Learning Library

← Back to Library

From Answers to Analysis: AI in Finance

Key Points

  • An MIT study found that copying decisions from ChatGPT (or similar LLMs) significantly reduces the amount of mental effort people actually use.
  • In finance and other high‑stakes fields, many users offload decision‑making to AI so they can claim credit for successes and blame the AI for failures.
  • Most people ask LLMs simple “answer‑only” questions—something Google already excels at—rather than leveraging the models’ analytical capabilities.
  • The current transition is from answer‑machine tools (like Google) to “thinking‑machine” LLMs, which should be used for high‑leverage, high‑value activities.
  • Effective use of LLMs requires structured, analysis‑oriented prompts that provide all necessary inputs, allowing the model to synthesize insights rather than just deliver generic answers.

Full Transcript

# From Answers to Analysis: AI in Finance **Source:** [https://www.youtube.com/watch?v=J95DmmvgjIE](https://www.youtube.com/watch?v=J95DmmvgjIE) **Duration:** 00:12:51 ## Summary - An MIT study found that copying decisions from ChatGPT (or similar LLMs) significantly reduces the amount of mental effort people actually use. - In finance and other high‑stakes fields, many users offload decision‑making to AI so they can claim credit for successes and blame the AI for failures. - Most people ask LLMs simple “answer‑only” questions—something Google already excels at—rather than leveraging the models’ analytical capabilities. - The current transition is from answer‑machine tools (like Google) to “thinking‑machine” LLMs, which should be used for high‑leverage, high‑value activities. - Effective use of LLMs requires structured, analysis‑oriented prompts that provide all necessary inputs, allowing the model to synthesize insights rather than just deliver generic answers. ## Sections - [00:00:00](https://www.youtube.com/watch?v=J95DmmvgjIE&t=0s) **AI Decision-Making Cognitive Offloading** - The speaker cites an MIT study showing that copying AI-generated answers bypasses mental effort, and criticizes how people—especially in finance—use ChatGPT for simple “answer” queries to shift responsibility for outcomes onto the AI. - [00:06:26](https://www.youtube.com/watch?v=J95DmmvgjIE&t=386s) **Evaluating LLMs for Market Forecasting** - The speaker outlines a real‑world experiment using strong prompts and statistical analysis of portfolio results to assess whether large language models can meaningfully analyze financial data streams, emphasizing that they are analytical tools—not financial advisers. - [00:09:42](https://www.youtube.com/watch?v=J95DmmvgjIE&t=582s) **LLMs as Rapid Scenario Simulators** - The speaker explains how language models let users instantly model multiple financial, housing, and career scenarios—acting as cheap digital twins—to improve decision‑making, and advises using structured prompts rather than generic queries. - [00:12:51](https://www.youtube.com/watch?v=J95DmmvgjIE&t=771s) **Prompt to Reflect, Toast** - The speaker encourages the audience to contemplate the discussed idea before concluding with a casual “cheers.” ## Full Transcript
0:00Do you remember the study that came out 0:03of MIT that talked about AI taking away 0:06people's brain power? It was it made a 0:08lot of waves. Basically, the TLDDR is 0:11that when it you copy and paste 0:13decisions out of Chad GPT or thinking or 0:15writing out of Chad GPT, it turns out 0:17not much of your brain gets used. Big 0:19surprise. There is a larger lesson 0:21learned here. In a lot of high-profile 0:24cases, and I'm noticing especially in 0:26finance, people want AI to take the 0:30burden of the outcome off their 0:33shoulders. They want to give it to AI 0:36and ask AI to make the decision for them 0:39so they can take credit for being smart 0:41and using AI when it goes well, stock 0:43goes up, or so they can blame the AI 0:46when the stock goes down. I have watched 0:48people mess around with chat GPT and 0:50they're not asking strong prompts. 0:52They're not asking analytical prompts. 0:54We'll get into how you do this better. 0:56They are just saying, "Give me answers. 0:58Should I refy at 6.2%. Is now a good 1:01time to refinance my mortgage? When do I 1:03sell my house?" Or, "How much money do I 1:05need to buy a house? If I want to move 1:07to this city, how much money do I need? 1:08I want to negotiate with my employer for 1:12a raise. How much should I ask for?" 1:14These are answer questions. They're what 1:16I call domain completion questions. 1:19Google is actually very good for this. 1:21This is what Google was designed for. 1:23Google was designed to work with your 1:25brain's propensity to ask for answers. 1:28It is the answer machine. And one of the 1:30things that we are all living through 1:32right now is a transition from answer 1:34machine like Google to thinking machine. 1:37Blat GPT claude Gemini etc. We need to 1:42make sure that we prioritize moving to 1:45thinking machines for high leverage, 1:47high value activities. Finance sure 1:50comes to mind. It seems pretty valuable 1:51because when you ask for answers from an 1:54LLM, the LLM is trained to be helpful 1:57and gives you a perfectly acceptable 1:59generic answer because your prompt did 2:02not give it the room to do what it does 2:05best. your prompt did not give it the 2:08room to analyze to actually dig deep. 2:12And so what I want to focus on in this 2:15video are the principles. I'm going to 2:17use finance as a lens because I think 2:19it's a high leverage high value 2:21activity. We all do it one way or 2:22another. Whether we're in a budgeting 2:24app or whether we're investing or 2:25whether we're negotiating compensation, 2:27we have to do with money. I want to talk 2:29about how you use LLMs for this because 2:31it's a lens into how we use them for 2:33high-v value thinking activities. And 2:35the frame I want to propose is that it 2:38is most effective if instead of using 2:41domain completion or give me the answer 2:44type questions, we change and we think 2:47of it as give me an analysis given all 2:51of these inputs that I'm going to give 2:53you. And we very carefully structure the 2:55prompt so that you can actually have a 2:58correct place for all the analysis that 3:01the LLM will need to complete in order 3:04to give you a reasonable overall picture 3:07of the decision you're contemplating. So 3:09the goal is that the LLM is a 3:11synthesizer. The LLM is a conversational 3:14partner where it can process inputs more 3:16efficiently than you. It can look at a 3:18discounted cash flow sheet maybe more 3:20efficiently than you, unless you're 3:21Warren Buffett. If Warren's listening 3:23from beyond the grave, hello Warren. But 3:26the point here is that most of us don't 3:28use it that way. I want to suggest that 3:31this is because we are very 3:34uncomfortable with uncertainty and using 3:36LLM this way extends the uncertainty 3:40runway to a degree that is difficult for 3:42most of us to handle. If you use the LLM 3:45to analyze your refinancing position, 3:48you don't get a decision back. All you 3:50get back is a set of options with a lot 3:54more color, a set of options with a lot 3:56more clarity around the details. You 3:58have given the LLM a lot of information, 4:00maybe your W2s, your 1099 income, 4:02whatever you have, right? Your current 4:03rate on the house, what you want to do, 4:05etc. And you're going to get a lot of 4:07options back. And the LLM may have an 4:09opinion, and that may be okay, and you 4:10may or may not agree, but it doesn't 4:13give you an answer because you didn't 4:15prompt it to be an answer completion 4:17machine. And so you have to sit with the 4:19uncertainty and the responsibility of 4:21the decision. This is the part where the 4:23lens zooms back into the wider world. We 4:26need to get a lot more comfortable with 4:28uncertainty and LLMs. LLMs are wonderful 4:31analysis tools. We need to take that 4:34analysis, own the decision and own the 4:38consequences of that decision. That is 4:40going to give us much much better 4:41results. It enables us to harness the 4:43incredible power for processing tokens 4:45that LLMs put at our fingertips and not 4:49reduce all of that power down to just 4:50one decision. You actually want all of 4:52that power on exploring optionality. And 4:54the prompts that I've developed for 4:56finance as a way of exploring this, do 4:58just that. And you can get them on the 5:00Substack and you can run them yourself 5:02and you can see they're designed to 5:04unlock analysis for various specific 5:06financial scenarios. But I didn't stop 5:08there. I thought it really has more bite 5:12to it if I actually run a live 5:14experiment. And so I am running a live 5:17experiment using a small amount of real 5:20money on Robin Hood and on Kshi the 5:23events market. In both cases, I'm asking 5:27three separate LLMs to formulate 5:30opinions, analyses, establish bets on 5:33specific trades that they want to 5:35execute. We will then run those trades 5:37and we will tell the LLMs that we will 5:40judge the results in 90 days. 90 days 5:42feels very short from a investing 5:44perspective, from a call sheet 5:46perspective, but it's also something 5:47that we can sort of get directional on 5:49and giving them the horizon gives the 5:51LLMs a chance to plan for short-term. We 5:53will see how we do. I've selected uh 03 5:56Pro for this, Opus 4 for this, and Grock 5:594. I'm going to write it all up. I'm 6:01going to have a copy of what they 6:02predict. We're going to track it. We're 6:04going to see how they do. The point here 6:06is not by the way pick the model that 6:08makes the stock go up. I know cases 6:12where companies are AI washing their 6:14financing and basically saying in AI and 6:17the stock will go up. Brr. It doesn't 6:20work that way. Stocks are not money 6:22printing machines in the hands of AI. 6:24Instead, they can process a lot of text. 6:26They can provide useful context. They 6:28can provide useful analysis to humans 6:30that make decisions. And so my question 6:32is actually do we have given a strong 6:35prompt good data from a real life test 6:39that helps us to understand how LLMs 6:42interact with data streams and make 6:45recommendations against real life 6:47markets with real life consequences and 6:50we're going to find out and at the end 6:51of the time like I fully expect that 6:53some of these will not work some of them 6:55will work. We are going to see if any of 6:57them actually end up coming out ahead of 6:59the ledger. And we are going to see 7:02whether there are substantial 7:04differences, meaningful differences 7:06between these models. And I may well run 7:08actual statistical analysis on the 7:10ending balance differences to see if 7:12they are within a normal distribution 7:14range from each other or whether they 7:15are actually substantially outside the 7:17confidence interval. I used to do a 7:18little bit of statistical analysis and 7:20that would be quite fun for me. That's 7:21the point, right? The point is not Nate 7:23is going to then pick this the the LLM 7:26that makes all the stocks go up and 7:27everyone's going to be happy. LLMs are 7:29not financial adviserss. I'm not a 7:30financial adviser. I am here to help you 7:33reframe how you think about LLMs and get 7:36you into an analysis space. When you 7:38look at the prompts, I want you to think 7:41about the prompts as tools for analysis. 7:45How do you take a tool that starts with 7:49here are the relevant inputs here is 7:51your role here is the silent reflection 7:54I want you to do hidden chain of thought 7:57so that you can understand what the task 7:59is whether you have all the inputs etc 8:01here is the output I want here is the 8:04success criteria here is your fallback 8:06or rejection criteria all of these 8:08things put them together into a 8:10structured prompt around a particular 8:12decision around whether or not you 8:14should sell your stock if you become one 8:16of those employees that get a stock 8:17event, right? Like it's a specific 8:18decision event, you can craft a prompt 8:20for it. Buy a house, you can craft a 8:22prompt for it. Start a new job, you can 8:24craft a prompt for it. And so part of my 8:26goal here is to basically lay out enough 8:27of these examples that you can start to 8:29take them and make them your own. You 8:31can start to take them and say, "Where 8:32do I need more analysis?" And maybe it's 8:35not finances. Maybe it's picking a 8:36college. Maybe it's picking an MBA 8:38program. I sometimes have people weigh 8:40in and they're like, "Nate, tell me the 8:42best AI program to take." And I'm like, 8:43it's hard for me to tell. Like, this is 8:45actually a great example where you 8:47should use a well structured prompt. 8:48It's a high-value decision. It should 8:50not tell you what to do. It should give 8:52you the tools to overcome decision 8:55anxiety if you can sit with a discomfort 8:57of working through an analysis. And so, 8:59what we're really asking for is not 9:02investment advice. It's really can you 9:04keep responsibility for the outcome 9:07inside you and ask the LLM to give you 9:11strong analysis for whatever the 9:13decision is. In particular, LLMs are 9:16very strong at analyzing wide ranges of 9:18textual input and they're very strong at 9:20developing alternatives and working 9:22through alternatives. So, you can get a 9:24wider picture than you would get from 9:26most humans because it will read more 9:28and it will look at more options if you 9:30frame your prompt correctly. And so one 9:32of the things that I think is sometimes 9:33helpful in these situations is you don't 9:35just write one prompt like you write the 9:37prompt and then you tweak some of the 9:39inputs. What if this scenario changes? 9:40What if that scenario changes? That is 9:42literally one prompt away. And it used 9:44to be nearly impossible to get. It used 9:46to be that if you sat down with a 9:47financial adviser, with a real estate 9:49person, with a career guidance person, 9:51you would spend the entire hour and 9:53whatever money you were going to spend 9:55working through just one scenario 9:57scenario at a of granularity than you 10:00can do with one chat in an LLM and now 10:03you can have 10 chats. Do you want to 10:05model a scenario where you put 25% down 10:07on the house, 15% down on the house? Do 10:09you want to model scenario where you 10:11take the job as a marketing manager at X 10:13salary or Y salary? and what that does 10:16to your budget. You can do all of that 10:18in a chat if you're willing to live with 10:20LLMs as analyzers. They free you up. 10:23This relates to what I've talked about 10:24with LLMs as digital twins. It's the 10:26same concept. You are using LLM to 10:29cheaply model future timelines. And when 10:32you can do that efficiently, it helps 10:33you greatly improve the quality of 10:35decision-making. And I've chosen 10:36finances to illustrate it because people 10:38find finances very tangible and it helps 10:40make it more real. And so, I've got the 10:42prompts out there. You can look at them 10:44if you want. If you just want to take 10:46this away though, I want to again 10:49challenge you to use a structured 10:52thinking framework in how you interact 10:54with AI. Do not do domain completion 10:57questions like Google. It is one of the 10:59biggest super tips I have for people 11:01with prompt questions. Don't prompt chat 11:05GPT like you prompt Google. It will not 11:08work well. You will not get optimal 11:10results. Instead, use the power of the 11:13AI by asking it to actually think. That 11:17means you have to retain and put on your 11:19thinking cap. You have to own the 11:21outcome and you have to challenge it 11:22with a structured thinking framework. 11:24And that's what my prompts are designed 11:25to do is basically lay out how you do 11:27structured thinking frameworks so that 11:29you can make the most of the power that 11:31AI requires. Yes, this is harder. You 11:33have to gather actual data. You have to 11:36sit down with prompt craft, which I can 11:38help with, but like everyone's prompt is 11:39going to be slightly different. And the 11:41analysis approach frontloads the 11:42uncertainty. You have to deal with the 11:44uncertainty up front. That's okay. You 11:47can get into that pattern now. You can 11:49get into it with finance. You can get 11:50into it with whatever frame you can with 11:52a big decision. And then it becomes 11:54second nature and you will use AI more 11:55and more that way. And that unlocks a 11:57ton of downstream benefit for you. If 11:59you are the person in your life using 12:01LLMs like this, you are going to 12:04progress faster because you can 12:05literally see more future timelines. It 12:07is like a hidden superpower that is one 12:09chat away. The goal is not to avoid AI 12:12for financial decisions. I don't want 12:14you to hear this and say, "Well, Nate's 12:16going to run a test on Grock 4 and Opus 12:184 and and on chat GPT, and we'll see, 12:21and if it goes well, then we'll use it 12:23for financial decisions, and if they 12:24lose money, then we won't." The beauty 12:26of this is that we learned something 12:28either way. The point wasn't the 12:30decisions. The point was the analysis 12:32and the relationship between the prompt 12:34and the results. The point is to use the 12:36LLM as a thinking partner. So that's my 12:38challenge for you. The choice is 12:40ironically very much yours. Can you use 12:43AI as a thinking partner rather than 12:46just using it for domain completion? 12:48Think about it. Cheers.