https://www.anthropic.com/research/project-vend-1
Fun aside, finance and code can both depend critically on small details. Does finance have the same checks (linting, compiling, tests) that can catch problems in AI-generated code? I know Snowflake takes great pains to show whether queries generating reports are "validated" by humans or made up by AI, I think lots of people have these concerns.
That’s a huge pain in the neck if you want to compare companies, worse if they are in different regulatory regimes. That’s the kind of thing I have found LLMs to be really good for.
It then _didn’t_ include a similar transaction (losing $7bn by exiting Brazil).
This was stuck in footnotes that many people who follow the company didn’t pick up.
Because I have seen Claude, as recently as a week ago, completely inventing and citing whole non existent paragraphs from the documentation of some software I know well. I only because of that, I was able to notice...
But even if a human went through the documents by hand and tried to make the analysis, they're still likely to make mistakes. That's why we usually define the scientific method as making falsifiable claims, which you then try to disprove in order to make sure they're correct.
And if you can't do that, then you're always walking on thin ice, whatever tool or methodology you choose to use for the analysis.
Show me the research supporting this argument. So far RAG and similar approaches is what limits hallucinations.
It's automatically retrieving information and adding it to the context. It's -in spirit- a convenience function so you don't have to manually provide it during the prompt. It's just a lot harder to pull off well automatically, but the fundamental practice is "just" context optimization
You're essentially saying "but that's not driving!!!!" After someone goes by in an EV, because it's ain't an ICE
That doesn't mean RAG isn't context optimization.
He must have passed this secret knowledge on, as they all say it now...
was exploring this idea recently maybe I should ship it
any articles to learn more about it?
However one of the major ways people know their model is correct is by comparing the final metrics against publicly available ones, and if they are out of sync, going through the file to figure out why they didnt calculate correctly.
Personally, this is going to be the same boon/disaster as excel has been.
>Does finance have the same checks
Nope. Closest is double entry system and that only prevents the most egregious stuff. It's the equivalent of you must close brackets in code...it's a constraint but the contents can still be hot garbage. For investment ideas that are literally zero guardrails, in fact quite the opposite as this demonstrates:
https://www.reddit.com/r/ChatGPT/comments/1k920cg/new_chatgp...
If you want some of the worst takes possible on emerging tech, come to HN
But if you want an obscure hack that improves garbage collection in java by 0.1%, also come to HN
-New models scoring higher on finance specific tasks
-MCP connectors for popular datasets/datastores including FactSet, PitchBook, S&P Global, Snowflake, Databricks, Box, Daloopa, etc
This looks a lot like what Claude Code did for coding: better models, good integrations, etc. But finance isn’t pure text, the day‑to‑day medium is still Excel and PowerPoint.Curious to see how this plays out in the long to medium term.
Devs already live in textual IDEs and CLIs, so an inline LLM feels native. Analysts live in nested spreadsheets, model diagrams, and slide decks. Is a side‑car chat window enough? Will folks really migrate fully into Claude?
Accuracy a big issue everywhere, but finance has always seemed particularly sensitive. While their new model benchmarks well, it still seems to fall short of what an IBank/PE MD might expect?
Curious to hear from anyone thats been in the pilot group or got access to the 1 month demo today. Early pilots at Bridgewater, NBIM, AIG, CBA claim good productivity gains for analysts and underwriters.
Finance and engineering both have a degree of verifiably. Building evals around finance is easier than, e.g., marketing work.
"Backtest this for me"
"Analyze this"
"Find a pattern"
"Beat the market"
Reading != Understanding"Whats the consensus in todays research about AAPL?" Out comes a distilled report with clickable links back to the ai slop Goldmans et al sent out this morning.
A summary with links back to AI slop is a _useful_ outcome? Why?
Saves the junior from coming in at 4am to spend 3 hours doing it. They can spend more time fixing the slide deck.
Much of the work is repetitive or formulaic or error prone. Plus it’s all digital.
We are using LLMs to analyze corporate filings/voice memos in real time to find anomalies/correlations. This works and was previously impossible. We also use LLMs for other financial stuff. And, no, LLMs don't make financial decisions, they only point us to check X.
Salaries are higher in Finance than other industries for the same job, as it is well known.
But also, budgets for everything else is also higher.
These companies will sign 3 year deals for support, have you onsite implementing and training + app and API subscriptions.
how can you ask this question, it literally called "financial". its screams money all over the place
The title is 'Financial Services' which is a broader sector.
The day he bought, everything went downhill in that particular company lol. But to be fair, he said that he just had this as chump change and basically wanted to just invest but didn't know what to (I have repeatedly told my brother that invest funds are cool and he has started to agree {I think})
Also don't forget all the people atleast in the crypto alt space showing screenshots saying that grok/chatgpt (since they only know these two most lol) are saying that their X crypto is underrated or it can increase its marketcap to Y% of total market or it has potential to grow Z times and it is the Nth most favourite crypto or whatever. Trust me, its already happening man but I think its happening in chump change.
The day it starts to happen in like Thousand's of dollars worth of investment is the day when things would be really really wrong
https://news.ycombinator.com/newsguidelines.html
(Submitted title was "AI ate code, now it wants cashflows. Is this finance's Copilot moment?" - we've changed it now)
Seriously, people on WSB have done some pretty crazy shit. Someone created an "inverse Cramer" tracker, another a "follow Cramer" tracker. And of course there's WSB trackers.
I'd be curious to know if anyone had used any of these successfully.
On a side note, Anthropic published a Claude Financial Data Analyst on Github 9 months ago that runs through next.js [2]
[1] https://github.com/search?q=financial%20ai&type=repositories [2] https://github.com/anthropics/anthropic-quickstarts/tree/mai...
Well, that's what I spend a good amount of time doing, and no, these things aren't going to spontaneously generate alpha and give "stock picks." Well, some of the deeper concepts can probably help do so, but then you're competing against hideously massive budgets in the same arena.
That said I do think that these tools could be a huge help to "daytrading". They could help with the screening and idea generation process. The concept of "factors" or underlying characteristics which drive correlation within certain baskets of instruments, is already well established in the finance industry. And indeed that concept can be widened out beyond the purely academic lens, so you may have a basket of interest rate sensitive names, or names that are one thematic hop away from a meme sector that is taking off. LLM style tools would be great there. Ex: I remember during COVID that for a week mask companies were taking off. One of these names also had a huge run up during the SARS epidemic. Pretty basic LLM style tools would be great at pointing stuff like that out, generating lists of equities which had unusual activity during pandemics within the last 20 years, etc. Much better than hard coding in filters to an old school screener.
Oh, I think machine learning is also being used in Nowcasting. That's where you take the current economic situation, compare it to previous regimes, and then sort of map out of probability distribution for likely forward paths. Good AI workload. I actually think it would be pretty cool to see something like that intraday (if large tech stocks are liquidating which of these smaller momentum tech names on my watch list have been resilient recently?). The thing is there's sort of the retail trading space, where most of the tools are fluff, and then the hardcore space where software engineers are working in OCAML and databases and have absolutely no need for more "presentable" tools. In daytrading, there is a big gap inbetween thet, and it's surprisingly empty.
In Global Macro/portfolio managent adjacent areas (ex: NowcastingIQ.com, was browsing that earlier today thus my thoughts on the matter) you can find humans who don't know how to code who want to use these tools and can afford $25,000 a year, but again in Daytrading - the actual intraday trading stuff that makes real money - there's less of an illusion that it isn't a robotic warzone.
As models got bigger and instruction following got better, everyone jumped on the general capabilities of the model + prompting
We're approaching wall that needs to be overcome with a completely new and unheard of breakthrough, otherwise we're going to have to go back to specialized post-training (which lends itself to vertical solutions)
I think people are seeing that now with stuff like Devstral being posttrained specifically for OpenHands and massively over-performing for its size at agentic coding
there isn't money or moat in this due to commodification.
Anthropic more than OpenAi are going for the integrations, verticals and MCP - I think that is the right play. "OpenAi Inside" can replace the "Intel Inside" sticker but their marketcap needs to go 1/100x
So what is the existing competition? what is JP Morgan doing already in house/Bloomberg offering?
Deepseek was made by a HedgeFund founder, so he is also well placed.
I feel that what was missing is exactly AI front office trade tools. The trading pros who wanted a black box investing style, i.e: the math says buy stock X so buy stock X, have had the option to do that with the knowledge that it's extremely effective based on the Medallion Fund returns. That's compared to a more traditional Warren Buffet-like style of valuing a business or even a more Michael Burry-like style of finding missed gaps for a collapse.
What was missing all these years is what this is. A way for someone who doesn't know much about investing (or doesn't have the time) to "just past data there and ask it is this a good investment" like other esteemed HN members mentioned they are doing.
> Using Vcaml and Ecaml, they wired AI tools straight into Neovim, Emacs, and VS Code.. RL Feedback: The system learns from what works, tweaking itself based on real outcomes.. Jane Street records the [developer] journey — every tweak, every build, every “aha!” moment. Every few seconds, a snapshot locks in the state of play. If a build fails, they know where it went south; if it succeeds, they see what clicked. Then, LLMs step in, auto-generating detailed notes on what changed and why. It’s like having a scribe for every coder.
It’s copy and paste hell and they’re just not solving it.
“Download all files” from a chat or git pull from a chat or sftp from a chat or something but please fix it.
The hamburger menu lets you select different artifacts, if there are several, and the "Copy" button has a dropdown that lets you either add it to your Project or download the file locally.
It needs “download all files”, as I said.
It is crazy to end up with 16 files listed in the hamburger file list and need to click download 16 times and keep track of what you’ve downloaded and then rename them properly.
As I said, Claude needs to fix this with sftp or download all files or git pull or something.
It's a slightly different modality for the application. Nothing AI does wasn't possible before. You could always "create a price performance chart showing a stock's movement with key events annotated since May". You could also always buy dozens of software that will not just give you all the charts you could possible think of, but any one that you could even dream of. Check tradingview.com or koyfin.com for a taste of what a "free" offering can give you. Then imagine what the 100k software gives you.
The difference is the interface. You'll 100% need someone onboarding on their 100k custom trading platform. It might take you months to master it if you never saw one of these things before. Once you have learned it though, your productivity and velocity is expected to significantly increase.
Now with the AI interface, you don't need someone onboarding you or months to learn. You can ask the AI to "build a benchmarking analysis against Velocity's athletic footwear comps" instead of learning how to learning how to use the software to create such a thing. Maybe you never saw financial analysis software before, but you spent the last 20 years analysing financials by hand (in 2025 for some reason) and now you wanna onboard to a financial software. You don't need to "learn" anything. Just describe your thoughts to the AI and it figures the interface for you.
How transformative was that for you? I don't know. Maybe your financial analysis tool is as big of a piece of shit as Reactjs is and it's mind-numbingly tedious to generate such report. "It's just a 75 clicks that you have to do" and the AI interface saves you from doing that like it saves me from using React's shitty interface (text editor) to write garbage react components that are all just a copy of each other.
Like there is no money in "GUI". There is a lot of work that each company wanting to build a good GUI app needs to put into their particular app. And the more specialized the app, the more custom and potentially complex and expensive that will be. But there are no "GUI companies", unless you count Microsoft and Apple as GUI companies.
My view is that AI, even if it is like a human, shares some of the weaknesses of a human in that it needs to be selective about relevant information. Frontends/UIs generally do this as well for specific use cases/workflows - there's a limit of what you can display on a screen after all. UI's aren't big data (humans can only see a couple of screens worth of summarized data to be useful).
This IMO at least in the short term affects the design of AI applications in general as well.
OpenAI & Anthropic would like to become huge on their “AI-UI” platforms.
Companies selling GUI toolkits in the 90s are all dead. No company made money on selling “GUI” as a technology. No one called Microsoft and Apple “the GUI companies”
Nothing any technology does wasn't NOT possible before that tech went mainstream. The point being tech saves time/cost and boosts productivity. For e.g. if you would have been able to find a webpage in an hour before, search made it easier to find that webpage. Similarly, AI synthesizes webpages and information for you.
That is the point of technology. If you could reach from point A to point B, using a bicycle, car, train or an aeroplane, each has its own use case at its own value and price point. Each such tech saves time/cost. To say that is is only a different modality, fails to capture the value add.
Every RDS database with a dozen of terabytes that's at the entire value of a business that's running it still comes with a "Delete permanently, skip snapshot" button and, believe it or not, accidentally clicking it is not THAT unheard of.
If AI is thought of as an interface for an application where the "destructive" functions are all explicitly and clearly represented to the user and all the other actions are safe to experiment with is acceptable.
Bad UX (be it GUI, CLI, TUI, AIUI or even physical) can cause catastrophic bugs. Remember the Cisco switch with a reset button above an RJ45 port? https://thenextweb.com/news/this-hilarious-cisco-fail-is-a-n...
Maybe they use it to help search but the search in my terminal is fairly bad