Chris Bell - Creating a Junior Accountant using AI (ChatGPT)

Chris Bell - Creating a Junior Accountant using AI (ChatGPT)

Episode - 015: - Where we learn from Chris, how Luka was born out of frustrations with the accuracy of finance data, how to automate workflows and searches, what types of questions you can ask the finance bot, how to prolong the life of your existing systems, the tools Chris uses to run his business and much much more.

After studying Aerospace Engineering, Chris first worked in Audit before moving into Data Analytics for KPMG. He then moved into finance where he eventually became Director of Finance & Operations for VenueScanner. Since then he’s become a CFO advisory to various companies and is now building AI into mainstream accounting platforms like Xero with his company Luka.

In his spare time Chris likes to run, go on hiking holidays and build tech products!

Audio Podcast Links

Show Notes

Where to find Chris & LukaHQ:

  • Try it for FREE

Example Questions that you can ask Luka:

Use Luka to surface Accounts Payable records:

  • What bills need paying currently?
  • Who do I owe the most money to?
  • Give me a list of all my creditors?

Use Luka to draft customer statements:

  • Draft customer statement for bayside club reminding them to pay invoices and show total amount owed
  • Do Bayside club have any unpaid invoices?
  • What invoices have been updated in last 7 days

Tech Mentioned

[Note] Tech for finance are a Notion affiliate, but we also use and pay for their services

People Mentioned

Other Mentions


  • [00:04:37] Accounting software accuracy can be frustrating
  • [00:08:50] Automating scheduling, email monitoring, and approval workflows.
  • [00:11:21] A future finance admin assistant integrated with Slack and Xero.
  • [00:16:38] Re-engineering ChatGPT with accounting questions
  • [00:21:00] Can ChatGPT prolong the life of existing finance software?
  • [00:25:58] What questions can you ask the bot?
  • [00:30:37] What tools Chris uses to run his business
  • [00:38:04] Why you need to be using Reddit
  • [00:41:06] Why we’ve not yet seen the paradigm shift with AI models
  • [00:42:45] The future of prompt engineering to make AI workflow seamless.
  • [00:48:13] The importance of AI training & learning + future advancements.



[00:00:00] Chris: So I guess more beyond just the accuracy jobs daily, there's a more general sort of job scheduler. So you could say, Hey Luka, can you every day post this journal?

So if a junior accountant asks the wrong question, Like there were themes in the questions that were being asked. You could then give a summary to the head of finance that would say, oh, Chris has asked about how to post accruals three times this week. Maybe they need a bit of training on that, or whatever.

I think the paradigm shift. I don't think it's happened yet. And I don't think pure chat will facilitate the paradigm shift into whatever the new world is where we use this AI models, religiously. Something that we really want to want to do is figure out what interface it is that makes the Open AI concept really valuable.

I don't think it's provided like game changing, relief to people's problems. I don't think it's provided a structural shift in how people operate. I think that structural shift will come with UI innovation.


[00:00:54] Adam: Hello and welcome to Tech for Finance, where we help finance professionals leverage technology to level up their lives. I'm your host, Adam Shilton, and in this episode we are chatting with Chris Bell, Co-Founder and CTO at Luka and CFO advisor to many.

After studying aerospace engineering, Chris first worked in audit before moving into Data Analytics for KPMG.

He then moved into finance where he eventually became Director of Finance & Operations for VenueScanner. Since then he’s become a CFO advisory to various companies and is now building AI into mainstream accounting platforms like Xero with his company Luka.

In his spare time Chris likes to run, go on hiking holidays and build tech products!

So thank thanks for joining me today, Chris. Good to have you.

Excellent stuff. So over to you. Then we'll start, you know, I know that Lukas's still kind of in build at the moment. It's all very exciting, isn't it?

It's very early stages for you guys. But do you just want to talk us through, you know, why you moved into developing tech products after being sort of a finance leader, I guess.

[00:02:07] Chris: Yeah, of course. So I guess, ever since my university degree, I did a few programming modules and after leaving university, I was always trying to figure out how to combine my technical skillset with these finance skills that I was picking up. I didn't really get time to work on the technical side of things until Covid hit. We all sat at home, which is probably a bit cliche if you talk to a bunch of founders around now.

But I sort of took it as an opportunity to learn Python, figured out how to build a stock analysis dashboard as a brief month where I was like, I'm pretty sure I'm onto the next million pound idea that no hedge fund has found. That then evolved into building my own accounting software. So I was like, Ooh, how do you debit and credit? Then I built my own general ledger. That over the last years has just kept on building and everything over the last few months has kind of come together with, I met my co-founder chat GPTs being released, and there's a few things that are combining that had made me take the leap to go full-time on the building.

[00:03:35] Adam: Very good. So, how long have you been building for?

[00:03:40] Chris: Well, so on and off for two years. Various different products. The most recent version, on top of ChatGPT since the start of the year.

So we've been going for six weeks or so now. Okay.

[00:03:54] Adam: And, I love this stuff and we could talk forever, right? But this is relevant to the audience, which is finance professionals who are curious to know a little bit more about some of these emerging technologies. We'll get to talking about how your solution works and why people should be building something like ChatGPT into the solutions that they're already using.

You mentioned there that you've seen it from the other side, SME finance and accounting, long months ends, lack of automation, that sort of stuff. Was that where this was conceived then? Did you take a step back and think, "Right, well, what problems am I looking to solve?"

So, do you want to talk a little bit more about that?

[00:04:37] Chris: Yeah, everything's born out of my own experience and I think the roots of my frustration and my thesis with how I'm building is that a lot of accounting products don't think about accuracy. They forget that the purpose of accounting is to record information accurately and they focus very much on the operational interaction with the rest of the business.

My core frustration really all started with month end. Every month I seem to be correcting the same transactions where people have miscoded or whatever. It baffled me why there wasn't something that just analyzed your accounts, checks what last month was, checks what this month was, and fixes everything based on that. That's where it's all started snowballing from.

[00:05:38] Adam: Just scratching into that a little bit further then. You said on the coding side of things, "Why am I still correcting the things that I've corrected a million times before?" Is that what you'd expect, more traditional checking like that, that credits and debits are in the right places and coded to the right jails, or is it a bit more advanced than that?

Especially with some of these more modern solutions, it's not just credits and debits, but it's the concept of dimensions, cost centers, that sort of stuff as well. So is it just debits and credits corrections, or are we looking at some of the non-financial data that overlays on that, I guess?

[00:06:17] Chris: I think ultimately the answer is everything, but I think you always just need to start with the simplest problems. And I think there is, certainly in smaller businesses, enough of a problem just looking at account code coding for invoices. Realistically, once that's solved, adding in dimensions and segments and all that other stuff wouldn't be complicated on top of that.

[00:06:54] Adam: So alongside month end corrections, automating that piece. What we're talking about there, are we talking about a validation that on point, something says that doesn't look right while you're doing that? Or is it retrospective? Are we getting to a period activity and then it's drawing our attention to after the fact data?

[00:07:18] Chris: So again, long term, the ideal would be point of entry, running checks and the like. The ideal would be to integrate with the auditors and get the auditors plugged in. And so every time someone enters a transaction, you fire something off to the auditors say, Hey, this has happened. Is it okay? They would say yes. And then when you get to the end of the year, everything's just checked and you move on. Because systems are the way they are at the moment, and it's difficult to integrate with a few of the existing accounting providers, it would need to be more on the side of running a job every day that checks all the transactions posted that day compared to yesterday, and go from there.

I guess the accuracy concept is where foundations come from, but I think from playing around with Open AI and ChatGPT, we feel that there's probably a lot more search functionality that we can build that is probably gonna be a much simpler problem to solve, but probably has some equal or more value in the short term anyway.

And again, long term, ideally you would solve everything and AI will solve everything in certainly the finance operations space. But I think search is probably what we'll start with.

[00:08:50] Adam: Okay. So we've got accuracy and running jobs to make sure that there's not any sort of outliers or anything going wrong. We've got querying. Are there any other areas of automation? I'm working with quite a lot of customers on stuff like AP automation and that sort of stuff as well. What is the remit of this? Are we quite narrow or is the potential not just global, but other areas of finance as well?

[00:09:30] Chris: Yeah, so beyond just the accuracy jobs daily, there's a more general job scheduler. So you could say, Hey Luka, can you every day post this journal? Can you message that person? Can you get approval for this? I think the scheduling stuff and interaction with email would probably come pretty quickly. Once we get the core functionality up and running, monitoring the accounts inbox and sending email chasers or just responding to very basic things.

Also, I think something quite powerful is approval workflows. If a salesperson wants an invoice changed, they could Slack Luka. Luka would then know who they need to get approval from. They would say Slack the head of finance, and say, Hey, so and so wants to change the invoice from a hundred to 200. Is that okay? Head of finance responds yes, and then the action is taken in Xero, and I think that approval flow is quite powerful. It's something that doesn't happen at the moment, but that creates quite a lot of stress, certainly anecdotally from the people that I know.

[00:10:52] Adam: So we're almost saying there that if you do have these technologies built into stuff like Slack and teams and those collaborative workspaces, you've almost got a finance admin assistant. That's kind of not a person that you're paying for. And again, we're encroaching in scary territory of replacing jobs, but we're not replacing, we're repurposing. That's what we always say, but that's definitely what it sounds like.

[00:11:21] Chris: Yeah. And that is a hundred percent what our vision is. Having someone that sort of in Slack who you send messages to and they will, Luke would then be integrated with Xero. Your email. It can write its own copy. It can, that's absolutely spot on what we're planning to do.

[00:11:43] Adam: So we'll get into the ChatGPT in the Open AI bit in a second. But I'm just curious, and this wasn't in the questions that I asked you before, but I've got a decent amount of experience, quite a lot, in finance and accounting software. And there have been efforts to build this sort of stuff into those systems previously. So, when you look at the SAP systems, they did have almost like a query bot that you could ask questions.

I can't remember what the technology was called, but it was pretty poor, right? The results that it surfaced were pretty awful. When we look at Power BI, you do have the ability to query and make that process easier.

I remember in a previous company as well, we also built a bot into teams that looked at billings for the resource that we delivered on jobs. But we got a bot to say, tell me the availability of this consultant and it had come back with their schedule and their bookings and the client that they were working with, and that was probably the most useful example of that sort of technology from a collaborative and not having to send a load of emails, for example.

What are you seeing is the fundamental difference between these new technologies that are intelligent and those previously existing technologies that were okay but kind of only good for really specific program purposes.

[00:13:33] Chris: Yeah. So I guess the relentless progressive technology really. And so since getting involved in this space, joining a few discord chat. it's just nuts. The how you like go to sleep. And the next day there's some new paper that's been written about someone testing this thing and someone testing that thing. And I guess the open source nature of OpenAI as well as a few other libraries that we're actually building on make all of these things very accessible.

And I think previously there was a tendency to build. So for SAP for example, they would've built that very much in-house, not connected to anything else trying. It's all proprietary. The openness of being able to plug into Xero, plug into Slack, plug into Open AI and just have something in the middle, like the orchestrator is something I don't think has been the environment hasn't been right for that up to date.

I would also add that the query language for Power BI, I wouldn't be surprised if that is able to do exactly what we're doing, particularly cuz Microsoft is buying Open AI. Kind of buying Open AI, and actually Microsoft, is it Microsoft Dynamics that will have a lot of this functionality in the not too distant future or they'll try and invent it.

[00:15:25] Adam: Yeah. And it's about making it fair game. Microsoft evidently wants to rule the world. But that doesn't help the people that aren't using Microsoft applications, right. So I think that's fair. Something that was quite interesting from back of my Dynamics days, and I suppose it was an element of ai, but in dynamics from a financial point of view, if you got an email from a customer, for example, saying, can I have an update on this invoice? And they gave an invoice number.

It turned the content of that email into a hyperlink so you could go straight to it in dynamics. That was probably the most intelligent I'd ever seen a system to get in terms of integrating with Outlook and Microsoft. But again, you had to be using Outlook and you had to be using Dynamics to fulfill that.

So to paraphrase you, I guess we're saying. We've moved from a territory whereby it's simple bots that are just surfacing data to something that's a lot more conversational, that understands a lot more from your written words. So you don't have to be as precise about your questions.

[00:16:38] Chris: Yeah. Understand. It's from playing around with OpenAI and there's this concept of an agent. Chat GPT is built on a model that essentially just recursively calls. So chat GPT recursively calls the base model over and over again to simulate a conversation. It's not just the model answering you directly and remembering your conversation. We implemented that concept with accounting questions. And it's mind blowing how good it is at picking what task to do.

So you say to it, Hey, you're an assistant with access to these five tools. One is search, one is edit, one is something else. And then you can ask, oh, can you find the legal entity for this? And then update all the invoices. And it will know to search Google, then get all the invoice numbers, then update each one individually.

And I think I wrote that program in probably like a hundred lines of code. There's perhaps getting a bit too technical, but I think previously in order to simulate that same functionality, you would've needed hundreds of lines of code managing each sort of way that someone would answer and everything like that. We are giving total control over to the AI and it's phenomenal the power that it has.

[00:18:09] Adam: So, so to the people that are a bit scared about that, right, because coming back to the accuracy piece and the following process and ensuring that only the stuff that should get changed gets changed, what, what does it look like from a, from a permissions perspective?

I'm assuming that we are getting to a point where we can give different seniority to change. So, so for example, you, you wouldn't wanna give a junior finance admin autonomy to, you know, ask ChatGPT to update supply bank details, for example?

[00:18:42] Chris: Yeah. Yeah, a hundred percent. And that's something that's ever present in my thinking when I'm building this stuff.

And version one, we've built a Slack bot sort of in the first three weeks of the year. And yeah, every time that we edited, we would not do it unilaterally. So someone would ask us to edit and we would hard code that you had to get approval for it. And we would only then take an action once we got explicit approval to take that action.

But, and I think the way that that is managed will constantly evolve over the next few years. But I think that'll be a common theme with AI. I think AI is actually surprising. It is much easier to implement at a higher level where the level of accuracy required is much lower. So it's quite easy to say, oh, make a decision about which tool to use. Cause if it picks the wrong tool, it doesn't really matter. But when you get to the stage of, oh, edit this, or do some fine admin thing, it's actually you need to go back to old school programming where you can explicitly define exactly what's gonna happen so that you can trust that it'll do what you need it to do.

[00:20:00] Adam: Well, going back to your point there about approvals, I guess it all comes down to the setup, but if you did have somebody junior that asks the bot a question or Luka a question, if that doesn't tally with Luka's permissions, maybe it could automatically, you know, ping a message to somebody more senior to say, somebody's asked me this. Are you okay for me to do it?

[00:20:23] Chris: Yeah, a hundred percent. And yeah, that's the communication, it's more than just approvals, right? It's just communication generally. And again, pie in the sky, it'd be great to sort of build in like some sort of training thing into it. So, if a junior accountant asks the wrong question, like there were themes in the questions that were being asked. You could then give a summary to the head of finance that would say, oh, Chris has asked about how to post accruals three times this week. Maybe they need a bit of training on that, or whatever.

[00:21:00] Adam: So, thinking out loud then, because what we often find, well, what I find in my world is that businesses grow, processes evolve, often get more complicated. Basic accounting and finance systems aren't fit for purpose anymore, so they have to go through this whole transformation piece to move on to the next level. But what I'm hearing here is that maybe the sorts of applications that you are building will increase the longevity of existing solutions. So, because they're able to build in automation to a platform that gives them access to something that's more scalable without having to completely migrate to a different system.

[00:21:45] Chris: Yeah. Yeah, a hundred percent. Going back a couple of years, when I was trying to build an accounting software, I was very conscious of the fact that there was this huge gap between QuickBooks Xero and then the NetSuite, Microsoft Dynamics, et cetera. In concept, an accounting software is really simple. It's just a list of transactions with different ways to interact with those transactions.

It was another thing that really frustrated me, why we need these bigger platforms. Very often, the only reason that you need the platform is for the other stuff, is for the ERP. Michael, my co-founder, always has this thing that NetSuite has more non-finance users than finance users. Because all they're really trying to do is sell you the ERP. They don't, the accounting software is just a way in. And so, yeah, absolutely building the tooling to connect an accounting software tool to all the other systems around. We should, in no reason that the ERP won't be obsolete soon.

[00:23:00] Adam: Yeah. I hope not. I do quite a lot with ERP. Yeah. Yeah. I mean that's, yeah, there are some limitations though, obviously, you know, there is a certain transaction limit in Xero, right? So, little bit, but by that time you'll probably get to the point where you are developing with larger ERP systems anyway. Right. I'm sure that's in your future timeline.

[00:23:22] Chris: Yeah. Yeah. Well, yeah. That, that, and maybe building our own general ledger at some point.

[00:23:26] Adam: Oh, right. Okay, fine. Well, maybe we'll have a round two at some point then. Yeah. But no, I, is it, yeah. And again, we could talk about this forever in a day because I, I, I do use ChatGPT quite a lot, not just to produce generic, boring AI generated content, but as a source of inspiration, as the foundations for certain types of content. You know, so I get it to write to a framework, but as soon as you start getting into prompt engineering, which I suppose is the buzzword at the moment of how do you ask better questions.

That's thing, that's where things become infinitely more interesting. So we are almost past capability we're into quality of the questions. So just from what you were saying there previously about asking it a question and getting it to go off and do something. I mean, that coming back to the point about systems, some systems do have limitations, especially if you've got a multi-company, multi-entity environment.

Yeah. Some systems do it, some systems don't. But for the systems that don't allow you to say, create a supplier in multiple companies in one go, this is where being able just to ask, look, hey, create me the supplier in all of my companies. You see what I mean? And it's that sort of stuff that's gonna really save. You know that sort of, because otherwise you're into building complex RPA procedures and all of that sort of stuff. And you've gotta get to the point where well does the time and effort on creating something like that warrant somebody just doing it manually? Yeah. But this kind of falls the territory in between the two.

[00:24:59] Chris: And yeah, so on the prompt engineering piece, our job will be figuring out what tools we need to give the bot or what we need to make the bot aware of so that it can make those decisions itself. Where previously, in order to build that functionality out, you'd need to connect to all these things and sort of write a separate piece of Yeah, you'd now, in theory, could just write one update tool and then tell the bot. Oh, if you need to update a contact, just tell the update tool to say contact with the new field, and then it'll just recursively call all those things itself. Which I, I think once the right formula is hit on, everything will happen so quickly because it'll, yeah, the bot will just start doing things itself.

[00:25:58] Adam: So we gave the example of creating a supplier in multiple companies there. What are some other questions that you are testing at the moment from a finance perspective when it comes to either that edit or query piece?

[00:26:14] Chris: So things like what are all the overdue invoices for this contact recently? what's our profit for the year to date? Can you add a PO to this? They're sort of the types of things we're asking, so okay. The way we've built the search functionality so far is totally flexible. What we do is pass the question into a SQL query, and then we've got a proprietary data model that we've built on top of Xero for that query to go into. The sky's the limit in terms of what you can ask. That's exciting.

[00:27:05] Adam: So when we look at Luka right now then, we've said you're still in build stages. There is an element of free access, correct me if I'm wrong at this stage, the people can go and have a look essentially. Am I correct in saying that?

[00:27:22] Chris: No, actually. So I think the website that you're looking at will be out of date by the end of the week, there should be a free thing that you can go to.

[00:27:30] Adam: By the time this comes out, okay. We'll be past that. Is that gonna be the Luka or is that a different URL now?

[00:27:39] Chris: Yes, it'll be LU K A

[00:27:43] Adam: 'Cause I'm looking at, and you can tell me off for speaking about the link that I've got at the moment. But I'm just looking and I'll share a screenshot in show notes, which you can approve before it goes live. What I'm looking at the moment is separated into dashboard, accruals, prepayments, deferred income, and so on and so forth.

[00:28:05] Chris: So I think that is out of date. That was a variance analysis tool, which a lot of the back end of that I'm still using. Some of error check and we'll use, and the data model that goes into that, we're currently using. But the output there, yeah, we've moved on.

[00:28:28] Adam: Okay. So by the time this comes out, assuming that we've got that next iteration available, what are you hoping that people are gonna be able to go in and have a play with?

[00:28:41] Chris: It'll essentially be a mirror of ChatGPT, but just with access to Xero. So you would log in, connect to Xero, and then just start asking questions. Depending on when this comes out, hopefully we'll also have a Slack integration, so you can ask the same commands in Slack. As well as, I'd love to have a beta test of an email monitoring to sort of try and extract actions from emails.

[00:29:18] Adam: That would be really cool. But, but at least at a foundational level, we're basically gonna have a ChatGPT for Xero in no time at all, which is really cool. And alongside that, if you're not doing it, maybe it's a thought. You're gonna have little prompts when people log in to say, "Here's an example question that you can ask." Is there gonna be like an FAQ where people can go and sort of learn a bit more?

[00:29:46] Chris: TBC. But yeah, we definitely will. We'll try. We're currently building version two. Version one was a pure Slack bot, where now we're doing a web app and it was much more restricted. But yeah, we definitely got that learning that if you don't tell people what to ask, they ask stuff that you really don't expect. We did not answer any questions. Fine.

[00:30:12] Adam: So I'll tell you what we'll do then. When we've got the show notes for the podcast sorted, you can give me some example questions. We'll have the link to Luka in the show notes, and then below that we'll just give people some example questions. They can click the link, copy and paste the questions, and they can give it a go.

[00:30:36] Chris: Yeah, sounds good.

[00:30:37] Adam: That. Perfect. So I'm gonna go off on a little bit of a tangent now.

Moving slightly away from ChatGPT for Xero because I'm more curious about the sort of tools that you guys use in a business to work productively, I guess, at the moment. So, do you have any project management apps or anything cool that you can show?

Because I love this sort of stuff.

[00:31:06] Chris: Well we use Notion. Notion is actually our primary project management app and it's phenomenally powerful. I guess with a lot of apps, when you try to plug it into an existing ecosystem, you often don't get to experience the full power of it because the task management is done in ClickUp, and so Notion is just for this or that, and then different people use things for different things.

But when you use Notion from day one, you can create a database over here of all your contacts, and then another page over here with a summary of this, and then import the contacts into the summary page. It's mind blowing. I think that's the main, and obviously from a programming perspective, there's a whole stack of stuff I could go into if you want.

[00:32:04] Adam: No, it's fine. We can save the technical stuff maybe for a conversation over a beer or something like that. That's fine. But no definite hat to Notion there. It's really good. I actually replaced Evernote and Trello with Notion recently.

Yeah. The only thing that I haven't quite got yet, and it is only a little thing, but I use an application called Spark Mail for my emails. It's like 40 quite a year or something like that. But I've got like three or four different email addresses now, so it just merges them into one.

But the tagline is "Spark will give you email superpowers" or stuff like that. And it's true, you know, it really does help you get to inbox zero because it's just clean. It helps you categorize, it filters all of your emails into these are emails from individuals, these are updates from the likes of HubSpot, Slack, whatever it happens to be, and then these are your newsletters. If you're not using that, I'd check that out as well.

[00:33:07] Chris: Yeah, we used Front at a previous place, which sounds similar, but actually that sounds much more.

[00:33:15] Adam: Yeah, Spark Mail is really good. But the only thing I haven't got yet is the link between Spark and Notion.

So I can't push my emails to Notion from Spark yet, but I hear that Notion is working on the next iteration of the API. Somebody on the last podcast as well told me that they're building almost like a ChatGPT team to Notion, sort of like a conversational thing, which blew my mind. I've not enabled it yet, but I've just reminded myself to do it, I think.

[00:33:44] Chris: I think some of it's already live. I think you can start writing stuff and then say "extracts actions" and it'll extract all your actions.

[00:33:52] Adam: I need to turn it on. I need to turn it on. Yeah. It was yesterday's podcast with a guy called Dante Healy. Really a good guy. He's worth following on LinkedIn. So, yeah, shameless plug for Dante there.

[00:34:08] Chris: Just on Notion actually. Yeah. An interesting, and not just Notion, but the concept of a company wiki in the world of these AI models that are text-based. A really well-updated Notion.

I think we're not quite there yet, but at some point in the not too distant future, being able to just chuck your Notion into an AI model will be unbelievably valuable because it will then have the context of "Oh, this is the product I'm building, this is..." Yeah. So there's Notion, I think in the not too distant future, will be pretty massive.

[00:34:45] Adam: Yeah. It's my second brain at the moment because I went from using OneNote, being in that whole sort of Microsoft ecosystem, which was fine, but it's pretty black and white. And when it comes to the task management, the Kanban sort of stuff, it's just not there. So I very quickly ended up using lots of different applications and then I used Microsoft To-Do, which again, not particularly good if you're going down that sort of whole Eisenhower, you know, do delegate, delete sort of stuff. So I spent quite a lot of time on that before eventually arriving at Notion, but Notion was actually a recommendation from another podcast guest, Soufyan Hamid, who's the presentation guy.

I found out about it because he shared a Notion page when we were doing a New Year's update. So he shared a Notion page and basically said all of your video snippets, just drop it into this upload piece. So from a collaborative perspective, I thought that was just absolutely ace and it was so easy.

So that's what got me hooked.

[00:35:53] Chris: Yeah, yeah, yeah, yeah. should I turn my light on or like, is this a visual thing or ?

[00:35:59] Adam: It's, it's fine. I can still see you. It's okay. okay, cool. So, so the question that I tend to, to ask at the end of these podcasts, and you, you've mentioned notion there as, as one, but the question is, what Apple gadget could you not live without?

And this is both personal and professional. A lot of people say it's their smartphones, but I prefer getting a bit more detailed than that. So do you have a specific app or gadget that you like to use that's a little bit more niche than just a smartphone?

[00:36:31] Chris: I spend a lot of time on Reddit. I get a huge amount of information off Reddit. There's an OpenAI subreddit where people share stuff and I'll scroll through this stuff often. Something valuable pops up and, yeah, both professional and personal, I couldn't live.

[00:37:05] Adam: Reddit is an interesting one. I've been getting more and more into it. For people that aren't with Reddit, just to demystify what a subreddit is.

So Reddit is basically an online community and software vendors, different applications, they've all got community groups. Those groups are called subreddits, hence why you've got a ChatGPT subreddit that is a list of community members talking everything ChatGPT.

You can subscribe to these sub-threads. The thing that's quite interesting is that they're quite heavily moderated. I've been seeing them come quite high in the top of search results at the moment because it's good quality, validated data. It's not just a webpage where anybody could write anything.

It's almost self-correcting because you've got an active community that is helping each other and making sure that stuff's accurate. I think it's a really good resource.

[00:38:04] Chris: Yeah. You get some really skilled people commenting on things. There are certain questions and you'll get someone responding saying, "Oh yeah, I actually just did a PhD in this". You obviously need to approach everything with a critical lens, but all the PhDs on there can't all be fake.

[00:38:34] Adam: They're moderated as well, so somebody is constantly making sure that people are following guidelines. We talked previously about the Microsoft relationship with OpenAI and the future prediction. I think they're already doing it with the advent of Bing chat, whereby they're building ChatGPT into Bing search.

Google is quaking in their boots at the moment because it's a proper threat for them. Unless there is the proper merge between the likes of chat and these search engines, Google search is going to reduce. So people are going to be using chat instead of just using standard Google searches, for example.

We have been looking at where to spend our marketing dollars. If people are searching in different ways, do we want to be heavily invested in Google for advertising? We did start looking into Reddit as an advertising mechanism, as silly as it sounds. We had a chat with the guys at Reddit and we quickly determined that actually, from a marketing perspective to build awareness of a solution, maybe not so good.

But if we were looking for consultants or software engineers, advertising for new jobs on Reddit, they're seeing a lot of success with at the moment. So, a bit of food for thought there. Maybe hedge your bets. Don't just use the LinkedIn recruiter and the LinkedIn job posting. Use Reddit for jobs as well. I think that was a fair shout.

[00:40:16] Chris: Yeah. I love Reddit, so use it for. Fine. No worries. So, we'll post all of the stuff in the show notes.

Are you okay with people to, to sort of follow you on LinkedIn and that sort of stuff?

[00:40:35] Chris: No, I guess I should probably make a Twitter account, shouldn't I?

[00:40:43] Adam: And your LinkedIn?

I'll give the link because it's something like Chris Dash be dash a 30 a Xero B 58 or something like that, which is hard to remember. So I'll just put the link in the show notes for everybody that's interested. And then the webpage, when it's released, will be Luka, l u k a, yeah, eight Fabulous.

And then what Google has been doing is making that interface interactive. So you get the dropdown, which gives you some more information, and then you can collapse that thing and sort of research. I think something that we really want to do is figure out what interface makes the Open AI concept really valuable because so far, I don't think it's provided a structural shift in how people operate.

[00:42:45] Adam: Yeah. It's about prompt engineering, and I think people have had a bit of a look at ChatGPT, signed up for a free account, given it a bit of a go, thought, oh, that's quite cool. Let's ask it some silly questions. But then not really invested the time into learning how to do the prompt engineering and to ask better questions.

And I think to come to your point there, that's one of the big hurdles. How do we make this easier? How do we build this into people's workflow almost without them really, without somebody thinking, I'm using ChatGPT? How can we make that workflow as seamless and as, you know, in a way that interrupts them the least? You see what I mean? Without them having to do a ton of research and effort in how to use it all. So, so I think you're right to point to that piece.

[00:43:34] Chris: Yeah, and whatever the future is, prompt engineering can't be part of it.

[00:43:43] Chris: The way I ask a question cannot impact the functionality that I get back or else it's not solving a generalized problem, it's solving a problem for a very specific set of people who know how to ask that specific question.

[00:44:02] Adam: That's interesting cause. And, and I, I've had this chat recently, so in the way that you and I are talking now, because we're humans and we've had lots of conversations, we subconsciously fill in the gaps, so I know the meaning of what you are saying.

Yeah. So, and it's easier for me to respond to a question because I can fill in the gaps. It's different with a bot because you've gotta tell it what the gaps are and you've gotta be very specific with how you ask it about context, how you ask it about the end result and that sort of stuff. But I think it's gonna be a challenge because how do you build in however many years worth of human interaction to the point where a bot can infer the meaning of something.

Yeah. Without having to receive a really detailed request. And I think that's gonna be part of, dare I say, art of programming these platforms in the future.

[00:45:27] Chris: Yeah, yeah, yeah, yeah. And doing the, I don't know if you know about fine tuning where you can take like a ChatGPT model and then teach it more about your specific use case. A way to solve that problem is to provide training data on a business level. So each business has their own fine tune thing.

But again, you run into a similar issue where you need so much context and at a very nuanced level. And you're constantly switching between high, low, it. I mean, like there's very smart people working on this, so I'm sure someone may figure out in the future, but it's a difficult problem.

[00:46:10] Adam: It'll come to the training, won't it? And this is when it's gonna come to what level of personal data do you want to give it? And not just personal data, but what information from a business perspective do you want to give it? So thinking out loud, and I'm sure people are doing it already, but you know, we think about Slack, we think about these collaborative tools like notion that are a massive repository of information.

So if you gave my notion or your notion, and people that aren't using Notion, you know, your notepad or your other note or whatever your data repository is. You throw an AI model and say, this is me, this is my language, this is the way that I speak to myself. Maybe that's a good training exercise for it to be able to help do that gap full.

And then why doesn't that, when you look at, you know, we as a company, use teams Yeah. And that has got an Absolutely. And we're only a small business. It's got an astronomic amount of data in it through direct message. Yeah, through teams chats, probably all of the chat transcripts from recorded videos and all of that sort of stuff.

So is the next phase to feed that into the AI model to say, this is how we communicate as a business. These are questions that we hear time and time again. So maybe that's an iterative process that points to learning.

[00:47:34] Chris: Yeah, yeah. And I think with all these new apps that will be sprung off ChatGPT, learning is intrinsic in it. We have a learning page where we can monitor people giving thumbs up or thumbs down to questions and review what's caused that.

In the medium term, we'll automate the process of feeding that back into the model to fine tune it for our use case, which is selecting accounting tools for accounting tasks. Learning is at the core of it all.

[00:48:13] Adam: There we go. You heard it here first. Luka HQ.

Wow. Fingers crossed. Alright, mate, well it's been an absolute pleasure. I am keen on having a catch up in 6 to 12 months to see how you're getting on. I'm sure where you're up to at that point is going to be a lot more advanced than where we are now. So I'm excited to see that.

For now, we'll give people the teaser. They can go and have a look and then keep updated for whatever comes.

[00:48:40] Chris: Awesome. Hopefully, I'm still plugging away. I've only been at this for a month.

[00:48:44] Adam: Anything can happen. Alright, mate. Excellent stuff. Well, good to have you on.

[00:48:48] Chris: Nice one.

Cheers Adam. Speak soon.


tech for finance

©2022 by Adam Shilton. Privacy Policy - Terms of Use

©2022 by Adam Shilton. Privacy Policy - Terms of Use