GPT-4 & Microsoft Co-Pilot
Has anyone seen the announcement last week from Microsoft on the integration of GPT-4 into Microsoft Office?
It looks like you can upload a word document to its new GPT-4 powered Microsoft Co-Pilot and it will convert it into a powerpoint deck for you. Imagine what it could do when you link this up to your bank's internal repository of documents / templates / emails. This among many other applications which are sure to expand rapidly has my mind running wild thinking of implications for the industry.
Check out this other video talking about GPT-4 is showing "sparks" of Artificial General Intelligence.
https://www.youtube.com/watch?v=Mqg3aTGNxZ0&t=1s
It appears that GPT and other LLM's are getting better by the day, with multiple new papers coming out showing rapid improvements in the technology. This seems like a complete step-function in technology and productivity.
What does everyone think about the near and long-term future of work and productivity with these LLM's? Will many junior level roles go obsolete almost immediately, or are there opportunities for this to just enhance the work we are already doing? Is AGI already here?
All thoughts welcome!
This industry at the junior level is about to become automated away. The way we think about work is probably going to be radically transformed at a rapid pace, aligned with the rate at which generative AI is improving at every iteration.
A lot of work in IB for juniors are manual tasks which entail adhering to outdated ways to pull together materials that can be automated in seconds. For the most part, the juniors are not adding any value on the decision-making side of IB, which truly cannot be automated due to the relational nature of the business. So basically, in the next two years, I can foresee banks seriously cutting down the volume of heads at the junior level. I'm gonna go look for a different job.
My bank hasn't even been able to integrate slack or similar messaging app because of compliance issues and budget, and this has been something they know they are lacking over the past 5+ years. Pretty unlikely that they'll integrate this AI tech in the next 2 years or let it just make decisions. LLMs are known to produce outputs which might be completely wrong, as long as they produce something. Who is going to be held accountable in front of the company during a pitch if the LLM model produced completely wrong numbers? LLM models still can't even comprehend simple inference logic. Sure, I think IB work can be automated in the next coming years, but don't think we are there just yet
I agree with you that banks are slow to adapt - but given the clear business use cases of this product, I see it unlikely that they will move very slowly. All it takes is a Microsoft Office subscription upgrade (I am sure you are using excel) to see tangible benefits right away. Additionally, with Github Copilot X, GPT-4 can essentially automate 55% of all coding tasks, making it much much easier for your bank's tech team to roll-out updates at a faster rate. GPT-4 is being integrated into all applications, so I am sure several of the third-party apps your firm already uses will roll out updates very soon.
The one issue I see is on compliance and legal, but as soon as a few banks do this, I figure all banks will rush to go. Happy to debate this point, it's a very interesting grey area right now. I just think that anyone who bans the technology will be left behind.
I hear you on the hallucination problem, but with chatGPT plug-ins that were just released late last week, it can search the internet and show you exactly what sources it is pointing to. Additionally, one of the co-founders of OpenAi, Ilya Sutskever, said he believes the hallucination problem can be solved, and that they are making rapid progress (GPT-4 is much better than GPT-3.5).
This evil little AI bitch is the worst thing to come out of 2022/23
I think if you were paying for ChatGPT Plus and get access to GPT-4, it would have given you a much better answer. There has been rapid improvement in the past year in these models, and it will only continue to get better. You may think that there is a ways to go, but that gap is being closed not in a timespan of years, but in months / weeks. ChatGPT Plus also has plug-ins, and I am sure a plug-in will be specifically tailored to pulling financial information. Additionally, since my original post, ML engineers at Bloomberg trained a LLM using bloomberg data called BloombergGPT which will provide accurate up-to-date information:
https://www.bloomberg.com/company/press/bloomberggpt-50-billion-paramet…
For everyone in investment banking, if I want to pull a multiple, or any financial information, I can just ask a chatbot to pull it for me instantly. I could then ask it to put that information in a table or graph, and it will do so. This technology is technically already here, and I would guess Bloomberg can probably roll this out pretty soon.
I fully understand that "boomers" aren't very adept at using technology, and there will always be need for juniors to handle this type of stuff, but what I find so powerful about this is that everything is done in natural language format. When my MD types out an email to me and asks me to pull an EBITDA multiple, he is performing his ask in natural language. Instead of emailing me, he can just email his chatbot.
I don't think its out of the realm of possibility in the next two years to train a model to interpret comments from an email (which is how most of books get turned) and then make those changes in powerpoint and excel. There are already instances of GPT-4 being able to use a regular mouse and keyboard to move around someone's graphic interface on their computer!
I am just thinking of very direct applications of this technology to current workstreams and workflow and not even considering how the way investment banking works could change entirely given how much more productive we could all be.
I could easily imagine a future where 1 associate can handle working on 15-20 accounts with an AI assistant helping them. Teams could be cut down to Associate, Principal and MD (similar to how PE firms run that lean). And the scary thing about this, is that type of future is not something that is a decade away at this rate.
Once the MDs learn to print a PDF, I am sure they can level up to using AI tech
This shit really does scare me fr. The argument has always been “front office finance roles will never be automated because it’s a relationship business. You’re a salesman”. Sure this is true at the senior level but the truth is analysts are not client facing. We are there solely to do grunt work. The vast majority of which consists of taking numbers from client docs and plugging them into model templates, then taking those models and popping them into some pretty slides with a few bullets that describe them. If you don’t think this is something that an AI will ever be capable of doing you are delusional.
Sure it will make mistakes. It won’t be perfect. But neither are human analysts. That’s why we end up turning 10 rounds of comments from our associates/VPs before anything makes its way to the MDs desk. The difference is GPT can turn a round of comments in 10 seconds that might take an analyst 30 minutes. Let’s be very generous here and say the GPT fucks up 3x as much dumb shit as ur average analyst. It takes the associate 5 minutes to scroll thru and give more comments. It takes the GPT a total of 15 min 30 sec to go thru 3 rounds of comments vs an analyst taking twice as long to get through 1. Obviously these numbers are made up and this is purely a thought experiment, but i find it hard to imagine a scenario in which the vast majority of “analyst work” as we know it is not automated within a few years.
Am I saying we’re all going to get fired next week? Of course not. But if you told me a year ago about what GPT-3 can do I would probably laugh at you. Who the fuck knows where we’ll be in 3 years when current college freshmen hit the desk? The job is going to be radically different at the junior level and this is undeniable. Maybe clients will come to expect exponentially more detailed AI - powered analyses that we can’t even imagine yet, somehow keeping demand for juniors flat. Maybe analysts will increasingly do more “associate” work off the bat and headcount will be reduced. But the day to day of a first year from the class of 26 will look radically different from someone who graduated in ‘22 and there’s no way around it. And frankly I’m shocked that more ppl aren’t discussing it seriously
I feel people jump to the binary option of everyone simply losing their jobs. And then others who feel that this is just advent XLS 2.0 and that it will just mean we will do more deals in the same amount of time.
I feel it is somewhere between the two. But first we need to understand what MSFTGPT or BloombergGPT mean.
Imagine if each bank was able to plug in their own drive / repository / library into GPT a la Bloomberg (google BloombergGPT), I think results could be quite interesting. With GPT having been able to ingest all this information and plugged into Microsoft Office, imagine typing: “Prepare a management presentation on X (private company whose PPT, PDF, XLS and other format information is saved on said bank’s drive and has been ingested by GPT) covering key investment highlights, market and industry overview, GTM strategy, personnel, financials, growth and M&A”. If you can get that very rough draft (i.e., 60-70% shell that is normally put together by a poor analyst using precedent decks over day or two or at least a few hours per section of CIM…* in 15 minutes*, I think that eases a huge pressure point. Imagine then being able to re-input this deck and saying, move certain pages around, add or subtract certain themes and so on. Huge game changer and the analyst’s time can actually be spent thinking how to structure it / narrative and checking key pieces of information instead of piecing the first cut of the poor deck together. Given how much times goes into making these vs how much time PEs spend reading these, this first step can be a huge time saver.
Similarly, if you are in a product team and often do decks on “Key Exit Considerations”, “Accelerated / Dual Track Timetables”, “Current M&A Environment and Mitigants”. Most such decks use precedents as starting points and are then iterated.
With a GPT that is plugged in, this can mean the first hurdle is improved or reduced in terms of time. The analyst effectively becomes a junior associate who reviews, corrects and fills stuff and can get comments processed in real time.
BloombergGPT or GPT trained on financials:
This is where it gets interesting. BBG will no doubt be able to get GPT to put together initial drafts of financials (at least for public companies) and basic models just by typing the assumptions in simple English: LBO with c. 4-5x leverage, unitranche bullet, PIK, 5 year horizon, transaction date 31/03/2023, sensitise around 15x entry, no multiple expansion.
And viola, you have your initial draft LBO in an editable XLS and you can iterate your assumptions in simple English.
The reason for going for assumptions in simple English is 1) this is what MSFT and BBG are gearing towards and 2) this is how instructions are often communicated to Analysts in first meetings. And this is on top of stuff like Kensho etc. that S&P already had been chipping away at.
Obvious considerations:
- These things will provide first drafts. Someone will have the fun, glitzy, glamorous job of checking the numbers pulled out are correct. Doesn’t sound like a front-office job for a HYP grad. Sounds more like an ancillary back office job especially when this thing gathers steam (~5 years).
Like another poster, absolutely do feel third-party vendors may start to specialise in selling GPT trained on certain relevant datasets which can then be sold to banks etc.
All in all, if this thing continues on this trajectory for the next 3-5 years then in terms of IB ranks my key questions are:
- Will banks having additional capacity (if they don’t cull ranks) mean they do additional deals? But this is dependent on macroeconomic conditions not just capacity. So basically to what extent are banks turning away deals right now (or in a normal year) that they would otherwise not turn away. Would the number of incremental deals outweigh the freed capacity?
- Will banks cull ranks and deal teams go to being MD, VP (combo of D and VP) and Associate (combo of ASS and ANA)?
- What will be the key differentiators in terms of deal making? Most banks will be able to churn out bigger documents at twice the pace (most marketing materials are useless anyway and models too detailed or pointless) so will PE be looking to do more “data driven analyses” by feeding in the crap churned out by the banks?
- What about fundemental L/S funds? Public equities. Public GPT. Analysts may have to wade through even more noise and crap. I feel like if anything it may even further depress the HF performances but I may be being pessimistic.