OpenAI Secretly Pays Ex-Bankers to Train AI for Financial Modeling
Pasting interesting details from the article first reported by Bloomberg:
"[NEW YORK] OpenAI has more than 100 ex-investment bankers helping train its artificial intelligence on how to build financial models as it looks to replace the hours of grunt work performed by junior bankers across the industry.
The group, which includes former employees of JPMorgan Chase, Morgan Stanley, and Goldman Sachs Group, are part of secretive project inside the startup that’s code named Mercury, according to documents seen by Bloomberg.
Participants are paid US$150 per hour to write prompts and build financial models for a range of transaction types, including restructurings and initial public offerings, according to a person familiar with the effort.
The company has also granted the contractors early access to the AI its creating that aims to replace entry-level tasks at investment banks.
...
The application process for Project Mercury involves almost no human interaction, according to the person familiar with the matter, who asked not to be named discussing non-public information.
The first step is a roughly 20 minute interview with an AI chatbot, which asks questions based on the applicant’s resume. The second phase tests candidates on their knowledge of financial statements. The final stage is a modelling test.
The job is flexible and contractors are expected to submit one model per week, the person said. Instructions include writing prompts in simple terms, then executing the model.
Participants receive feedback from a reviewer and are expected to fix any issues before their work is ultimately plugged into OpenAI’s systems, the person said.
Project Mercury has so far drawn participants who’ve previously worked at a variety of Wall Street outposts, including Brookfield, Mubadala Investment, Evercore and KKR, the documents show.
Some current MBA candidates at Harvard University and Massachusetts Institute of Technology are also participating in the effort.
Participants are asked to create their models in Excel and they’re also expected to follow industry norms for formatting the models, including for areas like margin sizes and italicising percentages."
Yea this was a very interesting read. The sense I get from this article is that the current financial models being put out by OpenAI are complete garbage. Especially when it comes to bespoke modeling and specific product modeling as seen in product groups. Matt levine had some great commentary on this on his latest money stuff column if anyone is interested. If they are doing training by hand with ex bankers, I expect it’ll be a couple more years before it takes some of our jobs. Even then, my mindset always goes to “if ai takes 50% of banking jobs, I just need to be in the top 50%”
This comment was a year ago, and there have already been massive improvements in AI's capabilities over that period, partly because of projects like this one.
But I can assure you that when people say the financial models AI produces are complete garbage, it's usually because their prompts are garbage. Classic garbage in, garbage out.
These models can already build extremely advanced financial models, including very niche ones you might only encounter once in your career. The prompts used in these projects are typically 10–30 pages long. So naturally, if you think you can just go to ChatGPT and ask it to "build a fully dynamic three-statement model and LBO, make all the assumptions yourself, and make no mistakes," you're obviously going to get something that falls far short of IB standards.
But if you spend 1–3 days writing a thorough prompt specifying every requirement, assumption, the logic a human banker would apply, structure, formatting rules, etc., instead of spending 2–4 weeks building a full model from scratch, I can assure you that the resulting model can meet IB standards after a few more back and forths/tweaks.
The issue is that many of the bankers working on these projects were either laid off or simply burnt out and quit banking. They then spent 70–80 hours a week on these projects learning how to train and use AI specifically for investment banking, particularly how to write effective prompts, evaluate the outputs, and understand the different grading mechanisms that will lead them to the same final output as the one they built.
Meanwhile, bankers who have been working 80–100 hours a week for the past few years simply haven't had, and probably won't have, enough time to experiment with AI and really understand how to make it work in the same way. My guess is that many never will, because when you're under pressure to deliver a model by a certain deadline, you don't exactly have the luxury of spending days on trial and error.
That's why I believe that, to get from where we are today to a point where AI can replace a significant portion of what bankers do, banks will either need to make this kind of AI training mandatory for existing bankers or hire people who have already worked on the other side, such as those involved in projects like this one. My guess is that, in a couple of years, some of the same bankers who were laid off and subsequently worked on training AI for investment banking will become incredibly valuable to the very banks that let them go. lol.
If I have to spend 1-3 days writing a fucking prompt, I might as well just run the model myself - then I will fully understand the drivers and know the numbers cold vs relying on an output that could still be slop
Thanks for responding to my year old comment. I would mostly agree with your analysis here!
Condolences on layoff
Current AI models are trash, and the bubble will burst before they make serious immediate headway on this. Not that this won’t be disruptive (read: helpful), but we’re a ways out.
They’re cutting out mercor who has been doing this for months already?
Mercor does staffing so mercor prolly doing it for them?
Yeah well Mercor supposedly already staffs for them - like when you’re reached out to Mercor the whole idea is you’re interviewing to work with OpenAI already and then hired by them as a contractor. This post confuses me because they’re either
1) just going public about that quiet partnership for marketing reasons or
2) OAI is actually skipping the Mercor staffing part entirely and cutting them out by hiring bankers full time on their side.
*If it’s #2, that’s pretty huge because that could have slashed Mercor’s valuation in half (at least) overnight.* They’re only valuable if all the major foundational labs use them for finding professionals to do evals. Tangentially related and not really mercor’s problem yet, but you could be very bearish and say companies like Uber getting into having drivers do “tasks” (which are a fancy word for finding people to label data/do evals) means other players are seeing the game and starting to enter this market. The foundational labs have run out of data - so they’re scrambling to get volunteers to improve their data or get access to private enterprise data to feed the models.
Basically the way this is framed I would think Mercor is involved but why not include them in the press as a strategic partner?
Omg - that’s so gross. Where?
How do I apply for this 😂
E
A year on, the labs' own job boards say something about liquidiot's question, whether OpenAI was cutting Mercor out by hiring bankers itself. I work on AITraining Jobs, which reads the career boards of the AI labs and of the vendors every day (posting data only, I haven't done these projects).
On 8 Nov, OpenAI's board had 10 open roles on its human-data teams, out of 816 open there in total. All 10 are salaried San Francisco jobs, $162K to $536K a year: program managers, evals people, data acquisition, and one "Strategic Sourcing Lead, Professional Services - Human Data". None of them is an hourly contract for a banker. Across the 25 lab boards we read, only 20 human-data roles state an hourly rate at all, from $18 to $100.
The banker hours sit with the vendors: 250 open roles for finance professionals across 26 of them today, typically $50 to $135 an hour, median $80, with Surge AI's investment banker posting at the top ($200 to $400). So the hiring fits SnappleApple's read rather than the cut-out scenario: OpenAI staffs the people who design and buy the work, including a lead whose job title is sourcing professional services, and the hours still come through staffing platforms.
For anyone here who did one of these projects: did your contract or onboarding ever name the lab, or only the vendor?
Yes you know exactly who you are working for and you work with multiple full-timers at that company once you sign a bunch of NDAs and start contributing.
bankers training their own replacements lol
Sit architecto sit molestias ut autem non. Deserunt quia qui omnis quisquam sit voluptatem. Et magnam aspernatur accusamus minima ducimus aut labore. Asperiores ullam distinctio ipsum nam provident numquam quam.
Dolores quisquam recusandae fuga est qui illo. Esse aut repudiandae et corrupti molestiae ut saepe. Magni voluptas veniam qui incidunt assumenda amet.
Iusto beatae aut minus dolorum a. Recusandae molestiae reiciendis fuga ut et adipisci aut. Quia praesentium asperiores sint nam earum atque autem consequatur. Sed vel dolorum distinctio molestiae porro possimus laborum. Ad enim quaerat quae ut tempora aut non.
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...