Fintech 🧠 Food - The future of Generative AI in Fintech
Plus, Goldman & Microsoft create a new network, Worldcoin wallet launches & why decentralization will power AI
Hey everyone 👋, thanks for coming back to Brainfood, where I take the week's biggest events and try to get under the skin of what's happening in Fintech. If you're reading this and haven't signed up, join the 30,670 others by clicking below, and to the regular readers, thank you. 🙏
Hey Fintech Nerds 👋
At Fintech Nexus this week, anything involving compliance, regulation, or Bank to Fintech partnerships was a sellout.
That tells you where we are in the industry.
There's a crisis of trust for banks, Fintech companies, and everything.
Those who can demonstrate it win.
Unless it's AI, we're at the shut up and take my money part of the cycle there.
Speaking of AI, I heard Sam Altman's fireside at Stripe Sessions and must say it was by far the best interview I've heard with Sam. It's a shame I can't find it online because his mainstream interviews always focus on AI-doomer takes, and the tech interviews are more utopian and flowery. The Sessions chat with John was as brass tacks and pragmatic as I've heard, yet it left me feeling optimistic.
We can and will use Generative AI; it's a paradigm shift, and there's a ton of noise and relatively little signal. All at the same time.
Of course, being there, talking to some Stripes who have an internal LLM available to every employee got me thinking about Generative AI Fintech.
This led to this week's 📣 Weekly Rant. The future of GenAI in Fintech.
PS. A massive thank you to Stripe for inviting me to Sessions. I hosted a panel about BaaS with some of the best Fintech nerds on earth. I know I said, "BaaS is dead." But I also think BaaS / embedded finance is today's biggest and most obvious revenue opportunity in banking and financial services. Growth is coming. If we fix the trust issues.
PPS. This week's Rant is the beginning of something I'd love to explore. If you're up for sponsoring a paper on Generative AI and Fintech, that could be fun.
Here's this week's Brainfood in summary
📣 Rant: The future of Generative AI in Fintech
💸 4 Fintech Companies:
Hadrius - AI-powered SEC compliance
Flagright - GPT Powered AML Risk Assessment
Ansa - Closed loop payments for Businesses
Bits - B2B onboarding orchestration
👀 Things to Know:
PacWest, First Horizon, and Western Alliance stock plummets then recovers as the banking crisis continues
📚 Good Read:
Weekly Rant 📣
The future of Generative AI in Fintech
We're witnessing a paradigm shift.
It's clear if you look at this week's Fintech companies, listen to tech company earnings calls, or use a product like Microsoft Teams or Notion; AI is everywhere.
It's becoming the default.
And it's one of the fastest growing platforms in history
Notion has hit $10m+ ARR from a new GenAI feature by charging $8 per month per user and is on track to hit over $100m by year-end.
That’s a ton of revenue from a feature.
But there’s something much more profound happening here.
From the mouse to the touch screen, history has driven us towards more intuitive interfaces. Using human language to communicate with a computer is the most intuitive way for humans and machines to communicate.
This "chat as the interface" could make LLMs and chat the new platform after the mobile, social, and cloud era.
If that's true, the entire tech sector flips on its head.
Platform shifts change who the winners and losers are.
But we’re still so early.
It is tough to tell the signal from noise in the GenAI conversation.
The "stack" is unclear, and we don't know the right long-term user experience. There's also that looming fear that AI one day wakes up, chooses violence, and wipes out the species.
But what does any of this have to do with Fintech?
Every VC I talk to asks me the same question "How does AI show up in Fintech?"
So this is my attempt to make it coherent.
I'm assuming *some knowledge* of Generative AI in today's piece, but if you want more of a start from 0, the previous Brainfood, "The Fintech Co-Pilot," may be worth reading.
The Generative AI stack
Early GenAI business models
Subscriptions as a direct revenue source (e.g. ChatGPT)
Enterprise partnerships (e.g. ChatGPT and Bing)
The product feature companies charge for (e.g. Notion and Superhuman)
The vertical GenAI company or GenAI for X industry
The future of GenAI
Web access makes GenAI current and more useful
Multi-modality lets GenAI take uploaded documents, images, videos etc
Open-source models unleash innovation and efficiency
Prompt chaining takes the output of one prompt to define the next to achieve a set goal
Maybe one day we get GenAI that creates new knowledge
How GenAI is being used in Fintech today
The productivity tool
The Vertical Fintech GenAI company
The Fintech SaaS adding GenAI as a feature
GenAI models for finance (e.g. BloombergGPT)
Where value gets added
Automating hard-to-replicate time-consuming but low-value tasks
Embedding content (like compliance rules) for your context
Assisting customers to do their busywork
Fintech business models
Subscriptions and API calls (SaaS)
The line item for the feature (e.g., Notion)
The Generative AI paradigm shift
1. Generative AI "stack." 👩💻
When you double-click on Generative AI, there's much more than ChatGPT or Dall-E.
Generative AI isn't one thing like the iPhone.
Although it is a new intuitive way to communicate with computers, a whole ecosystem of companies is making it possible.
This needs a picture.
I think in pictures.
The a16z stack from January seemed like a good a starting point as any other (you can find it here)
I've simplified it below. Some parts are self-explanatory, but it's a useful frame for understanding winners, losers, and how things might play out.
Hardware: Companies like Nvidia have specialist hardware for AI workloads
Cloud platforms: Most Generative AI workloads are happening in cloud platforms
App Ecosystem: B2C and B2C applications like Jasper, Github Co-pilot, Bing Chat
App + Proprietary Model: B2C and B2B applications like the ChatGPT service + application or MidJourney.
Closed Source Models: GenerativeAI models like ChatGPT exposed as APIs
Open Source Models: Generative AI models that are available to anyone if they can provide hosting and configure them to run
Model hubs: Platforms to share and host models and resources to manage and develop new models (e.g., Huggingface)
Immediately it's obvious that GenerativeAI is not one size fits all. You might use a closed-source model like ChatGPT or Dall-E today, but that could all change.
We're still early.
The ecosystem of companies and providers of infrastructure for this new world is also still emerging.
Below I added two new sections that will grow in importance.
GenAI developer tools
Vertical Generative AI (VGAI?)
Developer tools examples
Imagine you want to embed a model with a context like understanding financial services regulation. These "embeddings" need a place to live to save the user from having to prompt them every time from scratch. Companies like baseplate can store documents, images, or any content.
Then there's a whole ecosystem of companies that manage prompt engineering. How did your prompt perform, track, monitor, and improve?
Think about observability, too; how are the models performing for users? Are they getting the outcome they wanted?
We'll undoubtedly see many more categories emerge in the pure tooling space. These tools are already being used to bring various GenerativeAI solutions to market that go beyond what the raw models can do.
Vertical Generative AI examples
Any regulated industry has a context (e.g., Healthcare or Finance)
Specialists like lawyers will have similar requirements to each other
The ability to go deeper into the problem space in these verticals has been proven in SaaS; could GenAI have its own category of verticals? Perhaps.
It will depend on the business model.
2. Early Generative AI Business Models 💱
It is far too early to say which business model will stick, but there is no lack of them.
a) The subscription. ChatGPT charges consumer-grade users $20 per month for access to the latest models and API access. I buy the logic. If humans will pay $9.99 for a limited subset of movies and TV shows, why wouldn't they pay $20 a month for a team of research assistants?
b) The enterprise partnerships. We've already seen OpenAI power everyone from Stripe to Microsoft's ambition here too. The commercial terms aren't always disclosed, but I imagine this becoming like all enterprise sales in time.
c) The feature in other products. Notion and Superhuman have built proprietary Generative AI directly into their product experience to improve the experience of using the product. Microsoft consumes GPT-4 for its Office365 suite and into Bing Chat (with web access. Microsoft also consumes Dall-E in its chat product (ask it to generate an image). I can't find any big-name examples of companies using open-source models yet, but I suspect that is a matter of time.
The lesson from Bing and NotionAI is you need this feature. If not for the free marketing, then for the revenue. The harder question is do you buy build or partner to get there.
d) The Vertical Generative AI company. There is a SaaS business that solves problems for every sector of the economy. A SaaS for gyms, a SaaS for construction companies, and a SaaS for doctors. These sub-sectors have a shared problem set and complex workflows. By focusing on the vertical but distributing via the internet, the vertical SaaS business can profit from what would have otherwise been too niche. They exist because of enabling infrastructure like cloud hosting providers.
The same is true for GenAI. Companies can solve problems for a vertical, like reading specific regulatory documents or harder-to-get data sets to limit or improve GenAI outputs.
We need to see what comes next in the technology itself to speculate about what comes next.
3. The Future of Generative AI 🔮
To see the Future of AI in Fintech, we need to look at the future of AI generally.
a) Web access makes Generative AI up-to-date and able to cite sources. We've seen the power of this in Bing Chat, where asking it about recent events can drive different and better outcomes than GPT4. Bing Chat has become a handy research assistant in writing Brainfood, but I still find the core GPT4 model much better at wide-ranging queries. This will normalize over time, but web access could get more interesting if it goes beyond browsing.
But if web access becomes the default, we can rely on that via the API. Think of every app that relies on a data source like maps, weather, or stock prices. All of that becomes promptable.
For Fintech questions swimming through my head are things like;
What does web prompt + Open Banking look like?
Is there a Nerdwallet + prompting that becomes a feature, or is that a business?
What could we build with financial markets data?
There are endless threads to pull on here.
But so far, LLMs have been limited to text-based prompts.
b) Becoming Multi-Modal means a Generative AI model can read more than text prompts. Imagine uploading pictures, documents, or files and having the model do something with that file. Here's a great example. Ethan uploads a CSV to GPT4 and asks it to think of ways to visualize the data set as charts; then, he asks it to be prettier and exported as an MP4. The result can be viewed in the tweet at this link.
Imagine how much time that saved. Four hours? Ten?
This area is still a work in progress. Results may vary for a while yet. For example, a recent study showed that LLMs have some way to go when working with SQL for databases as an alternative to humans.
We’re still early
This feels solvable.
And when it's solved, the use cases here are unlimited.
Every data set just became something an intelligence could work with and manipulate. Things that took an engineer or data scientist to do historically are now available with the right prompts.
My health warning here is like all things Generative AI it's
Only as good as the skill of the prompter
Even with great prompts, it gets you to a 70 to 80% score; if you demand the top 10 percentile of performance, your best bet is still a skilled human.
However, when you consider that most finance infrastructure runs on flat files (think .txt), or Excel files (.csv mostly), being able to upload files is a massive win. Most of the value added from Fintech infrastructure companies has dealt with these file structures' complexity and admin pain.
Open Finance's value add is about fixing the terrible data stored in payment systems to make it more accessible and easier to read, unlocking use cases like cashflow-based underwriting.
Entire Fintech SaaS verticals exist to kill spreadsheets.
Multi-modal will be huge.
c) GenAI is trending to Open Source. Open-source models have gained massive traction in recent weeks because they can do more, be trained on less data, and it is unleashing innovation.
In August of last year, Stability AI launched Stable Diffusion, the open-source image generator competing with Dall-E and Midjourney. It can create "AI-generated art" or images from simple sentences like competitors.
However, it is credited with kickstarting an era of rapid development in AI-generated images. The differences are astonishing if you look at how far the models have come since August of last year.
(Attempts to replicate the Mr Bean actor Rowan Atkinson over time).
The same happens to generated text models. In March, Facebook's LLaMA leaked, Stanford University launched Alpaca, and in recent weeks Stable Diffusion launched StableLM.
What makes these open-source models different is their efficiency. Each achieves ~ a similar performance to GPT3 but is trained on fewer parameters. GPT3 was trained on over 175 billion parameters requiring a massive dataset. The "smaller models" average just 7 billion.
Being open-source has also unleashed a wave of tooling, allowing for things like
Running a model on your laptop
Running a personal model on your phone
Plugins and adaptions (web access or multi-modality added to models within hours of release)
This dramatically increases the development speed and reduces the cost of running a model and developing future functionality.
The downside is these models have no restraints. There's already an "unstable diffusion" model that can generate. Anything. Realistically. And I'll leave that there. If the cost of models decreases, this becomes a much more open field, and ChatGPT could be AOL, not Netscape.
The takeaway here is the bedrock hasn't settled.
d) Prompt chaining (AutoGPT) or the "do anything machine." AutoGPT is a program that breaks a goal into multiple prompts. It does this by chaining together responses from ChatGPT into a series of prompts to achieve your goal. So, for example, if you ask it to build you a dropshipping business, it will start to research categories, look for places to source materials and build website code. It's like a loop of outputs from LLMs where one feeds the next.
Today the skill of prompting is using multiple prompts to achieve a goal. Task serialization offloads that to the LLM. The creators describe this open-source project as experimental; it only works on GPT4 and is still buggy. Could entire variable workflows be completed from a single goal?
"Build me a dashboard for internal metrics." Could research the best dashboard output format, pull in web-based data, and use multi-modality to access internal data before building the dashboard.
"Build me marketing automation for an early-stage company." Could start researching how-to's and explainers, providers of SaaS software, and 3rd party staffing for human oversight.
Practical applications are likely far off here, but the pace of change in LLMs is incredible. So who knows. This may happen quicker than we can imagine.
e) What else? The north star for GenAI is generating new knowledge. To date, LLMs have remixed existing human knowledge with surprising effects. But they haven't developed anything genuinely novel. Their creativity is an impressive illusion, like how computers "generate a random number." They can't do random; they use the current date and time as a seed to generate a number.
A human can generate "random" from any stimulus that floats through their body or mind. For GenAI to do something similar is undoubtedly some ways off, but that's where we're headed. Sam Altman believes this is some way out, so I'm inclined to believe him.
So now we have a view of where LLMs are and where they're going.
Where is finance with LLMs today?
4. How LLMs are already being used in Financial Services 🤖
The internal productivity tool use case dominates, but "GenAI as a feature" and "GenAI for x" have emerged as categories.
To understand where it fits in financial services, we have to look at the history of software in the industry. The meta job of finance is understanding risk and deciding based on it (like pricing it.)
Understanding risk is a complex task that has high variability.
But it usually ends with a simple output and moves some numbers around.
Two customers rarely look the same.
Two underwriting tasks can shift dramatically based on tiny details.
Therefore with limited compute and storage at the time, we turned financial services software into a calculator + database.
We feed the machines the numbers, and they move the 1s and 0s around. But they're awful at complex and nuanced decision-making. So most of the value in Fintech software, infrastructure, and applications became about managing the complexity and making it fit into the calculator.
Robotic Process Automation (RPA) was the big hype of ~5 years ago, taking those repeatable workflows and turning them into tasks that could be automated.
But RPA is limited by the same problem. High variability.
A robot wouldn't know which emails and Teams messages must be flagged for compliance review or the suitability of new marketing copy. These somewhat subjective, laborious tasks can be done by an LLM, however (with human oversight).
This mental model of
Look for complex tasks
That is time-consuming
With high variability
Is a compelling way to go fishing for short-term opportunities (as an individual working in finance prompting and as someone looking to build or invest.)
Here's a rundown of some of the main themes I've seen
a) GenAI as the internal productivity tool. Bottom-up adoption is happening at scale and across the industry. Anyone with access to Bing or the internet can use an LLM for free (unless your employer blocks it), and at worst, its $20 per month to use. Many compliance teams have used LLMs to summarize large due diligence forms and documents. (I wrote more about this in the previous GenAI piece)
Things get much more interesting if the LLM can access company internal data sets.
Imagine if you had a team of researchers for any task who gave instant responses.
A Stripe told me a great internal use case that illustrates this idea. If you're a product manager, you can ask their internal GPT4 instance, "Who are my top 100 customers by spend in X GEO" and you'll get back a list.
Expect to see this become default as an internal tool at companies that can execute that sort of thing.
b) Generative AI as a Fintech Vertical. Every Y-Com Fintech company and AI-based startup pre-Series AI is adding "Generative AI" features. As a wedge, a marketing play, or the core of their offering.
In 💸 4 Fintech companies this week, two companies claim to be "GPT-powered" or Generative AI. Hadrius has inhaled the SEC rulebook to help investment advisors ensure marketing suitability or to scan emails and team messages to ensure compliance. This, but for every field of compliance, feels like a no-brainer (I'm not sure I'd have started with the SEC of all regulators, too 👀).
It's clearly a feature, not a product, but a great wedge.
Such a great wedge might we see generational companies born from this feature?
c) Generative AI as a feature in Fintech SaaS. If every company is becoming a Fintech company, why wouldn't every Fintech infrastructure company do GenAI?
I'm yet to see a massive Fintech company do a Notion or Superhuman and bet big on GenAI, but everyone I talk to is using it internally. At a minimum, their engineers for productivity and marketing on some level. But there's a wave of smaller companies adding LLMs as a feature in their broader offering.
But obvious use cases appear if you take GenAI and consider Fintech Infrastructure. There's a ton of work in finance that is looking for the needle in a haystack of paperwork. The GenAI is pretty good at finding most of the needles and arranging them in an orderly fashion.
AML investigations, can analysts have summaries of complex documents
Onboarding a business customer, can it neatly summarize 100s of documents against 10 different databases and spot weird things?
Payments orchestration, instead of orchestrating visually, why not describe the workflow you want?
I am fuzzy on use cases for B2B Vertical SaaS. Bit of a stretch here but I'm thinking about:
Vertical SaaS for construction, could it take email quotes from contractors and build a budget automagically
Chat is the interface for complexity. It's the UX to find needles in haystacks, and that's before it goes multi-modal or gets web access.
d) Generative AI models for finance. The big fear with GenAI is hallucination. That's incompatible with an industry that ends in an output expressed as a number in a database. One solution could be multi-modality, but the other is training the models directly on financial services documents.
Bloomberg recently released BloombergGPT, a model trained on 50 billion parameters that included English financial documents. This allowed it to significantly outperform competitive models in financial tasks (like sentiment analysis). Perhaps more impressively, it beat GPT3 on many general tasks too.
The possible advantage of owning the model means Bloomberg is not paying OpenAI to use ChatGPT. Like Notion or Superhuman, owning more infrastructure could make this a long-term revenue generator if it adds enough value to users.
The possible advantage of owning the model means Bloomberg is not paying OpenAI to use ChatGPT. Like Notion or Superhuman, owning more infrastructure could make this a long-term revenue generator if it adds enough value to users.
But if multi-modality and open-source takes off, none of us need our own model.
So much is unknown, but what we can start to see is how value gets added.
5. How value gets added to end users
It is too soon to have a comprehensive framework for where GenAI goes in Fintech because GenAI itself is still a primordial soup of development and chaos. But, no matter how it gets distributed, I'm narrowing in on some core areas of value creation. The value propositions are similar, whether from the model directly, via an app, or SaaS.
a) Automating hard-to-replicate back-office tasks. As an example, each money laundering investigation can be wildly different from another. Criminals are trying to hide their tracks. An investigation will require looking at countless data sources and bits of evidence without a particular sequence. Pulling all possible needles out of the haystack allows the investigator to focus on abstract pattern matching.
b) Preventing hallucination for your context. Whether it's Bloomberg's custom model or Hadrius embedding the SEC rulebooks into their SaaS offering, models perform better when the constraints of your context are automatically applied. With enough skilled prompting, this context can be added, but emedding-as-a-service saves time and improves outcomes.
c) Customer-facing is still danger unless it's assisting their work. Advising customers to buy a financial product or offering them a loan is a highly regulated activity. Regulators want Fintech companies to be able to explain how existing AI models work. Now imagine if a chatbot hallucinated and suggested the wrong product for a customer. Danger.
But if your customer is uploading a ton of documents to file their taxes, being able to pull out things they might want to look at or edit is useful. When you're assisting a customer to do something they had to do anyway, that's a different thing.
6. Future Business Models 👩💼
If the landscape of GenAI is still an open field of play, then the business model is anyone's guess. But here are some guesses.
a) The subscription or API call. Custom models (like BloombergGPT) or pre-embedded models like Hadrius could use the standard SaaS pricing as a direct distribution.
b) The GenAI SaaS fee for the feature. If Notion can charge $8 per user per month, why can't you? The short answer is probably usage. You'd want to be confident a user is getting value from that feature before you start charging for it. But that's no excuse not to be trialing it.
c) GenAI uses Fintech (e.g. cards) to power GenAI. If every SaaS business embeds finance, why wouldn't every GenAI company? If you're building GenAI tooling, is there a common problem set your user has, and where does finance touch that? This one could take a half decade or more to fully emerge. But just as Apple and Shopify got into payments, if GenAI is the new platform and interface between man and machine, why wouldn't payments live there too?
d) Decentralization for open source models. Historically, open source has been funded by corporates who use it either as a public good they benefit from or as a form of market capture for their moats. The reality is somewhere in between the two.
The interesting thing about crypto tokens and blockchains is that they act as an auditor and incentive mechanism. This can help solve our challenges with LLMs, like privacy or attributing and paying artists when the source material is used to train models.
In 👀 Things to Know, I discuss Sam Altman's other project, Worldcoin. His idea here is that if AI becomes big enough, it will run the economy so efficiently there will be nothing for humans to do, and we'll need a universal basic income (Worldcoin).
In 📚 Good Reads this week, there's a full piece on how tokens (the crypto variety) may become the business model for AI. This idea has been around for ~5 years, but what I enjoyed about this take is it's well reasoned by a data scientist and engineer.
Isn't it odd how AI and Crypto are at opposite ends of the hype cycle, and yet, the main use case for Cryptoassets could actually be AI?
That's a whole rant.
AI's default business model is still up for grabs, and if it's not advertising, then payments or Fintech will be in there somewhere.
7. Generative AI is a paradigm shift. ⏩
It is being adopted faster than cloud or mobile and has a shot at becoming the new platform.
It's too early to say who the winners and losers will be, but in Fintech, the low-hanging fruit is obvious in the back office.
Today consumer-facing is danger.
I want this to change. I want entrepreneurs to prove me wrong here.
Maybe multi-modality and the ability to inhale (embed) rulebooks and "guide" users could be the next UX frontier.
We've been struggling for decades to help consumers improve their financial health. Agents that combine AutoGPT and behavioral psychology can nudge us to better outcomes.
This topic will rumble on.
I didn't even begin to cover regulation, time horizons, or which existing Fintech verticals are the most impacted.
What happens to fair lending and explainability in a world of GenAI and LLMs as finance infrastructure?
Does that count as a human decision if staff use LLMs to automate some but not all of their tasks?
In the time between when I started drafting this Rant, Google Bard launched, and it looks actually good. If I had written this in January, we wouldn't have Midjourney v5, open source LLaMA, BloomgbergGPT, or a whole batch of Y Combinator companies that are GenAI af.
If you've got value from this Rant, I'd recommend going deeper into Things to Know and the Good Read this week :)
When things are moving so quickly, it's hard to tell the signal from the noise.
But I hope this helped you as much as writing it helped me.
This was fun.
4 Fintech Companies 💸
1. Hadrius - AI-powered SEC compliance
Hadrius automates compliance processes for registered investment advisors to save "up to 90% of the time" spent on tedious tasks. Examples include monitoring and flagging emails or Teams conversations, reviewing marketing suitability, suitability, disclosures, trade monitoring, and audit prep.
🤔 GPT is acting as RPA for tasks with high variability. Robotic process automation can only automate things that follow the same steps every time. It's hard to ask an RPA to "flag anything that might be high risk in this conversation based on SEC guidelines." Compliance is full of tasks that are hard to automate because they have high variability. It's the obvious starting point for this tech. I wonder if the SEC will have issues with the "explainability" of decisions at some point, though?
2. Flagright - GPT Powered AML Risk Assessment
Flagright provides transaction monitoring and sanctions screening for startups. Its attention-grabbing feature is "GPT powered" customer investigations to save analysts up to 2x of their manual work effort. Flagright compliments this with a rules editor, a library of existing rules, and analyst collaboration capabilities.
🤔 I've heard three times this week from banks, "We'd love an Actimize killer." Actimize is the incumbent solution in transaction monitoring. There's now a swathe of companies like Flagright that are modern, real-time, and focused on collaboration. For me, the key battleground is how much data you can collect and collaborate across fraud and AML. I sense that "GPT-powered" compliance will quickly become a feature, not a business. But it's a phenomenal wedge.
3. Ansa - Closed loop payments for Businesses
Ansa helps marketplaces, coffee merchants, and micro-merchants who often deal in low dollar amounts to reduce the fixed fees of a transaction vs. card payment rails. Building a closed-loop wallet is challenging due to the sheer volume of edge cases. Ansa solves this by making a closed-loop payment rail available as a headless service via an API.
🤔 Low dollar transactions are massively impacted by traditional payment rails. A fixed fee of $0.21 is a massive percentage of a $5 creator tip but a tiny fraction when purchasing a Macbook. Starbucks famously provided its customer a wallet and could drive more loyalty, lower fees, and store $10bn in customer balances. Love this. Starbucks for everyone else. I covered this theme in the future of loyalty a while ago if you want to go deeper.
4. Bits - B2B onboarding orchestration
Bits allows companies to onboard business customers by combining KYC, KYB, fraud, and AML data into a single platform. Bits support multiple markets and offers a decision engine, real-time risk management, and team collaboration.
🤔 If B2B finance is the key battleground, B2B onboarding is critical infrastructure. Per the QED / BCG report, the opportunity is international and B2B, so this is the right problem space. There are now so many companies that do this one type of orchestration its hard to know who will win and what the competitive differentiators will be.
Things to know 👀
1. Roundup of institutional Blockchain and Crypto stories
a) Canton Institutional Decentralized Network launched
Goldman Sachs, Microsoft, BNP Paribas, S&P Global, and CBOE have teamed up to launch the Canton Network powered by technology company Digital Asset. The network is owned by participants, creating a single golden transaction source. It can ensure regulatory compliance and privacy and runs on the smart contract language DAML.
🤔 $4 trillion in assets will be tokenized by 2030. Blackrock CEO Larry Flink says tokenization is the next generation of financial markets. The opportunity here is enormous. I'm surprised we don't have a bigger ecosystem of builders in this space. But perhaps that's the next opportunity. Bridging the public networks with these private ones. I know smart people have given that a ton of thought already. I literally try to shake people to see this opportunity at times. I guess if you know you know 🤷♂️
🤔 This is huge; it bothers me how little attention it might get because it's "TradFi." The Fintech world still massively undervalues the significance of institutional networks like Canton. While "blockchain, not bitcoin" was all the rage in 2017, many of those investments are now quietly paying off. Digital Assets (and their peers) had to solve substantial regulatory and technical challenges. The future of financial markets is tokenization, and programmable assets are the first step toward that.
b) Worldcoin Announces Worldcoin Wallet
OpenAI's Sam Altman's "other" project, Worldcoin, has announced a non-custodial wallet. The wallet has 500,000 users and wallet that offers "gas-free" transactions. The app focuses on proving personhood for individuals without government-issued IDs and offers non-government-based universal basic income.
🤔 Worldcoin was largely derided at launch because of its device, "the orb," an Iris scanning device. When you combine Sam Altman, Iris scanning, AI, and Universal Basic Income, you have the full bingo card of Silicon Valley tech utopian bullshit. Also, the "Orb Iris scanner" is straight from every dystopian sci-fi 80s movie. It doesn't take long to imagine a black market for stolen eyeballs as criminals aim to steal the universal basic income.
🤔 But the problem with that narrative is this team is really thoughtful. The device uses the Iris "because it is the richest biometric marker available and the hardest to defraud." Images collected by the Orb are immediately deleted, and all that is stored is a number representing Iris's uniqueness.
🤔 If Crypto is adopted at scale, it likely is the economic backbone of AI, and its first adopters will be the Global South. AI is the paradigm shift, and Crypto is the disruptive innovation.
🤔 But here's the problem I have. If I get hacked, I can't get a new eyeball as easily as a new password. This is massively over-engineering with very limited contact with the global south's challenges. It is very well-intentioned, but it's so tone-deaf it's agonizing. The global south needs proof of personhood as an alternative to government ID; we could solve massive social problems and lift billions from poverty if we succeed. But that needs to be guided by contact with reality. I'm not writing this off; this team is too damn smart to not achieve something. But just as we need regulation for AGI, we need a check and balance for proof of personhood.
2. PacWest, First Horizon, and Western Alliance stock plummets then recovers as the banking crisis continues
In the first week of May, PacWest, Western Alliance, and First Horizon Bank saw 40% stock price drops. PacWest announced it was looking at "strategic options" after struggling to raise capital, and First Horizon stock fell after canceling a planned merger with TD Bank. The outflow of deposits could force many banks to sell long-dated securities at a loss. After slashing its dividend, PacWest subsequently rebounded, regaining nearly all of its value. Still, market sentiment was that this was a bounce after the market went too far and that trouble remains ahead.
🤔 There's still trouble ahead for regional banks with deposit flight. The Fed may have pushed investors to be risk-off, leading to buying bank stocks, traditionally "Lower risk" and perhaps looking like a bargain given recent sell-offs. With the market still strong, investors are concerned that the Fed will have to continue hiking rates leading to a far worse recession. The winners are the M&A bankers and consultants. The losers are the entire economy.
🤔 Confidence is in short supply creating a run on stocks. Deposit outflow leading to the sale of capital at a loss began the doom spiral that took down SVB. The stock drop leads to more deposit flight, which leads the bank to need to raise funding, leading to an announcement, which creates stock drops and more deposit flight.
🤔 We're in a perfect storm for deposits evaporating across the industry. But let's keep this in context; deposits could be reverting to the mean. Banks had all-time high deposit balances following stimulus checks, and consumers had more free cash during the pandemic. Now consumers have a cost-of-living crisis, have burned through their savings and businesses are pushing the deposits into high-yield debt (treasuries) which offer more than most banks. This is creating a doom spiral that might not end soon, and likely will lead to lots more M&A.
🤔 "Regional" bank has become a dirty word; that's a misleading and bad thing. In any other market, these are solid banks. The US arguably has the most dynamic financial services market due to its diverse banking sector. Where Europe has regulators, and India has government infrastructure, the US market's regionals and smaller banks are the best counter to the large bank hegemony.
🤔 Regional banks do things big banks don't have to or can't. They're big enough to afford to take technology bets but small enough to not be killed by corporate sludge. First Horizon has launched a fully cloud-native, US-based subsidiary called Virtual Bank, enabling it to launch products at scale, integrate Fintech companies, and move at digital speed. Name a top 10 bank that is true for and where the product is live.
Good Reads 📚
1. AI and LLMs need decentralization
The most interesting contrarian take I've seen (and agree with) is that LLMs need decentralization. This piece makes a well-reasoned and experienced ML engineer's argument for why. The problem all LLMs like OpenAI, Meta, and Google face is they need to be open source, and open source AI needs a business model. That business model is tokenization.
LLM models have several core problems. Models are hard to reproduce because they have terrible software and data version control. Distributed, cryptographically secured databases are fantastic at version control. Data gets stale, models require incredible amounts of compute, and the incentives for owning and contributing data are all over the place.
🤔 As people point and laugh at Crypto, they're missing the biggest opportunity. Crypto is the business model for open-source AI.
🤔 LLMs need to be open source. We're worried about AI wiping out the species, and the open-source models are starting to outperform the closed-source models despite being trained on far less data. Open source is the fastest way to software robustness.
🤔 Open source needs a business model. Big Tech has successfully captured markets using open source (e.g., Chrome and Android). But if we're concerned about AI wiping us out and becoming controlled by an enterprise, we need it to generate income without being captured by a single organization. Decentralized tokens are a great way to do that.
🤔 LLMs have a huge data privacy issue, but decentralized ledgers act as a data provenance and data auditor. Every record is stored with a cryptographic signature. Those records cannot be edited; any changes leave a full audit trail. Wallets and signatures are a great way to verify permission to use data.
🤔 Isn't it weird that AI models burn a ton of carbon and somehow escape other technologies' ESG concerns? 👀
🤓 Extra Credit: The Internet Financial System
Tweets of the week 🕊
How memecoin trading works
That's all, folks. 👋
Remember, if you're enjoying this content, please do tell all your fintech friends to check it out and hit the subscribe button :)
Disclosures: (1) All content and views expressed here are the authors' personal opinions and do not reflect the views of any of their employers or employees. (2) All companies or assets mentioned by the author in which the author has a personal and/or financial interest are denoted with a * (3) Any companies mentioned in Rants are top of mind and used for illustrative purposes only. (4) I'm not an expert at everything you read here. Some of it is me thinking out loud and learning as I go; please don't take it as gospel—strong opinions, weakly held.