Can we please ban discussions of Claude or AI on this website unless a user meets certain requirements?
I am convinced that most of the postings about Claude or AI related tools are bots that spread misleadingly positive information about how good these tools are.
Most of these posts are ADVERTISEMENTS and I think the forum should ban this like they would other paid advertisements.
No, these tools are not useful on the job. They just aren't. Stop saying they are. You need training data to get better and importantly all training data is stolen from the internet. If it's not on the internet, they can't be good at it. If anthropic could steal data from Goldman Sachs they would, but they can't. This is why it is so good at competition math problems, but so bad at finance-related tasks even though the latter is "simpler".
Yes, you can use RL as an adjunct, but this is precisely my point. Coding data including code + text is in the 50 PB range and much if it is open source projects subject to what is in effect peer review. There are many efficient ways to steal this data from organizations like Mercor that focus on data laundering (no judgment here, just labeling what data labeling companies actually do). But each workbook that is 50MB might cost them easily $1,000 to $10,000 to steal. (Usually past employees simply leak work from past employers then bill hourly for turning them into evals for RL environments, but most of them essentially give the stolen goods for free). Note that this implies it would take 1 billion multiplied by a lower bound of $1,000 just to steal not even to truly re-create. If $1T sounds high to you, consider that this is the statutory damages that music AIs are being charged with but with music importantly the logistics of theft are much easier. The problem with finance is that truly the logistics are terribly inefficient. In the case of music AIs, the theft is relative straightforward (torrenting) and economically highly accretive (~$1T of musician data was able to be stolen for only ~$1B in settlement legal fines, which is incredibly lucrative), so I don't doubt that music AI can become good given the economics. Even if it were possible to steal $1T worth of financial memos and models, the problem is it is incredibly difficult to do peer review or anything proxying peer review. This is why it takes $1,000 to $10,000 -- the data laundering companies typically do undergo a peer review process (which is why theft is still so expensive). If we had torrenting sites or open source hubs for internal materials at Goldman Sachs and Morgan Stanley like we do for math papers, textbooks, then it would be a different story, but we don't.
The companies that are probably getting bots to post about this have literally no actual user engagement. People buy services like Hebbia for the same reason they buy modern art or a nice couch. It makes you look like a real hedge fund or whatever. When an LP comes by they have a question "Are you a real hedge fund or bank or whatever?" That's an important question to be able to answer and these services help with that. People don't actually use these services because that is not the actual value they provide. The most active user of Hebbia was a managing director at Blackstone. (Her daughter used it for her chemistry homework). Their user engagement is so low that they are I think selling off whatever is left of the company. I don't mean to single them out specifically. I'm sure the other companies are quite similar. They are absolutely valuable products, just like fine art or decor is valuable, but they are not valuable for doing work.
There are some business models that I do think will eventually make this work, but honestly its yet to be seen whether that will ever be possible economically. I do think that the most important vector to improvement is essentially the one being pursued here - e.g., brainwashing kids to get organizations to sign up for a useless product so that Claude can steal whatever data is on your computer. If you think they will not do this, ask yourself a question: outside of the microsoft office suite, how many popular pieces of software are actually desktop applications? All of them are cloud applications. Why make it a desktop application? Do people prefer that? Maybe its because there's something valuable on your desktop.
However, even though I do think they will be successful eventually at infiltrating and stealing files from our desktops (enough people will agree to this unknowingly), it is nearly impossible to capture enough context for this for training. Open source projects have unit tests, a commit history, descriptions and usage metrics (stars and forks). Excel files don't have this. Having just the "finished product" of an Excel file isn't really enough - you need the "commit history" like you have on Github (e.g., a display of the changes over time that have been accepted through peer review). This is not part of the standard professional practice of finance workers in a consistent way. In my view, the only way to solve this problem is with something like what Meta is doing (e.g., consistent and diligent surveillance of employees) since this can give you something like a "commit history" for Excel.
Most of the people on this site are kids that don't know any better. It's fine for people to post about this but perhaps they should have a certain amount of karma. The problem with this approach is that then the amount of bots will get even higher since the people posting about this will just use their little AI bots to post just enough to meet the karma threshold - making the site even more dead.
It's also a thing where like I don't WSO to become a site that is just bots. I already feel this way about other forums like Reddit. A really simple solution is just to ban this like you would other paid advertisements.
This guy gets it
This reads like massive cope to be honest. I agree that it’s not anywhere near being a fully independent analyst yet, but the stuff Claude has been able to do in ppt and excel when it comes to drafting stuff up is pretty insane. And I think on a 5-10 year time horizon it’s reasonable to expect the capabilities will be orders of magnitudes better. People should seriously start thinking about what skills they bring to the table in that type of environment and learn how to use AI to their advantage while humans are still in the loop. And this isn’t just affecting finance, this is literally affecting every white collar job, including the SWEs who work at Anthropic.
Agreed.
I'd love to see an example of the bot posts he's talking about. Yes I've seen some very obvious bot posts but in the wider discussions on "is AI useful", there are plenty of very obviously not bot accounts with long post histories evidenced by their banana count or whatever (unless you want to argue that people are selling "warmed" WSO accounts) that are making positive arguments about the usefulness of AI in their role. I tend to be decent at spotting hidden ads and bot posts, I can't relate to what OP is describing.
Also think he's taking far too narrow a view of when AI is useful.
How many “in 5-10 year” waffles do we have to endure? AI has been largely available since 2023 and yet no material progress has truly been made. Why is no one bragging about / showing exactly what the AI God did for them end to end with little to no input or claiming they cut their classes of juniors by 50% yet because it’s so great?
If you need a CS degree to get utility out of AI that’s a product problem, not a user problem. Coding ≠ the rest of the world’s tasks and what I see is a lot of brute forcing trying to make a square peg fit into a round hole. This will only endure while capital is available but the music will stop with AI too. Why is there no room for a more realist conversation?
"No material progress has been made".
You're telling on yourself here.
I don’t think you’re understanding the point I’m making so let me be clear: raw technological progress ≠ economic impact. The current AI trade requires the technology to be so advanced in every white collar domain AND easily adoptable (as easy to use as logging onto the internet to achieve value, as easy as it is to use a smart phone to speed up communication, making friends, finding information etc.) to make a profitable return on the TRILLIONS of dollars invested. That’s the key point: show me the money (outside of the big tech manufactured demand / financing circle jerk)
You cannot just write off how HARD it genuinely is to get AI output to be useful in comparison to the amount of $$$ in token spend to achieve a particular output for all fields outside of software engineering an abstract math. Mind you, we’re still in a period where tokens are grossly subsidized too and it’s still hard to declare real ROI. This is the whole game. If there is no product breakthrough to fill that chasm, you then are betting on the entire population to develop the near equivalent of a bachelors degree of AI expertise to optimize and steward these systems across prompting, caching, RL, SFT whatever the fuck to get the value that’s promised — which is unrealistic to expect at society scale.
I promise this isn’t just my old man yells at cloud thesis, economists at Stanford have just posted a paper describing this exact dynamic in detail, I suggest reading research outside of tech echo chamber to help temper expectations for the upside AND downside. That’s called being intellectually reasonable and it feels like if you don’t become a full sycophant, AI bros just want to write off this extremely valid concern off. They can do that when the money fountain is flowing, but don’t forget that investors will get that money back in blood one way or another, and they don’t have a 20 year window to see a smidge of return, you can only kick the can down the road for so long:
Paper: Past Automation and Future A.I.: How Weak Links Tame the Growth Explosion (Jones, Tonetti May 2026).
There’s even new research (commissioned by Ramp, so take with some grains of Salt) that also states the trend has been *increased hiring* rather than decreased hiring due to AI. Which sounds great! Nobody wants to advocate for job loss…but the return on investment is *predicated* on human cognitive labor replacement…a return funded by decreases in cost. if that cost saving is not going to happen…then the return needs to come from revenue explosion, which ultimately requires a sustained GDP north of what, like 4-5%? Don’t tell me when “AGI” is coming, tell me when we’re supposed to see signs of this required nationwide growth acceleration. Until that tide turns, I will certainly remain a healthy skeptic.
Paper: A New Look at AI’s Impact on Jobs, Ramp – Revelio: Firm-Level AI Spending and Workforce Adjustment (Ramp Labs, July 2026)
Trust, I’m not anti AI, I just don’t buy that it’s this Y2K Armageddon tech in its current form that everyone thinks it will be. Not LLMs alone at least.
.
You have good insight on the no-commit problem. But this entire take sounds like a sharp insight wrapped in a lot of motivated reasoning built to reach a predetermined conclusion
Most finance AI products underwhelming but that isnt evidence that the tech cant work. Its evidence that broad shallow wrappers are commodities and thin. The gap is product, not fundamentals capability. I recommend try building it yourself in claude code and see for yourself
“The gap is product” he says, as if the lack of product is not a multi-trillion-dollar problem.
I find Claude and the excel plugin to be helpful but not nearly as helpful as the boosters say. I think model performance has stalled for the past 3-4 months which has weighed on my optimism regarding the future of the technology.
I don’t know the context of the other posts, but to pretend Claude is not helpful is disingenuous. My group has opened up access to the plugins and web client and the amount of time it saves on inane tasks is meaningful. Using it with large datasets or to set up ad hoc excels can cut hours off of tasks.
No, it’s not replacing the analyst job, but, when used correctly, it does free people up to A) work fewer hours B) focus on more meaningful work. I also know of several peers whose banks have trained these models on internal data, and they are using these tools daily to great effect.
You will never convince me that @ojapar20 and @odog @digitalimmortality.com aren’t bots. All these dudes post about, on a finance forum, is AI.
Odog been a bot for 13 years?
A sleeper agents for AI?
@ojapar20 brings up good points about seeing the big picture beyond LLMs (news flash LLMs are used in computer vision for robots and self driving cars AKA “useful AI”). However, his/her has demonstrated some cultish behavior to certain personalities, which detracts from is perceived intelligence.
I write whatever I want on here.
Not a bot. I found this forum when I was a high school senior interested in finance. Switched to tech during college and have been lurking the forum since then since there's occasionally good posts about business and entrepreneurship.
I started posting because AI will fundamentally reshape our economy (just watch the below talk from Blackstone). If you want to ignore the most groundbreaking technology of the 21st century so far, go right ahead man no one's stopping you.
Can you reshare / post the title of the talk? Didn’t attach to the end of your message
Fairly confident this is from my post about what resources to ask my PM for. I promise you, I'm not a bot, just a very anxious fresh grad who doesn't want to be stuck at a dead end fund
Fairly confident this is from my post about what resources to ask my PM for. I promise you, I'm not a bot, just a very anxious fresh grad who doesn't want to be stuck at a dead end fund
Are you suggesting the WSO bot is useless and requires me to constantly scroll past large amounts of AI slop on almost every thread?
Good post — first structural bear case I've seen this cycle that arrives with formalism instead of vibes. And flagging the Ramp endogeneity + the conflict-of-interest timing is more reading than most people give to data that agrees with them.
I'd offer a practitioner's angle that sidesteps the bull/bear fight entirely: I don't have an edge on decade-scale TFP, so I stopped trying to underwrite either side. What Jones–Tonetti gives me isn't a sell signal on the AI complex — it's a map of where value migrates if they're right: to the non-automatable links. Judgment, trust, context, decade-long contracts. So my book caps the AI-capex factor at a fixed weight and owns tolls on weak links against it. Same thesis distribution, no hero call required.
The part I'd push on is the timeline of the test. The trade won't wait for decade-scale GDP data — hyperscaler capex language reprices quarterly, and late July hands us the Fed, the megacap prints, and the Q2 GDP advance inside the same 48 hours. So, honest question: what's your leading indicator? If the 4-5% acceleration never shows, at what point does "not yet" become "not coming" for the capital funding this?
Curious what would make you flip — that usually tells me more than the thesis itself.
Okay you're definitely a bot
real bro from argentina, cheers
This is a pretty good read on how the models work and their limitations.
The only bottleneck right now is training data. It will be difficult for a general knowledge, off-the-shelf model like Sonnet or GPT 5.5 to know the deep nuances of every company on Day 1. This is why you need someone to prompt it the right way or build a harness around it, both of which can still be unreliable.
However, these problems can be eradicated if every company fine-tunes the models on their data. If you take every problem you've ever worked on, turned it into clean, labelled data, and fine-tuned a model on it, that model will start to behave exactly as you do. It's like training a new hire to do your work.
You're saying this is uneconomical, but think about how expensive humans are. White collar employees cost anywhere from $100,000 to $500,000+ per year. Multiply that across the whole company, and you'll see costs in the millions, if not billions, annually. If a fraction of that money goes toward data labelling and fine tuning, you could have models that come extremely close to replicating these white collar employees. The models have already done this for coding tasks. There's no reason they can't do the same for other jobs.
This is the likely future of AI development, not just off-the-shelf prompting but baking in our deepest behaviors via fine tuning so the AI "becomes" us. As someone in this thread said, the fundamental capabilities are there. It's a matter of time before companies put this together.
This post has too many words, apologize.
Praesentium sunt maiores error qui. Neque est non nulla. Qui vitae amet fugiat et dolorem atque.
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...
Molestiae ratione quae in fugiat officiis voluptatibus. Quidem officiis architecto aliquid ut similique dolore. Qui pariatur fuga sint similique doloribus quo ut.
Illum et sit quis unde sint ut. Itaque sint ut vel tenetur magni. Sunt quasi nostrum aut optio. Debitis impedit debitis dolores deleniti ut. Corrupti qui itaque quo aut veritatis iusto.
Cum debitis omnis et dolorum et autem. Rerum omnis non omnis velit omnis iure deserunt. Libero cumque adipisci et dolores. Qui dolores impedit id minus dolor occaecati. Nulla quis quod sed aut qui nobis perferendis. Quisquam odit ipsa placeat adipisci aspernatur.