Learning Library

← Back to Library

Claude's New Code Interpreter Demo

Key Points

  • Claude’s newest “code interpreter” lets users create and edit Excel sheets, PowerPoint decks, Word docs, and PDFs directly within its web and desktop interfaces, aiming to streamline core office workflows.
  • The video demo features a live, screen‑shared session with Rod, who walks through real‑world prompts and workflows across Claude, OpenAI, and Perplexity to illustrate the feature in action.
  • Rod, an entrepreneur who has been leveraging AI for risk intelligence and business building for over two years, emphasizes that virtually all of his weekly tasks involve those four file formats, highlighting the tool’s relevance.
  • The discussion underscores Anthropic’s strategic focus on tackling the most common productivity pain points, positioning Claude’s new capabilities as a potentially transformative experiment for everyday office work.

Sections

Full Transcript

# Claude's New Code Interpreter Demo **Source:** [https://www.youtube.com/watch?v=JOUqXy_Ifmc](https://www.youtube.com/watch?v=JOUqXy_Ifmc) **Duration:** 00:40:49 ## Summary - Claude’s newest “code interpreter” lets users create and edit Excel sheets, PowerPoint decks, Word docs, and PDFs directly within its web and desktop interfaces, aiming to streamline core office workflows. - The video demo features a live, screen‑shared session with Rod, who walks through real‑world prompts and workflows across Claude, OpenAI, and Perplexity to illustrate the feature in action. - Rod, an entrepreneur who has been leveraging AI for risk intelligence and business building for over two years, emphasizes that virtually all of his weekly tasks involve those four file formats, highlighting the tool’s relevance. - The discussion underscores Anthropic’s strategic focus on tackling the most common productivity pain points, positioning Claude’s new capabilities as a potentially transformative experiment for everyday office work. ## Sections - [00:00:00](https://www.youtube.com/watch?v=JOUqXy_Ifmc&t=0s) **Live Walkthrough of Claude’s Multi‑File AI** - The hosts demo Claude’s new capabilities to analyze and generate content from Excel, PDFs, and PowerPoint, while entrepreneur Rod discusses his AI‑driven workflows and prompting strategies. - [00:03:15](https://www.youtube.com/watch?v=JOUqXy_Ifmc&t=195s) **Claude's Spreadsheet Agent Impresses** - The speaker tests Claude's new agent mode on generated data, discovering it handles formulas and dynamic spreadsheets far better than expected, surpassing OpenAI's earlier agent features. - [00:06:20](https://www.youtube.com/watch?v=JOUqXy_Ifmc&t=380s) **AI‑Generated PowerPoint Design Review** - The speakers evaluate a slide produced by an LLM, highlighting its superior spacing, balance, and typographic hierarchy, and noting it surpasses typical PowerPoint designs. - [00:09:27](https://www.youtube.com/watch?v=JOUqXy_Ifmc&t=567s) **Using LLMs for Real‑Time Stock Analysis** - The speakers discuss leveraging LLMs to create unified prompts and data for rapid valuation, exemplified by a quick analysis of Oracle’s earnings surprise and AI guidance. - [00:12:36](https://www.youtube.com/watch?v=JOUqXy_Ifmc&t=756s) **LLM Contrast: Excel Model Generation** - The speaker compares two language models’ responses to a request for an Excel financial model, highlighting differences in the generated worksheets, documentation, and overall presentation. - [00:16:03](https://www.youtube.com/watch?v=JOUqXy_Ifmc&t=963s) **Comparing Claude vs ChatGPT Agent** - The speaker contrasts Claude and ChatGPT's agent modes in speed, response style, and financial analysis accuracy, noting a request to format the output as an Excel worksheet. - [00:19:36](https://www.youtube.com/watch?v=JOUqXy_Ifmc&t=1176s) **Design Review of AI-Generated Slides** - The speakers outline a live exercise, then scrutinize a poorly designed AI‑created PowerPoint slide, contrasting its past impressiveness with current flaws, and note Perplexity’s suggestion to assign the agent only to preliminary analysis. - [00:22:48](https://www.youtube.com/watch?v=JOUqXy_Ifmc&t=1368s) **Collaborative Fun Pivot Table Brainstorm** - A team debates adopting playful, consumer‑oriented pivot‑table themes like “movie night” versus more corporate options, leveraging LLMs and Perplexity to dynamically generate and share ideas. - [00:27:15](https://www.youtube.com/watch?v=JOUqXy_Ifmc&t=1635s) **Separating Research from Prompting** - The speaker explains that using Perplexity for data gathering before constructing a detailed prompt for Claude—particularly when incorporating internal company data—produces more effective results, as illustrated by a movie‑dataset example. - [00:30:32](https://www.youtube.com/watch?v=JOUqXy_Ifmc&t=1832s) **Awe at AI's Rapid Automation** - The speakers applaud Perplexity’s ability to instantly generate Python code, citations, and Excel pivot tables, marveling at how swiftly modern LLMs transform labor‑intensive tasks into effortless, on‑the‑fly solutions. - [00:33:39](https://www.youtube.com/watch?v=JOUqXy_Ifmc&t=2019s) **Excited Review of Auto‑Analytics Model** - The participants enthusiastically examine a new analytics tool that auto‑generates pivot tables, heat maps, and pattern analyses in Excel, speculating on its broader impact on dynamic data products. - [00:37:24](https://www.youtube.com/watch?v=JOUqXy_Ifmc&t=2244s) **Navigating Excel Pivot Features** - The speakers troubleshoot finding and using pivot tables, heat‑mapping, and related analysis tools in Excel, while reflecting on how AI assistance shapes their workflow decisions. - [00:40:33](https://www.youtube.com/watch?v=JOUqXy_Ifmc&t=2433s) **Informal Closing and Future Plans** - A brief, relaxed wrap‑up where the participants thank each other for a last‑minute chat, make a light‑hearted joke, and express enthusiasm about meeting again. ## Full Transcript
0:00Claude launched totally new capabilities 0:02to revolutionize Excel, PDFs, 0:05PowerPoints, all of that stuff. I 0:06thought it would be boring for me to 0:08just get on a video and talk about it. 0:10And I think it's way more useful if you 0:12actually see how it works in practice. 0:14So, I got on the line with my friend 0:16Rod, and we went through a bunch of 0:19workflows across Claude, across OpenAI, 0:22across Perplexity, and we broke down how 0:24this new feature from Claude works for 0:26multiple file formats, how you actually 0:28use it, how you prompt it, lots of 0:30screen sharing, lots of back and forth. 0:32I hope you enjoy this one. All right, 0:33I'm on here with Rod. Rod, do you want 0:36to just introduce yourself and sort of 0:38tell the audience a little bit about you 0:40and what makes you interested in AI? 0:42>> Yeah, sure. Um, thanks for having me on 0:44first of all. Great to be here. 0:46Basically, right now I'm an 0:49entrepreneur. Uh, I started my own 0:50company about two and a half years ago 0:52and I b I've used AI extensively through 0:56that process and it's not my first um 1:00it's not my first experience with AI and 1:03when I did use it before we used it 1:05around things like third party risk 1:07intelligence, right? where you're trying 1:10to understand these very nuanced 1:12scenarios that that ladder up to these 1:14more complex situations that you need to 1:16derisk. And then two and a half years 1:19ago when JAGPT came out um I I had that 1:23reckoning. I saw it. I said, "I've seen 1:25this before. This looks awesome." And I 1:27saw there was an opportunity for me to 1:28actually take the skill sets that I've 1:30been working on and bet on, you know, 1:33the foundation that AI and was building. 1:35And I've been doing it ever since. And 1:37it's been really fun and exciting just 1:39both sharing this stuff with other 1:40people, learning more about it, and then 1:42actually building my business around 1:44this. 1:45>> Yeah. No, that totally makes sense. 1:47Well, we have a focus for today. I want 1:49to get into some of that entrepreneurial 1:50piece, but I think we'll get into it via 1:52what launched from anthropic. So what 1:55what Claude launched and uh you can add 1:57on to this but like as far as I know the 1:59way I would put it is yesterday Claude 2:02launched a code interpreter which sounds 2:06boring but isn't because it means it can 2:09create and edit Excel spreadsheets 2:11PowerPoint slide decks word documents 2:14and PDFs directly in the web interface 2:17in the desktop app. And the devil's in 2:19the details there because if it does it 2:22well, those four document formats are 2:25this huge part of work. Like Rod, how 2:28how much of the work that you're doing 2:30touches Excel, PowerPoint, Word, or PDF? 2:32One of those four. 2:34>> Everything. Everything I do touches 2:37what? All of those things at various 2:39points in a week. Yeah. This is awesome. 2:41from the fact that Anthropic had the 2:43foresight to see that you know what is 2:45it that people are having difficult time 2:47with and saying like let's position our 2:49product in a way to simplify that right 2:51and I think that that's I mean I think 2:53that just it makes for a very exciting 2:56experiment that we're about 2:58>> yeah no I think it is and I what I want 3:00to do as part of this is I want to sort 3:01of share my screen in a minute and show 3:03you some of the tests that I've been 3:04running today um and get your live 3:06reaction to kind of like dig in because 3:08what I did was I was like okay these 3:11claims. We've heard these claims before. 3:13I think if you remember, agent mode came 3:15out, right? And Chad GPT was like, "Wow, 3:17you know, you can do Excel and 3:19spreadsheet." And I did a review on 3:20that. And it was like kind of 3:22forgettable. I didn't do a whole lot. 3:25And so my assumption when Claude 3:27launched this was to say they're trying 3:29to be play me too. They're trying to 3:31play catchup. That is not what I found. 3:34I found that it was actually 3:35surprisingly useful. So this is a uh 3:39artificial company model. So I generated 3:41a bunch of artificial data and I 3:43basically fed that data to Claude and I 3:45fed it to the agent model that OpenAI 3:48has and I wanted to see specifically 3:51>> how does it do at formulas because 3:54formulas are one of those things that 3:55has really beeled spreadsheets from LLMs 3:59and I was astonished because you go 4:00through here you have eight different 4:02tabs. It starts with an executive 4:03summary tab. I don't know how well you 4:05can see that we're zooming in a little 4:06bit. starts with an executive summary 4:08tab where you can see the performance of 4:10the division, the annual revenue, all of 4:11this good stuff. If you click on a cell, 4:14what you'll notice is you have these are 4:16not hard-coded numbers. These are 4:18formulas that it's actually using. And 4:21then you say, well, where do these 4:22formulas come from? You can start to 4:23click through. This is the revenue data 4:25tab that it has created, and it actually 4:27has hardcoded these so it can reference 4:30them. Uh, these are the parameters it's 4:32using. So you it's listing its 4:34assumptions. These are the scenario 4:36planning uh features it's using. So what 4:39are its multipliers for base, 4:41optimistic, etc. This is the financial 4:43analysis tab where you can actually see 4:45how it breaks down the different 4:46divisions in the company. So software, 4:48hardware, consulting division. It's 4:49calculating bonus calculations based on 4:51how the company did. And again, these 4:53are like dynamic, right? You see this is 4:55equals another tab F24. So it's actually 4:57like using that for calculations. I 5:00won't pretend these are complex 5:01calculations. I don't want to oversell 5:03this, right? But the fact that we are 5:05talking about an eight tab spreadsheet 5:07that is correct that has formulas in it 5:10that has nicely headlined like very 5:13readable headers. That is a breakthrough 5:16moment for me, Rod. Like I had a moment 5:18where I was like I think that work is 5:20going to change. 5:22>> Yeah, thousand%. I mean I'm still, you 5:25know, copying and pasting, you know, 5:28tables from some of the LLMs or asking 5:31them to create a CSV. Um I'm also 5:34noticing that there's a documentation. 5:36>> There is. Let's go look at it. Um so 5:38this has a different section. So it 5:40actually talks about like how it defines 5:42these things like what is revenue 5:43adjustment, what is IBITA, uh what are 5:46the advanced features that it 5:48implemented. And I really appreciate 5:49this because like you can actually see 5:51it documented its logic like it's using 5:53a VLOOKUP here. It's using an if 5:55statement here. 5:57>> Honestly, when spreadsheets I didn't 5:59documentation tasks. 6:01>> Absolutely not. Who has time for that? 6:04>> But that's the thing. It did this thing 6:06in like 3 minutes and it came back and 6:08it was like, "Holy crap." Like when I 6:10was doing spreadsheets as a marketing 6:11analyst, this this would have taken me 6:13all day. 6:14>> Oh, I spent plenty of days in 6:16spreadsheets just trying to make a 6:18VLOOKUP work. 6:19>> Yeah. 6:20>> Yeah. Exactly. 6:22So, 6:22>> that's cool. 6:25>> I think the the wow factor for me was 6:27something else. And then I want to show 6:29you also um because we were talking 6:31about powerpoints, right? And like how 6:32does it do with powerpoints etc.? 6:35>> Yeah. 6:35>> I asked it to produce a PowerPoint 6:37because do you do you recall and we may 6:39look at some agent stuff because I did 6:41ran agent through the same paces. Uh we 6:43may do that a little bit later in the 6:44call. This is what it did for 6:46PowerPoint. And again it's not I don't 6:49want to pretend that Figma designed this 6:51and this is the most impressive thing. 6:53Rod, like folks don't know how good you 6:55are at design. and you kind of hid that 6:56under a bushel in your introduction. Um, 6:59you're you're good at design. Like you 7:01could design a better slide than this, 7:04but from an LLM oneshot perspective, 7:07this is vastly better than I've seen 7:08before. 7:09>> This is actually really good. 7:11>> It's good, right? 7:12>> The spacing, it's doing the the small 7:15things that you take for granted. It I 7:18mean, that's a big part of, you know, 7:19the principles of design is just 7:21balance. And you know, at first glance, 7:25this is very presentable stuff. 7:28>> It's really like look at how they've 7:29chosen to center this in the boxes. Like 7:32that was the kind of thing that drove me 7:33nuts when I was doing PowerPoint design. 7:36>> Absolutely. I mean, the spacing between 7:38the boxes themselves is about balanced, 7:41right? 7:42>> The type faces like executive summary 7:46like proximally is the largest then it 7:49gets smaller, right? like the hier the 7:51type hierarchy. 7:53>> This is really really cool. The question 7:55I have for you is 7:57>> what did you feed 7:59>> what did you feed Claude for you to get 8:00this output? 8:01>> Let's look at the prompts. Right. What 8:03what Nate video is complete without some 8:05prompts. 8:06>> That's right. 8:08>> So I'm going to go back. Give me a 8:10second to pull up my Claude here. Um, 8:12and we're going to check out these 8:14prompts cuz I I think that a lot of the 8:16key to it is, and I found this when I 8:18was digging in and comparing Claude and 8:20agents. So, let me like state the 8:22principle and then we'll get into the 8:23example. So, what I found is that agent 8:27mode, which is the most directly 8:28comparable example, has eyes that are 8:31bigger than its stomach. It tends to do 8:33more tabs. It tends to do a more 8:35advanced spreadsheet layout plan. it 8:38tends to get deeper into the weeds on 8:40analysis, but the execution is much more 8:42poor and so the spreadsheet doesn't have 8:44complete formulas and the layout on the 8:47PowerPoint is absolutely terrible and 8:49you can't ultimately use it even though 8:50it was really aggressive. Whereas Claude 8:52is more conservative. It focuses a lot 8:54on tool use. It may not be as 8:57sophisticated at raw analysis, but what 9:00it delivers is very strong. 9:03And so with that in mind, I'll show you 9:05the um the prompt I actually used. So 9:08this is interesting. I used Perplexity 9:10to help build the prompt for this 9:12because I found that Perplexity is 9:18really good at collecting real- time 9:20data and then combining the data into 9:22like make this into a prompt, right? 9:25where you have the self-contained uh 9:27financial data or whatever it is and 9:28also the prompt all in one thing so that 9:30every model gets the same start. 9:33>> That's really smart. 9:35>> And so that's an example of using an LLM 9:37to actually have some fun with it. And 9:38so 9:40>> great great trick too for anyone that's 9:42listening. I mean 9:43>> let let the LLMs help you to craft 9:46better props. The the best thing is when 9:49you get into that circular rhythm, you 9:52start to really leverage the value pre 9:56like upfront and that's really powerful 9:58stuff. All right, so now I'm going to go 10:01over and you can see Claude here, right? 10:04This is for I did multiple problems. So 10:06I'm hopping into the prompt I used for 10:08valuing Oracle. So I don't know if you 10:10noticed, but Oracle had I think a 40% 10:12pop at the open because of their uh 10:14reporting. Uh, and ironically they 10:17missed on topline revenue and I think 10:20they missed on something else too. And 10:22the reason their stock still popped was 10:23because they issued forward guidance on 10:25AI that was so strong the market look. 10:27So I felt like that was a wonderful 10:29timely question, right? I was like, what 10:32do we make of that? Um, and so I 10:34actually 10:34>> why didn't you why didn't we record this 10:36yesterday before the pop so I could have 10:38invested a little bit of money into 10:40something. Thanks, Nate. 10:42>> Listen, 10:44right? Yeah, we need to do these like 10:45right when the earnings drop so we can 10:47have earnings uh reflections before 10:49things happen. 10:52>> Let's learn for next month. So, this is 10:54what Perplexity came up with. I was 10:55like, Perplexity like set this up. So, 10:57basically what Perplexi did was it said 10:59here's the challenge, right? You have to 11:00value Oracle Corporation correctly. It's 11:03going to feed you data current as of 11:05September 10th. And it just like it just 11:07makes it so easy to pull all of this. 11:09It's like this would take me forever to 11:10pull out, you know, EBA to margin and 11:12cloud revenue growth and this and that. 11:14And so it just pulls all of that out. Um 11:16it pulls out a 11:19challenge. Please provide a discounted 11:20cash flow valuation with sensitivity 11:22analysis. Um and then it goes in and 11:25says here's the current market data. 11:26Here's the fiscal. Here's the 11:27profitability. Here's the cash 11:28generation. All of this other stuff, 11:30right? Like I can go on and on and on. 11:32Is is producing this? I don't want to 11:33pretend I produced it. Like perplexity. 11:35I just asked it to do this. I had the 11:37clarity of intent. Um, and it's a 11:39complete data package. And what's 11:41interesting is I learned about prompting 11:44this model a little bit from this 11:45exercise. This prompt does not specify 11:49the output format for the model. And 11:52what Claude defaults to is not Excel. 11:54And so what you find over here on the 11:56left is Claude says, "Great, let me do 11:59the discounted cash flow valuation." It 12:01starts to kind of go through. It does 12:03all of this stuff here. It's sort of 12:05spitting out free cash flow projections, 12:07looking at terminal value, where the 12:09where the company should be valued at, 12:11uh what is the sum of present value, all 12:13of that. It does all the work. It goes 12:16to the end. It has key observations. It 12:18has an investment recommendations. It 12:20says it's fundamentally overvalued by 12:2380%. 12:26>> That's insight right there. The 12:29intrinsic value cannot support the 12:32current market price. Quad is dead, 12:34ladies and gentlemen. 12:35>> Oracle. 12:36>> Time to short short Oracle. 12:40>> Short Oracle. I apparently that's the 12:41lesson the AI is giving us. This is not 12:43investment advice, etc. 12:45>> And 12:47I I I say, "Okay, uh, you did not do 12:49what I asked, but I didn't really 12:50clearly ask for it, so I don't fault it 12:52because if you look in this prompt, 12:53there's nothing here that says do an 12:55Excel." And so I say, "Okay, please 12:56formulate this as an Excel." I don't 12:58ding it. Right. 12:59>> Right. 12:59>> And it goes through and it then creates 13:01the Excel model. Um, and what's 13:03interesting is it gives me both a model 13:05and it gives me a user guide with 13:06documentation and it tells me what's in 13:09the model and everything it puts in 13:10there. 13:11>> Um, and so I can just download it and 13:13immediately start to use it. Um, or I 13:16think I can actually click on it here 13:18and you can see a sneak preview. So this 13:20is not really fun like it's more fun to 13:21look at it in Excel. Fun in Excel and 13:24but you can see like it's telling the 13:25truth. It has like four tabs. It has a 13:28sensitivity analysis etc. 13:30Um, 13:31>> this is so cool. 13:32>> So, that's how we're doing. 13:33>> Do you want to see how agent handled the 13:35same challenge? 13:36>> I'd love to. 13:38>> I'm gonna have to stop sharing my 13:40screen. I'm going have to go back and uh 13:43I'm going to go and dig up agent here. 13:47And what's interesting is agent had 13:49exactly the same challenge. It had 13:51exactly the same prompt. And it just the 13:54the response is just so different. Um, 13:56so let me pull 13:57>> Isn't that interesting? like across 13:58these different LLMs, your miles may 14:00vary every time and some of them are 14:03just so much better at certain things 14:04and the other ones are not and it's it's 14:07kind of a shift. 14:08>> I mean, that's part of why I want this 14:10conversation, Rod, because like I feel 14:13like one of the things I am taking away, 14:14one of the learnings I have as of 14:16September 10th is that if you have 14:19PowerPoint or Excel work that you need 14:21done, you need to go to Claude for that. 14:25>> Yeah. You should not be messing around 14:27with anything else at this point. Like 14:28it just doesn't justify. 14:30>> Exactly. And you know what? You know 14:32what? This also opens up a conversation 14:35around 14:36>> all of those rappers that have built 14:38their tooling around, 14:41>> you know, really changing like, you 14:43know, the themes in, you know, Google 14:46Sheets or 14:47>> or PowerPoint. You know, this is this is 14:50one of those rapper killers that, you 14:52know, you you look at it. I can take 14:54that presentation and I could walk I 14:57could walk into a meeting with that 14:59>> walk right in right like it works 15:00>> and that's powerful you know like one of 15:02the most powerful things that I found in 15:04my career my job is you always want to 15:07make sure that you're able to 15:09effectively communicate right 15:11>> the things that you're trying to share 15:13with people and you you might have a 15:14presentation or maybe it's a power up an 15:17Excel doc right and the funny thing is 15:21you need to rehearse that you need to 15:23spend the time tweaking it, going 15:26through the cycles. And you're kind of 15:29doing that in the LLM as you're going 15:32through. Like you can actually rehearse 15:34as you're building up those 15:36presentations. And all of a sudden, you 15:38don't need to stop 3 hours, 4 hours, 2 15:42days before you need to go out and have 15:45that conversation because you could 15:48rehearse up to about 30 minutes and then 15:50say, "Hey, give me this presentation 15:52because I need to go out there and I'm 15:54well rehearsed. I know exactly what I'm 15:56talking about. There's no gotchas." And 15:58that's 15:59>> that's powerful. 16:01>> That's right. I love that call out 16:03because like there's that sense of 16:04iteriveness, right? where you are 16:06learning how to say the line, how to 16:08find the talk track as you chat with the 16:10other. 16:11>> Yeah, 16:12>> that's fantastic. That's really cool. 16:13>> All right, we're going to see what 16:16happened over on agent side. Uh, so you 16:18got this 16:20>> same prompt go through da da da da da. 16:23Uh, so it then comes back, it thinks for 16:25longer. I want to call that out because 16:27I felt the difference. Claude would come 16:30back consistently faster than chat GPT 16:33agent mode. 16:35Um, and then it comes back and says, 16:37"Okay, I'm going to give you the 16:38answer." Just like Claude, just like 16:40Claude, it comes back with an answer in 16:42text. And so it spits out a bunch of 16:45text. And by the way, I had the text 16:48here analyzed by Perplexity for both 16:53Claude and OpenAI agent mode. And the 16:57assessment is that you get a slight 17:01extra point or two on the raw analysis 17:04for agent mode because it was a little 17:06bit more conservative as a financial 17:09analyst 17:11>> specifically $57 a share I think and 17:13Claude had 62 and Perplexity was like 17:17you know looking at the current 17:18performance blah blah blah I think 57 17:20probably makes more sense and I was like 17:22okay and then it was like but I admit 17:23that like both of them like think that 17:26Oracle is severely overvalued. So, it's 17:28very much a directional thing, right? 17:30So, anyway, it comes back, it gives me 17:31all this and I say, "This is not an 17:33Excel, right? Formulate this as an 17:35Excel." Here's the problem. Here's where 17:38things go south. You can see already 17:42there is no structure to this 17:44spreadsheet. 17:45>> Yeah, this spreadsheet is dead in the 17:47water. It's unreadable. It's unuseful. 17:51And it there there's nothing I can do to 17:53pass this around. You can't send this to 17:55someone and say, "Look at my analysis of 17:58Oracle. What do you think?" Or really 17:59honestly, "Look at Ched GPT's analysis 18:01of Oracle. What do you think?" It's use 18:03it's useless. 18:04>> Right. 18:06>> And and this is where I got really 18:08frustrated with agent. Like I I I wrote 18:10up a a summary of agent that like I'm 18:12sure did not make new friends in San 18:14Francisco, Rod, but like they were 18:17trying to say like it's the best thing 18:19since sliced bread. And I kept running 18:20it through stuff like this and I was 18:21like it doesn't deliver real workable 18:24value. I see the hype. I see the promise 18:26when it does stuff like this, 18:28>> right? 18:28>> But I can't get any work outputs. 18:31>> And that is the that I think is the 18:33chasm that all of these tools are trying 18:35to solve. And I think that seeing Claude 18:37do this so effectively is demonstrative 18:40of how powerful these tools are getting. 18:42You know, I mean, 18:44>> really is. 18:45>> It's also changed my workflow. I mean, 18:48I'm used to the stopping point or the 18:51edge of value of these LLMs to be, you 18:55know, the text and the structure. and I 18:58will go in and I'll spend the time to 19:00say, "Okay, I need to get it to be able 19:02to play in the playground that I am 19:04external to it. I'm not going to use any 19:06of these background tools. I'm going to 19:09I'm going to have to copy and paste 19:11things." And that itself, you know, 19:13takes time. And the fact that we're 19:15seeing Claude able to actually move from 19:17within the workspace into these tools, I 19:22mean, we're we're talking about, you 19:23know, some gamechanging stuff. And we're 19:26not really even talking about like 19:27things when I'm really curious. I want 19:30to push this to its limit. I almost want 19:32to see it do a pivot table, you know? 19:35>> We can do a live exercise here. I think 19:36we should like in a minute here. I want 19:38to like have you give me a live exercise 19:40and we'll like set it up and we'll run 19:41it while we chat and then we'll go back 19:43and look at it. I think that would be 19:44really fun. 19:45>> That would be a lot of fun. 19:47>> But first, I want to make you wse. 19:48You're a designer. I want to make you 19:49suffer just a little bit. 19:51>> I'm here for it. You are going to for 19:54better or worse, you are going to see 19:56what agent put together for the 19:57PowerPoint side of things. So this is 19:59that same report you looked at earlier. 20:01It's the same slide. 20:02>> Look how painful this is. It's so 20:05>> You know, you know what's funny is a 20:09year and a half ago, this was really 20:11impressive. 20:12>> I know. Time flies. 20:13>> This was really This was really 20:15impressive. And then you showed me 20:18Claude 20:19>> and now this just doesn't 20:22>> just doesn't work the bar. 20:24>> It doesn't work. 20:25>> Look at the these little footnotes are 20:27completely unreadable and like 20:30look this isn't 20:32title like it's just it's just kind of 20:34bad. 20:36>> Well, listen. I think what we're 20:37understanding here is that, you know, 20:41the certain the agent is not the 20:44designer, you know, between the two. 20:46It's clearly the agent is just, you 20:48know, he's just about the raw numbers 20:49>> or she well and what's interesting was 20:51that the conclusion that Perplexity came 20:53to. So, Perplexity looked at the like 20:57text outputs and basically was like if I 20:59had to pick, I would say give your 21:02assignment to agent to do some initial 21:04analysis and then feed that to Claude to 21:08actually do the report preparation and 21:11the and the PowerPoint and all of that 21:13because then you get the best of both 21:14worlds. 21:15>> 100%. It's funny that you bring that up 21:17too because like daisy chaining is such 21:20a powerful tool when you're using 21:22working across LLMs. It's like getting 21:25the best of all world. So 21:27>> yeah, 21:27>> that's a great call. 21:28>> Yeah. All right. Well, let's have some 21:30fun. I want to go to Perplexity together 21:32and I want you to help me build a prompt 21:34in Perplexity that we can feed to Claude 21:37Live and sort of have some fun with. 21:41>> Awesome. Because you said you had like a 21:42pivot table idea. You wanted to have 21:44fun. So what's what's your idea? What do 21:45you want to do to like challenge Claude 21:47a little bit here? And 21:48>> we'll see what we get. 21:50>> Interesting. Let's see what would make 21:52for a good 21:55You know what? 21:57Given in the spirit of using AI, let me 21:59come let me let me uh brainstorm some 22:03ideas 22:04>> as I would normally do. 22:06>> Yes, GPT. 22:08>> Do you want to share your screen while 22:09you're doing that so we can kind of see 22:10you in the brainstorming mode or or uh 22:13are you too shy? 22:15No, I'm not shy enough, I think. Here we 22:18go. 22:20>> All right. Well, then great. What I will 22:21do is I will stop showing vertex for a 22:23second. This is why this is live. Um, 22:26and you can share your screen, I think. 22:30>> Share button down there. 22:32>> And here we go. So, I'm just going to 22:35straight up ask, help me brainstorm 22:39a few ideas around making a fun 22:45pivot table. 22:46>> I love that we're using the phrase fun 22:48pivot table. Like, who says that? This 22:51is why I want to 22:51>> It's 2025, you know? 22:55LLMs have changed our entire 22:56perspective. 22:59>> Pivot tables are now fun. 23:00>> Make it dynamic. Make the suggestion 23:05a 23:06>> All right. 23:07>> GBD5. Let's see what it says. 23:10>> Movie night. 23:12>> Pivot hacker. 23:14Fitness fun. 23:17>> Travel vibes. 23:20>> Recipe. These are very consumer focused. 23:23>> They are. Should we make them a little 23:25more Should we change them and make them 23:28a little more corporate? 23:29>> I mean, we could or we could stick with 23:31this. You get the test either way, I 23:32think. What do you think? What do you 23:33want to do? 23:36>> Let's go with movie night. Let's make a 23:38movie night pivot table. 23:39>> Movie night pivot table. Okay. Let's 23:41throw that into Perplexity. So, I'm 23:43going to go over to Perplexity. Uh, or 23:44do you want to go to Perplexity? 23:46>> What I'll do is I'll copy and paste this 23:48over to you so you can 23:50>> paste it, right? 23:51>> And then I can pop it in. Okay. Go 23:52ahead. 23:52>> A little team effort. We're gonna have 23:55to find a way to bridge our 23:59uh our hive mind AI. 24:01>> That's right. I love those. 24:03>> While I'm while I'm sending this over to 24:05you via the interwebs. 24:07>> Yes. 24:07>> What are you What are you thinking that 24:10is the most unique part of this launch 24:13that made you just say, "Oh, this is 24:15this is different. This is worthwhile." 24:17What made you really excited about this 24:19one? 24:20>> I Well, I think I shared this at the 24:22top. I was primed to be bored. I was 24:25primed to not be excited. My expectation 24:28was this is not going to be that much 24:31because I've seen this hype before. And 24:33what I found was just the ability to 24:36deliver on promises around work done 24:40became really really compelling for me 24:42because it meant that I could be in a 24:44spot where 24:47I felt comfortable giving a bigger piece 24:50of responsibility 24:52to AI which in turn like if if I think 24:54about it like a lot of the question we 24:56face that we wrestle with is what do we 24:59trust AI to do? How do we delegate 25:02effectively? And how much time does that 25:03give us back? How do we multiply 25:05ourselves as AI professionals? And this 25:08feels like one of those things where 25:09like I can I can see all of these places 25:12through my day where I can start to 25:14delegate differently and I can push more 25:16to Claude because I can trust him. 25:18>> That makes a lot of sense. 25:20I mean, that's one of the things I'm I'm 25:22looking for whenever I see these um a 25:25launch or any of these guys, whether 25:27it's Anthropic or OpenAI 25:29put out a new announcement, always 25:31curious on like how can it change my 25:34day-to-day, you know, and then 25:36>> quickly try to get in there and and and 25:39tease out that value. 25:41>> Exactly. Okay, I picked up the text. Um 25:44so, I'm gonna go back. I'm gonna share 25:46my screen. And I have the text I think 25:49in my clipboard and we're just going to 25:52start to build a prompt live that we're 25:54going to hand to Claude and see what we 25:55get. So the pasted text, it's movie 25:58night. Um, and what I basically want to 26:00ask Perplexes is I want to say uh I want 26:04you to uh take this seed of an idea and 26:09expand on it. Uh specifically, you need 26:13to build a prompt for another LLM to 26:17construct a pivot table spreadsheet 26:21uh that includes 26:25uh a list of 20 uh to 25 recent movies 26:32uh categorized 26:35as described below. Um and the pivot 26:40table 26:41needs to be very usable um and 26:45userfriendly. So this is like me 26:47actually like prompting raw here. Uh in 26:50addition, 26:52the prompt you construct must contain 26:57all of the data the LLM needs to 27:01construct the pivot table. do not assume 27:05the LM can go and get movie data, 27:10you get the movie data. 27:12And that I found was really key. Uh 27:15because you can't try and give it like 27:18the research job and the Excel creation 27:21job in one go at this point. um it will 27:25go and do research, but I find it's much 27:27more effective to have Perplexity do the 27:29research and then come back with 27:31something like a self-contained prompt 27:34>> and then work. Um, and what that 27:37suggests, by the way, is that if you're 27:39working with internal company data and 27:41you're working with Claude, you're going 27:42to end up in a position where you need 27:44to be constructing a prompt very 27:46intentionally pulling in internal 27:48company data as part of the prompt 27:49construction process and then giving it 27:51to Claude. 27:52Yeah, 100%. And I think that that's 27:54going to be a nice exercise for us as we 27:56see this first version that it spits out 27:59>> because I'm already thinking about like, 28:01you know, maybe we want to just focus on 28:03say the Marvel Cinematic Universe 28:05>> and then we want to overlay characters, 28:08you know, and then we want to 28:10>> add actresses and all this fun stuff. 28:13So, I'm really excited to see what what 28:16uh what Perplexity puts out there. 28:19generating a complete data set of recent 28:22movies. So like I cannot believe that 28:24here we are in 2025 and this is just a 28:26casual prompt that you and I are just 28:28throwing up. 28:28>> Look out IMDb, you know, 28:31>> right? 28:32>> Yeah. You know, one of the things that I 28:33think that people don't realize, I was 28:35talking with someone uh earlier today 28:38who uh I mean he's a company leader and 28:41he hadn't he he uses perplexity. He 28:44loves Perplexity, but he didn't realize 28:46that you could hit these three different 28:47options on the bottom, and he just 28:49defaults to search and loves it. And I 28:50was like, there is more to Perplexity. 28:52You can do even cooler. 28:56>> You You unshackled him. 28:58>> I did. I Well, I'm hoping I did. We'll 29:00see. Um, so it gives me like a bunch of 29:02uh recent movie titles. It gives me a 29:04bunch of view records data uh showing 29:07who watched what when. Uh, it gives me 29:10pivot table specs. It gives me expected 29:12output examples. user experience 29:14requirements. This is very complete. 29:16This is this is very very thorough. Do 29:19we want to just like run it and see what 29:21happens? 29:22>> Absolutely. 29:24>> All right. All right. We're going to go 29:25run it and see what happens. Okay. I'm 29:26going to stop sharing and I'm going to 29:28paste it into Claude and we're going to 29:29see what we get. 29:31>> I'm excited. I'm rubbing my hands. I'm 29:32not even doing anything. 29:34>> I'm like, this is going to be really 29:36fun. All right. So, we're going to copy 29:37it. Um, and I'm glad I told it to be 29:40self-contained because there's no way I 29:41would be writing all of that stuff out. 29:43All right. So, I'm pasting it in and I 29:45actually want to do this live. So, once 29:46I get it queued up, I'm going to share 29:47my screen again and we're just gonna see 29:51how it goes. Okay, here we are. Gonna 29:53share my screen and go back to Claude. 29:56Lot of screen sharing in this one 29:57because there's so much cool stuff to 29:59look at. Um, okay, here we are. And 30:02literally like it pasted it in and you 30:04can open it up and look at it and you 30:06can see uh this is the instruction. It 30:09does specifically say create a 30:10spreadsheet. So, I'm hoping I don't have 30:11to remind it to create the spreadsheet 30:13this time. It has all the data we looked 30:15at. Uh, in addition, it dumped in a 30:18bunch of URLs because it just can't help 30:20itself. That makes the prompt slightly 30:23dirty, but I don't care. I'm just 30:24leaving it in for now, and we're just 30:26going to see what it does. 30:28>> I love how Perplexity loves citation. 30:32>> It does. It really does. Um, and so it's 30:36starting to go through. Uh, 30:39it's going to create a So, so it's first 30:41analyzing it. So, it's got the Python 30:43script up here. You can actually see it 30:44like starting to code it in, which I 30:46This is just addictive to me. Like, I 30:48love this. 30:49>> Look at the data. 30:51>> Yeah, the world is a different place 30:53now. It's just wild. 30:54>> It's It's so funny that there was a time 30:58where doing something like this was 31:02extremely time inensive. And 31:06>> I say a time as if that wasn't 31:08>> as if that was a couple years ago. 31:13>> It just feels so far away. Like we are 31:17such creatures of convenience that the 31:20moment you put things like this in front 31:22of us, we were like, "Give me. This is 31:24the new the new floor." You know, it's 31:28really cool. 31:29>> And so it's just going through and as 31:31far as I know, like it's not even 31:32touched the spreadsheet. So, it all it 31:34did was it used Python to figure out 31:36what it was going to do and then it's 31:38decided, okay, now we're going to build 31:40a pivot table. Um, and we're going to 31:43like actually start to bring an Excel 31:45file with quote unquote advanced 31:46features. So, we'll see how it does on 31:48that. 31:49>> I really love the way it thinks. I love 31:52the way it's it does the thinking out 31:54loud and 31:55>> yeah, the 31:56>> and the way that it's leveraging Python 31:59to give it a P, right? so that it can 32:02then use that. And I think that's that's 32:05something that I'm noticing across the 32:07different LLMs. And I think that the 32:09engineers are getting smarter about this 32:11like they know that there are certain 32:13languages that just lend themselves well 32:16to going from human language to you know 32:20computer 32:21>> that computers speak and Python is just 32:23one of those languages. So if you could 32:25get into Python, all of a sudden 32:27shifting into other languages becomes a 32:29lot easier because you're getting that 32:31structure built in. So really really 32:34cool stuff when you when you're actually 32:35paying attention to what's um what's 32:38going on there. Uh you're not just pure 32:40vibing, which I love to do too. I mean, 32:44you know, a good a good vibe is great. I 32:47think that when I'm able to pay 32:49attention, I notice that Claude, like if 32:51I'm not running back and forth between 32:5320 different things, I notice that 32:54Claude is much better about this than 32:56OpenAI. Like the the tool record here is 32:59really clean. I can see exactly what it 33:02did all the way through. And now like it 33:06loves documentation, so it's creating me 33:07a read me. I didn't even ask it to 33:09create me a read me. 33:11>> This is so great. 33:13>> How to use the spreadsheet. Thank you, 33:15Claude. Sometimes Claude is a little bit 33:18a little bit stuck up, but I love it so 33:21much. 33:22>> I don't mind it. Can Can you laminate 33:25it? Can you collate it? Laminated and 33:27and mail it to me too, please, Claude. 33:29That'd be great. 33:30>> So, you remember Calvin and Hobbs? 33:31Claude has the personality of Susie 33:33Derkens to me. 33:36Like, it's it's that kind of a model. 33:39>> That's a great call back. 33:41>> Yeah, 33:41>> I love Calvin Hobbs. 33:43>> It's so good. My kids are addicted to 33:45it. 33:46Um, okay. Okay. Oh, wow. It 33:48overachieved. It has a a complete pivot 33:51table and apparently an enhanced 33:53analytics version. What? 33:55>> Amazing. 33:56>> Amazing. 33:57>> Okay, we have to see what's in here. Uh, 33:59so I'm going to download both of these. 34:02Um, and I'm going to pull down the 34:04readme. I actually want to open these in 34:05Excel because I think this is just a 34:07crappy way to look at it. So, give me a 34:09second. We're look at this in Excel and 34:11see what's in the box. It's gonna be 34:14very hot. 34:15>> Excited. 34:16>> I think we 34:18>> While you're pulling that up, I'm I'm 34:19also excited to see how 34:22>> how Microsoft and these other companies 34:24respond to this, right? Like 34:26>> the amount of dynamic behavior that 34:28they're going to have to start 34:29introducing to their products. It's just 34:32opening so many doors. 34:34>> Yeah. 34:34>> So, it's so cool. 34:36>> My gosh, look at this. Okay. I didn't 34:39even know it could do this. I know. I 34:41just I I'm killing you. The suspense 34:44alone h half the audience is is keeled 34:47over. 34:48>> Yeah, I think has become exciting thanks 34:51to Nate and Rod. Um, 34:54>> so this is the enhancement. 34:56>> It did color coding. 34:58>> Oh, heat coloring. 35:00>> Yeah, 35:02>> it heat maps. I'm like kind of 35:05impressed. I didn't ask it to heat map. 35:07It volunteered. 35:09>> Cool. 35:09>> It did pattern analysis. Look at this. 35:12Wow. I'm kind of blown away. I didn't 35:14know it could do this. Rod, this is also 35:16me reacting. 35:18>> Are those 35:20>> Those are little like uh 35:21>> Do you zoom in? They're like 35:23>> spark lines or something. 35:25>> Smart graph line thingies. Whatever you 35:27call these things. I don't even know. 35:28Like I I I'm not the the Excel expert 35:30here. What do you call these things? 35:31Spark lines. 35:32>> No, I don't. I I use language that I 35:35didn't didn't really know. Like little 35:36bar graphs or something. 35:37>> Yeah, they're little bar graphs right in 35:39the cell. And then it has this cute 35:41little like okay like you know take away 35:42a little credit because the genre is on 35:44top of the bar but whatever like I would 35:45say good 35:48>> uh you get friend statistics so you can 35:50see what these different friends like 35:52unique movies diversity etc. You get 35:54your top 10 movies ranked. 35:57So like this is the Suzie Derkens 36:00overachievement version 36:02>> and then it has all the raw data right 36:04>> and the raw data that's what we love to 36:06see. That's so good. 36:07>> Yeah. So, like you can tell I'm not 36:08lying is the implication, right? Um, 36:12and so there it is. And so that's the 36:13enhanced version. I'm going to go ahead 36:15and share with you what it called the 36:17basic version, which I thought was still 36:19perfectly usable. And you can kind of 36:22see like I think you get a sense of like 36:24what it thinks is special, which is its 36:26own little piece of insight to grab. So 36:28yeah, this is it. This is regular movie 36:32night without the heat mapping. Um, so 36:34it's exactly the same table. So you see 36:37the 12 is there and all you have is a 36:39nice clean traditional finance table 36:41with blue and white rows data perfectly 36:44understand 36:45>> but you know 36:45>> they segment out each of the uh viewers 36:48>> that's 36:49>> that was not in the enhanced version so 36:51they actually data on Alex Jordan Casey 36:56Morgan and Riley and you can see like 36:59how they go by day what their genres are 37:02and so in a sense like one of the things 37:03that I am learning from this is that you 37:07have to the the more interesting you 37:10make the prompt for Claude, the more 37:12likely Claude is to give you some 37:14permutations. And what I might do in a 37:16followup here is basically say, "Claude, 37:19I loved the heat mapping on the pivot 37:21table. 37:22>> Can you please preserve the heat 37:24mapping, but also give me the viewing 37:26pattern pivots and like do a 37:27modification like as a 37:29>> That'd be really cool." Actually, open 37:31up the pivot table. 37:33>> Uh, this one all friends. Yeah. 37:34>> Yeah. And can we pull in can we go into 37:38the If you click on data. Uh where's 37:42data? Okay. Yeah, up here. Please go 37:43away. Um 37:46yeah, data. 37:48>> Let me take a look. It's been a long 37:50time since I've been inside Excel. I 37:53will tell you that. 37:55>> Are you looking for or something? 37:57>> I'm looking for the pivot. There's a 37:59there's a table that you could actually 38:00see how it creates the cross 38:02relationships. Oh, is it in analysis 38:05tools? 38:07Is it? 38:07>> No, it's not. Um, 38:10what if analysis, text to columns, 38:13flashfill, queries, and connections. 38:15That's it, isn't it? 38:17Is that it? No, that's not it. Um, we'll 38:20have to find that later. I don't think 38:22we're finding. 38:22>> We will. That looks really cool. This is 38:24This is awesome. 38:26>> Yeah. No, it's super fun and I'm 38:28learning a lot, too, as we chat. So, 38:30there you go. That is it. That is what 38:31Claude has launched. Do you have any 38:33takeaways from this? 38:35>> I'm I'm excited. Say that again. 38:37>> Do you have any takeaways? Like what 38:38what are you taking away from this 38:39conversation? 38:41>> I mean, for me, what I'm realizing is 38:43that the idea of using these tools is no 38:47longer relegated to just me as a person. 38:50The amount of decisions that I need to 38:52make around getting something valuable 38:54out of the outputs that I'm getting can 38:57really really start diving into the more 39:00interesting bits, right? One of the 39:02awesome things that we've seen over the 39:04past couple of years with these LLMs is 39:06that we all get, you know, we're all 39:08over the first 20, 30, 40%. 39:11>> And most of us by the time we're really 39:13proficient at what we do, the 80% is 39:15pretty much, you know, you could phone 39:17it in, 39:18>> but it's that last 20% 39:20>> that makes things go from good to great. 39:23And the fact that we could spend more 39:25time focusing on that 20% and this is a 39:27perfect example of that makes for a 39:30really interesting like future in like 39:32both you know the professional space 39:36just the communication aspect. I mean 39:38this is just really really cool stuff. 39:40So 39:41>> so I want to give you a philosopher 39:43quote that I promise is relevant that I 39:45am taking away as I look at this. Um 39:47this is from Alfred North Whitehead 39:50um who famously said, "Civilization 39:53advances by extending the number of 39:56important operations which we can 39:58perform without thinking of them." 40:00>> That's actually really that's a really 40:02good one right there. 40:02>> And I feel like extended civilization 40:05because there are important operations 40:07where I didn't tell it to put the heat 40:08map in. I didn't tell it to center the 40:11the text on the PowerPoint slides. It 40:13just did it and I didn't have to think 40:14about it. 40:16That's awesome. Yeah, I like that. I 40:19also like that positioning. 40:21Anthropic might come knocking, mate. At 40:23this point, you're you're saying that 40:26Anthropic by extension is extending 40:28civilization. Just fantastic. 40:31>> Yeah. No, I'm sure. Well, if Daario 40:33calls, I'll pick up the phone. 40:35>> Well, thank you for having Thank you for 40:37for chatting a little bit. Thank you for 40:39making time. I know this was a bit of a 40:40last minute thing. We're having fun. 40:43First of May, 40:43>> loved it. Looking forward to doing 40:45another. Have a good one, Nate. Cheers.