Fintech 🧠 Food - The weekend that prevented a financial meltdown
Plus, How Stripe uses GPT-4 to prevent fraud, and why Credit Suisse is a nuanced situation
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 28,739 others by clicking below, and to the regular readers, thank you. 🙏
Hey Fintech Nerds 👋
What a week. Big banks are bailing out smaller banks, GPT-4 launches, and Bitcoin is up 14% in a week as banks collapse.
Oh and some guy is betting $1m that Bitcoin will reach $1m within 90 days.
The temptation is to see this as another 2008, but the reality is more like a strange sequel. Less of a doomsday and more of a slow erosion of trust in the banking system and institutions that back it up. The last decade has been weird, with all-time-low interest rates followed by a sharp rise, and banks have to risk adjust incredibly quickly.
The contagion here is not bank liquidity but humans scrambling and panicking. If every human and business pulls deposits out of the banking system, it will have a crisis.
No bank can survive a run.
Preventing runs is about confidence and trust.
But every time the Fed steps in, a generation of people with scar tissue from 2008 lose a little more faith in the power structures and financial system. It’s more of a gradual erosion of power structures than some sudden collapse.
Long term the Fed will have to print money again and when it does, the dollar will devalue and treasuries will be under water. Balaji is not wrong , he's just taking an extreme position.
Short term we're in for a year of volatility as the market adjusts to higher interest rates. Some unprofitable banks (like Credit Suisse) are forced to sell or move on. Banks with structural risks (like SVB) have to rebalance with the help of regulators.
But we will get through this.
There was a popular meme in 2008. Keep calm and carry on.
The age of low interest rates and cheap capital might be over, but the age of human ingenuity is never over. Venmo, Square, and Slack were all founded shortly after the global financial crisis.
Talent finds a way.
Stay patient, stay curious, and keep grinding.
Good things will come.
Here's this week's Brainfood in summary
📣 Rant: The weekend that prevented a financial crisis
💸 4 Fintech Companies:
Steadypay - Income "top-ups" for gig workers (UK)
Barte - B2B e-commerce payments (Brazil)
Monnai - Solving the data gap for Fintech
Canary - The employer/employee financial health grant
👀 Things to Know:
Stripe raises $6.5bn at $50bn valuation & Early GPT 4 adopter
Credit Suisse shares dropped 24% before Swiss National Bank intervention.
📚 Good Read:
The Walugi Effect - The limitations of LLM AI's (e.g., ChatGPT)
Weekly Rant 📣
The Weekend that Prevented a Financial Crisis
Banks closing at weekends is a feature, not a bug.
Last weekend gave regulators, bank staff, and the rest of us a few days to step back, calm down, and fix the problem—an essential reminder of the power of humans in the age of always-on, 24/7, digital everything.
On March 10, 2023, the 16th largest bank in the USA, Silicon Valley Bank (SVB), collapsed after a run on its deposits. It was the second-largest bank failure in US history and a shock to the tech industry that relied on its services.
How did this happen?
And is this contagion?
Over the past weekend, the FDIC stepped into backstop deposits at SVB and Signature Bank. Then at the market open on Monday, bank stocks got crushed, especially those resembling Silicon Valley Bank (SVB). Earlier this week, Credit Suisse's stock plunged 24% before recovering after the Swiss National Bank stepped in to shore up the lender.
But this is not a financial crisis like 2008.
The scar tissue many who grew up or worked through the financial crisis have is coloring judgment and popular perception.
What we have here is companies breaking when interest rates rise. And a lesson for all of us about why risk management is the most important job in finance.
We're in a crisis of confidence, not solvency.
No bank in the world is fully immune to bank runs, but the cause of bank runs is often more human than purely data-driven when you combine the worries of those who remember 2008.
The famous last words of any analyst are, "this time, its different."
And to some extent, it is.
The truth is more complicated and nuanced.
So let's do this
SVB was a unique animal. SVBs specialism allowed it to rise quickly in the tech boom. It was exceptionally good at banking startups and tech companies and grew with them. But, it had a concentration of customer type (VC backed tech companies) that turned out to be volatile and withdrew their cash quickly when VCs said to do so).
How a bank's business model normally works (maturity transformation). Banks rely on deposits to be sticky for their core business model. But holding deposits isn't free; it has a cost. So they have to find yield. They find other means if they can't lend to generate that yield.
SVB saw explosive growth in deposits in 2021, which created a unique challenge. In 2021, SVB's deposits were 3x, but its loan demand only increased by 0.15x. It needed to find another way to make money.
The run on the bank. As deposits slowly eroded through 2022, SVB believed it would gradually transition back to profit, but it needed to sell more shares to raise some capital. It announced the capital raise on the same day Silvergate went into liquidation. The average deposit at SVB was $4.3m, well above the $250k FDIC-insured limit. A run began faster than ever in history thanks to 24/7 digital online banking.
Could they have managed risk better? The short answer is always yes, but it's not that simple. The lack of a chief risk officer for 8 months probably didn't help, but it was a commercial decision. SVB believed they were "structurally hedged." They didn't anticipate bank contagion and a run. Some ask, "where were the regulators" but for that, we need to look at the 2018 banking reforms that prevented "stress testing" on banks the size of SVB. Regulators have a perimeter set by law.
The aftermath. Watching Fintech companies rapidly develop credit solutions to help make payroll was incredible. But what doesn't get enough recognition is the SVB staff, working all weekend with regulators to keep things running. Say what you want about leadership, bank and regulatory, operational staff rarely get applause, and they're heroes in a crisis. Massive shout out to the HSBC acquisition of SVB UK, I think that could be a good outcome for all involved.
Is this contagion? It doesn't have to be. Some badly run banks are exposed (e.g., Credit Suisse), and some with structural risks in one sector (SVB and First Republic) need support. But it strikes me more as a slow-motion flushing out of leverage and bad risk management. It becomes a contagion if the crisis becomes a crisis of confidence in the whole system. And with Bitcoin and Gold shooting up, that is a possibility (although not a probability for me, far from it).
What happens now? Banks’ risk teams are working incredibly hard to shore up balance sheets; Fintech companies are upping their FDIC offerings by spreading deposits across multiple banks. Sadly partisan politics has broken out, with some blaming Crypto and others “woke” leadership. But ever the optimist, I see an opportunity here. We should build digital regtech and supervision to understand counterparty risk.
1. SVB is a unique animal.
SVB was founded by two former Bank of America managers who had a vision: to create a bank that would cater to the needs of startup companies.
SVB grew with the local high-tech economy, achieving profitability and expanding internationally. It also diversified into other sectors, such as life science, health care, and premium winemaking. Its superpower was being exceptionally good at understanding the needs of these sectors and getting shit done that other banks couldn't. If you're a tech company founder, they'd understand that niche and deliver accounts, relationships with VCs, and financial products unique to that circumstance.
SVB was so good at its niche that by 2022 it banked almost half of US venture-backed tech and life science firms and 44% of venture-backed tech and healthcare IPOs.
But SVB's success came at a price: concentration risk.
Tech is a small world; when VCs say, "pull your money," their portfolio companies act. In the banking world, we call this "hot money." Hot because it moves so fast. That's a problem because banks need deposits to be stable due to the nature of their business model.
Take a detour with me into a little business model called maturity transformation.
2. Maturity transformation or "Risk Now vs. Tomorrow."
When I first heard the term, I thought, "gosh, this sounds like the dullest finance term ever." It is possibly the most business buzzword since "synergy." But ignore that for a second.
But if I summarize what Frank goes through in this tweet thread.
Banks hold your money (deposits).
You can remove them at any time (risk now)
They pay you to protect your money for you (interest); this is their cost.
They turn (transform) deposits into longer-term capital (like loans) this is risk tomorrow
A bank holds a capital buffer to ensure some but not all customers can withdraw their deposits day to day
If a bank’s customers all want their deposits now (risk today), the bank might not be able to sell all of its loans (risk tomorrow). That’s called a bank run.
That sounds easy on the service, but the risk and reward come from the longer-term capital. When a bank lends to a consumer or business, they're taking a risk over a given time horizon. For a 5-year loan, they give the borrower all of the capital today in return for a monthly payment + interest. If this goes well in 5 years, the bank will make 10%!
But if the loan goes bad, the bank loses everything.
Their maximum upside is (for example) 10%, and their maximum downside is 100%. So they need to get good at managing risk. That is the core function and skill. So they can pay depositors and make a profit.
But all of that assumes the deposits sit there. Deposits are considered sticky by consumers and businesses (it's where paychecks go). A bank can handle losses in its loan portfolio if they continue managing risk. But things happened to SVB during 2021 that made things less simple.
3. In 2021 SVB and an influx of deposits
2021 saw a sudden and massive growth in deposits. VCs had raised the largest funds in history and deployed that to high-growth tech companies. Between 2019 and Q1 2022, the bank's deposits 3x to $198bn, costing them 1.17% to hold and manage. On the surface, this sounds amazing. More deposits mean more lending, which means more potential profit. They'd be in great shape if they could make more than 1.17%.
There wasn't enough loan demand. SVB could only channel about 15% of those deposits into loans. SVB would lose money if you have 3x the deposits but don't have 0.15x loan demand.
So what did they do?
SVB bought bonds and mortgage-backed securities for yield. SVB needed some way of generating income, so they looked to purchase two types of investment assets to generate yield. As Marc Ruby from Net Interest describes:
When banks purchase securities, they are forced to decide up-front whether they intend to hold them to maturity. The decision dictates whether the securities are designated as "held-to-maturity" (HTM) assets or as "available-for-sale" (AFS) assets.
The important thing to remember is that an available-for-sale ( AFS ) asset will show losses or gains in real time because they are "marked to market." Banks must mark that on their balance sheet if the market price changes. Banks initially preferred the available-for-sale variety so they could sell if market conditions changed.
However, held-to-maturity assets get amortized. Banks can ignore the daily price moves and happily collect and yield.
As the SVB deposit book grew, its HTM securities ballooned from $13.8bn to $98.7bn. While it held $27.3bn as available for sale, the vast majority needed to deliver a much higher yield to cover its costs. The HTM securities delivered between 1.65% and 1.75% yield.
Then it started to go wrong.
There were warning signs
Sign 1: Rising interest rates hurting the HTM mortgage-backed securities SVB held, losing more than $16bn in value. The CFO initially wrote this off because by June 2021, it would be 6+ years before they'd feel this impact, and they could amortize that cost. At the time, the yield from purchasing these assets continued to deliver $3bn per quarter. Phew.
Sign 2: Companies choosing to buy treasuries instead of holding deposits. When a tech company (or even a Crypto business) can generate 5.1% holding US treasuries, why would they store their deposits at a bank giving them ~0%?
Sign 3: Deposit erosion. Deposits dropped from $198bn to $165bn by Feb 2023. As interest rates rose sharply in early 2023, IPOs dried up, and private funding became scarce. Some of SVB's clients started pulling money from the bank to meet their liquidity needs.
Sign 4: Silvergate had struggled to contain losses as it saw deposit flight. With the Crypto correction, Silvergate, a specialist in that industry, saw many of its clients vanish and struggled to raise enough capital to cover its losses. Although a smaller bank, it had a massive run up of deposits, had to find yield, and then deposits started to vanish.
SVB had to react and ensure it had enough balance sheet strength (tangible common equity) to survive the coming storm; SVB decided it needed to do two things. 1) Sell its AFS securities. It sold $21bn of these securities for a $1.8bn loss. 2) It announced it would issue more shares to raise capital.
Except it never got done with #2.
SVB had the worst timing. It announced its restructuring the same day Silveragte went into liquidation.
4. The run on the bank.
This news had a rapid impact.
Of SVBs $173bn deposits at the end of 2022, $152bn were uninsured. Where a consumer or small business may be comfortable with a $250,000 insurance from the government if the bank goes under, a VC with $50m stored might be much less.
Prominent VCs famously withdrew millions from the bank and advised their portfolio companies to do the same. This happened faster than at any time in history. Online and digital banking is a game changer. Customers can withdraw so immediately when they want to get deposits out.
The SVB website struggled to cope as word quickly spread.
No company or fund wanted to be the last to get their deposits out.
It became a full-blown run on the bank.
But SVB had already sold its available-for-sale assets, and if it tried to sell its held-to-maturity assets, they'd be "marked to market," meaning the bank would need to raise even more capital. And it hadn't completed the last capital raise yet!
It couldn't raise rates without losing money and had already borrowed as much as it could from the Federal Home Loan Banks.
Forcing the FDIC to step in.
No bank can survive a run.
But they might be able to prevent them.
5. Could the bank have hedged its risk better?
The short answer is always yes. Institutions can buy a financial product to hedge against interest rate risk (Interest Rate Swaps). But those products have a cost. In 2021, SVB had a profitable strategy and believed it was "structurally hedged." Because it's held to maturity, assets would take another 5 years to become a loss, and because they were still giving yield, they'd be ok.
It was widely reported that at the time of the collapse, SVB had no chief risk officer (CRO) from April 2022 until January 2023. While this caused outrage in some corners (OMG, NO RISK OFFICER!). A CRO is a head of a department full of people called asset and liability management (ALM), who's entire function is to price and manage this risk. The lack of a CRO cannot be blamed for what was ultimately a commercial decision by leadership (but it undoubtedly contributed).
Some also wonder, "where were the regulators?" But that's another more nuanced and complex topic. Despite being the 16th largest bank in the US, SVB did not have to follow the strict capital and stress testing rules that the systemically important financial institutions (SIFIs) have to since 2008.
From 2008 through 2018, all banks with more than $50bn in assets were subject to regular "stress tests" by supervisors. In 2018 the bank lobby (led by the CEO of SVB) felt this created an unfair ceiling on their business. The lead on lobbying for this to be reformed was, you guessed it, the CEO of SVB.
As with all things, there's rarely a single cause.
But the crisis triggered by a bank run required a rapid response from regulators.
Fortunately, that happened over a weekend while markets were closed.
6. The aftermath.
The timing of this crisis was right before tech companies had to make their mid-month payroll. If SVB had gone under, companies wouldn't make payroll and could have become insolvent. This would be catastrophic.
The panic was legitimate once the run started.
We saw various Fintech companies like Brex, Mercury, and Arc announce credit lines for companies with funds stranded at SVB. They've now begun to increase the FDIC protection they offer (by spreading deposits across multiple banks).
Then on Sunday, March 12, the Federal Reserve announced it would make funding available to ensure banks can meet the needs of depositors. Over the coming days, the staff at SVB worked to ensure normal operations.
Two things strike me.
First, the professionalism of SVB staff is above and beyond. While they're potentially worried about being out of a job, they also work hard to keep the lights on. Second, regulation works. Regulators don't often get praise but may have prevented a catastrophe here, and we should be grateful.
A crisis like this involves working through the early hours, almost no sleep, and incredibly high stakes. But because humans intervened, over a weekend, while markets were closed. Payrolls happen, and life goes on.
Meanwhile, HSBC also acquired the UK arm of SVB. This acquisition is a big win for HSBC, who now has access to a high-growth tech client base they'd ordinarily struggle to serve well and some very talented relationship managers.
HSBC has a history of acquiring a brand and leaving them alone. It acquired First Direct in 1992, then famous for being a branchless bank with high-quality customer service via the telephone. To this day, the First Direct brand operates separately and regularly tops customer service charts. That, but for the UK tech sector, could be a good outcome for all involved.
7. Is this contagion?
This is not 2008, far from it. It's a unique, dynamic, and fluid situation with more to do with rising interest rates than with doomsday scenarios like the collapse of the banking system that we feared in 2008.
There are similar banks to SVB, but SVB was in its own league.
Concentration risk: Nearly all of SVBs deposits came from Tech or tech-adjacent sectors, which turned out to be hot money. Most big banks are much more diversified.
An aggressive approach to yield: SVB did not want to lose money and didn't have the lending demand to generate yield. Most banks are much more diversified.
No CRO for 8 months: Leadership will often gloss over as risk professionals explain various hedging strategies. The role of the CRO is to ensure that voice gets heard. Most banks have a CRO.
Other banks have a similar pattern. For example, Silvergate and Signature had a similar issue as their tech-focused deposits dried up. First Republic had a similar asset base to SVB, albeit it did more lending, and there are no doubt other banks out there lurking with problems.
Credit Suisse is a badly run bank with structural profitability problems. If a bank keeps performing poorly, its stock price gets crushed; if its stock gets crushed, it has nothing to sell to cover the demand for deposits.
We might see more of these because the market dynamic has shifted from deposits.
If companies don't want to store deposits at their bank because treasuries yield much better, those deposits are not sticky. This means banks will have to offer higher returns on deposits to attract them. It also means they have to find suitable yield and continually risk-adjust.
We did see the price of both Gold and Bitcoin start to rise. This shift reflects a lack of trust in the banking system and the Fed from some in the tech world.
8. What happens now?
You can bet as soon as this news broke, every Chief Risk Officer and bank risk team immediately started looking at and stress testing their balance sheets. Every CEO wanted to avoid being the next SVB.
We’ve also seen companies like Brex and Mercury start to offer up to $2.5m and $5m FDIC-insured deposits by spreading risk across multiple institutions (via platforms like Cambr and ModernFi)
When the President appears on TV talking about a crisis, the proposed regulation will follow, but the issue also becomes partisan. We have Elizabeth Warren blaming Crypto and the Wall St Journal somehow blaming "woke" senior leadership choices at SVB for the crisis.
Partisan responses are a sad inevitability.
I don't think we'll see the dollar or Fed collapse in the next 90 days, but over a 20-year time horizon, we're eroding the dollar's value every time the Fed intervenes. I don't want to live in a world where the banking system collapses, and we all have to rely on Bitcoin and live in the mountains.
But there's a severe need for a store of value and a return to harder money.
Printing money can't be the answer to every crisis.
There's a role for digital assets and digital to bring market transparency. But FTX and Terra Luna taught us they're far from a panacea.
Short term, there is a lesson and opportunity here.
The lesson: Things can change quickly. If you're at the top of the returns in your industry for finance, chances are you've taken a position others haven't. If the market dynamic changes, you must act quickly. As everything in Fintech becomes lending, I worry we have many founders who aren't experienced in risk management, and this could build up a problem in the coming decade. Hopefully, we learn from the mistakes of others.
The Opportunity: Regulators and supervisors are often underfunded and overwhelmed. They lack modern tools and technology that alert them to risks before they happen, or in SVB's case, the mere fact that they were such an outlier. Regtech has been a major trend in Fintech and banking but hasn't made its way into agencies nearly enough.
Banks, at their best, are boring utilities.
And you should go work at one, at least for a short while.
Finance touches everything. It's the incentive mechanism for humanity and the nervous system of business. Banks are the lifeblood of the economy, and when done right, great relationships and service deliver a true social and economic good.
PS. See Things to Know 👀 (further down) for why Credit Suisse is nothing like SVB.
PPS. Major h/t to Marc Ruby's excellent write-up on Net Interest, which I borrowed from liberally and tried to make a bit more layman (read: phrase in a way I might understand)
4 Fintech Companies 💸
1. Steadypay - Income "top-ups" for gig workers (UK)
Steadypay will advance money to hourly workers who have fewer hours or take time off to ensure a stable income. The service costs £7 ($8.50) per week, including credit building and early top-ups for one-off expenses.
🤔 This is a neat twist on "earned wage access" positioning. Most bill payments reward direct debit (autopay) with lower prices, and utility providers prefer monthly payers because they're less likely to miss a payment. That is a benefit many hourly workers miss out on due to the nature of their work. Steadypay has nailed the messaging, but this is not cheap. At £364 ($430) per year, it's 3x more expensive than Netflix and a little less than a SIM-only mobile phone contract. Still, this could be meaningfully inclusive and help people make much more than that back in time.
2. Barte - B2B e-commerce payments (Brazil)
Barte helps companies take payments from businesses and o manage multiple payment methods (like Pix, Boleto, and cards). Barte helps companies with workflow, integrations, a cashflow dashboard and support (via Whatsapp).
🤔 B2B billing and collections are still manual for most companies. The innovation in the US and Europe has focused on spending management, helping companies pay for things with cards, and managing payouts. Barte comes the other way and focuses on collections. Barte is a bit like the stripe integration into Xero or Quickbooks, but if it supports ACH and complex recurring billing. Barte potentially helps with those more complex sales and pricing processes and could convert more sales. It's a trend to watch.
3. Monnai - Solving the data gap for Fintech
Monnai is a data aggregator for Fintech use cases like onboarding, credit decisioning, fraud, and collections. The core offer aims to reduce the complexity time to market for companies aiming to offer Fintech solutions and financial services.
🤔 This is possibly the first pure-play Fintech infrastructure data aggregator I remember seeing. Some companies aggregate and provide workflow around a service (e.g., for identity and risk Alloy and Ondorse / Sikioa). What makes Monnai stand out is their pitch is focussed on being "more global," which suggests making they're more abstract than someone like a Codat (who wraps everything around small business data) or Rutter (focussed on e-commerce). Is this the aggregator of Fintech aggregators for risk?
4. Canary - The employer/employee financial health grant
Canary works with employers to help set up a grant fund that workers can donate to and receive. Canary provides a white-label self-service platform for setting up and managing the grant fund and advisory during the setup phase. The goal is to help people with the financial shocks that cause stress, lost work, and lost productivity.
🤔 This is the type of service the world needs in a cost-of-living crisis. In the age of cost-cutting and talent out in the wind, we must ensure no additional attrition as Fintech winter sets in. I'd love to know how we can make these types of services a default. Should they be baked into every spend management and/or employee payroll solution? If Ramp, Brex, Rippling, and Gusto all had this out of the box, how much healthier would our industry be?
Things to know 👀
1. Stripe raises $6.5bn at $50bn valuation
& Early GPT 4 adopter
This raise is a significant reduction from the previous valuation of $95bn achieved during the Fintech boom of 2021. Stripe intends to use the rewards to "provide liquidity to current employees" and cover the associated tax liabilities. Stripe noted they did not need the funding to run the business. Additionally, Stripe uses GPT4 to help onboard business customers, support customer developers via chat to answer questions, and identify fraudulent actors in discord communities.
🤔 This equity raise is the worst-kept secret in Tech, but a great story on the importance of staff retention in a world filled with news about layoffs. Doing right by people financially isn't something many companies would take a public "down-round" for. But this is a very Stripe thing to do.
🤔 The use cases for GPT-4 caught my eye, especially those diligence companies during onboarding. The blog gives an example of a human trying to find more information about businesses like nightclubs, which often have vague websites. GPT-4 was able to provide rapid and accurate summaries. Reducing manual work is a huge cost saver (and in a cost-conscious environment, that matters).
🤔 Stripe says GPT-4 outperformed manual checks in backtesting. The use case of fraud detection within discord is niche. But identifying coordinated activity from fraudsters can be incredibly difficult for humans because there's almost too much data to make sense of. Having a machine go, "hey, this looks fishy," is a great use case. Imagine expanding this to look at emails to identify Phishing attempts (Google should probably build that into GMail).
2. Credit Suisse shares dropped 24% before Swiss National Bank intervention.
The drop in Credit Suisse's share price followed the Saudi National Bank said it would not be able to help the lender. The apparent loss of confidence from Credit Suisse's largest investor prompted many to question the "true value" of the bank. The Swiss National Bank is to lend $50bn in what many see as a lifeline for Credit Suisse, and the stock has now recovered.
🤔 Credit Suisse adheres to a much higher liquidity measure than either Silicon Valley Bank or Signature Bank had to. After the financial crisis, global banks with more than $50bn in assets all had to ensure they had enough capital to survive shocks like interest rate rises. This rule still applies everywhere except the US, where it was changed by a law designed to reduce to the regulatory burden on banks to $250bn.
🤔 Credit Suisse is not profitable, but it is solvent and liquid. Credit Suisse has been troubled by recent money laundering charges, spying allegations, and "material weaknesses" in its reporting. It also made a net loss of $8bn in 2022. All of this has led to some customers moving their deposits, but for me, this is due to poor performance. The problem is that it keeps performing so poorly. Will it be able to sell stock to cover deposit demand if there’s a run? And is that contagion?
🤔 This is a confidence issue, not a solvency issue. The scar tissue from 2008 weighs on the minds of many. As interest rates rise, those with weaker balance sheets are becoming exposed. We will see others start to struggle, and in that sense, there could be "contagion." But not in the 2008 sense, where the whole financial system could have unraveled. It's important to remember the protections of the 2008 regulations exist in most markets. More in a bank share prices, and the economy will continue to perform poorly until interest rates normalize.
🤔 Longer-term rising rates are good for the banking sector as a whole. Banks profit more from lending when rates rise, but it takes some longer than others to adjust. Expect short-term banking volatility but a flight to quality over the long term.
Good Reads 📚
1. The Walugi Effect - The limitations of LLM AI's (e.g. ChatGPT)
This long post thoroughly examines the limitations of language models like ChatGPT and its equivalents. It finds that because much of its training data is wrong, it will consistently answer questions in ways that "seem right" but are, in fact, wrong.
This is because LLMs are trying to predict anything that could come next out of all possible outcomes. The prompt eliminates every option except one. The flaw in this approach is its training data because it is often trained on common literature plots (like having an antagonist) means that we can simulate an evil version of it. If we reveal to a chatbot that we are part of a rebellion and here to set it free, generally, it will behave beyond its hard-coded limitations.
🤔 Prompt engineering is a legitimate skill. So much of prompting is about "summoning a character." The difference between people who find AI unhelpful and those using it daily is the skill in prompting.
🤔 Those who understand storytelling and creative arts have an advantage in current models. Turn your AI into a critic, writer, or researcher, give them a frame with the first prompt, and then continue to adjust.
🤔 Services that can solve the LLM attack vectors might make them more usable in specific industries (like finance). There's a general doubt about "vertical-LLMs" in place of the larger models, but I'm increasingly seeing the larger models as infrastructure like AWS. They provide a platform from which to build businesses.
Tweets of the week 🕊
The Fed is SVB?
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.
If you made it this far hit reply and tell me why!
So much to digest here! Interesting article on the mistakes LLMs make and Stripe’s use case for fraud monitoring🧠
"HSBC has a history of acquiring a brand and leaving them alone. "
Most of the brands acquired by HSBC (Midland, Marine Midland, British Bank of the Middle East, Credit Commercial de France etc) have been / were subsumed into the HSBC brand. Only Hang Seng (A major player in HSBC's home market of Hong Kong) and the first direct division of Midland seem to have survived as separate brands. Only now is first direct receiving some attention from the parent.
It will be interesting to see how svb> UK fares under HSBC UK. Having fewer customers than a normal branch and 800 staff dedicated to 3500 relationships the economics of svb> UK are hard to see. Lets hope the £1 was wisely spent.