Future of the Industry

It's not quite that simple, as just replacing people. The good thing about human capital, in any industry, is that it scales quite linearly. You can bring in someone new whenever you need new stuff done. Like, if you bring need to fill X amount of capacity, you can reasonably assume that Y amount of people can handle that. This is very much a solved problem. MDs and hiring managers know how many people they generally need to fulfill any task. 

Ai is a bit different. When you replace 5 people with Ai, it isn't that expensive. But replacing the next 5 comes at a significant cost increase to that of the first 5. Again, o(n^2) complexity. It's also generally the case that Ai can speed up the productivity of a person, but outright replacing that person is something we haven't seen quite yet. Companies will say they're laying off folks due to Ai, when the reality of things are that they'd be laying those people off anyways and trying to replace them with Ai is a simple stop gap to do things.

For example, the biggest hedge fund right now trying to build out a full Ai system is Man Group. They've been working hard on agentic ai to try and replace quant traders and researchers. But this seems like less of a "We're doing this because it's the best decision" and more like a "We've fallen really far behind and need to lay people off, so we'll say we're using Ai to replace people so investors don't panic". It still seems like most people pushing Ai as a replacement for actual human work are doing so in an effort to mask their inefficiency as a company. 

There is also another element, which is future thinking for these companies. For example, at a lot of IB firms or tech firms(bringing tech in because the entry level argument is particularly strong here) the entry level candidates don't actually make up for their salary. You have to recruit them, train them, and hope to get a decent amount of return out of an entry level banker or software engineer, or really in most other areas. Entry level workers aren't going to produce something amazing that will make a ton of money. For the most part, entry level workers are a cost center. So, why do firms even bring in entry level people? Well, simply put, because they need senior people, which do not occur naturally in nature. They grow from juniors. A lot of MDs at Goldman started as analysts at Goldman, and they make 10x for the firm what they would've cost the firm during their analyst stint. Most firms are willing to pay hundreds of thousands to an entry level person to do menial work in the hopes that they manage to hold on to a few people who make genuine impact on the firms bottom and top lines. It's simple long term planning. All the firms cutting headcounts at the entry level will end up losing out on a lot of future talent pools. This is why "useless" interns still exist. They're not being used to run any business. They're purely there so the firm can pick the few they think might make good impact on the business. Hell, part of the reason most banks have been okay training entry level talent for PE for so long was because once those people became senior in their PE firms, they would bring business into the banks they worked for. It's all about bringing in good future business. That's the game for a company. Prepare for short term and long term. 

So the whole idea of all entry level work being dead is likely overblown. Eventually political and business leaders will realize that replacing juniors with Ai would mean no senior people, and that's the death blow for any business. Ultimately, there will always be work no one wants to do that you can pay someone a couple hundred thousand to do. It might not always be spreadsheets and powerpoints, but it might be monitoring models in the future. 

I like to think of it this way: Back in the 80s and 90s, the hottest job on wall street was as a trader. Everyone thought trading was the hottest job out there. Then, the dot com really came into its own, and all of a sudden there was a real fear of automation. And for a large part, that did take over. Trading desks at banks are less populated. S&T is second to IB in many people's minds. But, that automation made way for a new kind of trader: Quant traders. Without this automation, prop firms from the 80s, 90s, as 2000s never would've been able to grow into the behemoths they've become. Now, quant trader is probably the hottest job in finance. It's come full circle. Same thing with IB. Automation of some things, sure, but what quants do is monitor the models and puppet with them, along with working with greater data sets and alternative data. Perhaps, we'll see the same thing begin to take shape with IB. More and more quantitative desks and firms pop up, utilizing Ai and human talent to create an ecosystem that can service clients in a more efficient manner. But that's all theorizing. Anything could happen really. 

.

I don't agree in general. For one, that's just not how math works. Right now, CoPilot and other LLMs are limiting people to how many responses they can use, and getting more and more constrained on costs. Why? Because, like I said, things don't scale linearly. They scale quadratically. And a big part of that is also training costs. Models get more and more expensive to train, and so to improve a model a small amount costs potentially billions of dollars. 

I am more than sure that the costs of AI are way cheaper than a junior. I mean, $200 (2.4k p.a.) not enough based on what the average user does, so let's say what might be acceptable per year? 30k? 50k? 70k?

You're comparing the costs they charge one individual with limited input tokens across multiple context windows to an enterprise where the context window will be massive, the input size massive, and the need for safer training and security massive. One prompt right now likely burns a hole in OpenAi's pockets. Eventually, these firms will need to charge the actual price of the models. How much is that? Well, OpenAi states they need to 40x their revenue over the next 4 years to become profitable. This is assuming no major competition, no major regulation, no major changes in hardware. That estimates revenue at 125B to break even, which is more than 2 Oracles worth of revenue to break even. Keep in mind, model costs and training data are becoming harder and harder to improve on. And also, keep in mind, OpenAi also probably wants to profit, so they'll want to see more than that. How much does that really mean for enterprise customers? Well, we know that currently, no one is really using enterprise. Everyone is using copilot, because it's being forced down their throats. Realistically, we don't know how much Ai will cost an enterprise, but we can generally say that JPM or GS will want a whole model build around a customer with custom training data, and as a result you run into that O(n^2) problem again. 30k a year is nowhere close to what it will cost to make these models really strong enough to truly "replace people"

Besides, eventually it will become cheaper. Everything is expensive to set up, the infrastructure/energy costs, because the focus was making the idea work. Once the idea is solidified, the focus will shift towards how to be more cost efficient. It's a business. It's just a matter of time.

Not how it works. Training AND usage are both incredibly expensive. Infrastructure is costly, but I feel like you just completely ignore anything I said about the primary cost of Ai having a quadratically scaling multiplier, meaning every new input of data requires n^2 new GPU compute power. It doesn't scale the same way other tech businesses do. It's just not the same. 

AI will replace juniors. It's just a matter of when with a cap on max. 10 years.

When has it ever been the case that technology wipes out people? IB has been around since the civil war. In that time, we've seen technology boom to the point where now, you can load a 10k into a DCF and have it basically filled out in seconds. You can do that without complex Ai. You can do a lot of the things "AI" is great at, with very simple tools. Hell, if it was all about costs, you'd outsource everything to a sweatshop in Hyderabad and never worry about a thing. That's just not how businesses run. There will 100% be work to be done by entry level people. Otherwise, there wouldn't be senior people to run things, and most companies know this. Also, if bankers do get AI to speed up workflows, it means they can do more deals faster, meaning more deals get done, meaning the company has more money and capacity to do deals, meaning they can hire more people. That's a simple principle that has played out throughout history. There are so many examples: The cotton gin, the printing press, the car, the internet, computers. All of them supposed to change humans and make it so that people no longer had to work. 

I'll try to respond as best as possible to each of these statements

One thing I'm not sure you acknowledge is the rapid cost reductions (not dissimilar to what were seeing on the hardware side of humanoid robotics). For AI, the LLM inference prices have dropped meaningfully, dropping as much as 10x per year, as is evident from the A16z chart below. I think it is safe to say that if  this trend continues, cost may not be so much of the bottleneck that it is right now for more widespread adoption. Transparently, I don't have the slightest clue if this trend continues, but I do think there is a >50% chance it does.  

A big reason behind this cost reduction was "easy pickings" or "low hanging fruits" as massive leaps from the initial cost downwards were primarily just due to software optimization and minor model tweaks. Also, using a constant metric like MMLU score of > 42 will of course lead to improvements when benchmarked, but the MMLU benchmark is just one benchmark. The main issue for these LLMs is, however, still the quadratic nature of LLMs. You can model it out like 
np^2​+∑(i=1)(nr)​​(np​+i)1
which looks weird but I promise is more simple than it seems. It's modeling out how input tokens scale quadratically and outputs scale slightly more linearly(which is why output is always going to be cheaper). For a simple example, let's use 10 million tokens, 20 million tokens, and 30 million tokens. 10 million tokens will come to a relative cost of 1, while 20 million to a relative cost of 4, and 30 Million has a relative cost of 9. If we started to see the real costs of Ai get charged, that could be a real barrier for large corporations that would likely need billions of tokens. As of now, we're still seeing the early days of the Ai business model taking form, where no one has figured out how to make it profitable. 

AI is likely pretty far away from being able to replace IBD analysts, but I think revisiting this topic in 5 years might be worthwhile. Sure right now we are replacing call centers with chat bots, but I would guess (based purely on anecdotal evidence) that we might be further along at handling complex tasks in the near term. Not sure if you have ever used AlphaSense, but you can use its AI tool right now to search for a public company, and have it analyze all SEC filings within the past year to highlight a topic of interest, say R&D expense trends. Do you think engineers will continue to increase the efficiency at which these types of tasks can be handled?

So, the biggest barrier for something like AlphaSense(which I have used a few times, it's neat and well put together) is going to be figuring out the best model to use. For something like looking up a 10k and summarizing, it's no issue. The chatbot is already trained on that data, so it already has those weightings built in. That is a linear cost. The problem comes about when you try to build out full teams working off the same chat or idea. A big problem with Ai is again, the quadratic nature of scaling. So, you can try an SSM model, but that has the major issue of an index page that can lose data, or have important stuff overwritten. Say, for example, a small passing comment is made on an investor relations call, the Ai doesn't weigh it properly, doesn't add it to the index sheet, and that thing ends up being a major problem of some sort. Well, then your blackbox model has spit out something inaccurate. So there's a lot of issues with these types of companies and Ai models because outside of researchers, most people cannot understand wtf is going on behind the curtains. But more importantly, a major issue for these models remains hallucinations. What's the point in having a whole Ai that can get key points from a 10k if your analyst needs to re-read it anyways because the Ai kept hallucinating? 

I believe that at some point the benefits of corporate adoption of AI to replace jobs in many departments will outweigh the costs, do you agree? If so, what is your rough estimate at when this will occur? I have heard several corporations talk about back office automation over a 5 year time horizon. Admittedly, this is far from automating IBD analyst roles, but it should serve as a benchmark for AI improvement and job displacement. 

Okay, so, generally speaking, the only areas I see getting full automation right now are in areas of very menial, repetitive tasks. But also, I made this point elsewhere, which is that Ai has been around for 3-4 years and we're already seeing companies talk about how they'll use it everywhere, but, well, you can hire people for the same cost as a full Ai system might cost. They're called Indians. There's millions of them that would take 1/10th the paycheck to work at a bank, and they'd be more than willing to work the crazy hours too. Generally speaking, IME, entry level people almost never produce enough to make up for their salary. Even very talented people aren't worth it to many firms, because they know they'll have to spend months training and getting someone fully up to speed before they're really ready to "take the reigns" and by then, they might've left already. The biggest reason you hire entry level people is because that's the best way to get future leaders of your company. Not just CEOs, but MDs, VPs, and all other sorts of people. You really don't want to replace all your entry level people, because then you have to pay a crapton in the future to bring in experienced talent, and experienced people also cost more to train, AND experienced people tend to leave at higher rates. When I was at a prop shop for 3 years(SIG, DRW, etc.), almost no one left. Extreme loyalty to the firm, and the firm had great loyalty to them. I'm sure we could've outsourced a lot of the work, and I'm sure we could've hired less entry level people, but almost all the people I know who joined the firm out of College have remained there. 

I'd also wager that most companies committing to Ai are still in the 80% of the 80/20 rule. Here's the reality of things: JPMorgan and Goldman Sachs are never going to have good internal Ai systems, because they offer 1/3rd the pay top tech and quant firms do, all while having a culture most Ai researchers won't join. They're currently working with second tier engineering talent building out systems, and while right now they're getting those easy pickings, these infrastructure projects need constant supervision, updates, training, oversight, and risk management, all of which will cost increasing amounts of money. So while OpenAi or Google(probably Google, they're putting out stuff WAY better than OpenAi is right now) can probably figure these things out, banks won't. It's a major disadvantage to them, but that's besides the point

And speaking more broadly to many points made by you, they're not necessarily wrong in any way. Ai has improved a lot, and I'm reading a lot of the new papers on reinforcement learning that has me in awe of what some of these researchers are accomplishing. But speaking more broadly, I really think Andrew Ng is right here. I've met him one time, brilliant dude, and generally speaking his take is that menial work will change and enhance workers in a lot of ways. His example was the internet, and how something like spreadsheets didn't eliminate accountants, it moved their tasks, and that's generally how I'd see things too. A lot of work will be automated, but I don't think IB, or for that matter most jobs people actively want will be.

It's important to segment out your tasks for your job. There are a lot of tasks you have that can be automated without your job being replaced. Ai is great at tasks, less great at jobs. So try to embrace it early. Figure out how to be the expert on Ai. Don't fear it, fear leads to rejection which leads to worse outcomes. Try to understand what is happening with these models. 

The points you make are fair and accurate for the present moment but you are not accounting for the fact that the models are improving at an exponential pace. All the literature I have seen suggests they will be fully capable of making the kinds of judgments and analyses juniors in banking are tasked with within a decade, leaving only senior roles / relationship managers as having any kind of utility.

Idk what papers you're reading, but I've seen far too many papers come out of top research labs putting doubt on this. Most research papers I'm reading suggest slow progression in most areas, and the more recent "illusion of thinking" paper was extremely damning. Also, you're comparing utility to capacity. Yes, these models could have the capacity for extremely in depth analysis and thought, but in the real world utility matters so much more. When requests are ambiguous, Ai struggles. Also, this point, of "Leaving only senior roles" makes no sense at all, because senior people are not naturally occuring. They become senior after years as a junior. So expecting a business to operate smoothly when you replace your entire talent pipeline is like the Nets trading away their entire future first round haul for Kevin Garnett and Paul Pierce. Short term, it works. Long term, it's terrible. 

Hell, we're seeing companies that previously committed to using Ai have to switch back to people. Klarna is a real world example of this, and over the long term you'll see more and more. 

In my humble view, I believe public markets investing will belong completely to the pods who will have models informed by algos and AI that make beating their performance impossible. One asterisk is for the exceptional single managers (e.g., Tepper, Hohn, Ackman, etc.) who have demonstrated they can beat the S&P 500. Unclear if they will also die off over time if they struggle to match the performance of Citadel, Millennium, Two Sigma etc. 

This is probably false. Every Ai system would be trained on the same data, same papers, same researchers, and so they would likely make the same decisions as other Ai's. If you're going for a more rules based system, then you've already got that in algorithmic trading with a flare of ML. Markets always adapt around what is getting the most capital inflow and not. Right now there's more money in the short term markets than ever, and this means there's a ton of mispricing in the long term. If you put a ton of money into an Ai system, it will mean anything Ai considers bad based on training data and biases will automatically have a massive mispricing. 

Private equity is less clear to me - it's a strategy that depends on capacity for underpaying and over-levering, but given the amount of capital all chasing the same opportunities, I don't think the same returns that were achieved in the past are going to exist in the future. Can you live a comfortable life as senior MD at XYZ MF, definitely (assuming you survive the grind and aren't pushed out for whatever reason). Can you become Schwarzman over the next 50 years? Almost definitely not. No one knows if rates will go back to the level of 2010-2020 in the future, but everyone experienced who I have spoken to about this believes the entire industry is cooked for the forseeable future; carry is just not going to be there for some time and market is way way way too saturated with talent / powder.

I don't disagree that PE is in a rough spot right now. Far too many people entered in the 2010s, and right now PE firms are trying to hold onto as much capital as possible due to lackluster exits and such. But that's a short term thing. If people start to leave PE, and PE shrinks, it means less private buyout deals, meaning more opportunity, meaning more people enter. It's market cycles. 

I think there will always be a need for two types of people in finance: senior MDs in investment banking who are good sales people and are trusted by other human beings to give good financial advice and senior principals in PE who need to make ultimate judgment on whether to buy an asset at X price or not, regardless of what the technical analysis might say (after all, a lot of what we do in PE is just backing into things to make a partner who has already made their mind up happy, one way or another). 

This is a misinformed view IMO. For one, both senior MDs and senior principles didn't just magically get there. They got there from decades of experience at the lower levels. But also, your assumption is just that these models are set and forget, no issue, never hallucinates, when that couldn't be further from the truth. It's far far more likely that analysts work with Ai models, rather than being replaced by them. 

usernameladiesman217

I'd push back a bit on this notion. Not the notion of security, that much is a guarantee, but that these LLMs in house can't catch up. Sure, at big banks they'll 100% stay behind, as tech firms regularly pay more for better talent, and finance firms are seen as "second class" compared to tech firms for the top AI talent, but that's not who would win. It's much more likely, at least IMO, that a firm like SIG or Jane Street could open up an advisory arm, and use their extreme technical prowess and talent to develop AI systems that CAN actually compete with top research firms like Anthropic or OpenAi. 

I've thought for some time that the big banks are losing due to their inability to adapt. They're losing MM volume to trading shops, research volume to external firms and specialty firms, lost banking volume to the EB, asset management AUM to passive index providers and pure play asset managers, and are actively losing client volume and younger retail banking clients to fintech firms and neo banks like Chime and SoFi. 

The next revolution in banking isn't coming from Goldman or Morgan, it's coming from some nerds teaming up with former bankers to make a system that works better. Being a diversified giant isn't as appealing these days. 

100% agree that the smaller, more nimble shops have much more room to work with. In fact, I already know that there’s smaller shops not regulated by FINRA that are letting their people have free reign over AI use, especially the research oriented ones. My friends at some of the FAANG companies are freely allowed to use any LLM they want at work too. I just think anything that has to answer to FINRA is never going to progress to the point where AI will replace jobs. They cant outsource their modeling work to third parties either because they would require inputs from confidential client information.