Latest / QB Power Hour Podcast / Become a Promptologist with Knitwit [9.8.26]
Transcript
- 0:23Dan DeLong: So we are gonna be talking about a, a, term that I coined a couple months ago called being a promptologist.
- 0:31Dan DeLong: And we're gonna be talking about a new a new offering called Knitwit that Joel has, has created.
- 0:39Dan DeLong: And so we'll just be talking, that through.
- 0:46Dan DeLong: Oh, if you did join late yeah, those those links won't be there, so they should be in the, chat there.
- 0:54Dan DeLong: So apologize for that.
- 0:58Dan DeLong: And okay.
- 0:59Dan DeLong: So Sharrin is not here today either.
- 1:03Dan DeLong: She is recovering from a procedure, and so we wish her well, with that.
- 1:09Dan DeLong: So flying solo today, so appreciate you joining me on this on this journey here.
- 1:16Dan DeLong: Okay.
- 1:16Dan DeLong: So let's see.
- 1:18Dan DeLong: We've got our spotlight sponsors, so appreciate checking out checking out them to support the, the QB Power Hour and keeping it free for you guys.
- 1:28Dan DeLong: If you want to join or, connect with Sharrin or myself in our respective communities, there's the QR codes for, that as well.
- 1:39Dan DeLong: So Joel is the CEO and founder of Timesheets.com which is a competitor to QB Time.
- 1:49Dan DeLong: So if you are looking for a an alternative to to QB Time, Timesheets.com is is a great place to start.
- 2:00Dan DeLong: But he created a he created a a self-serving tool called Knitwit, and that's what we're gonna t-talk about today because these days with AI being so prevalent a lot of people are unable or maybe just don't know where to begin when it comes with, learning and, utilizing AI.
- 2:31Dan DeLong: Or you might be-become super...
- 2:36Dan DeLong: I'm trying to think of the word to describe it.
- 2:37Dan DeLong: I'm sorry, I'm a little lost for words here today, so I apologize for for that.
- 2:43Dan DeLong: Let's just jump into it here, and we'll hopefully get to the get to the meat and potatoes of what we're talking about here today.
- 2:52Dan DeLong: So what we're covering is highlight some of the topics of why large language models, LLMs, disagree.
- 3:01Dan DeLong: The cost of AI because we've gone from this Phase season that going from everything was free when, ChatGPT was first came on the scene, everything was free, and now we're getting into this point where it's that investor money that is that is plugged into those, those companies is starting to run out.
- 3:30Dan DeLong: So they're they're trying to figure out a way to monetize or cost charge for, these things.
- 3:39Dan DeLong: So Joel actually in part of my communications with him he was telling me about the under mech- the mechanisms of tokens and why things c- are, starting to cost so much.
- 3:56Dan DeLong: So we'll talk a little bit about that.
- 3:58Dan DeLong: We'll talk about how Knitwit helps with that.
- 4:01Dan DeLong: And then we can't do a live demo.
- 4:04Dan DeLong: I guess we could, but I was hoping that Joel would do said live demo.
- 4:11Dan DeLong: So hopefully we'll...
- 4:13Dan DeLong: he'll be joining shortly with that.
- 4:16Dan DeLong: But just so that we, Joel and I are going to be doing a deeper dive demonstration of Knitwit next Wednesday, and I'll, have links for that on, on our the School of Bookkeeping workshop.
- 4:33Dan DeLong: So if this intrigues you and you wanna see a little bit more insight into that please join us for that, and I did put them in the links in the chat.
- 4:46Dan DeLong: All right.
- 4:47Dan DeLong: So let's launch the first poll question, which will be Sharrin would be doing.
- 4:54Dan DeLong: So just wanna get an understanding as to what you use for AI, the large language models that you use the most.
- 5:02Dan DeLong: This is, should be a multi-select, so if you, use Claude and Grok and Gemini and you can mark them all off there.
- 5:14Dan DeLong: I know ChatGPT was the, large language model that kinda came out first, and that's for me ChatGPT is just I, tri-- I used it and I tried it and thought it was pretty amazing, the things that it could do.
- 5:36Dan DeLong: But I've, I never really adopted paying the fee for ChatGPT and, going in, going down that route because other AI tools kinda came in secondary.
- 5:50Dan DeLong: So I'm more...
- 5:53Dan DeLong: I, resisted Claude for a while because I didn't wanna learn something new.
- 5:58Dan DeLong: And but since I have adopted and looked into Claude Claude has been the, the forefront for me.
- 6:08Dan DeLong: So I'm gonna go ahead and end the poll here.
- 6:11Dan DeLong: It looks like most people have done it, and just take, take a look.
- 6:15Dan DeLong: Yeah, it looks like fifty-nine percent is using ChatGPT, followed closely by Claude and Gemini Copilot, maybe something else.
- 6:27Dan DeLong: Two percent for Grok.
- 6:29Dan DeLong: I, think I, don't really use Grok that much or I haven't, really investigated, the, Grok model.
- 6:37Dan DeLong: Mostly because of who's behind it.
- 6:41Dan DeLong: But that's just me.
- 6:44Dan DeLong: Okay.
- 6:45Dan DeLong: So let's see.
- 6:46Dan DeLong: Oh, next next slide here.
- 6:48Dan DeLong: So let's talk a little bit about why different LLMs give different results, right?
- 6:53Dan DeLong: You can have the same prompt with different models.
- 6:56Dan DeLong: Why do they wildly differ?
- 6:59Dan DeLong: So if you joined us for our previous QB Power Hour I think when, was that?
- 7:06Dan DeLong: Oh in June we did People versus Robots, and Sharrin had did a great exercise where she took the exact same prompt and put it through four different LLMs, and then evaluated each of the outputs, right?
- 7:24Dan DeLong: So surprisingly, or maybe not surprisingly, is that those results varied, right?
- 7:32Dan DeLong: And, it she had to go through and scrutinize the output, and she gave it a letter grade.
- 7:39Dan DeLong: So there's a link there in the slides.
- 7:42Dan DeLong: If you didn't join us for that one you can watch the, recording from Zoom.
- 7:50Dan DeLong: But really the the, difference is the training data, right?
- 7:55Dan DeLong: So the, what that large language model is trained on, right?
- 8:02Dan DeLong: Like the data set that they have to, work with its architecture and size, right?
- 8:11Dan DeLong: If, you're using Claude you have a choice of which model it's using off of that training set that, we're talking about.
- 8:20Dan DeLong: And that alignment tuning because each company tunes its model's voice and those guardrails.
- 8:28Dan DeLong: It's, not really so they-- tho- those outputs are fine-tuned to that particular model, right?
- 8:39Dan DeLong: And then randomness, right?
- 8:41Dan DeLong: And we'll talk a little bit about tokens in, a, moment here.
- 8:47Dan DeLong: So which will hopefully shed some light as to why things are interpreted differently.
- 8:56Dan DeLong: But the, the bottom line is none of this re- means that the model's broken.
- 9:01Dan DeLong: It just really means the results are a matter of process, right?
- 9:06Dan DeLong: So it's like cooking, right?
- 9:09Dan DeLong: There, there's recipes that are out there.
- 9:12Dan DeLong: There's fourteen thousand different ways to make a meatloaf, and they all come out little, bit different.
- 9:21Dan DeLong: So it really just de- depends on that kind of prep and mixing of those ingredients.
- 9:29Dan DeLong: So your results may vary even in the same chat or the same inter- iteration of conversation that you're having with the model.
- 9:42Dan DeLong: So why does this matter to you, right?
- 9:44Dan DeLong: So if you have inconsistent outputs you're gonna have inconsistent results, and those results could be fragile.
- 9:56Dan DeLong: We did a, an, an interesting workshop cohort that we did at School of Bookkeeping where the example was how many Rs are there in strawberry?
- 10:09Dan DeLong: And it was confidently saying that there were two Rs in strawberry when there are-- anybody who can count Rs would be able to see that there's three of them.
- 10:21Dan DeLong: So it, gets to the point of where if it has a confident hallucination that those outputs could be challenging and, throw that, Ish-- A, a potential issue of vote of confidence type of thing.
- 10:43Dan DeLong: And looks like Claude was just updating here.
- 10:46Dan DeLong: If you caught that.
- 10:48Dan DeLong: Oh, let me sh- stop sharing the results here, the qu- the poll.
- 10:53Dan DeLong: And then you don't have a repeatable process, right?
- 10:56Dan DeLong: If you're gonna have a different output each time it's not gonna it's not gonna be a shampoo thing of lather, rinse, repeat, right?
- 11:08Dan DeLong: So you need to really make sure that you're thoroughly rinsing so to speak So we wanna talk a little bit about the cost of AI and the cost that the monetization of AI.
- 11:26Dan DeLong: We've de- depending on where you sit on the spectrum there could be a lot of all these things about data centers and all that stuff.
- 11:37Dan DeLong: We're not gonna talk about that.
- 11:38Dan DeLong: We're just talking about the the cost of tokens what you're actually paying for and, that sort of thing.
- 11:46Dan DeLong: So really, what is a token?
- 11:49Dan DeLong: So a token is a portion of a, a word or a portion of a word basically.
- 11:59Dan DeLong: And this was something that that when, I spoke to Joel about Knitwit that really s- shed, shed some light to me.
- 12:09Dan DeLong: So the way that...
- 12:13Dan DeLong: In, in, in its fundamental sense, a large language model is taking a word or a portion of a word, turning it into a token, and then that token gets run through the model.
- 12:29Dan DeLong: And then what it will do after thousands of lines and millions of lines of code will try to predict the outcome in token form and then reassemble that, right?
- 12:45Dan DeLong: So that kind of explains a lot of these hallucinations, right?
- 12:48Dan DeLong: Where it's really taking words or characters of words, turning them into a, a mathematical equivalent, doing math behind the scenes, and then reassembling that and then putting it back in a, a language, which makes a lot of sense to me why if you've ever had an email or a text exchange with me I misspell things horribly.
- 13:17Dan DeLong: I tried Grammarly, and it made it worse because it was making things harder for me because, it would e- it, it wouldn't, didn't work for me.
- 13:29Dan DeLong: But I, I didn't quite understand at that point, like, why why it understood my misspellings, right?
- 13:38Dan DeLong: And, this kinda sheds that, to light to me for how it's taking a, a, a word or a portion of a word, figuring it out what that word might be, turning it into a token, and then giving that output.
- 13:56Dan DeLong: And so those tokens are, in and out.
- 14:00Dan DeLong: They, turn into word...
- 14:01Dan DeLong: Your words turn into tokens, and those tokens are, used on both sides of the lang-- the, LLM's processing in the back end.
- 14:16Dan DeLong: So they charge for both in and out.
- 14:19Dan DeLong: They get you coming and going with, regards to what, a token is.
- 14:23Dan DeLong: And it's every message.
- 14:26Dan DeLong: So every long...
- 14:28Dan DeLong: The long chat histories it may have to reread some of that if you have an ongoing chat.
- 14:37Dan DeLong: And more importantly, it may lose the context.
- 14:41Dan DeLong: If you close the chat window you may need to re-explain because, oftentimes what we do sometimes is we we-- That experiment of how many Rs in, strawberry, you close the chat window if-- even after it realized its error, it's oh yeah, I counted it again, there's actually three Rs in strawberry.
- 15:07Dan DeLong: Then it you close the chat window and come in and ask the same question, and it's now back to two.
- 15:14Dan DeLong: So that memory of...
- 15:16Dan DeLong: a-and the re- the recollection of its mistake, is, lost in the, content or context, excuse me.
- 15:26Dan DeLong: So where costs in tokens tend to sneak up, right?
- 15:30Dan DeLong: So if you paste in a sixty-nine page PDF every power every character becomes a billable context, right?
- 15:41Dan DeLong: So best, better option is to summarize and then paste only the relative parts.
- 15:48Dan DeLong: Those never-ending chat threads because the model rereads the whole history each time.
- 15:55Dan DeLong: So the fix for that would be starting a new chat.
- 15:59Dan DeLong: I've noticed with Claude, a lot of times it will summarize the chat.
- 16:04Dan DeLong: As you get pretty long it'll re-summarize so that it does not have to reread the entire chat You know, vaguely vague wordy prompts the lot of back and forth where you have to go back and forth with the chat.
- 16:21Dan DeLong: Being as specific as possible and then those tones up front, can, help with that.
- 16:29Dan DeLong: Re-explaining instructions the same context typed over and over.
- 16:34Dan DeLong: Obviously, every time that you type something, that turns into a token and will ultimately be a billable event so to speak with, regards to how they're going to ultimately be charging for these things.
- 16:51Dan DeLong: And then using your priciest model for everything.
- 16:54Dan DeLong: So I have-- I'm not able to keep up with all of the terms for the, models because they're, changing so quickly.
- 17:04Dan DeLong: Claude has Opus and Sonnet and I'm not quite sure what Google Android had, suites with all of their, operating systems.
- 17:17Dan DeLong: They had Ice Cream and those types of things.
- 17:19Dan DeLong: So I don't know if they have a theme for the long language lar- large language models.
- 17:26Dan DeLong: But it's really hard to keep up which with, one, each one.
- 17:30Dan DeLong: But whatever the latest version of the, of model that they're using those are going to cost the most as far as tokens are concerned.
- 17:44Dan DeLong: So you can change the model, and so using, using, matching the model to the task is, really what it boils down to.
- 17:52Dan DeLong: If you're just having a, a chat trying to understand a concept maybe use the low, end of that, or for research.
- 18:02Dan DeLong: And then when you're ready to actually do something like Claude Cowork or Code or something like that, then you use the high the, the high-end one to actually do the work once you've processed what it is that you're actually gonna start talking with the with the model So four ways to cut the token toss token toss, token cost.
- 18:30Dan DeLong: Sorry, can't say that right.
- 18:33Dan DeLong: Tighter prompts, right?
- 18:34Dan DeLong: More more concise about your length, the format, tone before you actually, start the, chat.
- 18:46Dan DeLong: Reuse your prompts instead of and then don't rebuild them.
- 18:51Dan DeLong: Maybe sa-saving your best prompts as templates.
- 18:54Dan DeLong: Managing those, contexts and again, matching the model to the task.
- 18:59Dan DeLong: So the goal really isn't to spend the, the least it's, spending it on the right things, right?
- 19:05Dan DeLong: So right tool, right job, right?
- 19:07Dan DeLong: So there's plenty of examples of tools aligning with what the best way to do these things.
- 19:17Dan DeLong: You don't have to use a a screwdriver for everything.
- 19:22Dan DeLong: So y- there, there are different tools that do different things for for a reason.
- 19:29Dan DeLong: So what a lot of people end up doing is they wanna spend less, right?
- 19:34Dan DeLong: So they cut corners on prompts to, save a few cents or accepting good enough skipping the context redoing the same scratch task from scratch the next time.
- 19:48Dan DeLong: That, that's like a, a trap because you end up doing the same things over and over.
- 19:55Dan DeLong: Spending smartly, right?
- 19:56Dan DeLong: If you write one clear, complete prompt the first time using the right model for the complexity of the task saving the, a, a working prompt as a template so that you don't have to go through that conversation again to get the the temp-- the, output that you were looking for.
- 20:22Dan DeLong: Trusting the output because you built a repeatable process, right?
- 20:26Dan DeLong: By being able to craft your prompts ahead of time and knowing what that output is, then you spend, end up spending less time scrutinizing the results All right, so let's throw up that second poll question Have you noticed different AI models give you the same different answers to the same questions?
- 20:52Dan DeLong: That, that kind of goes back to what we did with the People versus Robots webinar where Sharrin gave the exact same prompt and asked for the e-exact same output to different models.
- 21:08Dan DeLong: And and it ended up giving, different outputs While you're answering that, I'll go ahead and continue on.
- 21:22Dan DeLong: So this is exactly why Joel created Knitwit to for, just this reason.
- 21:32Dan DeLong: And he could probably tell you better than I could how how he came up with this with this idea.
- 21:43Dan DeLong: But basically it is that front end.
- 21:47Dan DeLong: It is a tool to help with that front end type of thing with regards to how crafting the, prompts so that you so that you get the right output or can do something with that output all in one place, right?
- 22:07Dan DeLong: So we'll talk a little bit about that.
- 22:08Dan DeLong: So it's a practical way to evaluate the, the, the outputs.
- 22:14Dan DeLong: It, you can check the facts, you can check the fit and, check the repeatability.
- 22:21Dan DeLong: What Knitwit does and, I'll go into my account wherever it happens to be so we can kinda take a look at it is, Oh, that's, I think it might be on the next slide.
- 22:33Dan DeLong: So really what it is, it's a workspace.
- 22:35Dan DeLong: So at, at the center of it, Knitwit is a, a workspace that that takes advantage of all of the wasted space that is typically in a large language model.
- 22:51Dan DeLong: If you go into ChatGPT, right?
- 22:56Dan DeLong: There's a lot of real estate on the screen, especially if you have a 30, 36-inch monitor or something and you full screen it there's a lot of white area, right?
- 23:08Dan DeLong: So what he, did with with Knitwit is to be able to use all of that space proportionately, right?
- 23:16Dan DeLong: So it's it allows for prompt staging so that you can stage your and craft your prompt and work on your prompt before you actually send it to any of the models, right?
- 23:31Dan DeLong: And then when you do send it, you can send it and compare it to multiple models.
- 23:37Dan DeLong: And you have an option of doing it fast, right?
- 23:42Dan DeLong: So it's really just a toggle button to say,"Hey, just use it, do it quickly or cheaply," or,"I need a little bit more knowledge," or"I want, the latest," so smarter, right?
- 23:56Dan DeLong: So you have fast or smart.
- 23:59Dan DeLong: And then it will allow you to compare each of those models.
- 24:03Dan DeLong: And as the, the name of Knit for Knitwit implies, you can stitch all of them together.
- 24:13Dan DeLong: And I think I just got a- I got a reply from Joel that he's joining, but I don't see him yet So we'll so we should be able to get him on here to, to do the demo when we're actually there, but he should be joining shortly here.
- 24:37Dan DeLong: So do, do.
- 24:40Dan DeLong: Okay.
- 24:40Dan DeLong: So the benefits of, so you can structure your prompts to yield those better results.
- 24:48Dan DeLong: The, the models are flexible so that you can choose which one which ones that you wanna, wh-which you wanna send it to.
- 24:57Dan DeLong: Because as you'll start to learn some models are better for certain things, certain tasks.
- 25:06Dan DeLong: So you can send it to the right one as you start to be able...
- 25:11Dan DeLong: If you send it to all of them you'll be able to see how each one has responded to, to said prompt.
- 25:18Dan DeLong: And then the, the prompts are able to be combined where it can or the outputs, I'm sorry.
- 25:26Dan DeLong: The outputs can be combined so that you you can put together the commonality threads and see where they differ, and then you're able to edit that, that output to, and share that with with whomever, right?
- 25:44Dan DeLong: You can share it with your client, you can share it with your team member.
- 25:49Dan DeLong: Let's see here.
- 25:51Dan DeLong: He's having trouble.
- 25:52Dan DeLong: Oh, he-- it went to the wrong email for him, so give me one second here.
- 25:59Dan DeLong: Okay.
- 26:00Dan DeLong: Bear with me just a moment while I get him invited.
- 26:08Dan DeLong: Yeah, I did, Sorry about the, the delay.
- 26:13Dan DeLong: And I'm in one location here, not on my normal workstation, so everything is in the way.
- 26:23Dan DeLong: So while I'm doing that, I'm gonna share the results so you can see what other people are saying.
- 26:30Dan DeLong: Second is Link All right.
- 26:35Dan DeLong: So hopefully he'll be popping in here shortly here.
- 26:40Dan DeLong: Okay, here's where we would normally have Joel do the live demo.
- 26:46Dan DeLong: Let me find out where mine is.
- 26:48Dan DeLong: Here we go.
- 26:49Dan DeLong: Over here See if let me give him a, a minute or so to see if he can't join us here Reading form.
- 26:59Dan DeLong: I'll see, take a look at the questions that have been in here because I can't talk and look at the chat at the same time.
- 27:08Dan DeLong: See if there's any questions here.
- 27:13Dan DeLong: Yeah, James yeah, Perplexity is, also one of those models that, that, you could use.
- 27:24Dan DeLong: Da, Okay.
- 27:26Dan DeLong: Somebody else was asking about the handouts Let's get to them Is all sorts of things from Zoom in the way, and I'm trying to get these things out of the way here.
- 27:39Dan DeLong: See Let me get him invited here so that There it is.
- 27:43Dan DeLong: Okay Hopefully when he joins in, we'll be able to do the...
- 27:48Dan DeLong: Let's see here.
- 27:51Dan DeLong: Where are we?
- 27:53Dan DeLong: Throw this over here All right.
- 27:55Dan DeLong: Let me make this bigger so we can see.
- 27:58Dan DeLong: All right, so this is Knitwit which is, again it's that, dashboard of And, utilizing the the, the space that's available, right?
- 28:12Dan DeLong: So across the top you have you, have your chats where you can start working with with the, the interviewer, and it will help...
- 28:28Dan DeLong: It'll start to create that, that chat for you.
- 28:32Dan DeLong: There is a composer which is creating a conversation or is, creating a, a, a staging, right?
- 28:42Dan DeLong: So if I click on it here, Why can't I click on these?
- 28:45Dan DeLong: Let's see here.
- 28:48Dan DeLong: Right here.
- 28:48Dan DeLong: So there's the composer which is you have your instructions here which you can start to, to type in something in here, but then there is a prompt staging area.
- 29:01Dan DeLong: And none of this is sending it to your any of the AI models, right?
- 29:07Dan DeLong: So so the, this was, A- and these are stages.
- 29:14Dan DeLong: I- in the staging area, you can reorganize.
- 29:17Dan DeLong: So you can drag and drop those stages around.
- 29:22Dan DeLong: You can edit them so that it has that sort of it puts that in that order.
- 29:30Dan DeLong: And then when you submit it to to AI, you have this, this option over here to audit the prompt.
- 29:40Dan DeLong: There's your...
- 29:41Dan DeLong: This is the faster versus smarter option here at the top.
- 29:46Dan DeLong: And then so the GPT, the Claude Haiku, Gemini Flash, and then the Knit, which is both of them, right?
- 29:57Dan DeLong: Or all three whichever ones you, send it to.
- 30:02Dan DeLong: And then when you send that stage so let's, this is the thread itself here in the middle, right?
- 30:10Dan DeLong: Which is everything that has that is in this thread or in this stage is, gonna be sent to to the model whi- whichever that happens to be.
- 30:27Dan DeLong: So you're able to craft and create and edit those prompts before you send it to any any AI model.
- 30:39Dan DeLong: And then if you send it to multiple ones, then you can stitch them together.
- 30:44Dan DeLong: So here I have two different outputs from from both both of those models.
- 30:55Dan DeLong: And then if I want to have the insights, I can run the Knit and...
- 31:00Dan DeLong: Oh, now it's...
- 31:03Dan DeLong: Why is...
- 31:04Dan DeLong: Oh.
- 31:04Dan DeLong: See, I think my credits are low, so I can't do anything So I need to talk to Joel about, that.
- 31:12Dan DeLong: So I'm unable to, actually stitch it together.
- 31:16Dan DeLong: But if you...
- 31:17Dan DeLong: Yesterday I had 1,000, so I'm not sure where they all went.
- 31:23Dan DeLong: But when you run the, Knit, then it will look through both results and isolate those, those outputs and then allow you to do c- combine those together.
- 31:38Dan DeLong: Got another email here.
- 31:43Dan DeLong: Let's see He's trying to join in, but for some reason he can't he's having some trouble joining the doing the link here, so apologize for, for that to invite him under his other email All right so hopefully he can join in shortly.
- 32:07Dan DeLong: But this is the general idea, right?
- 32:09Dan DeLong: So my son actually did this without using Netwit.
- 32:17Dan DeLong: He did the same thing where he created a...
- 32:21Dan DeLong: Here's Joel.
- 32:23Dan DeLong: All right, I'll let you turn on your camera's there.
- 32:29Dan DeLong: All right.
- 32:30Dan DeLong: Joel, are you...
- 32:31Dan DeLong: Can you, can we hear you?
- 32:34Joel Slatis: Hopefully.
- 32:34Joel Slatis: How are you?
- 32:36Dan DeLong: Good.
- 32:37Dan DeLong: So we are in process.
- 32:40Dan DeLong: I've been struggling to do this without you.
- 32:44Joel Slatis: I'm sorry I couldn't join.
- 32:46Joel Slatis: I've been trying for 20 minutes.
- 32:48Joel Slatis: The problem was I didn't have a Knitwit a Zoom account with Knitwit because- it's a new business, and I had to register for an account and it was confused with my other account.
- 33:01Joel Slatis: It's been an- Ah annoying nightmare to try to get in here.
- 33:03Joel Slatis: So anyway, I apologize- Got it but we're starting in 20 minutes, right?
- 33:06Joel Slatis: I'm not, or I'm not late?
- 33:08Dan DeLong: No, we're we've been live for about 40 minutes now Yeah.
- 33:15Dan DeLong: Oh, good lord.
- 33:15Dan DeLong: So I've been, I'm, I've been going-
- 33:15Joel Slatis: I, had it on my calendar for 10 o'clock this morning.
- 33:18Dan DeLong: Ah, okay.
- 33:18Dan DeLong: Okay?
- 33:19Joel Slatis: I'm so sorry.
- 33:20Joel Slatis: Here I am.
- 33:22Joel Slatis: Okay.
- 33:23Dan DeLong: All righty.
- 33:24Dan DeLong: So yeah, so I was going into the, the the live demo.
- 33:29Dan DeLong: So I've already gone up to this point.
- 33:33Dan DeLong: So folks joining or folks who are watching us banter back and forth here.
- 33:41Dan DeLong: So.
- 33:43Joel Slatis: Okay.
- 33:43Joel Slatis: I hopefully kickstart things for us.
- 33:46Joel Slatis: So we were here to talk about mainly prompts.
- 33:49Joel Slatis: Obviously Knitwit is an important product, but let's start with giving your users some important information about prompts.
- 33:55Joel Slatis: Yes.
- 33:55Joel Slatis: Now if, I I, don't know if you want me to just launch here or,
- 34:02Dan DeLong: Yeah I was, going through...
- 34:03Dan DeLong: So if you can I'm gonna stop sharing, and then if you wanna share your screen and just give the over- overview architecture of of, Knitwit I, think.
- 34:16Dan DeLong: And people help me out in the chat if I did the slides justice up until this point so that we're all aware of what we're trying to solve for with with Knitwit.
- 34:32Joel Slatis: Okay.
- 34:32Joel Slatis: Perfect.
- 34:33Joel Slatis: So let me start with And, again, I apologize this was postponed from last week.
- 34:41Joel Slatis: I had it on my calendar for 10:00, guys, so I'm really sorry I'm late.
- 34:46Joel Slatis: But, No worries bear with me.
- 34:48Joel Slatis: I'm just gonna share my screen here.
- 34:52Joel Slatis: And again, I thought I had a few minutes to prepare, but we'll just launch here.
- 34:56Joel Slatis: Okay.
- 34:57Joel Slatis: So I am...
- 34:58Joel Slatis: Knitwit is a new product, and I created it because I needed to be able to build long prompts.
- 35:05Joel Slatis: And they're the ChatGPT is just basically inadequate, right?
- 35:10Joel Slatis: You've got this tiny little text area, which they actually finally made a little bit bigger, but it's basically just a useless, not, doesn't work very well kind of thing.
- 35:20Joel Slatis: So we're gonna, we're gonna improve on that, as you probably figured out with Knitwit.
- 35:27Joel Slatis: So let me get into the app again.
- 35:29Joel Slatis: I just wanted to give you a contrast.
- 35:31Joel Slatis: Now there's three sections.
- 35:33Joel Slatis: There's the composer, the threads, and the response.
- 35:36Joel Slatis: The composer is where you write long prompts or just any prompt.
- 35:40Joel Slatis: So you can send your prompt directly to the compo- to the AI, or you can build on it and create something big.
- 35:46Joel Slatis: So if I switched, for instance, to I'll just look at our, Our context, this is what a long prompt might look like when you're actually doing something useful with AI.
- 36:03Joel Slatis: But obviously, this is something that you can't send to the ChatGPT's tiny little area there, and you can't edit it either.
- 36:11Joel Slatis: You can't make changes to it, and you can't really use it for anything useful.
- 36:16Joel Slatis: Now, when it comes to accounting there's a lot of things that you guys do every month that would be really useful for this.
- 36:23Joel Slatis: And again, I'm just launching into this, so I apologize.
- 36:26Joel Slatis: But let's suppose we wanna create a new prompt for something that you guys do often, which might be, Dan, I don't know, help me out here, like analyzing a profit and loss statement from the previous month, or maybe looking at some- Yeah
- 36:39Dan DeLong: that's, probably the easiest use case for the the, accountant community.
- 36:44Dan DeLong: Taking something as an output and then putting it have it, having it be analyzed through the model.
- 36:53Joel Slatis: Okay.
- 36:53Joel Slatis: So I'll walk you through that use case'cause it's really an interesting one, and then we can talk maybe a little bit more.
- 37:00Joel Slatis: So the first thing I'm gonna do is create...
- 37:02Joel Slatis: Now, Knitwit is not just a place to build and save better prompts.
- 37:06Joel Slatis: It's also a way for you to organize all of your chats because if you're like me, before Knitwit, you just had a million new windows open randomly to chat here and to chat there, and they-- it's water under the bridge.
- 37:21Joel Slatis: After you're done with that chat, you close that window, it's gone.
- 37:23Joel Slatis: Here, we have a foldering system, so basically folders and subfolders.
- 37:28Joel Slatis: So I'm gonna go ahead and create a new one, and I'm gonna call this Profit and Loss Reusable.
- 37:35Joel Slatis: Okay?
- 37:37Joel Slatis: And now we have something called a prompt builder.
- 37:39Joel Slatis: Dan, I don't know if you've seen the prompt builder yet.
- 37:41Joel Slatis: Have you seen it?
- 37:41Dan DeLong: Yes.
- 37:42Dan DeLong: Yeah, you showed me that last week.
- 37:43Dan DeLong: I used it yesterday.
- 37:44Dan DeLong: Okay.
- 37:45Dan DeLong: It was-
- 37:45Joel Slatis: Oh, how'd you like it?
- 37:47Dan DeLong: I, did- It's pretty cool, right?
- 37:47Dan DeLong: I did like it, yeah.
- 37:48Dan DeLong: It was just a, a nice little interview and it asked me pertinent questions for what it is that I wanted to do, and then ultimately created that stage for me.
- 37:59Joel Slatis: Now, for people who don't know how to write a good prompt, this is built right into Knitwit, so it will help you, and I think you'll see the value in it right away.
- 38:06Joel Slatis: So I'll say, let's just say,"I want to create a report that I run monthly that looks at the previous month's P&L and identifies interesting things for me to, for, for me to look at." Okay?
- 38:34Joel Slatis: All right, so I send that.
- 38:36Joel Slatis: Now, it's actually gonna analyze that and come back with a list of pertinent things.
- 38:40Joel Slatis: What types of things should the report highlight most?
- 38:43Joel Slatis: Significant budget variances, unexpected cost spikes, revenue.
- 38:47Joel Slatis: And you'll notice these things have underlines under them, and I actually can make this bigger.
- 38:52Dan DeLong: There you go.
- 38:53Joel Slatis: Yeah, sorry about that.
- 38:55Joel Slatis: I'm a little flustered, so I forget- my basic stuff here.
- 38:58Dan DeLong: I think
- 38:58Joel Slatis: we all are.
- 38:59Joel Slatis: What type of...
- 39:00Joel Slatis: Yeah.
- 39:00Joel Slatis: What type of findings should the report highlight?
- 39:02Joel Slatis: Significant budget variances.
- 39:03Joel Slatis: So it gives you a list, right?
- 39:05Joel Slatis: So you can select the ones you care about.
- 39:07Joel Slatis: I care about unexpected cost spikes and margin con- contractions.
- 39:12Joel Slatis: So I can easy- I can select these, or I can type in myself and et cetera, right?
- 39:19Joel Slatis: But you can just add it, and then you click Send, and it's gonna continue the interview, and it's gonna ask me a few more questions.
- 39:25Joel Slatis: Now, I can quit the interview anytime I want.
- 39:27Joel Slatis: I can build the prompt whenever I want by just clicking Build Prompt, or I can wait and see if it gives me any other good ideas that I might wanna use here.
- 39:36Joel Slatis: So what format should the report use to present these insights?
- 39:39Joel Slatis: A bulleted executive memo, a structured markdown table, annotated variance commentary, or a side-by-side briefing or outline or something like that.
- 39:47Joel Slatis: So let's just say a side-by-side outline.
- 39:49Joel Slatis: We'll send that, and we'll do one more just to see what it comes up with, and then we'll build our prompt.
- 39:54Joel Slatis: So this'll only take a second.
- 39:55Joel Slatis: What primary constraints should guide should guide the depth and length of this outline?
- 40:03Joel Slatis: A strict three-slide executive limit, et cetera, et cetera.
- 40:06Joel Slatis: So you get the idea.
- 40:08Joel Slatis: So I'll make a choice, and then I will...
- 40:12Joel Slatis: Oh, I have to send that last one, and then I can build the prompt.
- 40:16Joel Slatis: So here is my prompt that it just built.
- 40:19Joel Slatis: So it gives me a goal.
- 40:20Joel Slatis: Now, I can edit any of these, right?
- 40:22Joel Slatis: This is an outline.
- 40:23Joel Slatis: Think of this as a scaffolding.
- 40:25Joel Slatis: So the goal is to analyze the monthly profit and loss to detect notable financial trends, anomalies, et cetera.
- 40:31Joel Slatis: You're a senior financial analyst.
- 40:33Joel Slatis: It gives you a role.
- 40:34Joel Slatis: It talks about the key points, surfaces material variances note and noteworthy financial patterns for leadership prioritizes anomalous expense surges and unexpected cost spikes and presents the output is presents a ready slide briefing.
- 40:51Joel Slatis: And let me just reduce this a tiny bit so we can see more of it.
- 40:55Joel Slatis: So this is your prompt, any of which can be edited Also, you can add new stuff, like for instance, the P&L.
- 41:04Joel Slatis: So here you would just upload file.
- 41:07Joel Slatis: You would go to your downloads directory, and you might select your profit and loss by month, and that's gonna upload the file and drop it right in here.
- 41:19Joel Slatis: So now I have my actual file uploaded along with my prompt, and now we wanna make sure that you guys are saving money and that you're working faster.
- 41:29Joel Slatis: So we built a feature to audit the prompt.
- 41:31Joel Slatis: That's what this button is down here.
- 41:33Joel Slatis: I'm gonna click it, and it's going to perform an audit on this prompt and tell us if we're missing anything important that will help us get a better result.
- 41:43Joel Slatis: So you don't have to be a promptologist.
- 41:44Joel Slatis: Knitwit helps you become one, to use the term that we used.
- 41:50Joel Slatis: So let's let's see what the audit comes back with.
- 41:52Joel Slatis: Now, we designed this audit so it comes in as a checkbox list.
- 41:55Joel Slatis: So as you check stuff off by improving your prompt, you can just knock it off the list.
- 42:01Joel Slatis: Now, by the way, for those while it's working, Knitwit has been in development for nine months.
- 42:06Joel Slatis: It's not perfect.
- 42:07Joel Slatis: We're releasing it to beta.
- 42:10Joel Slatis: There are bugs.
- 42:12Joel Slatis: There are things that will go wrong.
- 42:14Joel Slatis: For instance, this audit is taking a little bit too much time.
- 42:18Joel Slatis: I'm gonna run it again because I'm not sure what's going on.
- 42:22Joel Slatis: It's not working right.
- 42:23Joel Slatis: But there are little things here and there that are not gonna work right, so hopefully you'll bear with us.
- 42:28Joel Slatis: However, because today this group is our I guess worldwide launch for Knitwit.
- 42:37Joel Slatis: You are literally attending our launch party.
- 42:39Joel Slatis: We're going into beta today.
- 42:42Joel Slatis: We we really appreciate any feedback that those of you who wanna sign up for and check out an account will give us.
- 42:50Joel Slatis: And anyone who says that they were in this meeting today gets a lifetime 30% discount on the software.
- 43:00Joel Slatis: All you have to do is say that you were in School of Bookkeeping with Joel and Dan when you, sign up or talk to anyone about it.
- 43:08Joel Slatis: I'm not sure what's going on with the audit.
- 43:09Joel Slatis: It's it's playing hard to get this morning.
- 43:12Joel Slatis: The guys are- were working on some final touches for today's meeting so that we come in here and be strong and show you guys a great product, and they don't know that the me- Oh, here we go.
- 43:21Joel Slatis: It's working now.
- 43:22Joel Slatis: Okay.
- 43:22Joel Slatis: So basically, it's gonna tell us for clarity, right?
- 43:27Joel Slatis: This is the Clarity section.
- 43:29Joel Slatis: Don't define, I didn't defi- I didn't define material variance numerically.
- 43:34Joel Slatis: In other words, are we talking about 10% months over month, less than 25, or greater than$25,000 in absolute dollars?
- 43:42Joel Slatis: It would give me- it would give the AI a better result if I defined what I meant by material variance, right?
- 43:51Joel Slatis: That's an example of how I can make my prompt better.
- 43:53Joel Slatis: So let's say I decide to do that.
- 43:55Joel Slatis: I find where it says material variance in here, and I- or I can just add something.
- 44:01Joel Slatis: A material variance is more than$10,000, 10K.
- 44:11Joel Slatis: So I can just add that to the stage.
- 44:12Joel Slatis: Now that's part of my instructions, and I can knock that off my list as having been done.
- 44:18Joel Slatis: And it also, so you can read through this list, not everything in here you're gonna wanna do, but it- but these are just a lot of good ideas about how to make your prompt better so that when you actually submit it to the AI, you get more for your money.
- 44:30Joel Slatis: So let's now submit to the AI.
- 44:33Joel Slatis: So I'm submitting.
- 44:33Joel Slatis: All I have to do is click Submit to Stage.
- 44:36Joel Slatis: It adds my prompt here, and I can view it.
- 44:39Joel Slatis: So here's my prompt.
- 44:40Joel Slatis: It'll come through collapsed, and it's gonna generate a title.
- 44:43Joel Slatis: So later I can see what it is.
- 44:44Joel Slatis: And it's generating a response on GPT Mini.
- 44:48Joel Slatis: And you see here, I can switch back and forth between models.
- 44:53Joel Slatis: I don't know if you got this far, Dan.
- 44:55Dan DeLong: Yeah.
- 44:55Joel Slatis: Did you get this far?
- 44:56Dan DeLong: Yeah, I was just about to talk about the d- the different models.
- 45:01Dan DeLong: Somebody did ask a, a question do you need to have accounts with all of these if you're using Knitwit?
- 45:11Joel Slatis: No, actually, we give you access to all that.
- 45:13Joel Slatis: So if you're paying for, ChatGPT, and if you're paying for Claude you wouldn't need them.
- 45:20Joel Slatis: You're getting that you're getting access to those through here.
- 45:25Joel Slatis: So default for t- for the, for any new, for any account basically is GPT Mini, Claude Haiku, and Gemini Flash, and these are all set to default to the latest version.
- 45:37Joel Slatis: So when GPT Mini updates to from 3.1 to 3.2 or something like that behind the scenes you can be assured that these are automatically updated because these are using the latest versions through our, integrator who we go through, and they guarantee the latest versions.
- 45:55Joel Slatis: That's how it's set up.
- 45:56Joel Slatis: So now, if I switch over to the smarter models, which most of the time you don't need, but if you're really doing something deeply deep thinking, important, et cetera then you switch over to the smarter models, and that gives you the latest GPT, which is 5.6.
- 46:12Joel Slatis: It just came out.
- 46:13Joel Slatis: Claude Opus and Gemini Pro.
- 46:16Joel Slatis: And and here, if I go back to GPT, you can see the response.
- 46:22Joel Slatis: Okay?
- 46:23Joel Slatis: Now, here's something that you can't do on GPT.
- 46:27Joel Slatis: If I-- let's suppose I'm looking at my results, and it came from this file, right?
- 46:33Joel Slatis: It looked, it r- read through my profit and loss.
- 46:36Joel Slatis: It gave me some interesting information about it.
- 46:39Joel Slatis: I can edit this response And I can delete the parts of the response I don't care about.
- 46:46Joel Slatis: Maybe I don't care about slide five, so I'm just gonna go ahead and highlight all of slide five, delete.
- 46:53Joel Slatis: It's gone.
- 46:54Joel Slatis: Now it's no longer part of the response.
- 46:55Joel Slatis: I might need to do something silly like renu- renumber this one to slide five and this one to slide six, and you guys know how this stuff goes.
- 47:03Joel Slatis: But the thing is, I can also go up here to the top, and I can leave a note.
- 47:10Joel Slatis: Here's here's the report with my comments, right?
- 47:17Joel Slatis: So you can curate.
- 47:19Joel Slatis: The point is, that you can curate the response.
- 47:21Joel Slatis: Because you're saving it, it has a place inside of Knitwit where it lives.
- 47:27Joel Slatis: And so every month you can have this report, it gets saved, it has your comments, and it's been curated down to get rid of whatever GPT put in there that you didn't wanna necessarily see.
- 47:38Joel Slatis: So I'll save my response- Miriam,
- 47:40Dan DeLong: Miriam is asking can you use your own subscriptions?
- 47:44Dan DeLong: If you already have a subscription to Claude for their example how do you use your own token usage?
- 47:54Joel Slatis: You would go to y- you would go to account settings, and under models, these are the smarter models.
- 48:04Joel Slatis: You can see they're all GPT latest, Opus, Claude, like I was saying.
- 48:07Joel Slatis: These are all the faster models.
- 48:09Joel Slatis: These are the default models.
- 48:11Joel Slatis: If you wanted to add your own GPT model, you could.
- 48:15Joel Slatis: So basically you just click add new you select your provider from the list you add your API key, and then you can select a specific model that you want.
- 48:27Joel Slatis: Right now this is set to I forget what this is set to.
- 48:31Joel Slatis: And then you can give it a nickname.
- 48:33Joel Slatis: So you might say,"My model," right?
- 48:36Joel Slatis: So currently, I actually have mine for testing,'cause I have a GPT account, like probably most of us do.
- 48:43Joel Slatis: And and I've got that in here, and I can add it, and now it shows up in my list as GPT 6L.
- 48:50Joel Slatis: And I can change this label, because this is the newest one, so I put in 6L to remind myself that we're on six, right?
- 48:57Joel Slatis: But I can change that label.
- 48:59Joel Slatis: So when you put that in, it shows up in the composer here.
- 49:07Joel Slatis: And I didn't use it.
- 49:08Joel Slatis: I can actually query it.
- 49:10Joel Slatis: I use GPT Mini to query, but I can switch over to my model and I can turn these off.
- 49:15Joel Slatis: I don't even need to see them if you don't wanna use them.
- 49:17Joel Slatis: So the answer is all you have to do is put in your API key and you, can configure Groq or your own GPT model or whatever you want.
- 49:27Joel Slatis: If you were using GPT and you wanted to keep your own license, that's fine.
- 49:34Joel Slatis: There's really nothing wrong with it at all if you're a heavy user.
- 49:38Joel Slatis: Now there are some cool things that you can do to save costs also.
- 49:41Joel Slatis: You can turn off web search.
- 49:44Joel Slatis: You only need to use that when you're looking for a P&L analysis, you don't need web search.
- 49:49Joel Slatis: So having it on is just something you can just leave it off like I do.
- 49:55Joel Slatis: It's on when you're using GPT and there is no toggle for it.
- 49:58Joel Slatis: You're just paying for it.
- 50:00Joel Slatis: Compare is a cool feature.
- 50:02Joel Slatis: I don't know if you covered this, Dan, but compare- No, not yet.
- 50:05Joel Slatis: If I-
- 50:05Dan DeLong: Yeah, go ahead
- 50:06Joel Slatis: if I submit my stage It's gonna submit to every single model at once, all at the same time.
- 50:15Joel Slatis: And it's gonna recalculate all these.
- 50:19Joel Slatis: That's what the compare feature is.
- 50:21Joel Slatis: It's a useful feature for when you wanna say, see what three or four models all say at the same time.
- 50:27Joel Slatis: And that leads us to one last really cool and interesting feature, which I'll show you as soon as some of these other models come in.
- 50:35Joel Slatis: Gemini Pro came in first.
- 50:37Joel Slatis: It shows me that it took- it cost me one credit to actually do this search.
- 50:45Joel Slatis: These are still working because I, did it using the smarter models.
- 50:48Joel Slatis: I should have actually said- Set these to faster.
- 50:51Joel Slatis: If you use the smarter models, it's gonna go out and really think about these things- Yeah and really deep think.
- 50:58Joel Slatis: Now, once these are finished thinking and they come in, I may be able to use Knit Insights.
- 51:03Joel Slatis: Now, Knit Insights compares all the models, and it shows you what all those models, where they agree, where they disagree, and unique contributions.
- 51:17Joel Slatis: In other words, ChatGPT said this but none of the other ones said it.
- 51:23Joel Slatis: Opus said this, but none of the other ones said it.
- 51:26Joel Slatis: So while this is working, let me switch to, I think, just a different group where I may have a Knit
- 51:33Dan DeLong: Insight- and since we're coming up on the, top of the hour I know people sometimes need to bolt out because they've got a- appointments or meetings.
- 51:42Dan DeLong: I just wanted to touch base and let people know I'm putting back in the chat all of the links for today.
- 51:50Dan DeLong: Knitwit's website.
- 51:52Dan DeLong: Joel is going to come on the school bookkeeping workshop next Wednesday, where we're gonna dive in a little bit more in detail.
- 52:01Dan DeLong: And we are going- Joel and I are going to be doing a two-week- Sure Cohort where we will have four different sessions where we will, you know, r- basically giving you a, a guided tour of everything that, that Knitwit has and breaking it down into a, a little bit more- consumable pill to swallow.
- 52:29Dan DeLong: The cohort is$147, but if you have a Knitwit account, it's free.
- 52:34Dan DeLong: So you can pre-register for the cohort.
- 52:39Dan DeLong: If you like what you're seeing here please join us for the workshop if, we'll, redo, basically redo today next Wednesday with Joel.
- 52:51Dan DeLong: And then if you wanna join the, the cohort at the end of September that would be wonderful to have you.
- 52:57Dan DeLong: So we will probably run over a f- a few minutes here.
- 53:02Dan DeLong: If you can stick around, great.
- 53:03Dan DeLong: If not that, that's totally fine, too.
- 53:06Dan DeLong: Hopefully you join us next Wednesday on, on the workshop so that we can take a, a deeper tour of the, of Knitwit, hopefully when we're both on the same time schedule.
- 53:19Joel Slatis: Yeah, on time.
- 53:20Joel Slatis: Yeah.
- 53:21Joel Slatis: And, I'm so sorry, Dan.
- 53:23Joel Slatis: I feel awful about this.
- 53:25Dan DeLong: My
- 53:25Joel Slatis: giant event or
- 53:26Dan DeLong: large event I had a part in that with rescheduling the whole thing, and and it was a little hectic last week with death in the family and everything.
- 53:35Dan DeLong: So I totally understand.
- 53:38Dan DeLong: And everybody here on, on today's webinar had been so gracious and and understanding, so I really appreciate the audience helping us out
- 53:47Joel Slatis: with, understanding.
- 53:47Joel Slatis: Yeah, and I was really hoping to have a much more useful substantive conversation over how to actually write better prompts, but we can do some of that stuff next week.
- 53:56Joel Slatis: Yeah.
- 53:56Joel Slatis: The good news is, with the prompt builder and the audit feature the, the group can y- write better prompts without actually knowing the inside nitty-gritty about how LLMs work.
- 54:12Joel Slatis: It gives you already- Yeah As you can see on the screen here, a scaffolding to work with.
- 54:19Joel Slatis: And that saves you time, and it gives you that's the other big thing about Knitwit, is that it saves your conversations and keeps them in a nice...
- 54:28Joel Slatis: If you have a quick chat, you can jump over to the chats feature, right?
- 54:32Joel Slatis: So here's a a, a, quick way to just do what what is the best way to cook spaghetti, right?
- 54:42Joel Slatis: You can actually, to p- to cook noodles, right?
- 54:46Joel Slatis: You can ask that to ch- to chat and create a new chat or do whatever you need to do.
- 54:52Joel Slatis: The composer is really for building you know- More complex stuff that's work-related that you wanna keep AI generate in a knowledge base, right?
- 55:05Joel Slatis: And that's what this is all about really, is to try to get our arms around that.
- 55:11Joel Slatis: Can I show you this now?
- 55:15Dan DeLong: The and, Yeah just talking about- Yeah some people are asking about the pricing and how do they get the the, discount those types of things.
- 55:23Dan DeLong: So, just talk a little bit about that
- 55:26Joel Slatis: Sign up for a free account and, yeah, sign up for a free account, and the discount, because it is write this down 30%, Joel said 30.
- 55:35Joel Slatis: We're gonna have to work with you to get you that discount.
- 55:39Joel Slatis: So we'll create a discount code for you.
- 55:42Joel Slatis: It's all the we're going through Stripe, everything's ready to go.
- 55:46Joel Slatis: I don't have the discount code ready.
- 55:48Joel Slatis: That's how new it is.
- 55:50Joel Slatis: But I wanted to give you guys something that was gonna be even a little bit better than the nonprofit 20% discount that we're gonna offer clients.
- 55:59Joel Slatis: 30%.
- 56:00Joel Slatis: And the reason I'm doing it is for two reasons.
- 56:02Joel Slatis: One is I'm really excited.
- 56:04Joel Slatis: This is our launch event as uncoordinated as I have been for it.
- 56:11Joel Slatis: But also I want all of you to realize that hopefully that I'd really love to have your feedback.
- 56:18Joel Slatis: So go, you can sign in with Google.
- 56:20Joel Slatis: You really don't need any information to get an account.
- 56:23Joel Slatis: You can start using it right away.
- 56:25Joel Slatis: Oh, and let me show you really quickly since this came in now, this is what a knit looks like.
- 56:31Joel Slatis: So it tells you agreements between all the engines.
- 56:35Joel Slatis: So I have an answer here for GPT.
- 56:37Joel Slatis: Did I get one from-- No, I have an answer for GPT, I have an answer for Opus, and I have an answer for Gemini Pro.
- 56:44Joel Slatis: All of them took a look at the same P&L.
- 56:46Joel Slatis: The knit is now the comparison.
- 56:49Joel Slatis: So now you wouldn't necessarily need to do this for P&L, but there are a lot of things where you're, for instance, have a question about no tax on overtime's new law, where you wanna double-check your work, and that would be a great place to do a knit.
- 57:02Joel Slatis: So you can see it agrees on all this stuff.
- 57:05Joel Slatis: Here you have unique contributions.
- 57:08Joel Slatis: So GPT said reconcile first framing with bounded in illustrative scenario.
- 57:13Joel Slatis: So GPT presents the information this way, and Claude presented this information, and Gemini gave these kind of th- feedback that, and these are unique contributions.
- 57:25Joel Slatis: So Gemini gave it, but Claude and GPT did not.
- 57:29Joel Slatis: Does that make sense?
- 57:30Joel Slatis: And then it would show you any conflicts finally.
- 57:33Joel Slatis: Whether the P&L puts all the central factual disagreement.
- 57:38Joel Slatis: GPT and Claude both independently sum up sum the displayed expense lines to fifty-six thousand one hundred and fifty-one against the reported total of that blah, blah, blah.
- 57:47Joel Slatis: It just goes through and it tells you where they disagree.
- 57:49Joel Slatis: You can read it later or you can do your own, really.
- 57:52Joel Slatis: Yeah.
- 57:52Joel Slatis: It's fun to try when you're doing something useful.
- 57:55Joel Slatis: But that's what the knit-- That's what the Knit does.
- 57:59Joel Slatis: So we tried to build special features like the prompt builder, the audit, and the Knit insights tab to, really make this work for you without a whole lot of extra work, and that's what we've been trying to do.
- 58:13Dan DeLong: Somebody asked in the chat, is it available anywhere, right?
- 58:16Dan DeLong: It's, it, you don't, it's not regionalized to Canada versus the US or is there, a location-based challenge?
- 58:26Joel Slatis: For the moment, the United States and Canada are the places that we're looking for business.
- 58:30Joel Slatis: Our terms of use say that you really shouldn't use it if you're overseas because there are some GDPR requirements that we're not ready to say that we can support because there's just so many.
- 58:44Joel Slatis: We're working on a SOC 2 certificate.
- 58:46Joel Slatis: We're working on GDPR.
- 58:47Joel Slatis: There's quite a bit of work that we're doing to try to get ready for all of this as a startup, a nine-month-old startup that's only going into beta literally today.
- 58:58Joel Slatis: But if
- 59:00Dan DeLong: you do- But I, superseded your sharing find it from Europe, I'm
- 59:02Joel Slatis: not going to kick you out.
- 59:05Dan DeLong: I superseded your sharing, so people were asking how much is this?
- 59:09Dan DeLong: So I have the slides back up here about the pricing.
- 59:14Dan DeLong: This when we talked about it you were gonna offer a 10%, but now you're offering 30%.
- 59:19Dan DeLong: So-
- 59:20Joel Slatis: I, really wanna get feedback from people.
- 59:22Joel Slatis: I was gonna offer, I couldn't remember the number, so I- So since I was sitting here, I was gonna discuss it with you before.
- 59:28Joel Slatis: But since I was sitting here, I was like, you know what?
- 59:30Joel Slatis: This is our launch event.
- 59:31Joel Slatis: Let's get people on.
- 59:32Joel Slatis: Yeah.
- 59:33Joel Slatis: Let's get people using it.
- 59:34Joel Slatis: Let's get some feedback.
- 59:35Joel Slatis: Let's make it nice for them.
- 59:37Joel Slatis: So I, I grabbed a big number on, in- Yeah intentionally.
- 59:41Joel Slatis: So, so it's gonna be te- 30 instead of 10, everyone, and that's just me saying it right now, so that's what you'll-
- 59:47Dan DeLong: Yeah, so we'll provide we'll provide Joel with the list of attendees that we had here today, and he'll be able to cross-reference that with anybody who registered.
- 59:57Dan DeLong: So please use the same email that you registered with, the Power Hour so that we so that we, can do that cross-referencing pretty easily, or just just make sure that you get a message to Joel to say,"Hey, I saw you on the, Power Hour.
- 1:00:18Dan DeLong: Want a discount here." But you have different tiers that, that come with monthly credits.
- 1:00:24Dan DeLong: Of course, the question is what if you run out of credits?
- 1:00:27Dan DeLong: Can you top'em, top off?
- 1:00:29Dan DeLong: How does that work?
- 1:00:31Joel Slatis: Yeah.
- 1:00:31Joel Slatis: You could just buy more credits at, face value that you just go through the shopping cart.
- 1:00:37Joel Slatis: You can buy$10.
- 1:00:38Joel Slatis: I think we're gonna put a minimum of$10 on it.
- 1:00:43Joel Slatis: But that's not I don't think anyone's gonna really balk at that if you're using so many credits that you run through your whole allotment.
- 1:00:50Joel Slatis: And by the way Knitwit gives you the quick and easy ability to change over from the smarter to the faster models.
- 1:00:57Joel Slatis: And mo- for most of you the smart the smarter models are not necessary for the most of the things you guys are gonna be doing.
- 1:01:06Joel Slatis: And so you can save a lot of money, a lot of money by just using the faster models, and it's just a toggle.
- 1:01:13Joel Slatis: So I would recommend for, this group that's to just use the, faster models most of the time.
- 1:01:21Joel Slatis: And- Got it Now for those of you who want, like I said, we're gonna do a whole thing.
- 1:01:25Joel Slatis: We got use cases.
- 1:01:26Joel Slatis: Some of the use cases are really amazing, the things you can do with this.
- 1:01:30Joel Slatis: I don't get to have the fun that I was looking forward to today.
- 1:01:34Joel Slatis: So definitely come to the Get a free account right now.
- 1:01:37Joel Slatis: Go to Knitwit.ai.
- 1:01:39Joel Slatis: Just log in with Google.
- 1:01:40Joel Slatis: There's no cost to log in and play with it.
- 1:01:42Joel Slatis: Come next week, it's free.
- 1:01:44Joel Slatis: And then we'll be able to show you some of the more interesting use cases.
- 1:01:47Joel Slatis: Yeah.
- 1:01:48Dan DeLong: Yeah.
- 1:01:49Dan DeLong: And then we're gonna be doing a, a four session, hour long each session, Joel and I, where we'll break down just navigation in Knitwit, stitching, use and various use cases through, through the cohort.
- 1:02:05Dan DeLong: So as long as you have a, an active N- Knitwit account- It's free.
- 1:02:10Dan DeLong: So you can register for that as well so yeah, so just wrapping this up here.
- 1:02:16Dan DeLong: We're a little bit over time.
- 1:02:17Dan DeLong: We appreciate all the grace that you've extended to us with the us being Knitwits literally with the timing of everything.
- 1:02:31Dan DeLong: But yeah, the...
- 1:02:32Dan DeLong: What this does is it allows you to work smarter, not necessarily harder to get the to fine-tune your prompts and your outputs.
- 1:02:44Dan DeLong: Either you're taking those outputs and sharing them, with your clients, or you're taking those outputs, sharing them with your team, taking those outputs, putting them in your own model to generate something, because that's something that Knitwit doesn't do, is it doesn't necessarily generate a video or a picture or things like that.
- 1:03:02Dan DeLong: But this will help give you a fine-tuned, crafted pr- prompt to give to those models to get the output that you're looking for so that you s- spend wisely your token usage.
- 1:03:16Dan DeLong: So next week...
- 1:03:18Dan DeLong: I did launch the last poll question already.
- 1:03:21Dan DeLong: So next time, it's actually gonna be next week we are gonna be doing a, an app showdown which is the AI co-pilots that that are out there.
- 1:03:32Dan DeLong: So Double has, extended into the, the AI cleanup and the AI work inside of inside of QuickBooks Online and other GLs.
- 1:03:45Dan DeLong: And a new player in the market is Accru.
- 1:03:48Dan DeLong: So we are having them to come on, and we are gonna be comparing those two next week, not next time.
- 1:03:57Dan DeLong: So I- what I will do after today is I will change the the frequency and the cadence back to its normal every other week.
- 1:04:07Dan DeLong: So that will get everyone back on the same cadence because I did get a lot of concerned emails saying,"These are the wrong dates." So that will take care of itself, and that was my fault because I had to reschedule last week to this week and and that, that changed all sorts of things because Zoom doesn't allow you to choose specific dates and in your webinar series and, and allow you to keep them that way.
- 1:04:35Dan DeLong: So so thank you for joining us, and thank you for, putting up with our shenanigans with with trying to figure out the, what time things actually started and all those things.
- 1:04:48Dan DeLong: And Joel, we will see you next week at School of Bookkeeping on the, workshop.
- 1:04:54Dan DeLong: And hope everybody has a great week ahead, and we'll see you next time on the Power