLM as an assistant actually might teach us how to better specify tasks for human assistants as well!

In the video, I use a publicly available dataset of securitized auto loan performance to explore the capabilities of AI tools for analyzing data. You can see how the tools plan analysis, write code, generate graphs, and explain their conclusions. I even show some of the code underneath LIDA to reveal how such a tool ‘prompt-engineers’ the LLM to produce high-quality results. Interestingly, interacting with an LLM as an assistant actually might teach us how to better specify tasks for human assistants as well!

I hope you enjoy this video and find it educational and thought provoking. If you have any feedback or ideas on other technologies you’d like to see explored here, please let me know. Thanks for reading/watching, and if you like what you see, please share with a friend or colleague!

Share

One of the enduring themes of this newsletter has been both the importance of constant learning given the incredible pace of AI development but also what the study of AI can teach us about how humans encode and use knowledge. As AI models and the applications built on top of them achieve ever more impressive feats, we feel compelled to ask deep questions. Should we reconsider the way we think about intelligence and creativity? What does the concept of intellectual property mean in the age of generative AI? More specific to our focus here, to what extent will AI reshape the landscape of financial services, and how quickly? What opportunities are there to use these new technologies to drive economic growth and opportunity?

Today’s newsletter covers a wide range of content. Meta’s LLaMa2 has the potential to spur plenty of follow-on, domain-specific innovation, particularly with its commercial license. Speaking of licensing, Ben Evans wrote an excellent and historically-informed primer on the issues surrounding generative AI and intellectual property. While Mint.com debuted in 2007, the search for what’s next in personal financial management still continues, with LLMs offering a potential step-function increase in usefulness and personalization.

Many large companies are hiring a “head of AI”, though the role differs massively from firm to firm. Speaking of jobs, there’s a chance that AI will create a whole new set of roles that won’t require a four-year degree, somewhat aligning with Noah Smith’s Revenge of the Normies thesis. There’s a chance a new breed of PE firm will drive massive efficiencies through buying companies and then using AI to massively restructure their operation. Read about all this and more below, as I link to 11 articles or papers and offer a brief summary and commentary on each one.

September also brings the beginning of conference season. I’ll personally be at Finovate in NYC (Sep. 11-13), Debanked in San Diego (Sep 21), and Money2020 in Las Vegas (Oct. 23-25). If you’re around for any of those, let me know!

As always, please share your thoughts, ideas, comments, and any interesting content. If you like this newsletter, please consider sharing it with your friends and colleagues. Happy reading!

Latest News & Commentary

Meta launches Llama 2, a source-available AI model that allows commercial applications - ArsTechnica

Meta recently released a new open-source version of its LLaMa foundational language model - LLaMa2 - which now comes with a license that permits use in commercial applications. The earlier releases of LLaMa, which were not available commercially, were a boon to researchers and open-source developers, as they enabled relatively permissionless experimentation. With the new release, developers can use LLaMa2 models in actual product development, and they can even download and run the model on their own hardware (llama.cpp, for instance, makes it possible to run the model on an M1/M2 mac). Of interest is Meta’s restriction on use of the model in applications with "greater than 700 million monthly active users in the preceding calendar month", presumably to guard against its use by BigTech competitors. While the new model appears not to be as performant as GPT-4 on many tasks, the open-source architecture and permissive license are likely to enable plenty of innovation for the development of domain-specific applications.

Generative AI and intellectual property - Benedict Evans

The potential for generative AI to upend established conceptions of intellectual property has been a consistent theme in this newsletter. Ben Evans’ newest essay does a typically elegant job of laying out the issues and discussing which are already well understood by existing law or practice and which result in completely new questions. He references parallels to various debates around creativity and authenticity over the past 500 years, from Raphael and Durer to Taylor Swift and Spotify. The debate over intellectual property in the age of generative artificial intelligence is both very old and very new. Either way, it’s far from over, and your perspective will be enriched by Ben’s analysis.

FinTech Lending with LowTech Pricing - Fisher College of Business Working Paper

In this wide-ranging research paper, the authors demonstrate how loan pricing from leading fintech lenders mostly relies on traditional methods of risk assessment, namely FICO and other widely-available consumer credit metrics. They argue that a clearly-observed discontinuity of interest rates at FICO 660 (generally considered by regulators and industry as the threshold for ‘subprime’) even for borrowers with otherwise similar levels of predicted delinquency risk amounts to a transfer of value from subprime to prime borrowers. In their words: “The pricing distortions result in substantial transfers from nonprime to prime borrowers and from low- to high-risk borrowers within segment.” While many fintech lenders tout advantages in the use of non-traditional data and advanced statistical techniques in underwriting, this particular analysis shows that FICO is still the most important determinant of interest rates, even when other data are available that would result in a more precise assessment of risk. I often tell my colleagues: “Think of the most sophisticated fintech lender you know - they are less sophisticated than you think, and there’s a ton of opportunity for innovation in risk management.”

The new pursuit for a Mint alternative: powered by AI - Ntropy Blog

In yet another must-read blog post for hands-on AI practitioners in fintech, the team at Ntropy discusses the potential for a new generation of personal financial management tools, built with AI. This problem has existed for some time. In 2007, I remember walking around telling my colleagues at American Express about Mint.com and how impressive it was. 16 years later, not too much has changed! Beyond the initial value of connecting your accounts and displaying them in one place, online financial management tools still struggle to dispense truly personalized and useful advice. However, the reasons for this - complexity of bank transaction descriptions, diversity of personal financial circumstances, incomplete and fragmented data - might make LLMs just the right technological advancement to break the stagnation. Robin at Ntropy goes through the opportunities and challenges here, discussing not only “modern Mints” such as Monarch Money but also Ntropy’s own open-source financial assistant proof-of-concept, Cookie. It’s still probably a while before people have access to truly widespread, reliable, and cost-effective financial advice, but it’s exciting to think about the potential for AI to make a real difference here.

Promise and peril: Fairness in AI-based lending - Bank Automation News

In this incredibly long and thoroughly-researched article, Victor Swezey reviews the potential for AI approaches to decrease bias and increase access in lending (warning: paywall). Victor interviewed many leaders in AI and fintech for the piece, including me. He included a few quotes from my remarks, such as: “When you can make quick, data-driven decisions that don’t flow through the potentially biased black box of the human mind, then you increase the diversity of people you’re able to lend to.”

The hottest new job is “head of AI” and nobody knows what they do - Vox
Apparently, tons of companies are hiring a “head of AI”, though the role differs massively from place to place. In some firms, such a person has a tech/data background and oversees technical efforts to integrate AI into products, while at others, the individual is more of a cross-functional evangelist for the integration of AI into everything, including branding or recruiting. As with any job that carries a fancy title, the actual influence and impact of someone in the position is likely to be a function of everything but the title itself.

Paving the Path for “Blue-Collar AI” Professionals - Myra Roldan

While the blazing-fast progress of LLMs, computer vision, GPUs, and other AI-enabling technologies owes much to PhD researchers in industry and academia, the proliferation of AI within various fields may actually create a whole new class of ‘blue-collar’ AI jobs that don’t require even a bachelor’s degree. This is the thesis put forward by Myra Roldan, who argues for the ability of community colleges to train a diverse population of future AI professionals. The vision is compelling, especially considering that many firms will really be adopters of AI rather than developers of it, more focused on the maintenance and operation of systems than their design. Interestingly, her list of example jobs falling into this category includes roles such as Machine Learning Engineer, which most tech people certainly wouldn’t call blue collar by any stretch. Perhaps this isn’t actually a matter of blue vs. white collar, a distinction that is rapidly losing meaning or relevance. Arguably, plenty of so-called traditionally blue-collar jobs are safer from AI disruption than white collar ones! There is more potential for AI to accelerate the emergence of a collar-blind economy, where the economic rewards and social status of a given job are not as tightly linked. 

BarbAIrians at the Gate: The Financial Opportunity of AI - A16Z

If AI has the potential to remake companies and industries, there will be massive opportunities for investment, some more obvious than others. In this well-structured essay, Alex Rampell at Andreessen Horowitz discusses the potential for AI to substantially impact productivity in ‘bits’ businesses before those involving ‘atoms’ and how since there are tons of even non-tech ‘bits’ companies, the potential for AI-driven transformation is enormous. Interestingly, he lays out the case for a new wave of private equity investing. As opposed to the earlier wave of leveraged buyouts driven by financial engineering and operational efficiency, this one would be driven by the potential to run companies more profitably with the judicious application of AI. In this vision, a PE group would buy an established company and then use the best AI technologies to radically reshape its product, operations, and workforce, thereby making it far more productive, profitable, and valuable. I fully expect this to happen. The question is which investment groups will be best qualified to execute an “AI takeover”. Will existing PE firms build AI practices, or will there emerge a brand new type of “AI PE” firm with a different combination of talent, technology, and capital?

Fintech, AI, and the Challenges with Compliance, Fraud, and Risk - Cowboy Ventures

Promise and pitfall, benefit and backlash, revolution and risk. Indeed, the outlook for AI in financial services is full of great upsides as well as potential downsides, particularly in the areas of fraud and compliance. While new technologies may streamline compliance processes, detect fraudulent behavior, and reduce the role of humans and their potential bias in critical financial decisions, they also create new types of risk. These include nefarious uses of the technology (e.g. deepfakes) or challenges to established practices of data governance or model risk management. This post discusses some of these risks and opportunities, many of which will surely be tackled by a wave of AI-native startups targeting the financial services market.

Checkout.com launches AI-powered ID verification solution

Checkout.com, a major online payments provider, has rolled out an identity verification solution that uses facial recognition to corroborate user identities over real-time video. In addition to capturing a static photo from an identification document such as a driver’s license, this capability prompts the user over a video connection to match identities and establish the person’s legitimacy. Ecommerce in general is a constant battle between convenience, purchase volume, and fraud prevention. As AI-enabled fraud vectors such as deepfakes proliferate, companies need to use these same AI capabilities to detect and prevent malicious use.

How Wells Fargo is dominating AI transformation - Axios

In this article, Axios writers reference a study from Evident, a bank-focused research firm, which studied various banks’ investments in AI. Apparently, “JPMorgan Chase & Co. is leading the field in research, Capital One is leading in patents, and Wells Fargo is leading in investments.” The speed at which the largest, most established industry players are rushing to integrate AI into their businesses is a contrast to previous technology waves (e.g. cloud, mobile) where much of the innovation was pursued by smaller firms, while the larger ones lagged behind. AI is an area where size and budget may confer real advantages, particularly if management can align on concrete goals rather than chasing trends. Right now, every large financial institution is making big promises with AI and putting serious money behind them. It will be interesting to see in a year how these investments turn out and whether anything really transformative will have been achieved.

AI is built by, used by, and used for the benefit of humans. While fundamentally a human creation, many people seem to easily forget this. It’s a bit like how people think of a ‘corporation’ as some alien entity separate from, and perhaps even at odds with, people, when in reality, a corporation is owned by people, run by people, employs people, and produces goods and services for people. It’s just a useful abstraction. Indeed, many of the most important questions surrounding AI involve humans, and a lot of them concern employment. What kinds of jobs will skyrocket in popularity and value when the AI revolution hits financial services? “AI engineer” might be the hot new tech job of the next decade. Some companies - perhaps finding that the availability of AI talent lags behind the size of their AI ambition - are also retraining and up-skilling their workforces to take advantage of this new technology. 

As countless new companies in fintech emerge with modern AI at their center, there’s a yet-unsettled question of where AI is best used. Furthermore, there is vigorous debate about where value is likely to be created and what types of companies can build a compelling business and a defensible moat. Below are links to the thoughts and frameworks published by 2 venture capital firms. There’s room here for a vast marketplace of ideas, not only about which types of companies are most likely to succeed but also about the capital structures best suited to funding them.

As always, please share your thoughts, ideas, comments, and any interesting content. If you like this newsletter, please consider sharing it with your friends and colleagues. Happy reading!

Latest News & Commentary

Databricks Strikes $1.3 Billion Deal for Generative AI Startup MosaicML - WSJ

In one of the largest deals so far in the AI space, Databricks, a leading data infrastructure platform provider, purchased MosaicML. Mosaic allows companies to train their own LLMs and other modern generative AI models using their own data and in their own environment. These two companies seem to have a natural fit, as many of the large firms that use Databricks as a data warehouse and analytics platform are also interested in building proprietary models using that data. For instance, the WSJ piece mentions Replit, who already used Databricks for their data pipelines and who then used Mosaic to build a proprietary code-generation model that accelerates and amplifies the work of developers. Specific to the adoption of AI in financial services, past issues of this newsletter have addressed many issues that are likely to be relevant, including data ownership, model governance, and security. The Mosaic approach seems to be quite compelling given these concerns. It will be interesting to see if financial institutions like TD Bank or Capital One, who already use Databricks, will use the combined company’s platform to train and tune proprietary models on their own infrastructure that comply with their model governance and regulatory approaches.

Sharing Our Field Notes: The State of Generative AI in Financial Services - Bain Capital Ventures

The team at BCV has been putting out a ton of high quality content lately, and it’s a fantastic gift to the ecosystem. This recent post is a 10-part guide to their ‘field notes’, a thorough exploration of the state of the industry with respect to generative AI along with a variety of well-considered observations and valuable frameworks. There is so much useful and thought-provoking information in this piece. For this post, its length is an asset, and it’s totally worth a read. So sit back with a drink during your nightly AI learning hour (this is totally a thing, right….?) and read the whole thing!

The Rise of the AI Engineer - latent.space

Many past eras in technology have seen the emergence of a ‘new hot role’, a job that didn’t exist before but that suddenly becomes the most highly-sought position. These roles require specialized knowledge, are highly competitive, and demand high compensation. Some established practitioners shift into these roles through their work, while others train for them specifically. They are often ‘of the moment’ and subject to their own hype cycle. For example, in 2012, Harvard Business review called “data scientist” “the sexiest job of the 21st century”. This piece makes the case that “AI engineer” is the next hot role, that it requires a unique combination of skills and will be in massively increased demand. According to the author, an AI engineer is somewhere between an ML engineer or data scientist and an application developer. This person is someone who doesn’t build the models or machine learning architectures but rather ties AI capabilities together to build applications, autonomous agents, or AI-enabled tools. My sense is that experienced software developers and ML engineers already have the talent required to do ‘AI engineering’, as long as they spend some time learning a new set of tools and frameworks and shift their mindset to AI-specific applications. It’s likely not a bad career move for the right people to target this niche, and there’s a good chance that every company will want to make sure it has sufficient AI engineers to realize its AI-related ambitions. 

Latent Space

The Rise of the AI Engineer

Thanks for the many comments and questions on HN and Twitter! We convened a snap Twitter Space to talk it through and >1,000 AI Engineers tuned in. Playback here…

Read more

3 years ago · 159 likes · 10 comments · swyx

Citizens Bank retrains employees for AI - Bank Automation News

Rather than (only) hiring AI experts into new roles, Citizens Bank is retraining its own workforce to utilize AI capabilities for the benefit of the bank and its clients. According to the article, this can take the form of training branch employees to be more advice-oriented (perhaps using AI tools, though it doesn’t specify), and it is also developing employees on the technical side. In fact, they may be offering people with computer science degrees who work at the bank but not as developers the opportunity to retrain and take on new, technical roles focusing on AI. The article also makes the point that while some companies may be speaking of AI as a reason to cut headcount, the reality is that AI adoption also requires a lot of work from people. It’s good to see this kind of retraining effort being discussed. An explosion of new technologies, tools, and use cases creates compelling career opportunities for individuals, and organizations looking to use AI to their advantage will need to have a well-considered talent strategy.

Why customers don’t want chat bots - Creative Good

Every few years, companies seem to get really excited about…chatbots. Previously, however, these bots - whether in the form of text or ‘interactive voice response’ (‘IVR’) were extremely poor substitutes for human customer service. In the absence of real data, I’d still wager that the highest-frequency statement uttered to a voice-response customer service bot is: “I WANT TO SPEAK TO A HUMAN!” However, now that LLMs have provided examples of truly useful AI using a chat-based interface, the chatbot concept is getting a major reboot. In this piece, Mark Hurst argues that nobody wants a chatbot and that these bots only will exist to create efficiencies for the companies that deploy them, rather than to improve service for customers. There’s quite a bit to agree with here, though I’d make the case a bit differently. First, rather than simply deploying chatbots to just replace the work done by human agents, I’d hope that companies would use AI within their operation to automatically solve the problems that require such assistance in the first place! Next, for a finite (but expanding) set of tasks, AI agents may actually perform better than some proportion (again, expanding) of human agents, which would reserve the human agents for the most critical, most complex problems. Perhaps the role of human customer service agent will become a much lower-volume, higher-skilled, higher-paying job given the focus solely on the most difficult and important issues. Finally, whether or not service-providing companies use bots for some of all of their customer service interactions, there’s a major use case for customers using bots to handle their side of the interaction, as the author alludes to with DoNotPay. 

Founder Framework: Identifying AI Opportunities in Financial Services - Emergence Capital

Everyone in the tech and venture ecosystem is trying to develop a point of view on the best applications for AI and to hone a framework that will help them identify companies with a unique and differentiated ability to win in the market. This piece from Emergence Capital looks at multiple functions or ‘jobs to be done’ in financial services and rates each on 3 dimensions: the extent to which AI can incrementally enhance existing processes or business models, the urgency of customer ‘pain’ or need, and a function’s tolerance for risk. After assigning a high/medium/low rating for each dimension for each of the functions, they arrive at 5 areas of particular promise. These are: fraud, accounting, insurance, payment risk, and financial research. Finally, they consider what infrastructure needs to exist, and how trust and regulatory frameworks will continue to evolve. We’re only at the very beginning of the AI fintech era, and it’s hard to say which of these arguments will pan out, but it’s great to see people in the field putting their ideas out there and engaging with builders about what could be most valuable. 

Fintech Job Fair - This Week in Fintech

Our friends at This Week in Fintech are hosting a Virtual Job Fair on Aug 7-9 to help candidates find and connect with companies hiring in the fintech space. Details & registration are available on the Fintech Job Fair page.

Today’s newsletter contains a wide variety of content, along with my context and commentary for each link. Generative AI has the potential to massively disrupt the structure of the U.S. banking system, and CEOs of big businesses also need to consider how to lead their organizations through AI-driven transformation. Innovations like OpenLLaMA extend the power of highly-capable foundation models to additional use cases, potentially unlocking a huge wave of bottoms-up innovation the same way open source software did earlier. 

Several financial services providers are debuting AI-powered chatbots, and a few examples are covered below. While we may wonder whether a chatbot would be worse or better than human assistants, perhaps that is the wrong question. There is a massive range of variation in human skill, task-related performance, and the quality of service provided by a human employee of a client-facing business. AI chatbots may not now or for quite some time be comparable to the best human reps, but they might perform better than some percentage of reps, even today. That number will certainly increase over time, and if done in a responsible manner with high attention paid to accuracy, it will expand access to good financial products and advice at more favorable costs.

Speaking of accuracy, all models require data, and generative AI models require data at massive scale. A couple of the pieces below consider the data that underlies all these technical advancements, including a deep dive into the world of human data labeling. In a world awash in data, much of it is still messy, hard to access, poorly labeled, inaccurate, or not legally viable.  Those developing AI-powered applications for financial services particularly need to have data availability, quality, and accuracy as top priorities. This is easier said than done!

As always, please share your thoughts, ideas, comments, and any interesting content. If you like this newsletter, please consider sharing it with your friends and colleagues. Happy reading!

Latest News & Commentary

Generative AI Is the Attila the Hun That’s Going to End the Banking System as We Know It - Matt Harris @ BCV

In this boldly-titled piece, Matt Harris of Bain Capital Ventures makes the argument that generative AI will drastically reshape the U.S. banking system, making the regional bank failures of the last several months look small in comparison. The United States is highly atypical among advanced, Western economies, with thousands of banks, many of them being small local ones. Attributing small-bank customer stickiness to inertia rather than loyalty - a point I’ve made for some time - he states that generative AI will overcome this inertia by making financially-optimizing recommendations to customers and lowering the cognitive load of switching institutions. He predicts the transition from 4,500 to ~100 banks will be ‘ugly’ but recommends that regulators ease the process by facilitating mergers and acquisitions to land at an equilibrium of 10 or so ‘too big to fail’ banks, along with 50-100 more specialized institutions, all well-capitalized. Change in the structure of the banking system is clearly going to take place, and the prospect of GenAI as the catalyst makes sense. It’s pretty clear that today’s system is sub-optimal, and it will be up to banks, their customers, and policymakers to consider options for the best possible future state. 

CEOs may not realize it, but they already know what to do about A.I. - Ken Chenault & Sam Palmisano in Fortune

Two of the most accomplished American business leaders make the case that CEOs must actively lead their companies through AI-driven transformation. In doing so, they reference past transformative technologies, the impact they had on business, and the consequences for those who were caught flat-footed. AI requires not just technology investment but a full-on management approach, and they cite 3 specific recommendations: 1) expand the scope of data governance to include AI; 2) train employees in all functions (not only technical ones) to be AI-literate; and 3) evolve corporate culture for the age of AI. It will be interesting to see how large companies, particularly those in financial services, are able to transform their cultures and businesses for the AI era. While there’s probably not one Fortune 500 that isn’t talking about its ongoing application of AI, there’s likely a substantial gulf between superficial toe-dipping and serious transformation. Consider how long it has taken so many banking institutions to undergo “digital transformation”. Perhaps today’s big companies will learn the lessons of past technological shifts and move quickly and decisively. In a couple years, it will be very clear who took which approach.

OpenLLaMA: An Open Reproduction of LLaMA

In February 2023, Meta AI released LLaMA, a powerful large language model (LLM) designed to achieve state of the art results while requiring less computing power than other comparable models (e.g. GPT-4). Thanks to Meta’s public release of LLaMA, many researchers have been able to use it as a foundational model to fine-tune for particular purposes. Indeed, many interesting AI papers in recent months - including some covered in this newsletter - have used LLaMA as their base model. While LLaMA is a fantastic achievement and powerful tool for the research community, it is not licensed for commercial use. Now, Xinyang Geng and Hao Liu from Berkeley AI Research have released an open-source reproduction of LLaMA, called OpenLLaMA. The model is built using the same procedure described in Meta’s original paper but is trained on a different, open-source dataset known as RedPajama (itself designed to be a reproduction of the LLaMA training set). According to the evaluation metrics published, OpenLLaMA performs comparably to LLaMA on a variety of benchmarks. Most importantly, this version is released under the Apache-2.0 license, which permits commercial use. If AI is indeed “having its Linux moment”, as many have stated, open-source foundation models like this will enable people and companies to fine-tune and build for their own specific applications. It will be interesting to see how models like this get adopted in fintech and financial services, particularly with the ability to run such a model without sending any data to 3rd parties. It’s not too hard to imagine the combination of a powerful, open-source base model and a massive proprietary dataset resulting in some truly compelling applications.

Digital bank One Zero to debut generative AI chatbot - Finextra

Almost every financial institution has some form of automated question-and-answer system. These come in the form of “interactive voice response” (IVR) phone menus or web-based chatbots and are typically powered by some form of pre-coded decision tree. One thing they all have in common: customers hate them! If you frequently find yourself shouting at the anthropomorphized IVR-bot, “I just want to speak to a human!!!”, banks are hoping that generative AI chatbots may solve the problem. One Zero, a digital bank in Israel, is piloting a GenAI-based chatbot with a small cohort of customers in advance of a widespread general release in Q4 2023. In doing so, it hopes to provide: “...an exceptional customer experience that closely resembles one-on-one human interaction, with zero wait time.” The bank partnered with AI21 Labs to build a chatbot that would not need to send data to a public cloud service and would be able to competently handle sensitive information. My assumption is that every single bank is looking at and considering this type of technology, yearning for its benefits while simultaneously grappling with important concerns, including data privacy but also how to mitigate the risk of a bot providing incorrect or misleading information to a customer.

Monarch AI Assistant

Monarch, an online consumer financial management platform, released an AI-powered personal financial assistant. The demo and examples on the website look interesting and informative and presumably rely on several already-existing capabilities of the Monarch software, such as transaction categorization, investment tracking, and goal-setting. According to the FAQs, Monarch uses GPT-4 to take a user’s questions and provide relevant answers in the context of the data in their account. For example, “how does our restaurant spending compare to others like us” or “what are our top recurring subscriptions”. While I don’t think an AI financial assistant is ready to give truly unique advice or handle particularly complex portfolios, innovations like this are all about access. If Monarch and services like it can use a generative AI application to provide relatively standard prudent financial management, budgeting, and investment advice to consumers with uncomplicated portfolios, that is likely superior to what a non-trivial portion of American households could otherwise obtain.

Adobe is so confident its Firefly generative AI won’t breach copyright that it’ll cover your legal bills - FastCompany

In another example of generative AI’s potential impact on intellectual property law, a new subscription for AI-based services in Adobe products will come with protection against copyright infringement claims. As discussed in a previous issue of this newsletter, there is mounting concern about intellectual property issues around the use of large language and computer vision models, particularly if the models were trained on data that included content subject to copyright. Adobe has been a leader so far in this arena, training their “Firefly” image generation model only on images to which they own the rights or on those that are in the public domain. They are so confident in their approach that for enterprise Firefly users, they will now “...compensate businesses if they’re sued for copyright infringement over any images its tool creates.” This is an interesting approach to potentially-unclear or fast-changing areas of law, whether in design software, financial services, or otherwise: doing what you believe is the right thing and then offering to indemnify users against claims to the contrary. 

AI Startups Have Tons of Cash, but Not Enough Data. That’s a Problem. - WSJ

While generative AI startups have raised billions in funding to pursue seemingly massive and multiple opportunities across a wide range of industries, many are struggling due to a lack of unique or proprietary data. This article in the WSJ tells the story of startups who are building AI-based products that target a particular use case but who face difficulties gathering the data they need to train task-aligned models or build a competitive moat. Some have tried to partner with more established firms for access to this information, but often, the large companies who possess the largest and likely most useful datasets are reluctant to provide access. Security and privacy concerns are top of mind, as are ownership issues around the resulting models and their eventual output. One investor compared the drive to obtain access to useful data to an “arms race”. This is clearly an important issue and likely familiar to anyone who has attempted to build a business that relies on data. It’s difficult to obtain data by partnering or convincing other companies to share what they perceive as a precious asset. It’s highly preferable to be organically in the flow of data for a product-related, operationally-critical reason. Models are easy; the hard part is data. If you’re building a product based on generative AI, or investing in someone who is, it’s important to consider if there is a differentiated ability to obtain high-quality, relevant, and usable training data. 

Are we there yet? - Amias Gerety, QED

In this essay, Amias at QED explains the “hackathon mode” that is so common around novel uses for LLMs, why this is exciting, but why the best AI-driven solutions require going beyond “hackathon mode”, particularly in areas like financial services where accuracy is paramount. “Hackathon mode” is fantastic, particularly when it allows a single developer, or even someone non-technical, to use LLMs to quickly build an experience that might have otherwise taken weeks or months, and more than 1 person. While the ability to enable creative and productive work with greater efficiency is an impressive and useful capability of LLMs, there are industries, financial services in particular, where most of the relevant training data are not publicly available, and the particular business logic required is not accessible to general-purpose LLMs. Amias gives the example of 2 companies where he is an investor and board member, Ntropy and Ocrolus (where I work). These companies recently announced a partnership that is, in his words, “not only a triumph over ‘hackathon mode’ for each of them, but also allows their customers to get the benefits of proprietary data and fit for purpose models.” With highly skilled machine learning and data science teams, both companies could perhaps build “slide-ware” versions of each other’s products fairly rapidly, but lending and other financial use cases require a much higher standard. Ocrolus and Ntropy have prioritized accuracy and real-world relevance over “hackathon mode”.

AI Is a Lot of Work - The Verge

Models require data, and the latest, most impressive generative AI models require incredible quantities of it. This long and detailed article explains the world of data labeling and annotation - in other words, the work humans do to create the data used to train AI. Tasks performed by human annotators range from the simple and repetitive (e.g. “label the elbows in this photo of people”) to the complex (e.g. “talk with this chatbot about science fiction and mathematical paradoxes all day, and rate the quality of its responses”). This type of work is done by many companies, across many geographies, for a wide range of hourly compensation that tends to be a function of skill required and geography. Models are only as good as the data on which they are trained, and this is a good read for anyone seeking a greater understanding of some of the (often surprising) work that goes into developing today’s most impressive generative AI systems.

Ocrolus Uses AI to Automate Document Capture, Improve Accuracy, and Reduce Fraud for Lenders - Cardrates

Mike Senecal at Cardrates interviewed me about how lenders are using Ocrolus’ AI-driven document automation platform to make high-quality decisions with trusted data, unparalleled efficiency, and a top-notch customer experience. I enjoyed speaking with Mike about a wide variety of topics on this theme, including how Ocrolus has combined best-of-breed AI products from Google, OpenAI, and Amazon with its own in-house models trained on proprietary data. We discussed how small business, mortgage, and consumer lenders rely on accurate data and insightful analytics to make important credit and fraud risk decisions. 

Share Fintech AI Review

Something fun about studying and experimenting with generative AI is that it can inspire us to consider aspects of human cognition we take for granted. Computers and humans clearly operate on different hardware, but as large, pretrained language and computer vision models become capable of human-level creative output, they can cause us to think about both the mechanisms and meaning of human thought and creativity.

Intellectual property, in concept as well as law, is likely to be affected by an accelerating AI wave. A couple of the articles linked below discuss aspects of this issue. While there is widespread concern - and even a few lawsuits - about copyrighted data being used in model training, it’s an interesting thought exercise to consider a completely different way of thinking about this. If a human learns by reading a lot of copyrighted material, it would seem extremely strange to subject the independent creations of that human’s mind to some form of IP scrutiny. That’s because the human is not memorizing and regurgitating specific material but rather encoding a representation of knowledge into one’s own unique brain. To go a step further, Jon Stokes has an excellent blog post which conceptualizes every single ‘file’ - i.e. any text, image, video, or audio that has ever existed or could ever exist - as an integer on an infinite number line. When a user enters a prompt to a generative AI model, the model is using that prompt to search through its latent space to output something that can be represented as a number, a number that, metaphysically at least, already exists. Taken ad absurdum, this idea pretty much invalidates all intellectual property law, which is extremely unlikely and undesirable, but it’s good food for thought.

It’s also interesting to contrast the operational and financial aspects of generative AI with previous innovation-enabling technologies, such as open source software. Generative AI is comparatively expensive, relying on specialized GPU hardware that is in the midst of a supply crunch and also costly to operate in terms of electricity use. Costs will undoubtedly come down with scale, but for the time being, startups building AI solutions may face pressure to monetize earlier than their peers in previous years. This is likely positive, as it could drive a wave of companies with practical, demonstrable real-world value, catalyzing even more investment and innovation in the sector. One of the links below is to a post from Lightspeed Venture Partners discussing the companies and teams most likely to succeed in Fintech+AI and what they can do to build a meaningful moat.

As always, please share your thoughts, ideas, comments, and any interesting content. If you like this newsletter, please consider sharing it with your friends and colleagues. Happy reading!!

Latest News and Commentary

Consumers are paying for AI - Bhoram Cho

New technologies inevitably spur countless ideas and birth ambitious new companies seeking to build products that best take advantage of their capabilities. For a while, many of these technologies came in the form of open source software, which was free to acquire and cheap to run on ever-more-efficient commodity hardware. In contrast, much of today’s generative AI boom is powered by companies building solutions on top of large language models offered by larger companies at a cost. GPT-4, for example, is priced at $0.03/1K input tokens and $0.06/1K output tokens. There are 2 ways to look at this. The first is that it’s incredible to be able to access such impressive, world-changing technology for literal pennies with almost no barrier to entry. The second is that if a company is using these technologies in a production use case, the pennies will start to add up quickly. Bhoram Cho - an entrepreneur, product leader, and former co-founder of a company I really wish still existed (KitchenSurfing) - wrote a post mostly focusing on this second point. Companies building AI-based products on top of commercially available large language models are going to either need to monetize sooner than those in the previous wave of startups or control costs through the use of smaller, local language models. 

Lenders court Gen Z with ChatGPT-like tech - National Mortgage News

Plenty of financial products are commodity-like in nature, so it’s interesting to think about what makes consumers choose 1 provider vs. another. For example, if 2 lenders offer a 30-year fixed-rate mortgage on the same exact terms/rates, how can 1 of these lenders differentiate itself via a unique customer experience? This piece in National Mortgage News considers how lenders might utilize various communication channels, including using AI-chat-assistant-style experiences to personalize borrower interactions. One technology vendor quoted in the article makes the point that younger consumers are not necessarily opposed to speaking with a live human; it’s just that they only want to do so if more automated methods have been exhausted, making it absolutely necessary. Of course, the mortgage sector is highly regulated, so lenders using chat AI technology will have to be extremely focused on accuracy and on making sure that an AI agent doesn’t do anything that can only legally be done by an NMLS-licensed rep, such as rate quotes or preapprovals.

Name, Image, Likeness — But Make It Gen AI

If copyrighted data is used in training a large, pre-trained generative AI model, can new outputs of the model potentially be considered derivative works and therefore subject to copyright restriction? Serene Papenfuss explores the implications of generative AI in copyright law in this interesting post, focusing in particular on the concept of “name, image, and likeness” (NIL). While explaining the details of a supreme court case involving a Vanity Fair-commissioned Andy Warhol painting based on a photo of Prince, she quotes a portion of the U.S. Constitution’s Intellectual Property Clause: “To promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries.” In other words, part of the reason for the existence of copyright protection is not to restrict, but to encourage, new creative work. The use of textual or image data in AI models is an interesting application certainly not considered by the framers! On one hand, if protected intellectual property is used in training a model, there’s a decent argument that the owners of that IP are entitled to some consideration. On the other hand, if large, pre-trained models mimic human learning and creativity, we may need to think about this a different way. For instance, a person might learn by reading a lot of copyrighted content but then store mental representations of it in ways that are unique to that person’s own brain and inextricably linked to other experiences. It would be laughable to consider the knowledge in that person’s brain a copyright violation, and an output of that person’s brain would need to be quite a blatant copy to be considered a violation. Generative AI represents a new frontier in intellectual property law, and we can expect a lot of activity here over the next few years.

How generative AI is enabling the greatest ever theft/opportunity - Diginomica

In this wide-ranging piece, George Lawton discusses many potential intellectual property issues brought to the forefront by the ascendance of generative AI. Many of these examples were perfectly possible, even if perhaps in different form, before the current AI wave, but all the attention has brought increased scrutiny. For example, Getty Images is suing Stability AI for allegedly using its content in training data. While the post has a bit of a glass-half-full perspective, it’s worth reading. He’s certainly right that generative AI has triggered a new wild west that will require people, businesses, and governments to think differently about IP.

Big banks are talking up generative A.I. — but the risks mean they’re not diving in headfirst - CNBC

There’s practically no company, big or small, that isn’t talking about the potential for generative AI in its business. Executives at large financial institutions have been aggressively touting their organizations’ experimentation in this area, testing the technology’s ability to do things such as providing financial advice, automating existing processes, or aiding in fraud investigations. This article recounts conversations and presentations from the recent Money2020 conference in Amsterdam, demonstrating an incredible amount of enthusiasm but also quite a bit of cautiousness. Banks in particular are sensitive to concerns around data privacy, accuracy of information, and the many sensitivities around customer interaction. This of course comes as no surprise, and large, heavily regulated institutions are smart to exercise some restraint. However, the winners are likely to be the firms who are able to conduct many experiments and then actually deploy targeted solutions in a rapid and compliant way.

Fintech x AI: The Lightspeed View

This post from Lightspeed Venture Partners discusses the differences between traditional machine learning and generative AI for fintech applications. They make a case for the combination of predictive AI and generative AI depending on the use case, particularly in cases where accuracy matters. It also displays a useful market map of the firm’s existing AI investments as well as a view on trends they are seeing today. As in many areas, the best founders at the intersection of fintech and AI bring deep subject matter knowledge as well as technical expertise to truly understand the problem being tackled and the factors that will lead to a valuable and enduring competitive moat.

Share

Toy examples are easy. Building something usable, cost-effective, and valuable for real world use at scale requires serious knowledge and effort. In financial services, where accuracy and compliance are top priorities, this is even more true. The news linked below demonstrates this in a couple ways. 

A super interesting post on the Ntropy blog shows how GPT-4 when thoroughly-prompted is quite good at transaction classification but that its accuracy can be matched with 100x lower cost and latency by a purpose-built solution, which of course requires significant expertise. The CFPB released a report on chatbots in consumer finance, which have a lot of potential for improved customer service if done well but could also provide a poor experience if used only as a cheap, drop-in replacement for human agents. Intuit is apparently using its vast financial dataset and codified knowledge of accounting standards and the tax code to develop purpose-built applications with generative AI.

The greatest useful innovation will occur at the intersection of deep technical expertise, large quantities of highly accurate and relevant training data, and a well-informed, highly-focused orientation to solving particular problems.

As always, please share your thoughts, ideas, comments, and any interesting content. Happy reading!

Latest News & Commentary

We asked GPT4 to categorize our financial data. A glance into the future of transaction enrichment. - Ntropy Blog

Ilia Zintchenko, CTO of Ntropy, a financial data enrichment platform, wrote a detailed and fascinating article on various methods of extracting meaning from financial transaction data. If you’re in the world of payments or lending, you’re likely familiar with the difficult-to-parse and highly unstandardized nature of bank transaction descriptions. Categorization of these transactions is intuitively useful for many applications, such as underwriting a loan, authorizing a payment transaction, or understanding the expenses of a business. However, categorization is a truly hard problem, in large part because categories depend on context (consider how the early directory-based web search engines failed, in part because there is no universally-relevant system of ontology that works for every user in every context). Ilia explains the difficulty and tradeoffs of multiple approaches to solving this problem, including human expert labeling, large language models, small language models, and rules-based systems. Most interestingly, this post shares detailed benchmarking of Ntropy’s transaction tagging vs. human tagging as well as multiple GPT-4-based approaches, comparing accuracy, cost, and latency. The code to run the benchmarks is even available on github. It’s really nice to see Ntropy share both theory and practice in such detail. 

One takeaway for those interested in generative AI for financial services applications: GPT-4 with a long and well-crafted prompt is capable of impressive accuracy. However, because OpenAI charges by the token, long prompts can be quite expensive. Similar or perhaps greater accuracy can be achieved with a custom solution, as Ilia demonstrates, at orders of magnitude lower cost and lower latency. Of course, this requires technical skill, accurate training data, and real investment. 

Inside The Rise Of A Fintech Startup Using AI And Human Insight To Fight Fraud - Forbes

Fraud is an unfortunate reality in financial services. If you’re in the business of lending money, someone out there will likely try and steal it, and the criminals are often smart and highly motivated. It takes time, money, and technology to stay a step ahead of the bad guys. Fortunately, there are companies like Sentilink, profiled in this Forbes piece by Jeff Kauflin. The article tells the story of how its founders - who are both incredibly smart and nice people - started the company after identifying surprising and novel fraud patterns and have now scaled it to over 300 customers. It details examples of how the company uses its AI models to detect and identify synthetic ID fraud, as well as how it uses human fraud analysts to review cases and spot previously unencountered fraud techniques, which they can then build into their algorithms. 

Chatbots in consumer finance - U.S. Consumer Financial Protection Bureau

The CFPB released a new report on the use of chatbot technology in consumer finance. Chatbots are now in widespread use, and the report shares findings that all of the top 10 U.S. banks have some form of chatbot, and 37% of the country’s population interacted with a bank’s chatbot in 2022. Some of these bots are quite simple, using rule-based systems and question-and-answer hierarchies, whereas others use more recent innovations such as large language models, often supplied by third-party vendors such as Kasisto. The report uses information from the CFPB complaint database to identify several risks posed by the use of chatbots. These include: inability to solve more complex problems, giving sub-optimal advice, wasting customer time if unable to resolve a dispute, or potentially revealing unauthorized personally-identifiable information. Of particular concern is the tendency for LLMs to provide incorrect information in a highly confident tone. Of course, many of these issues exist with human customer support agents as well! The warnings are valid and timely, though as with many uses of technology, the thing to be regulated should be the activity itself, not the technology used to accomplish it. Specifically, banks need to comply with all existing laws, and it shouldn’t really matter whether a violation is committed by a human or non-human agent of a bank.

Intuit Introduces Generative AI Operating System with Custom Trained Financial Large Language Models

In this press release, Intuit announced their creation of what they call GenOS, an in-house platform for the creation of product experiences powered by generative AI. This appears to be a set of capabilities for Intuit’s developers and not (yet, perhaps) available to the outside world. The release doesn’t go into a lot of detail on the specific product use cases, but it does make sense that Intuit is in a very good position to develop vertically-specific AI, powered by their vast treasure trove of consumer and business financial data. If done right, there’s a real opportunity to help streamline the process of understanding and managing one’s finances. For example, things like understanding your personal credit, managing your taxes, or doing bookkeeping for a business would conceivably be helped by a specialized, well-trained LLM agent that also had access to reliable and accurate data, as well as an overlay of law, accounting standards, and tax code. Making these activities easier and less stressful and error-prone will be compelling to a large population, and it will be interesting to see how Intuit’s GenOS materializes in their customer-facing products.

Why AI Will Save the World - Marc Andreessen 

In a passionate, logical, and well-argued essay, Marc Andreessen outlines his optimistic vision for why AI has the potential to drastically improve the human condition and why the doomers and opportunists spreading panic about AI are misguided and wrong. If you’re reading this, you’ve probably already read Marc’s piece, but if you haven’t, you should. The case he makes is incredibly convincing, listing the many opportunities for AI to solve the world’s problems and debunking some of the most common fears with real historical data. He contends that the real risk is the U.S. not sufficiently pursuing AI and ceding AI dominance to the Communist Party of the PRC, with its dark and dystopian vision. It’s compelling, inspiring, and the most important essay Marc Andreessen has ever written.

On one hand, AI tools that are in production and widely accessible today are capable of truly impressive feats, not only automating routine tasks but also performing work previously considered to be solely the province of human creativity. On the other hand, generative AI is a new and rapidly-changing field, and there are plenty of things it just doesn’t do yet. Language models can make up information that has no basis in fact, and they can often be wrong in surprising ways. 

This week’s newsletter dives into several examples of the opportunity for AI in financial services, but also the cost, uncertainty, and potential pitfalls. Nvidia’s CEO spoke about how Generative AI is already decreasing the distance between ideas and creative execution across a variety of fields. AI has the potential to expand access to financial services, and it has multiple straightforward potential applications in the insurance industry. The new wave of AI may even usher in a new era of human rights, as has occurred with previous technological revolutions. 

There’s also concern about a lack of AI developer talent, whether smaller financial institutions can afford to keep up with a potential technological arms race, and how large language models will (among other things) create a lot of new and interesting ways to make mistakes in finance.

At a time where there is legitimately both hope and hype in equal measure, it’s valuable to gain a deeper understanding of the technology and how it works. Those exploring the use of AI in fintech will find it useful to understand the current capabilities of GPT models, and there’s a link to a video from Andrej Karpathy, who explains it better than anyone I’ve seen.

As always, please share your thoughts, ideas, comments, and any interesting content. Happy reading!

Latest News and Commentary

Who's going to work on all these AI projects at the big banks? - Business Insider

Executives of large financial institutions are understandably wide-eyed about potential use cases for generative AI. While there’s no shortage of compelling applications and projects being proposed, one of the biggest issues for banks seems to be competition for the talent to actually build them. Given the rapid pace of AI advancement, it’s no surprise that the relatively few engineering practitioners truly on the cutting edge have their choice of employment prospects, making both recruitment and retention challenging. It will be interesting to see how many of big banks’ proposed AI initiatives come to fruition, and which don’t due to lack of talent. In my experience, however, there are some incredibly talented engineers in the financial world, so perhaps banks would do well to up-skill their top talent on the latest AI technologies. I’d guess that prioritization, autonomy, and access to computing resources and data will be the actual determinants of AI project success for big banks, more so than lack of people.

'Everyone is a programmer' with generative A.I., says Nvidia chief - CNBC

Nvidia CEO Jensen Huang described generative AI as the most important computing platform of our generation. He spoke of the ability to use natural language prompts to accomplish what could previously be done only through hand-written code, as well as the ability to use multimodality to understand more than just text and numbers, and how this will unlock a huge new chapter in human creativity. Already, tools like ChatGPT and Midjourney are able to generate truly impressive work from fairly simple text prompts, and we’re really just getting started. Generative AI is collapsing the distance between idea and creation, as so many tools have done across human history (though now with an adoption curve many orders of magnitude steeper). “Everyone is a programmer” is one way to think about this collapsing distance, though I think the more concrete implication is that the nature of what it means to be a programmer will change. The combination of creative problem solving, logical reasoning, and ability to think at multiple levels of abstraction is still a valuable human skill. Perhaps generative AI, and LLMs specifically, make these traits even more productive on their own without requiring as much syntactic fluency. Some hope for the banks that can’t hire enough AI developers…

State of GPT - Andrej Karpathy @ Microsoft Build

Elite AI developer Andrej Karpathy, one of the founding engineers at OpenAI, gave a fantastic talk at Microsoft’s developer conference on the state of GPT. It’s 42 incredibly informative minutes of content, presented in a way that’s accessible even to non-experts. He discusses how LLMs are trained and the various phases of fine-tuning required to build a helpful ‘AI assistant’, including the vast amount of human labor involved. He explains helpful ways to consider how LLMs ‘think’ and how that compares to human thought, and he shares recommendations for how to most effectively use GPT models in practice. The talk is so educational, accessible, and well presented that it basically eludes summary. It’s very much worth your time to watch the whole thing.

AI's impact on the banking industry: Association president says the 'jury is still out' - Fox Business

National Bankers Association president Nicole Elam opined on the opportunities and risks of AI in banking at the recent Milken Institute Global Conference. In general, she touted the potential of AI and digital banking in general to increase access to capital and better meet the service expectations of consumers. She expressed concern, however, about the cost of adopting newer technologies, particularly for smaller institutions without large technology budgets, as well as potential compliance issues that can arise from newer and less-well-understood tools. Overall, automation has been shown to enable greater access to capital, and institutions adopting automated processes end up serving more diverse borrower populations. In fact, I’m a co-author on a paper that demonstrates this using data from the Paycheck Protection Program. The cost issue is certainly a concern, as the U.S. has a large and highly-fragmented banking system, with thousands of banks and credit unions. While large banks typically have massive technology budgets and correspondingly good digital consumer banking experiences, many smaller banks do not. Given the cost and talent required to adopt AI for productive use, there’s a risk that AI adoption in financial services will be highly uneven, perhaps leading to more consolidation and making the biggest banks even bigger.

Generative AI is Coming for Insurance - A16Z Fintech

A16Z Fintech partner Joe Schmidt muses on the applicability of large language models in various insurance use cases. He outlines opportunities in underwriting/decisioning, as well as in sales and servicing, and discusses why LLMs may prove effective in areas where previous machine learning technologies have not. The examples make a lot of sense, and many are likely possible with today’s generally-available AI technology, particularly in less complex areas of insurance like auto and home. It will be interesting to see how existing insurance incumbents adopt AI tech to improve parts of their existing businesses vs. how a new, vertically-specific insurance startup could build in an AI-first way from the beginning.

Technology Will Change The World - Will The World Change With It? - Forbes

Nik Milanovic, fintech aficionado and founder of This Week in Fintech, wrote a comprehensive piece in Forbes explaining the potential transformation from AI in the context of 5 prior technological platform shifts over the past couple hundred years. As Nik notes, every technological revolution has made some jobs obsolete, created others that nobody previously envisioned, and has materially improved standards of living across the globe. Most interestingly, he considers the potential for utilizing the benefits from AI to create a national program for Universal Basic Income (UBI). UBI is controversial, and I’m by default skeptical of the concept for many reasons. However, Nik makes an interesting and compelling case that, “Major economic shifts throughout history such as the industrial revolution and mass electrification led to corresponding shifts in the basic rights and protections afforded to people in those societies as a result.” This is a thought-provoking read on one of the many policy questions that will undoubtedly arise as the applications of AI materialize.

Generative AI enhances alternative data lending opportunity - Fintech Nexus

Isabelle Castro Margaroli from Fintech Nexus interviewed Ocrolus (disclaimer: I work at Ocrolus) CEO Sam Bobley about the use of AI to power the company’s document automation platform, used by hundreds of lenders across multiple verticals, including small business and mortgage lending. Sam explains how Ocrolus has used multiple automation technologies since the company’s inception in 2014, including proprietary AI derived from its massive document dataset and ability to produce high-quality labeled training data in-house. It combines this technology with best-of-breed AI tech from Amazon, Google, and OpenAI to offer an end-to-end document automation and analysis platform that helps lenders to make high quality credit and fraud decisions with trusted data.

Money Stuff: AI Bots Are Coming to Finance - Bloomberg

In his typically witty style, Matt Levine comments on several examples of AI being used in finance, including in risk modeling, financial engineering, trading, and customer service. He references a statistic that at some large banks, 40% of listed job openings are for AI-related hires. He also details a few examples of questionable judgment in the use of AI. For instance, if a bank uses generative AI to suggest information that needs to be reported to regulators, can a bank accept those recommendations? Will regulators accept it? Also, LLMs are known to occasionally ‘hallucinate’ (i.e. confidently state incorrect, made-up information). Of course, humans often exhibit the same behavior! But if an AI system does this and then makes financial decisions based on it, it’s interesting to think of the implications. He somewhat hilariously considers all the interesting new ways of making mistakes in finance that will be introduced by AI: “New ways to be wrong!” These all provide good evidence for following one of the recommendations made by Andrej Karpathy in his GPT keynote: “Use them as a source of inspiration and suggestions, and think co-pilots, instead of completely autonomous agents…it’s just not clear that the models are there right now.”

Share

It’s been another exciting week in the world of AI, with product announcements, compelling content from industry thought leaders, and thought-provoking research papers that inspire visions of the potential for AI in fintech. For me, reading the content linked below led to 2 particular areas of thought.

First, with any technology, it’s useful to consider multiple phases of adoption and maturity. In one phase, it may allow us to do things we do today, just better. In another phase, it may present opportunities that never could have been possible or even imaginable. Generative AI, and specifically LLMs like ChatGPT, are so immediately useful even today, that people and businesses are already integrating them into the first category at a blistering pace. In Ben Evans’ excellent writing, he sometimes describes machine learning as “infinite interns” or “one intern with super-human speed and memory”. What can you do with this super-human intern? A lot, it turns out. This is also a reason why it’s important for policy makers and those who influence them to focus on outcomes and use cases for AI rather than the technology itself. There’s such a massive set of knowledge-work tasks where this type of technology can be useful that we may actually spend a lot of time in the first phase, especially in regulated markets like financial services.

It’s even more exciting to consider what happens when we use generative AI to achieve things in fintech that are completely impossible with earlier technology. This is many orders of magnitude more difficult, and even that is probably an understatement. However, I suspect that many of the innovations in this category will come from a realization that so much of what we call ‘intelligence’ or ‘knowledge’ can be thought of and interacted with as language in ways that were hard to imagine before we all started spending so much time with chatGPT.

Second, just as the concept of ‘language’ takes on more meaning, the concept of a ‘model’ used in financial services needs to evolve as well. For financial institutions, model governance is an incredibly important function. Any model used to make decisions or to manage a portfolio of assets needs to fit into a well-documented process and pass the scrutiny of many stakeholders, including regulators. Typically, model governance entails a highly-traceable data lineage, tightly-specified conditions for model training and evaluation, and clear explainability. In the world of generative AI, however, most projects start with a giant pre-trained model from OpenAI, Meta, Google, Anthropic, etc. and then fine-tune it to align to a specific application. In this paradigm, it’s extremely hard for a model governance process to understand all the details of large-model pre-training. Perhaps some financial institutions will only use models that they can train in-house completely from scratch, even at the cost of performance. Alternatively, reviewers and regulators may eventually consider some of the foundational models to just be part of the stack - like an operating system or language runtime - and not really subject to observation (I think this is less likely). I may ask some of my friends in the model governance world about this and share the results in the next edition.

As always, please share your thoughts, ideas, comments, and any interesting content. Happy reading!

Latest News and Commentary

Is Generative AI in 2023 as Transformational as Indoor Plumbing Was in 1920? - PYMNTS

Karen Webster of PYMNTS and QED Partner Amias Gerety had a thoughtful and engaging fireside chat about just how transformational generative AI can be. In doing so, they reference Robert Gordon’s thesis that the technology innovation between the civil war and 1970 had a much greater impact on productivity than anything in the 50 years since. Can AI potentially be as transformative? They discuss how some of the greatest potential from generative AI may come from its ability to lower the cost of exploration for new ideas, thereby potentially leading to more breakthroughs in technology, science, and health. Referencing previous platform shifts (desktop→mobile, on-prem→cloud) generative AI may be more likely to be on that level than other recent technology waves (such as voice or crypto), though it’s only truly possible to judge a real ‘platform shift’ after the fact. Amias also shares his perspective on evaluating the (many) AI-related pitches that VCs now receive, including how if GPT is available to everyone, it still comes down to whether a founder/team has a unique experience or advantage that makes something easy for them that is hard for others. In addition, Karen and Amias had some great thoughts around sensible areas of focus for AI policy and industry self-regulation.

LIMA: Less Is More for Alignment

In this fascinating and somewhat surprising paper, researchers from Meta, Carnegie Mellon, USC, and Tel Aviv University described how they took a pre-trained large language model - LLaMa, open-sourced by Meta AI - and fine-tuned it using only 1,000 hand-curated, very high quality examples. They then compared the quality of its results on a set of prompts to those of 5 well-known language models, which were generally tuned with far more intensive techniques (e.g. DaVinci003 from OpenAI was tuned using a massive amount of RLHF). Amazingly, humans rate LIMA as just as good or better more than half the time!* The authors use the results to emphasize that almost all the knowledge in LLMs is learned in pre-training, and that it takes only a limited amount of high-quality tuning to make the model very effective at particular tasks. The paper is a pre-print and likely needs some clean-up. However, the concept has some thought-provoking implications. If nearly all knowledge comes from pre-training, what exactly is knowledge, and does it change how we think about domain-specific intelligence? This is somewhat different from traditional statistical model-building, where there’s great benefit to having the largest dataset. It’s essentially a computer version of what we already know intuitively about humans. If you find someone really smart who has a ton of foundational knowledge and knows how to learn, it doesn’t take too many examples to teach that person a new task, and if the instruction you give that person is very high quality, it may not take much for this smart person to start doing high quality work too.** 

Eight critical approaches to LLMs - Gradient Ventures

Anna Patterson of Gradient explains 8 key principles for companies building B2B applications using LLMs. This post is a great read for founders building something new and also for those looking to build new AI applications within an existing business. She explains each principle in detail using specific and thoughtfully-curated examples. Enterprise technology, whether using AI or not, needs to solve real business problems and demonstrate quantifiable value in a particular domain. One of the great things about the present moment in AI is that one can easily envision numerous valuable uses for the technology, and the tools to explore it are generally here today. 

Ramp unveils impressive new suite of AI features

Ramp announced a set of impressive AI-powered features across its products. The capabilities include contract analysis, automated vendor negotiation, smart transaction coding, expense automation, and a smart AI assistant known as “Copilot”, which lets finance teams ask natural-language questions and receive answers based on their own data, even suggesting potential actions and workflows. The offerings appear to be incredibly useful and quite smoothly baked into the product experience, and I’m sure CFOs are clamoring for early access.

Multi-Modal Deep Learning for Credit Rating Prediction Using Text and Numerical Data Streams

In this paper, researchers evaluate an alphabet soup of deep learning techniques in the context of predicting corporate credit ratings. This is a different domain than most of the visible work in fintech, which often involves consumer and small business lending, rather than the credit ratings of large corporations. In the paper, the authors gather a large volume of structured (e.g. bond performance, financial ratios, market data) and unstructured (e.g. earnings call transcripts) data. The results overall are encouraging and do vary quite a bit by the types of models used and the techniques used to combine them. Of particular interest is that text-based data outperformed numerical data in model prediction, suggesting great potential for the use of unstructured data in corporate credit rating, especially as LLMs evolve.

ComplyAdvantage Introduces AI-Powered Solution to Combat Payment Fraud - The Fintech Times

ComplyAdvantage, the UK-based financial crime detection technology provider, announced a new suite of AI-powered fraud detection capabilities. The solution appears to cover over 50 forms of fraud through a diverse set of machine learning techniques. They list major fintech companies such as Holvi, Novo, and Realpage as users of the new product. Fraud is an unfortunate reality of the real world, and smart financial services companies need to constantly evaluate new technologies and data sources to stay a step ahead.

AI-driven personalization-at-scale: The key to boosting fintech customer engagement and revenue - VentureBeat

Leaders at Chime, Tricolor Auto Group, and Envestnet discuss opportunities for AI-driven personalization in fintech. The article provides a good overview, and the video is available for viewing on-demand. For most of history, personalization and scale have been natural opposites. Only with sufficiently advanced technology and abundant computing power can companies personalize experiences in a scalable efficient way. Perhaps generative AI can take personalization in fintech to new levels, increasing access to financial services and getting the right products to the right customers, on the right terms, at the right time. 

*If instead of humans, you ask GPT-4 to be the judge, it gets similar results. Weirdly, GPT-4 even prefers LIMA over itself 19% of the time.

**This is part of why I’m so grateful my wife and I were able to send our son to Montessori preschool, with its emphasis on hands-on, foundational learning.

This newsletter aims to be at the forefront of these advancements, the implications in financial services, and what it means for humanity. At what feels like the beginning of such a potentially dynamic era, it is easiest for some to consider risks, downsides, and the potential for what could be lost. However, the potential for positive change is even greater, and this newsletter will have a generally optimistic and tech-positive orientation. 

At first, this newsletter will chronicle the most interesting Fintech/AI news from the past week, along with some concise commentary. In the future, we may expand to essays, interviews, and other media formats (e.g. podcasts, video). It will be free for the foreseeable future - please consider subscribing. I’m excited to explore fintech, AI, and visions of our shared future together!

P.S. Any opinion I express in this newsletter is solely mine and does not represent that of my employer or any other party. Furthermore, my strategy is to integrate a wide range of content and perspectives, including those with whom I may not partially or fully agree. Reading and engaging with multiple perspectives makes us all stronger! Links do not equal endorsement.

Latest news and commentary

I Cloned Myself With AI - She Fooled my Bank and my Family - WSJ

Joanna Stern at the WSJ worked with AI tools like ElevenLabs and Synesthesia to create highly-credible impersonations of herself that successfully bypassed Chase’s voice biometric authentication system. These tools are only going to improve, and every financial institution needs to proceed with the assumption that this is already happening. 

Sam Altman on the End of the World or the Dawn of a New One

Bari Weiss interviews OpenAI CEO Sam Altman on the Honestly Pod. While not directly fintech-related, it’s a wide-ranging, thought-provoking discussion from a very good generalist interviewer with the man at the center of the present moment in AI.

How Consumers Want to Live in a Conversational Voice Economy - PYMNTS

Incredibly detailed and thoughtful piece by Karen Webster on the long arc of voice as a medium for financial services transactions. It’s a good example of “living in the future and filling in the blanks” and is really worth a close read. I’m encouraged by all the possibilities for high-quality voice AI, from the transformational to the mundane. I know I’d pay for a bot that solves what I call the “I just want to pay you” problem, wherein a consumer has to spend hours on the phone with a bank, hotel, airline, etc. to solve a seemingly routine matter made complex by a large company’s own systems. 

Outsourcing Your Financial Life to GPT-4 - DoNotPay

DoNotPay’s Josh Browder set up AutoGPT to automate portions of his financial life. The bot apparently identified and canceled unused subscriptions, negotiated with Comcast to lower his bill, and identified opportunities to improve his credit score though writing disputes to the bureaus. Personally, it would be a while before I’m comfortable using a hallucination-prone LLM to automate financial moves of any real consequence, but the examples here provide some interesting potential use cases. On the other side of this, it will be interesting to see if companies do anything to safeguard themselves against consumers who use a bot to dispute every single transaction in the hopes of getting a discount/refund.

Ocrolus Adds OpenAI GPT Embeddings for Deeper Automation in Financial Document Analysis

Ocrolus (disclaimer: I work at Ocrolus) announced the incorporation of GPT embeddings from OpenAI into its extensive suite of document automation technologies. In addition to tech from Google, Amazon, and now OpenAI, Ocrolus has developed its own computer vision and NLP models, trained on its extensive document dataset of hundreds of millions of pages combined with its in-house data labeling capabilities. This combination of off-the-shelf models plus purpose-built AI is powerful when building solutions for particular B2B verticals, and Ocrolus has used this to cater to lenders in small business, mortgage, personal, and other markets. 

5 AI Mortgage Tools That Have Launched Post ChatGPT - National Mortgage News

Last time I got a mortgage, my contact at the mortgage company congratulated me for being so organized with my documentation, that underwriting took only……45 days! This is a market that can benefit from automation in a massive way, and it’s good to see a roundup of various companies experimenting with or integrating AI into their products.

Embrace Generative AI to Unlock Value - Goldman Sachs

Matt Lucas from Goldman shares a short video about how most value will be created by humans working with AI, rather than AI actually replacing their jobs.