How soon will AI fuck our industry?
What’s your take on how soon headcount across the analyst and associate level will decrease simply because of developments in LLM’s? Further, with AGI possibly on the horizon, is the industry facing a paradigm shift that will effectively wipe out the industry as a whole save a few rainmakers at each firm for business, 1 or 2 VPs to guide AGI to do the job, and then maybe 1 or 2 analysts or associates that do jack shit but that the firm artificially keeps to preserve a pipeline? Thoughts?
No.
lol would love more thoughts… what makes you say that?
Because we still haven't invented a fucking spreadsheet that can save in under 30 minutes when data tables and circulars are involved. And my god damn Remote Desktop still freezes tf up whenever more than 20 people are on a Zoom together. I'm definitely down for this scary breakneck-speed advance in technology that we've all been waiting for, for the last 5 years.
Riveting analysis.
"near the singularity; unclear which side." iykyk
Headcount won't be materially impacted any time soon. It will be like 30-40 years at a minimum. AGI isn't anywhere near the horizon. LLMs all have the same fundamental problem (insofar as it means "fucking" the IB/PE/VC/etc industry): they are nowhere near capable of being able to handle open-ended tasks with multiple potentially valid answers based on "bespoke" knowledge that the typical analyst/associate has to do.
Technical competence is much harder to learn than merely being able to anticipate the correct word in general English. There is a reason why AI providers in technical fields such as the law understand that the hardest part of getting a useful "ChatGPT-style AI" for their fields isn't things like hallucinations or the lack of data but how the AI is able to interpret the data. To go back to law, knowing what a court ruled is one thing. LLMs can do a moderately good job at this. But that is not really the thing you pay a lawyer to do. What lawyers really get paid for (at least in regard to this narrow part of practicing law) is synthesis. Is what the court said dicta that can be ignored? Was this overturned in a different case? Are the facts slightly different in the current case which means we can ignore this ruling? Is there a new law that changes the ability to implement this ruling? Those are things that LLMs are trash at now. If you look at legal AI providers, they know this. THIS is the biggest thing those tech companies are working on. THIS is the real gating item for LLMs taking over jobs in fields like IB and such. It isn't regurgitation of information but the analysis that LLMs simply can't do at any meaningful level.
Until this gets solved, LLMs are going to be a tool that helps facilitate work in fields like IB. LLMs are not going to be materially replacing any jobs.
I know someone at one of the big AI startups and the consensus belief there is that AGI will happen by 2030 if not earlier. You are living under a rock or are willfully ignorant of the advancements in this space if you believe that AGI is 30-40 years away. One could make the argument that AGI is already here depending on how you define it. Mass technological unemployment is already starting to happen in my opinion. Even if this technology only makes people more efficient and increases their productivity 1-2x, that will still result in massive layoffs - just look at what’s happening in Big Tech.
In my experience, the only people using AI are the junior employees that this technology is supposed to replace. I think layoffs in an industry like tech aren’t as comparable where everyone in that industry is tech savvy (plus one could argue those layoffs aren’t tied to AI but over-hiring). We are still dealing with MDs that are living in the stone ages that struggle to adopt new technologies and want things printed on paper.
I think adoption in knowledge based industries outside of tech may not roll out in the same manner simply due to stubbornness in their old ways. I see it rolling out more in a manner where the VPs and below today that are using AI slowly take over from the old guard and those future MDs need less associates and analysts. Not sure I really see mass layoffs happening, just fewer future job openings. Alternatively, while each individual company may need less people, it will also be that much easier to start companies with the help of AI and therefore more companies are launched offering more jobs.
In any case, banking is still an industry that requires you to convince someone to buy something (sales) so I don’t see it replacing humans entirely. Certainly puts less of an emphasis on technical skills and more of an emphasis on people skills… modeling skills effectively become valueless.
Who knows, if I’m wrong on all of this and get canned and can’t find a new job because of AI I’ll just go get my bartenders license and live out my dream in a small ski town.
(1) "I know someone at one of the big AI startups and the consensus belief there is that AGI will happen by 2030 if not earlier."
This is a fallacious appeal to authority which is an illogical argument. Illogical arguments are definitionally wrong. Your argument is not even wrong. If you are going to make the argument that AGI is coming soon, can you please make an argument that has the ability to be correct? At this point I am not even asking for a correct argument, just one that has the ability to be right. Please.
(2) "One could make the argument that AGI is already here depending on how you define it."
No. AGI is not here. The only way you can say that AGI is already here is if general intelligence is a vacuous concept.
(3) "Even if this technology only makes people more efficient and increases their productivity 1-2x, that will still result in massive layoffs."
(i) What do you think "massive layoffs" are?
(ii) What do you think happens if it leads to ~30% increases in productivity?
(iii) What do you think the adoption curve will be in the fields that we are talking about?
(4) "just look at what’s happening in Big Tech."
Business cycles are real. Big Tech acting as if business cycles don't exist and then having to deal with the consequences of them existing is not the argument you think it is
The people who have tons of unrealized equity and investors who need exit liquidity at some point all believe that a transformative technology is on the near horizon? And these [people/companies also are burning through cash at record rates and need to justify that? Yes, I am sure they will offer an unbiased view.
Would you recommend doing medicine?
Okay I'll bite because posts like this drive me insane. I am one of those people who worked in the type of AI startup you're describing. I can tell you from personal experience, the guy you responded to hit the nail on the head. If I were to simplify the analogy for you: if I give you A, B, C, you can think of the alphabet, you can think of cryptography, you could even suddenly remember your tax ID number for whatever reason. A computer needs some baseline logic on how it should interpret the A, B, C. It cannot dynamically come up with its own frameworks on how to see the world.
In our case, the models we custom built would work only after massive amounts of learning was done. And then we'd have to plan on a data refresh and recreating our base set of variables used in our models to account for inherent biases that would sprout up over time because surprise, the world isn't static. I actually left the company in part because I could see that this was a massive, massive bubble (at least in the way people are hyping it up).
Your expectations are way too high for AGI, especially when it comes to LLMs. Try using them for a while and their luster dulls significantly. The founders promising the world in 5 years if you’d only give them more money are either naive or predatory. Tech approaching AGI probably won’t be an LLM.
Garbage in, Garbage out
Prompting is hard, but you can completely mitigate all the cons that you've outlined by piloting the AI workflow correctly.
AI can currently do a great job at everything except provide those insights or synthesis. What AI can do very well right now is create entire roadmaps and workflows with some prompting so that any average joe can follow the steps that are outlined. Although you need to guide the AI through tougher problems that require some human touch.
It gives you a complete problem solving roadmap for any situation. That skillset is completely devalued now. All you need is someone with enough experience, or in the case of coding someone who understands the code base on a deep level to pilot the AI through areas that need a bit of guidance.
In my opinion, tech is one of the hardest knowledge areas to apply AI in, and there are already large production quality applications at the level of giant companies being created right now with just a few people
JL12: LLMs are "nowhere near capable of being able to handle open-ended tasks with multiple potentially valid answers based on bespoke knowledge that the typical analyst/associate has to do.... what the court said dicta that can be ignored? Was this overturned in a different case? Are the facts slightly different in the current case which means we can ignore this ruling? Is there a new law that changes the ability to implement this ruling? Those are things that LLMs are trash at now." (Internal quotations omitted).
Analyst 2 in IB: "Prompting is hard, but you can completely mitigate all the cons that you've outlined by piloting the AI workflow correctly."
No, you literally can't. To stick with the legal example, you can't prompt your way out of being able to identify what is dicta using today's LLMs. You physically can't prompt your way out of the issue with future courts overturning the ruling with today's LLMs. They just are not able to do this with any time of meaningful accuracy.
You can't mitigate the LLM not understanding what it is looking at merely by "piloting the AI workflow correctly." It is a fundamental design issue. No amount of prompting will get these LLMs to recognize a Supreme Court ruling in 2023 limits a 1960s appellate court standard that applies an 1860s law. The information is just not sufficiently connected for LLMs to draw those connections. Modern LLMs pretty much require you to already know this information so you can correct the LLM. That doesn't save time nor mean that the LLM is a threat to replace the actual human any time soon.
And this is not me merely stating this. All I am doing is telling you what published research says.
buddy, you work in corp dev. you don’t have to worry about AI taking IB jobs hahaha
Okay...?
This ad hominem would only happen on WSO lol
Too many finance employees are in denial about this because they have too much to lose, so you're unlikely to get a good answer on here.
Purely technical skills will just become devalued in the marketplace IMO (as is already happening with the number of unemployed computer science grads). This will be a bit of a reversal of the trend of the past few decades (where technical people studying com sci/econ/finance/engineering were rewarded handsomely while non-technical graduates were left unemployed) so it's not a catastrophe.
AI's impact is not going to happen overnight, however. It will be a steady drip of employers increasing their power over technical workers, as more workers compete for fewer jobs. That's how labor markets always work.
If you have good people skills and work in a job which requires them, you're far less prone to automation so you will face a better labor market and more people will be willing to pay more for your labor. Simple as that.
What do you think the current state of LLM's ability to analyze information is?
Excellent. Better than most humans, and it won't be long until it's better than 99% of humans, and then 100%.
AI agents are the real thing which will lead to job losses. Think of the number of back office and middle office staff in banks and financial institutions whose job it is to build simple financial models everyday or look at financial ratios or dashboards. An AI could probably do 90% of their jobs already, you just need the right tech and software integrations.
If you still don't think this would happen, think of banking 20 years ago. Many finance jobs - bank tellers, data entry clerks, accountants - have disappeared. The next wave is coming.
We are only seeing a wave of unemployed computer science grads because it is the “hot new industry” (+\-5 years).
All of these CS grads are going through the “average” finance major trying to break into IB experience. The majority don’t make it and are shouting failure. Reality is the average finance major can work back office at any company in the world and does not expect IB pay outside of IB roles, while the average CS major cannot work at any company and expects FANG pay everywhere
AI has not replaced CS majors. Also I can tell you from multiple tech deals, AI is not a feature displacing people until 2030+
This
Yes the denial is insane. Can't believe they're in high finance yet lack critical thinking.
agree with the other commentor. you are all wrong lol. junior employees in finance are gonna be wiped out lol
this
Working at a BB and it is already happening. Shareholders don't care if MD is comfortable or not about it.
Can you elaborate?
Once admins, customer service reps, legal assistants, etc are replaced with AI chat bots, then you can start to worry. Next will be mid-skill jobs like paralegals, then you can really worry. But as long as the paralegals and mid skill jobs are there “supporting” the high-skill jobs, the high-skill jobs should be relatively non-impacted.
I have GPT premium and once tried to use it to scan some spreadsheet screenshots and convert them into excel... it f***ed up every other cell and sometimes just completely hallucinated a whole new column or row. I just don't see it doing modeling for a very, very long time.
One thing I've learned in my short career is to never bet against technology. Already happening at my fund. Solutions are crude but I have ChatGPT open on one monitor 24/7 and it gives me at least an hour back every day. Agentification/tool use will eliminate the need for current analyst tasks, but I'm sure we will find other uses for them.
Some Examples:
1) I feed 100+ page contracts into GPT4o and get perfect summaries of the terms, where I would have expected counsel to take 1-3 days to review
2) I am trying Gemini Deep Seek for market sizing and it has come in shockingly close to our consultant
3) Perplexity saves me countless hours trying to source companies and look at funding histories
This tech replaces cognition, which has only been augmented up to this point. We have fully replaced most non-cognitive routine labor (auto assembly, big agriculture, etc...) and it has made those jobs trend more towards the gaps where labor mechanization does not 100% work.
Much in the same way, the "touchy-feely" parts of finance like sourcing, term sheet structuring, human capital, vendor selection, will become the main basis of the role. Juniors will pick up the slack where AI doesn't have great reliability like modelling.
+1 SB
LLM tools give me a 30% increase in productivity. I think you can apply that 30% increase in efficiency across all knowledge work.
I am not a software developer. I am a technical architect. But with LLM, I can write production level code with only requirements. I can feed LLM an architectural diagram, ask for an implementation in Python / Terraform and ship my client ready-to-run artifacts before the end of the business day. I've never been a developer but with my knowledge of system design and data structures / algorithms, I can generate production-grade code. LLM is a revolutionary technology in my industry.
Can you share example prompts for 1, 2 and 3? I am very interested here.
We have similar feedback from technical staff on the coding front - they are able to push some smaller new features in a day or less that would have otherwise taken a week or more. Some companies also using C++ machine translation on their stack that would take literal decades for a human. Not a great look for CS grads from my semi-target school who are complaining (at least on Reddit) that they cannot find any jobs...
As for prompts, they are all different depending on the situation. With o1 I feel like I can write 1-2 sentences, and it "thinks" about the gaps required. For 4o or Perplexity Pro, I'll usually write closer to a paragraph and tell it exactly what I am looking for. Very important to start a new chat every time you refine the prompt, or it gets confused.
Some heavily truncated examples:
1) "find the (working capital provision, earn-out structure, venue/forum, any other unusual items) in this contract/term sheet/NDA" or "find the effective post-dilution share price for the Class A/B/C shares after this transaction" or "find the dilution/accretion on existing shares after a transaction with X company with the following terms"
2) "Find the market size of a product in X market with Y pricing tiers and Z ARPU, that differentiates using A, B, & C features - and here is their product roadmap and our tech diligence report [I'll paste those in here in plain text if working with o1...] and claimed market share",
3) "Find me every company in XYZ space and their total capital raised, including potential majority transactions, and then find some similar companies and their total capital raised " [this generally is a task I use perplexity for and have found it to be more reliable than Crunchbase in a lot of searches].
These are not perfect, but they are much faster than the hours I would have otherwise spent on each task and usually give me something that requires 50-70% more polishing than an analyst output.
Look at your MDs and you tell me if they would be capable of putting a book together with AI
Can AI run a call with 25 people from differing organizations who are tense, stressed, and lashing out at each other over merger terms.
Can it run a process with 100 buyers with differing asshole natures(and jerk off everyone's egos)
Can it motivate a mediocre sunsetting hired gun portco cfo who needs to answer 450 diligence questions and keeps putting it off (and likewise keep the MDs happy that cfo is making progress).
Can it figure out the right context in an email to cc the right people, leave off the right people, and deliver the perfect email that doesn't cause seniors blood pressure to explode with deal stress(for large cross border co advisor deal teams with like 100 people involved)all while intuitively knowing who the real "decision makers" are?
I can go on. I can't see AI doing anything I do.
Lol, most of this is internal politics, and exactly why AI should come in handy.
With it, the sunsetting CFO may focus on the 20 key questions, your 100 buyers could be pre-qualified to the 5-10 most relevant, and I can ensure you that your "perfect emails" already rub some people the wrong way.
I can see AI do a lot of what you do.
this
Not sure if this is deliberate or not, but I will bite.
Majority of things mentioned here are not what the “AI” is being touted to achieve, and frankly are much more about “managing” personalities.
The AI can certainly help with:
- producing first cut of the merger model so that your analyst and associate can brain storm the trends (with you or before showing you) instead of wasting hours putting together the file and linking / adding in the manual inputs manually [time saved and better spent on analysing / checking the work + easier to iterate instead of having to wait for the juniors to process the next batch of comments]
- Update existing PPT templates for your firm’s merger model outputs so that you can spend time before your call proofing the outputs instead of waiting for your juniors to have a last minute fire drill
- give you an overview of the takeover code in different jurdistictions and what your options may be in different scenarios related to offer timelines, changes to offer terms etc.
- provide you with precedents for similar mergers where deals fell over or got over the acceptance thresholds (much harder to do without Dealogic or solid internal databases)
I could go on. But an efficiency gain of 20-30% itself is pretty solid.
Personally, I am quite keen for decent quality roll-out of things like Mosaic, Endex, Tracelight, Shortcut etc. to see if we can recalibrate from doing pointless work to more meaningful analyses
For it to actually do work unsupervised we need to have good confidence in accuracy . For the sake of argument assume it’s 99% (which it’s clearly not). That means it will on average make an error every 99th prompt. Not good enough for the financial sector where such hallucinations can be extremely costly. To get it up to 99.9% it’s a ten-fold increase in accuracy (from 1% error margin to 0.1%), which is far from guaranteed and we don’t yet know whether Moore’s law really applies. “Exponential curve” might just be a shot in the dark for now as feeding it double the training data in the future won’t necessarily bring crazy increasing returns to scale (might happen, I’m saying that it doesn’t necessarily have to happen due to some law of nature, which many claim). Tech companies are incentivised to invest heavily in it and publicly state that they believe in its transformative power because otherwise they’re left behind if this does indeed happen. If it doesn’t, then they lose the invested capital, but so does everyone else in that scenario. They have much more to lose by not believing in it. Basic game theory.
We actually know that there is a fundamental design "issue" that can't be scaled out of being an "issue."
A great example is of this is with legal AI. The biggest problem legal AI providers are having with LLMs is not hallucinations or anything like that. It is having the models understand things like a 2023 Supreme Court ruling modifies a 1960s state appellate court standard that applies an 1860s law. And legal-specific LLMs like this actually have a tremendous advantage. They get some of the base databases in the world to train on. Westlaw and Lexis catalogue functionally everything and provides incredibly valuable labeling that makes it (relatively) easy to trace information. There are not many databases in the world on any topic that are as useful for training as the American legal databases. So we know it is fundamentally NOT a data issue. It is with the design of the actual AI itself that is the key constraint.
While feeding these models more data will definitely make it better at certain things, it needs to be good enough at everything. Being better than 99.999999999% of people at one thing doesn't mean you achieved this type of AGI if you haven't done the minimum for everything.
Not everything can be "solved for" with more data
So much cope in comments
you need great MDs who the F500 can blame if deal goes bad and a few hotshot analysts who will become next rainmakers. Mid tier VP processing can be done by LLMs. Gg
The banking industry will always need humans. GenAi isn't a differentiator, it's table stakes and everyone will have access to it. Your customers will expect it. Your clients will continue to play you off each other and demand more. So banks will still need smart humans to find competitive edge.
Though, with any hope, your jobs will be completely overhauled so that you are put to work doing things a machine can't do. Then we can extract maybe extract some real growth from the banking industry.
Love this line of thinking. My views are somewhat similar. Keen to hear yours on mine below:
- Assume most banks get access to the same AI platforms. This means (big? assumption) that the amount of time spent on “data gathering” aspect of modelling, marketing materials can be cut short with mid-level employees focussing on the “value add” stuff
- However, no senior MD will just let their mid-level staff go home early. Either the idea then is for mid-level staff to see what value add they can provide to the data or brainstorm more scenarios (either on existing deals or potential new pipeline deal analyses of strategic options) or something else to keep them occupied
- Banking MDs too will be running their own “scenarios” if they don’t need to wait for days for analysts / associates to run the analysis and feed this back to them (i.e., something like Mosaic or Shortcut) where the MDs do “high level” analysis directly and only then ask juniors to delve deeper if numbers directionally look “OK”
- On the client side, HF and PEs have access to the same stuff. They know that all of a sudden the bankers don’t actually “need” (need being the key word) to pull 3ams to put together a model or CIM. They will expect “more” differentiated insights or will put downward pressure on banking fee (remember PE has its own headwinds to contend with) if they feel the “effort” required does not match the “service” or “product” (they have the same AI; they can easily train it own their own internal respositories to get a good cut of the CIM based on Commercial and Financial DD reports and past investments too - a process not dissimilar to how banks will be running their updates potentially)
Quite a few assumptions in my comment here and happy to discuss them more but want to avoid a longer essay just yet 😅
Wont need to worry anytime soon - the back and middle office roles that literally a monkey could do will be replaced first… then it’s time to worry. AI can prob automate 80% of what we do if I had to guess, but there are still judgment calls and people need to define the inputs.
A monkey can do your job too :)
As a math dude now doing a part-time computer science degree, I think you won't get a very meaningful answer here. People on this forum have at least some confirmation bias. And >95% probably have no idea how an LLM works or what that really is.
this
What do you think the answer is?
It’s not just LLMs. People are being incredibly myopic here (unsure if that’s due to a selection bias or ignorance… possibly a combo, or maybe I am just off my rocker)
Similar background to yours. Don’t think most people know what’s coming
I decided to pop this question into ChatGPT 4o to hear its opinion directly from the horse's mouth:
I find this reasoning difficult to argue with, and it's more or less what I came up with beforehand. Given that the rate of technology development is exponential because new technology allows us to develop additional technology even quicker than before (a concept I first learned from Luigi Mangione himself about 9 years ago, no joke), it would not surprise me if the hybrid future described at the end of ChatGPT's answer happened even sooner than 10 years from now.
If you are skeptical, try asking ChatGPT 4o to perform common analyst tasks like building a DCF in Excel for a particular stock. It's still very "stupid" / limited in certain regards, but I think you will be surprised at what it's capable of already, especially regarding how quick it is. You can even give it comments like you would an analyst, and it will turn them instantly (but it will get hung up on some of them). I don't think it will take too much longer (maybe a couple of years) before it's at a point where it can build perfect LBOs or DCFs in a matter of seconds. Of course, humans will still need to adjust assumptions and review the models, but the time savings will be immense. Instead of an analyst only being able to handle a few live deals at once, they could probably be on 10 or more at once. Therefore, banks won't need to hire nearly as many junior bankers as they do today, and they may have to lay some off depending on how quickly AI advances relative to natural banking attrition.
To be honest, it is a very interesting conversation and it is great to read the arguments on both sides of the scale.
I don't have a well educated view on this, other than reality is very complex. Models (human developed, at the core) can give a good abstraction, but remain limited on certain constraints. If the risk derived from errors (which btw humans also make, and a lot!) remains high (e.g. in law, as per above comments, and chip in medicine, some areas of finance, etc.) I think there still be a degree of human interaction (supervision or otherwise). That being said, the potential for improvement in these and many more areas is there.
What worries me more is the extend to which humans may disinvest in the skills/learning that are relevant precisely for these supervisory work or discerning of outcomes. If all trainees and Jrs don't get much work to do, or just very narrow tasks, at some stage there will be less ability to critically question the output from models. I do hope the degree of knowledge for humans gets expanded (we can learned better, faster the basics and move to the frontiers) by the use of models, not the other way around.
In terms of AGI: probably this will evolve as a form of interconnecting specialised models.
Inference remains an open challenge, having "mastered" deterministic/probabilistic and make good strides into "reasoning".
Beyond there, the ability to cope with reality and its inscrutable "laws", "exceptions", "black swans" and hidden mysteries is a trait humans have developed over millennia. It is important to note that however, we are limited by our own weaknesses, including learning processes, mortality and egos and on and on.
Interesting times to be alive, I just hope the future will be brighter, prosperous, stimulating, safer...
I’m going to offer a very different view. I work at a large cap PE firm and the efforts are already underway to automate key functions of the junior role. I think the two primary questions are: (1) When will AI be capable of doing key junior tasks and (2) when will firms feel comfortable implementing it?
My base case based on what I see at my large cap PE firm is that key junior tasks can be done by software in the next 2-3 years and these solutions will gradually be adopted in 3-5 years. I think this begins to affect class sizes over the 3-5 year horizon.
I think the interesting change will be how PE (or investing roles more broadly) will change in terms of the skills they look for in recruits. I don’t think truly good investors have ever been defined by their ability to model, but if AI can model in seconds and all you need to do is tweak assumptions and interpret the outputs, than modeling abilities essentially become commoditized.
Of course, even today the best recruits are the ones that are really good at understanding business models and what creates value. But remove the need for truly great modeling skills and I think there will be a greater emphasis on other traits. Things like interpersonal skills, ability (or potential) to source and actually close deals, etc. Of course, understanding how to use AI to your advantage in the first place will be important as well.
Agree with everything jl12 said in this thread.
Until all lawyers get completely replaced, you don't have to worry one iota. Let's be clear, in one sentence the technology we have today is basically a efficiency tool that requires constant iterating and knowledge sharing from the user itself. Sure, this saves 20-30% in efficiency but in no way can supplement full on tasks, much less jobs.
Jobs are made up of thousands of connected tasks, and even completely disrupting a few of these unilaterally does not mean the job itself is at risk. If anything, i would argue that 30% efficiency, because it becomes table stakes and capitalist societies are always competing with each other, you actually have no net change in job counts unless it's a positive one. New jobs like prompt engineers will actually replace mundane office assistant or remote customer service gigs.
You can see this in more closely competitive industries like the hedge fund space. If everyone has access to something new like credit card data, eventually it's table stakes and although a nice addition to consumer investing, not at all a key driver of differentiation in the never ending quest for alpha generation. But when these new insights or tools or data points came out, some people thought it would mean the end of primary research. If anything, differentiation became even more of a skillset that top analysts get paid more for.
I think what will happen is that you'll have a higher skill gap between the Have's and the Have-Not's. If the average American doesn't get on board, they'll easily be displaced by efficiency and productivity gains from the Chinese, Japanese, Europeans, or what have you. Even without external global competitive pressure, domestically this will inevitably create larger income disparity gaps. In tech we all know about a 8x engineer. Now it'll be a 800x engineer as a top 1% human and a LLM will always beat out a a shitty human without an LLM.
The most productive thing someone can do is to continue learning and adapting with new technology. As throughout all of mankind from 4000 BC to 2025 AD, this is the best way to continue being relevant in contemporary society. Screaming and fear mongering about job loss is just as silly today as it was when the discovery of petroleum wiped out the #1 job in the mid 1800s at the time, the whaling industry.
I agree with much of what you said and you offered some very insightful points, but I disagree with you about the timing. As I mentioned in a previous post, the rate of technological development is exponential. I believe this is especially true when it comes to AI because it is a technology that constantly and automatically improves on its own over time regardless of whether humans are actively working on it. It would not surprise me if AI in just a few years is dramatically (if not incomprehensibly) more capable than it is today.
I also disagree with your analogy about the whaling industry both because it is factually incorrect and because it is a poor parallel. The only place I’m aware of where whaling was the #1 job during the mid 1800s is Nantucket. Otherwise, most people were farmers throughout the world during that time. Moreover, replacing a single industry (regardless of its size) with a new, more universally beneficial industry is not the same as the advent of an ever-evolving technology that will fundamentally alter all industries. Will humans still be needed in all industries to use and guide AI for the foreseeable future? Of course, but the number of humans needed to do so will be dramatically reduced.
(1) "the rate of technological development is exponential."
You know that technological development isn't some black box, right? We know quite a lot about how technological development structurally occurs. We can actually make real predictions instead of just posturing vaguely at hoping a generic historical (edit: trend) is applicable to the technology being discussed.
(2) "I believe this is especially true when it comes to AI because it is a technology that constantly and automatically improves on its own over time regardless of whether humans are actively working on it."
How mechanistically does this work and how effective is this type of improvement? In other words, you made a claim (a fairly reasonable one), and I am interested in knowing what the basis for this is.
Very well said
I hope not I just got started lmao
when you have people like jl12 writing paragraphs questioning AI's abilities you should really think about it again
the question is, PE funds are always looking to cut costs and increase margins, so if you have AI who can do the work of aligning ppt contents creating simple models understanding email contents, why would you be paying $200-400K for someone with 2 years of experience? PE/IBD at the end of the tunnel is just a sales job.
Honest first thought on the matter is that AI as a field might just create new job titles in finance, like quant subgenres
AI cant even align a fucking powerpoint, do you really see yourself on that level?
Banking is going to get decimated. They are too wound up their own asses to understand this. This is a terrible forum to ask this question.
I don’t think banking/dealmaking will be gone completely, but in terms of the need for junior analysts and comp, both will go down.
What might happen is that this IB/PE scene is gonna become even more competitive, since there is need for experienced seniors to guide the firm/direction, but very little number. To all those people that think current LLM can’t automated spreadsheets and powerpoints, you are just wrong. Companies are just not focusing on that. Why focus on sth so meaningless and small when AI can solve bigger issues like what it’s doing at Palantir, etc
Basically there are two camps: 1. it will take a long time given human ego is involved in every deal making (especially M&As) or 2. Wallstreet is decimated and IB/PE will be gone in next few years.
As always the answer will probably lie somewhere in the middle - theoretically if IB/PE were to be replaced with AI, then AI is sophisticated enough to replace all office worker's jobs as well as any jobs related to investments (i.e., if an AI can do better complex dealmaking / provide better IRRs vs. humans, then it can replace HF and other AM jobs). I think BigLaw will probably be hit first before IB/PE gets hit and then HF.
However, it rules out few important aspects 1. AI will be better and quicker as time goes on, and it will be able to do portion of analyst's jobs. However, given the finance industry is heavily regulated with many emphasis on who to blame when things go wrong, most likely human intervention is still required 2. I think AI will allow analysts and above to most likely focus on human aspects / qualitative aspects of dealmaking rather than just crunching numbers. Negotiating SPA/SHA will still require human interaction, and whether you are selling a company or buying one, you still need to meet the management team and have a MP session to see if you are comfortable overall.
I agree with the previous comment's ChatGPT response - in the short term it will increase workforce's efficiency, but in the long run full replacement will be tough as M&As are done not through just numbers, but with human interactions.
[Comment was bugged - replying to jl12]
I do think that we agree on a lot of stuff, and I really appreciate your condescending tone. I believe it furthers the discussion to be talked down to.
1) "This is correct. But to make it clear what I am saying, since I thought this part of the argument was clear but your obviously I was wrong, what I am saying is that LLMs are the mechanism through which AGI will be achieved. That the technology that will first cross the AGI threshold is going to be an LLM. Since these LLMs are nowhere near AGI, AGI is nowhere near happening."
While LLMs are necessary to achieve AGI by definition, that doesn't necessarily mean that LLMs will be the first to cross the AGI threshold. A great example is with autonomous vehicles. I think it is hard to put in perspective if you don't own a Tesla, but Tesla's ADAS technology is now as safe as humans and will only get better by orders of magnitude with planned updates in the near future. This is a similar exponential curve. As autonomous vehicles drive more, they generate more data, which trains their AI, which makes it safer so more people use them, which generates more data, etc. What you define as AGI in this space is different for everybody, but I think it is safe to say that autonomous vehicles will have AGI within 3 years. Now, the adoption rates of these vehicles are a different conversation.
3) (also touches on 4)) "Correct. I am glad you agree with me. And do you know what the actual scaling laws are? Or do you just know this buzzword? I ask this because we have a very good understanding, structurally, of how innovation works. So if you want to talk about this, let's talk about it. Otherwise, stop using buzzwords you don't understand."
Recursive self-improvement is a real thing that puts AI training on an exponential curve. A great example is with synthetic data. For example: An AI uses itself to generate an essay on the topic. The AI then grades that essay objectively. Then, the AI uses both the essay and the evaluation to refine its own training. All that an AI needs for synthetic data is compute. You talked about how energy is a limiting factor in the timeline of the scale-up of these data centers (needed for the necessary compute). You know more than I do about the timelines, so I trust you there. But what if compute becomes exponentially cheaper / faster during that same timeline? I touched on it before, but we have continued to advance compute beyond Moore's law through advanced packaging techniques (3d packaging, etc.), system-level platforms (rack-level compute, etc.), and networking advances (photonic interconnects, RDMA, etc.). These aren't innovations that are linear in their impact... these are orders of magnitude better than their predecessor technology. Maybe I'm optimistic, but based on the rate of innovation in the past, the same rate of innovation in the future would yield much more powerful compute at lower costs, requiring less energy and, therefore, more powerful models with more training (either on real or synthetic data).
4) "As I said in my response above, LLMs are going to be the mechanism in which we first achieve AGI. Two of the biggest problems is that a fundamental part of intelligence is being able to pull together existing information and make something new out of it and being able to handle ambiguous, open-ended situations... Design issues are not something you can scale out of for intelligence."
This may be true; you know more about legal AIs than I do. To my understanding... the power of an AI model is a function of three things: algorithm, compute, and data. Advancements in algorithms are more consequential (like the transformer in 2017 1000x+ AI power) than compute and data. However, it is still a function of all three. Nevertheless, let's say the algorithm is truly the limiting factor and creates an asymptote in the "AI power" function.
Enter: quantum computing. i) Quantum computing is coming sooner than people think, and ii) Quantum computing will accelerate the transition to AGI exponentially (and even describing it as "exponentially" is doing the rate a disservice).
i) Google's release of Willow cuts the expected timeline of viable, commercially available quantum computers by years. The advancement they made is not to be overstated. The biggest problem in the space was quantum error correction (these computers giving the wrong answer). Essentially, because of the nature of these qubits, keeping the system coherent becomes harder and harder as you add more. They reached a threshold where now you add more, and the errors become less and less through a new algorithm. Some advancements still need to be made, but if you talk with anybody who knows what they are talking about, this innovation seems inevitable within the next decade... max (compared to before when it was an "if" question). Additionally, China (Zuchongzhi 3.0) is right there as well.
ii) To be clear, I am talking about gate-based quantum computing (Google, IBM, etc.), not annealing-based (DWave). I'd be more than happy to explain to you the mechanisms behind how quantum computers work... but here's the TLDR for its impact on AGI. Quantum computing excels at 1) Optimization 2) High-dimensional spaces 3) Simulation. AI models could leverage quantum computing in optimization: optimizing the weights on parameters and optimizing the hyperparameters (model configurations). Also, transformer neural networks (which is the AI architecture we are talking about) operate in high-dimensional spaces. Quantum natively operates in high-dimensional spaces (Hilbert spaces), enabling it to capture more complex data relationships (like your beloved AGI law example). While the previous two would make the model's orders of magnitude better in their training, function, power, etc., it still does not solve your perceived need for a better overall algorithm/architecture to achieve AGI. Quantum computers, excelling at simulation, would do meta-optimization (create and simulate new neural network architectures, find the best one, and optimize it). These are architectures that classical computers couldn't even conceive of. AGI would be achieved extraordinarily quickly after a scalable, error-correcting, commercial quantum computer becomes available (again, quicker than you think...).
I recognize that I am optimistic about the future of technology, but this is all rooted in the historical pace of technological advancements. By stating that "AGI is nowhere near a thing today," you are saying that there will be no more major technological advancements in the near future, which I believe is false.
Ai can’t handle “pls fix”
I actually think AI will help the industry because it can take care of repetitive tasks like data crunching and helps analysts uncover insights faster, so teams can focus on big-picture strategies. I'm just entering the industry but AI has helped me a lot already. I used this app called FinPrepAI to help me prepare for my technical interviews so it's going to raise the bar of candidates.
AI should be stopped globally. If I were in charge, I'd make it illegal until we understand exactly how it's going to impact society (and make sure it's not to the detriment of most of the population).
The way it's going, will lead to global misery and mass revolts in cities.
0 regulation or thought has been put into rolling out this completely unecessary and tyrannical technology.
Save this post.
Actually agree with this
I've been thinking about this a lot lately because of Excel.
One threat could be Python replacing Excel as a consequence of how LLMs may make Python more accessible. Unless MS incorporates more robust LLM tools into Office we could see something like this take place in the enterprise market without a large amount of friction.
I might certainly be wrong but if you're already on a management track at the moment it wouldn't hurt to think about. It isn't disruptive as much as it might be iterative.
Every time there has been technological innovation including revolutionary innovation such as excel or the internet, the pace of deal making has increased and there has been a push for more bankers, not less. There’s always more work to do. Even if all the current junior tasks were completely outsourced, guys like me will create new work to get a competitive edge. There’s so much more we can be doing if resource wasn’t constrained. If I had AI doing everything my anakysts were doing, I’d find a whole lot of other things to do.
If a company can not make my FactSet update take 2+ fucking hours it’s be worth a billion dollars today.
I'm not concerned, seniors will always have to scream and get mad at someone, AI won't cut it.
Your synopsis is now.
Top BBs are developing this own LLMs
Mid tier BBs will be chastity belted by regs and too stupid to develop their own, they will wait for factset s version, while EBs and boutiques will suddenly seem to be employing generic ultility LLMs that make them seem like they have three times as many employees.
Let’s face it—some of you are in total denial. Banks are all about cutting costs, and AI is a big part of that strategy. It’s not going to replace every job, but it’s definitely going to shrink the workforce.
If your MD isn’t comfortable using AI, trust me, someone else will be.
AI is a game-changer because it frees up time for the real value-add tasks.
If your company doesn’t get on board, you’ll fall behind, lose your edge, and eventually your market share. Harsh but true: if you don’t figure out how to use AI to your advantage, someone else will—and they’ll take your spot.
That's why the US should work to ban it globally, so nobody gets an "edge", until we understand the full ramifications of rolling out this technology.
The clock isn't "ticking". It's being made to tick, for no good reason.
Laughs in DeepSeek
The point is only a fraction of the value added tasks that could be done are getting done, I’d love nothing more than AI to take the necessary but low level work off my juniors plates. Believe me, I have a lot of things for them to do.m after that.
Hmm, what are some new things they (juniors) could look at?
This has been said every time there’s an innovation
”why do we need an analyst program? We have lotus” - said by the then head of IB at Lazard
”this factset thing is amazing. Do we need all of these analysts when we don’t have to run comps or manually update share price graphs.” - my MD when he discovered I had moved all my comps to factset
Were just scratching the surface in ways we can make junior bankers work 80 hours a week
AI and large language models are not replacing investment banking analysts anytime soon. Valuation side of things is too subjective. Models can spit out comps, DCFs, or precedent transactions, but they can’t weigh the nuances of why one deal comp might be more relevant than another or why a certain multiple "feels right" for a specific situation. Analysts connect the dots, tailor the pitch, and handle all the curveballs that come with transactions. AI doesn’t have the instincts or judgement.
you do realize S&T got decimated due to introduction of electronic trading exchanges and many people said the same thing during that time.
It's all about rate of change.
When I was a kid I played a lot of Doom then later Starcraft 1. When I look back at the graphics, they were pitiful. Compare that to recent video games which are nearly photorealistic. I had a lot of free time about a year ago and played a lot of PS4 games. In a lot of cases the graphics today are pretty much cinematic and look like a movie. The same is true with VR. I bought an Oculus rift when it first game out back when it was nearly impossible to use. I had a headache for nearly 24 hours because I misconfigured some of the settings initially. Today, the most recent version is photorealistic.
In college, I was mesmerized by elsewhere.org's postmodern text generator. It wrote about Foucault or Derrida just as well as any English major. It of course couldn't write anything other than postmodern English papers. More than anything, it convinced me that a lot of academics in the discipline were full of it. Around when I graduated the early predecessors to the GPT models were released. Early GPT models would get confused by literally anything but it was remarkable it was a general purpose version of elsewhere. And of course it wasn't useful for much except entertainment. GPT4 was suddenly useful and actually could help with very basic programming tasks. Then now Claude Sonnet is actually a helpful assistant. If you've talked to a software engineer that's used Cursor many will sheepishly admit it is making them less literate at programming. They will all tell you it can never replace them, but many of them find themselves unable to do their jobs effectively without it now.
Look at how far these AI models have come just in recent memory and think about the rate of change. There's a lot out there on the horizon that just isn't widely known yet. Most people haven't used o1, o3-mini or Claude Sonnet, learned how to effectively do prompting. These aren't even cutting edge. There's a lot on the horizon that isn't visible to you unless you are obsessed with the space and tinker quite a lot on your own with open source projects.
Right now you're right - they can't do Powerpoint and Excel. But how long will that remain the case? Just a year ago, I thought banking analysts would be protected by the cost of local GPU inference, but I don't think this is the case anymore. A year ago I tried running Llama on my Macbook and it crashed my computer. But now the models are quantized and distilled. Also, M2 chips on Mac products have enough GPU RAM to run these locally. Deepseek is frankly smarter at simple short-response questions than most Ivy League grads. And Deepseek is free and involves no privacy issues if run locally. Relative to the general population, it is hands down smarter for brief questions.
Saying "it's not taking our jobs anytime soon" is missing the point. Of course: these models can't do any economically valuable work on its own quite yet. But I don't think that will last for all that long. Sure, maybe the timelines of AGI in a few years are optimistic. And also maybe AGI is different from actually taking people's jobs. But what I do know is this: if you increase the supply of something the price of it goes down. These models at the margin increase the supply of knowledge-work available on the labor market because they empower less skilled workers to compete more effectively. What this means is that workers are more fungible with one another than before. By empowering everyone individually, everyone is collectively disempowered. It's ironic, but that's how economics works. (Think about how right now junior bankers work insane hours given how much competition there is. Imagine now more competition.)
I don't see a scenario where this is not true and I do actually think young students fresh out of school are hurt by this the most. At the margin, would you rather hire 5% more junior talent or use your existing talent more effectively with AI? Probably the latter. I wish I had a solution but I don't think there is one that's easy to point to since I actually think most high paying work will be affected by this in some way. Broadly, my best guess is that being a laborer will simply be much worse than being an employer (or self-employed) for the next generation. Aside from this, the other solution is probably to invest a lot of your savings in AI stocks as a hedge. I don't think it happens right away, but it will happen at some point during the careers of most people on this forum.
This
Good, nuanced take. +1
I wonder how views on this thread have changed in the 6 months. Quite the curve of change with this tech
Shortcut...
+ Tracelight, Mosaic, Endex, Model ML too 😏
Haven’t even mentioned the OpenAi, Anthropic, Gemini stuff lol
Bump
I ended up contracting for one of the foundation model companies and commented on this thread before.
I have changed my view after seeing how the sausage is made.
The issue is that AI models are impressive largely because they are based on absolutely gigantic amounts of stolen data. It’s easy to forget this because I mean generally there’s no point belaboring this aspect. But keep in mind that maybe $100B or $1T or so of intellectual property was gifted to these companies for free because we decided it was ok.
Why do I mention this? Because in areas where you do not get this legal subsidy the AI models will not work. Shortcut for example (the AI that does excel) is a very severely retarded product. AI is not magic and it’s important to understand that it does not work without an unprecedented amount of theft. If you do not steal billions of dollars of IP you will get a retarded product like shortcut that will be trained to pass one test for a demo but literally be retarded everywhere else.
Like I said I contracted for these foundation model companies and they basically have some scheme to basically steal data to train finance models (basically paying current finance employees to gift them work they did for a past employer but in a legally innovative way). It’s quite clever but tbh the issue is they have to pay for it which will make it economically difficult. It’s possible it will work though if it turns out you don’t need much data. But I suspect they will need just a lot of data meaning this new innovation will not work. So I do think actually finance jobs will probably be safe largely due to the sheer cost of theft.
one big problem is that let’s say you make Harvey or Shortcut or whatever and you pull classic Silicon Valley playbook of change terms of service to steal data? The problem is in this case the customer of tech company is not a person it is a company with lawyers. This means tech company cannot easily steal data like you can like a normal tech company. A lot of tech companies trying to work in the space here all assume that somehow they will get around this but I haven’t heard a concrete answer as to how. Probably for any bank to sign a SaaS deal with you the contract the bank will sign will need to say tech company can’t steal data. And I think in this case it will likely be enforced. Or at least enforced enough that it will make it expensive to steal enough data to train these models. I think stealing the worlds collective books, literature, blogs, code, scientific papers (collective work of most of humanity) is mostly fine etc because big tech co is doing it but I think probably stealing Goldman equity research probably is a crime? A workaround is of course to get the bank to sign the agreement then steal it anyway but I actually think tech companies likely won’t do this? I’m not sure since it’s arbitrary where you draw the line but I think the law will respect the IP of banks since I think they matter to the law? these folks are all knaves but think tbh actually finance ppl are safe actually do to this analysis.
Ad animi similique voluptas totam accusantium aliquam. Et soluta molestiae voluptates architecto tempore aut harum. Dolorem assumenda cupiditate nihil accusamus dolorum eos.
In atque qui sunt voluptatibus sit. Ut magnam veritatis repudiandae sint voluptatum. Et repudiandae numquam eveniet laudantium explicabo. Enim voluptate aspernatur velit dolorem ut.
Nam natus sit id consequatur odio hic. Nemo aliquam ducimus eius quasi corporis labore. Facilis sapiente repellendus reiciendis reiciendis iusto eum. Odit cupiditate deleniti laborum maxime et. Pariatur voluptatem repudiandae odit doloremque error nemo ullam. Aut magni omnis nihil non ab maxime aut voluptatem. Et magni eos et.
See All Comments - 100% Free
WSO depends on everyone being able to pitch in when they know something. Unlock with your email and get bonus: 6 financial modeling lessons free ($199 value)
or Unlock with your social account...
Architecto nostrum iure soluta molestiae nemo. Rerum quibusdam sunt unde alias commodi. Quo ex qui impedit expedita est sit.
Perspiciatis et vel fugiat ipsa qui accusantium architecto corporis. Ipsum tempore aut saepe voluptatem. Odit ducimus asperiores nihil et tenetur sunt.
Quia expedita animi a. Architecto aspernatur inventore eum itaque unde. Consectetur ut animi magni dolor quasi perferendis reiciendis.
Aspernatur consequuntur dicta omnis cumque debitis amet asperiores doloribus. Voluptas aut voluptas consequatur maiores. Molestiae nesciunt quod qui ab nam ad eum. Qui magnam commodi dolorum sit.
Alias reiciendis atque unde culpa accusamus dolores et. Et et aperiam at nostrum sed autem neque.
Voluptatem nemo doloremque temporibus rerum ut. Maxime facere debitis necessitatibus. Repudiandae iure quasi rerum sequi. Voluptatem similique beatae error ut. Eos repellendus iure ut et et voluptas hic. Laborum tempore impedit ut voluptatem rerum dolorem. Placeat dolor magnam vitae dolore omnis et dolores.
Sed ut sunt minima debitis. Quas itaque qui quaerat quas et minus qui. Sunt repudiandae nesciunt eaque odio nisi deserunt sint. Quia impedit maxime unde explicabo in eos. Commodi harum cupiditate in nobis et rem sint. Possimus fugiat et nemo ab soluta.