AI Is Accelerating Faster Than Humans Can Keep Up: What Should Finance Pros Do
In this episode of Future Finance, Paul Barnhurst and Glenn Hopper discuss the accelerating pace of AI development and why trying to keep up with every new model is becoming a losing battle. They explore what finance professionals should focus on instead, including AI workflows, process design, data, governance, and automation.
Paul and Glenn also discuss how AI is changing the way finance teams work. As tools become more capable, they believe the lasting advantage will come from knowing how to structure processes, give clear instructions, manage AI-powered workflows, and apply human judgment where it matters most.
In this episode, you will discover:
Why finance professionals should stop chasing every new AI model.
Which AI skills are likely to remain valuable as tools continue to change.
Why process thinking, data, and orchestration are becoming more important.
What risks finance teams should consider when using vibe coding and automation.
Why governance and human oversight still matter in AI-powered workflows.
Paul and Glenn explain that finance professionals do not need to constantly switch tools to gain value from AI. Instead, they should focus on building strong processes that can work across different platforms as the technology continues to evolve.
Follow Glenn:
LinkedIn: https://www.linkedin.com/in/gbhopperiii
Follow Paul:
LinkedIn: https://www.linkedin.com/in/thefpandaguy
Disclosure: Portions of this episode (such as the introduction or promotional segments) use AI-generated voice narration produced under human editorial review.
Future Finance is sponsored by QFlow.ai, the strategic finance platform solving the toughest part of planning and analysis: B2B revenue. Align sales, marketing, and finance, speed up decision-making, and lock in accountability with QFlow.ai.
Stay tuned for a deeper understanding of how AI is shaping the future of finance and what it means for businesses and individuals alike.
In Today's Episode:
[00:00] - Welcome
[02:21] - The Race for AGI
[08:28] - AI Is Moving Fast
[17:38] - Stop Chasing Models
[19:43] - Durable AI Skills
[25:03] - Data & Orchestration
[28:08] - Vibe Coding Risks
[32:18] - AI Governance
[34:42] - The Cost of AI
[42:07] - What to Focus On
[44:35] - Closing Thoughts
Full Show Transcript:
Host: Paul Barnhurst (00:55):
Future Finance is brought to you by qflow.ai, the strategic finance platform solving the toughest part of planning and analysis. B2B revenue, align sales, marketing and finance seamlessly, speed up decision-making, and lock in accountability with qflow.ai. Welcome to another episode of FP&A Unlocked, Financial Modellers. Oh wait, Future Finance. Glenn, I can't even remember what show I'm recording today. What's wrong with me?
Co-Host: Glenn Hopper (01:33):
This is part of the reason I had to let go of FP&A today. The whole universe right now, Paul, is too much to keep up with. I have whiplash.
Host: Paul Barnhurst (01:42):
And it is so easy, whether it's good or not, it's so easy to create content nowadays.
Co-Host: Glenn Hopper (01:48):
Yes, content, but should we put a qualifier on that? It's easy to create content, but maybe not much easier to create quality content.
Host: Paul Barnhurst (01:57):
I would agree. I mean, I think it's a little easier in that AI can help you with images, it can help you with video. If you are the one that's going to put in the time, it can make you a lot more efficient, but unfortunately there's definitely a qualifier. Most of the contents being released now is not the standard it needs to be. LinkedIn, our finance people, who's scrolling LinkedIn and you're like, why is everybody thinking an AI generated comment is helpful or post or et cetera? All right. Well again, welcome to Future Finance. As y'all know, we are sponsored by QFlow. If you haven't checked them out, check them out. Great platform. We're going to talk about today the fact that the Race for AI has just increased incredibly. So we have a few things, going to share a little bit of a tech stack survey and then kind of talk at the end what you do about it.
(02:53):
But where I want to start, Gwen, is you haven't seen yet a video is released this week that I think sums up this whole race for AGI in a humorous way really well, but it has so much depth in this 30-second video. So we're going to watch it. So if you're listening to this, I encourage you to go to YouTube, watch this part. If not, you can still hear the video. And then I want your take since you haven't seen it. So it's titled The Race for AGI. You ready?
Co-Host: Glenn Hopper (03:23):
Yep.
Host: Paul Barnhurst (03:24):
We're going to bring it. All right, I'm going to hit play here. Here we go.
Co-Host: Glenn Hopper (03:37):
You traitor. Cautious. It's different. Oops, wrong trend. Run damn it run. Eat my lawsuits nonprofit boy.
Host: Paul Barnhurst (04:01):
GPUs for all your starving models. We can do tremendous things together.
Co-Host: Glenn Hopper (04:17):
We need to slow down.
Host: Paul Barnhurst (04:30):
All right. First time you've seen it. Thoughts
Co-Host: Glenn Hopper (04:33):
There,
Host: Paul Barnhurst (04:33):
Glen.
Co-Host: Glenn Hopper (04:34):
Yeah, pretty funny and pretty amazing. Every time I watch AI generated video, I think about, do you remember from years ago the Will Smith eating spaghetti video, just how -
Host: Paul Barnhurst (04:44):
Yes, like
Co-Host: Glenn Hopper (04:45):
Three years ago. It wasn't even like uncanny valley. It was just nightmare fuel. And thinking about how far AI video has come, it's amazing. But that aside, people ask me all the time about the AI doomers or the sort of Ray Kurzweil camp of the post AGI and super AI. I don't know where this is going. I do think, and just don't take this as, I'm talking out of both sides of my mouth here, that's why it's taken me so long and I'm stumbling. But here's what I'll say.
Host: Paul Barnhurst (05:16):
Glenn is speechless from the video people. This is the first. Enjoy it.
Co-Host: Glenn Hopper (05:22):
Here's what I'll say. There is something about how much money all these frontier lab companies have raised and they have to get a return for their investors. We saw that OpenAI has delayed their IPO. I think Anthropic is still moving forward for this year. I think that talking about the dangers of AI because it's so smart and can do so much is probably ultimately good for them from an investor standpoint, unless they go too far and government steps in and puts blocks in that keep them from advancing. I'm not saying that's right or wrong. I'm not going to opine on that. I certainly don't think the current administration is going to do that, but I think that a lot of the fear mongering comes from trying to hype how good their AI is. At the same time, we do need some kind of guardrails. It doesn't have to be super intelligence that it could get into the hands of a bad actor and do something really bad.
(06:21):
I mean, there's hackers out there, the white hats and black hats for years have been an issue, and now everybody has nuclear weapons for their hacking. So we do need to have some guardrails and some safety and all that. The whole idea of AI running amuck and taking over, that to me, and Andrew Eng wrote about this this week or last week, but the whole idea that AI is escaping and doing things on its own, to me, that's not a knock against AI. It is, "Hey, Anthropic and OpenAI, how about you build better sandboxes that they can't break out of?" To me, that's where the failure is, not in the technology itself. But no, a hundred percent. I mean, we're racing towards this cliff. We don't know what's on the other side. We know if we pause that the potential, the existential threat, the geopolitical, whatever you want to call it of, well, what if China gets there first?
(07:08):
And I get all that too. And also, Paul, we're old. We've been through a lot of technology changes in our life. None has been like this, and we don't really know what the other side looks like. So yeah, I get it. I don't know. I don't know if I answered your question at all, but a lot of mixed feelings on that video, but I think it perfectly encapsulates where we are.
Host: Paul Barnhurst (07:27):
Yeah, I think everybody's some mixed feelings and tonne of humour. I love the llama. He's sitting there. Oh, wrong trend, and jumps on it and the lawsuits and great laugh, but it makes you think, where are we heading? Is this a cliff? What does it all mean? Where does regulation China fit in? And the reality is none of us can clearly answer it. I think we all agree there needs to be guardrails. The AI needs to do a better job of their sandboxes. And where that line is, how we cross it, what happens with AGI, none of us can answer that. But I think it encapsulated so many things that are going on in one short 30-second clip that I thought it'd be fun to start with. A little bit for a laugh and a little bit just to remind everybody that Glenn and I are here learning with all of you.
(08:23):
And today, what we really want to go through is Glen wants to share something that he shared with me earlier that just shows that we're accelerating at such a pace, and I think I'm going to say this right, that it's impossible to keep up. Fair assessment, Glen, let's bring your screen up and we're going to walk through this and I'll let you take it where you want, but I think this is a really interesting thing. This is an artefact you created, it looks like, with Claude. And so go ahead, check out on YouTube again. Glen's been fancy with his Claude building and he's a little
Co-Host: Glenn Hopper (08:59):
Journey.
Host: Paul Barnhurst (09:00):
Just
Co-Host: Glenn Hopper (09:01):
Firing the AI slop cannon with my show notes.
Host: Paul Barnhurst (09:05):
Slop cannon with Glen, brought to
Co-Host: Glenn Hopper (09:07):
You
Host: Paul Barnhurst (09:08):
By... No, I'm just kidding.
Co-Host: Glenn Hopper (09:11):
So here's the thing. I've met a lot of AI experts since say December of 2022, and these AI experts are really, really good at keeping up with the news and the latest models. And you don't have to know anything about what's going on underneath them. You can just know all the models exist and what they do, and you can have your max subscriptions for ChatGPT, for Claude, for Grok, pay $200 a month to each of them and talk about how you're automating your whole life. I'm not discounting that, but the problem with that is if you're just spending all your time keeping up with whatever model is winning the latest benchmarks and doing the best, then are you really being productive or are you just chasing the latest and greatest syndrome? Yeah. And it's been hard from the beginning. If you think about the pace of how AI models have rolled out compared...
(10:13):
So it seemed fast then, but let's look at some numbers on here. So I was thinking about November of 22, ChatGPT 3.5 comes out. It took all - What
Host: Paul Barnhurst (10:24):
Was it? Didn't they add two million users within 48 hours, the fastest growth ever quicker than Instagram when it launched? Wasn't that the kind of watershed moment?
Co-Host: Glenn Hopper (10:35):
Yeah. And you know what is funny here though? I think about the adoption rate and the hype cycle and the adoption from mine and your viewpoint from a consumer side is one thing, but from a business side is another. And I don't have good stats on the business side, but certainly that's accelerating. But just from the technology itself, it took from November of 22 until March of 2023, 104 days to go from 3.5 to GPT-4. And GPT-4 was a significant step upgrade from 3.5. I remember. And then - People
Host: Paul Barnhurst (11:14):
Weren't told to put rocks in their pizza and all kinds of crazy stuff. Or that was more Bard, if you remember that disaster.
Co-Host: Glenn Hopper (11:24):
Yep, yep. And we'll talk about Google in a minute in this too. And then we go from March of 23, by April of this year, we had ChatGPT 5.5, but we went from 5.5 to 5.6 in only 76 days, and then 5.6 to six in 56 days. And then they rolled out Astra and then Sol and Luna, the version six of it within 19 days of each other. Granted, same model, just the smaller versions of it. But 2023, it was every few months you saw a new model and it was hard to keep up then. 2024 and 25, we were adding incredible capabilities, reasoning, computer use, coding agents. And now this week I was doing a live presentation with a big enterprise level company and we were going through some of their workflows and I was talking about the difference between chat and work in Anthropic.
(12:27):
Well, that morning -
Host: Paul Barnhurst (12:29):
Cowork.
Co-Host: Glenn Hopper (12:30):
They merged the two. Yeah, sorry. Yeah, you said
Host: Paul Barnhurst (12:32):
Cowork. Okay. Just to be clear.
Co-Host: Glenn Hopper (12:34):
Yeah. Well, when you go in the app and you're choosing between, do I want it to just use the chat interface or the whatever Anthropic calls it? Yeah, or work for me,
Host: Paul Barnhurst (12:43):
The cowork interface. Yeah.
Co-Host: Glenn Hopper (12:45):
Yeah. Then the next day they dropped the new 5.5 in the middle of it. But prior to it, and I was big on ChatGPT out of the gates and stuck with them for most of my work, really until around January of this year. And I though, I can't believe it. I'm actually going to switch to Anthropic. I feel like the whole professional work world switched to Anthropic and that was the case.
Host: Paul Barnhurst (13:10):
When 4.5 came out, it started the switch, Opus
Co-Host: Glenn Hopper (13:14):
4.5. And that was the case really until August, September, really until Astra six came out or ChatGPT-6 Astra, whatever. It's hard to even keep up with the names anymore, but then it was like, oh, wait a minute. Because I probably said on this show, I was telling everybody, I think Anthropic is going to be the business platform and ChatGPT is going to be what's left for the consumers. And boy, have they proven me wrong. I'm blowing through so many ChatGPT tokens right now. It's doing incredible work. I'm on the max plan with that. So are you
Host: Paul Barnhurst (13:44):
Finding
Co-Host: Glenn Hopper (13:45):
Out
Host: Paul Barnhurst (13:45):
You think Astra is better than Claude right now?
Co-Host: Glenn Hopper (13:48):
Yeah, right now it is. Yeah, honestly. I'm still using, I have so many coding projects in Claude and I've heard, and you can now, you can integrate them. Transitioning between them is not nearly as painful as it used to be. So right now I've got ChatGPT and Claude talking to the same folders and I've got this whole elaborate workflow as all of us AI geeks do, but I'm using them back and forth. But in my anecdotal, I don't have an official benchmark, but right now ChatGPT's winning and I haven't said that since the end of last year. But the problem is chasing all this, switching back and forth and what conversation did I have where? And it's like I'm trying to use it to be as productive as possible, but now people are lighting me up about Muse. I haven't had two seconds to look at Muse.
(14:37):
I should read tech. These are from just friends who know I'm the AI guy, so they're asking me these questions and they're dropping all these crazy... I'm like, I don't even know what you're talking about. And I spent all day every day doing this, but I guess the point to me that's coming out right now, well, so let's just look at just this month. So the beginning of the month, okay, we've got Fable 5.1 and Mythos 5.1 from Anthropic and Google's kind of doing their things. They've been really focused on their flash models. And then this Meta, the Muse product comes out, which I still haven't used. And then, oh, September 3rd, now we have Astra. By 8th, Muse is going along more. Then we've got DeepSeq with their latest drop. Grok 4.7. I don't know anybody really that's using Grok for anything.
Host: Paul Barnhurst (15:26):
I will add something on Grok. Microsoft is adding it to its model. So in the Excel agent, I think it's beta right now, there are some rollouts where you can select Grok. They're adding 4.6 and 4.7 in addition to having the GPT and the Clauds. And now in some situations, Microsoft really hasn't announced this much, but underneath Copilot, they're sending it straight to their own AI. They're not sending it to Claude or Gronk or Copilot, they're using Microsoft AI. So in this announcement time, in the last little while, they've kind of quietly released their own AI in all this, and I'm sure it's a cost savings thing to tune their models for certain cases so they don't have to pay someone else that can just use their server bandwidth they already have. But I thought that was interesting. So keep going. But I mean, just the Gronk one reminded me of that, which just shows as you're stacking all this in there, there's even other things.
(16:25):
I'm not seeing on here... Oh, you do have Opus 5.5, right?
Co-Host: Glenn Hopper (16:30):
Right. And then an hour and a half later, OpenAI drops their Saul and Luna update to get them to GPT-6. And truthfully, maybe that's going to help me on some of the tokens. It's helping me a lot right now using ChatGPT. So 5.5, I haven't had a chance to test this for Opus. Opus, one of my biggest problems with Clog was the copy that it was writing. When it talked to me, it was fine, but if I was trying to put together a presentation or a client deliverable and have it help me do that, I don't know. It knows I'm a consultant, so I guess it was trying to do McKinsey speak, but it talked like a crazy person. Who says these words? I don't say that. Why are you saying that? And then so I was in ChatGPT six was better at writing. It's frustrating to think you're going to get AI to help you with something and it ends up taking you longer because then you have to rewrite content.
(17:26):
Anyway. So that just happened yesterday. So 5.5 drops and then the Soul and Luna models drop within an hour and a half of each other. And then today, because everybody's wondering what is going on with Google, they do announce that their latest flagship model is in post-training right now, and it's going to come out here in the next couple of months. The head of DeepMind said it's going to ship much earlier than your end. So I think it's going to be interesting to see what Gemini is doing. But I guess my whole thing here is I have to follow it because I'm training on it and I'm using it in actual client implementations, but this is a losing battle just trying to chase what's out there. And I've got a lot more that I researched. I mean, it's pretty interesting to see where we are right now and how much it's accelerating.
(18:11):
And I've got a lot of other notes here that we can talk about on the show, but maybe let's pause there. Tell me what you're thinking. How do you use it in your workflow? What are your clients yet? The
Host: Paul Barnhurst (18:19):
Rate of acceleration at this point, I think I go back to Secret CFO. I think many of our audience follows them. I know you're very familiar with him. When Fable came out a couple months ago, he dropped a note. Fable 5.0 came out, everything you need to know, and he had an arrow. Then down at the very bottom of the post it was nothing, get back to work. It was banned a couple days later by the government. Humour, got a tonne of comments, but the reality is don't chase the shiny toy. Build good processes. Some of the best software, and I think that the softwares that's going to win, are the ones that are agnostic and can allow you to switch between tools and they make decisions on the backend because as a human, you're not going to be able to keep up. And if your goal is to always fine tune and get that one extra percent or save that tiny bit of tokens or whatever, it's not worth the effort.
(19:17):
It's like the person that has 20 credit cards in their wallet and tries to keep track of which one to use everywhere to maximise their points.
Co-Host: Glenn Hopper (19:25):
That is a great analogy.
Host: Paul Barnhurst (19:27):
Less than 1% of all people can do that,
(19:30):
And the time they spend is probably not worth the savings, even though they're like, "I'm the cool guy who got a billion points." It's more of a challenge. So just stop trying to do the challenge, stop trying to keep up. Don't be the points guy. That's my take. At the end of the day, I think we all agree, in many ways, these models are going to become kind of a commodity in the end. They're going to be very similar, and it's what we can build around them. It's what we can do with them. There'll be more and more open source. So focus on building good processes that can work with AI and keep an eye on it. You do need to stay current. You can't stick your head in the sand, but don't try to keep up with every change. Just work on building good processes, check every few months, and pick your horse.
Co-Host: Glenn Hopper (20:17):
Yeah, that really, really great point. And let me bring my screen back up again, because this is where I'm having a... It's a bit of a conundrum to know what to train people on. And to me, it's not about... If you're training on how to use a model, yeah, I get it. It's like if you've never seen Excel before, you need some training to understand, to get started moving in there. And most of it though, you're going to learn from... Once you learn a little bit how to use Excel, you're going to see some other formulas people did. You're going to figure out how to do it for themselves. You're like, "What's a slicer? Oh, that's cool. Let me figure out how to make that," or whatever the latest the tech is. So if I'm training how to use a tool and the interface is changing that often, I think, "Well, have they wasted their money on..." I'm arguing against myself, my own business, but have they wasted their money on training?
(21:11):
And I would say no, because you need to get the basics and then you can come in and once you have the basics and you're using it, you start to figure out on your own. So you've got to enter somewhere. So if you haven't been using the tools, you need some basic, this is how to. But what I'm more interested in, the durable skill, what's going to last out of all this is how do I change the way that I work is how do I change the way that I work with AI? So to think about that, I just think when ChatGPT 3.5 came out, it was you upload a spreadsheet or whatever. Well, 3.5 couldn't really do spreadsheets, so you know what I mean. When the models first came out, before we got into reasoning and copilots and all that, it was, "Let me ask you a question, you help me do some analysis," and it's just a chatbot back and forth, and that's great.
(22:03):
We found a lot of great uses for that. Well, then we started getting copilots. We started getting AI, whether it was companies that came up or built into from the SaaS tools that we use, it was, "Hey, now we've got AI inside Excel in our email and our ERP. We've got faster keystrokes, faster ways to build models, to find errors and models and all that." So we started seeing usefulness in that. Then we get projects and skills that it's like, "Oh wow, if I have a project and I can have shared context and I share this project, I don't have to come and re-explain everything every time I come in here." And then skills, "Oh wow, if I do this task once, I can turn it into a skill and do it every time." And then you get scheduled tasks and you think, "Well, I had the project, I had the skill, now instead of me having to do this, I can schedule it to run every Monday morning at 8:00.
(22:51):
And then, oh, if AI can't handle something, I can vibe code a little app to do it." Now, for those watching on YouTube, the agentic harnesses and really full automation by building in code and AI, I feel like this may sound like hubris and maybe somebody can prove me wrong. I have a hard time finding any sort of digital task. I can't fully... Well, me and my team, but we could build in a couple of days, whatever your task is from three-way matching to variance analysis to bank reconciliations, anything that you have a repeatable process in. So what does that mean for our jobs? What are we going to be thinking about? And the way I look at it is we need to become managers. So everybody had management training for managing humans, but I think that the durable skill that comes out of this is understanding how the AI works.
(23:44):
And then yes, all of our finance and accounting skills, those are table stakes. And then from there, we need to learn process thinking and where to apply generative AI, where to apply basic coding, how to architect these, how to put the governance and the guardrails in. And that's what's going to last out of this, but I don't know what does it mean as these AIs... And I'm not going to be a doomer because I think every technology that's come before has led to more jobs, not less. But at the same time, these jobs are going to change a lot. This is like robotics coming to manufacturing. Yeah.
Host: Paul Barnhurst (24:23):
I mean, I think I'm going to be pump the brakes. Change takes way longer than when the technology's ready. Always have -
Co-Host: Glenn Hopper (24:33):
The human part, yeah.
Host: Paul Barnhurst (24:33):
Always will. If you can bring that back up for one second.
Co-Host: Glenn Hopper (24:36):
Oh, yep, yep.
Host: Paul Barnhurst (24:36):
I'll have that back to the stage here. Ever feel like your go-to-market teams and finance speak different languages? This misalignment is a breeding ground for failure, impairing the predictive power of forecasts and delaying decisions that drive efficient growth. It's not for lack of trying, but getting all the data in one place doesn't mean you've gotten everyone on the same page. Meet qflow.ai, the strategic finance platform purpose-built to solve the toughest part of planning and analysis, B2B revenue. QFO quickly integrates key data from your go-to-market stack and accounting platform, then handles all the data prep and normalisation under the hood. It automatically assembles your go-to-market stacks, makes segmented scenario planning a breeze, and closes the planning loop. Create airtight alignment, improve decision latency, and ensure accountability across the team.
(25:46):
So you got chat, copilot, project skills, you have scheduled tasks, via coding, agentic harness, full automation. Great. If you have full automation, I think autonomous finance is a myth in the sense of, hey, you could just have it fully autonomous. I think there's always human in the loop until we hit AGI and we'll deal with that when we're there and what that means, but at the moment. But I think for the training, I think for most people where the focus needs to be is understanding skills and not the technical size of skills. Skills are how to write instructions, learning how to write instructions, when to turn it over to AI, how to do the scripts. It's really about learning orchestration and orchestration is skills and tasks. And honestly, if you learn skills, tasks, you learn the scripts, agentic harness is just kind of wrapping it all together to me.
(26:40):
It's not like it's some whole new thing you have to learn. So that's why I think I believe the most important skill going forward, and it should have always been for FP&A people, and not all finance people, but some finance, is you have to understand data because you got to have the schema, you got to have the context, you have to have a solid data foundation. And all we have to do is look at any Excel model to see many people think in the financial and the accounting ledger and not in a data way, how many people use tables, how many people normalise data? There needs to be better understanding of data, and then you need to understand design thinking. How do I design processes? How do I manage workflows? Whether it's a markdown file, whether it's a slightly different thing. I think from there, the rest will kind of naturally build out if you focus on those things.
(27:31):
So that's kind of my thought on this whole thing and what people need training on. When we train somebody, we showed them how to do projects, we showed them skills, we showed them how to build instructions in Excel. We gave a couple different examples, but really focused on if they can learn these basic things, over time it will take off. So that's kind of what I focus on. I'm not going to say that's right, but that's how I think about it. And frankly, that's where I'm at. I'm not using any agents. Yes, have I scheduled tasks? Yes, do I get emails every day in my inbox? Yes. Do I use ChatGPT, Claude, Copilot, whatever, on a daily basis? Yes, I do. Am I behind? I think so, but compared to most people, I'm not. If I look at what you're doing, I know I should do a lot more and I need more time and there's things I could automate.
(28:18):
And so I think if you learn how to do those fundamentals, you'll be okay. And to me, those fundamentals is understanding really the skills, the instructions, the workflow, the orchestration, because that will guide you toward an engineer. Yes, eventually you need to figure out some basic vibe coding, but that's really, again, if you know how to give good instruction, you understand skills, you understand workflow, you can do vibe coding with natural language. You need someone to validate your code if you haven't learned some coding or you need to learn some basics, but I don't think that's a whole new level. Am I off base?
Co-Host: Glenn Hopper (28:50):
No, no. No, you're spot on. My only thing now that I'm seeing with vibe coding is amazing power, but with that amazing power comes amazing responsibility in that if you are building an app or coding anything and you're putting it in your GitHub repo or direct on Vercel or whatever you're doing, wherever you're posting and you don't know about the basics of cybersecurity and things about OAuth and role-based access and don't put your API key in a public repository, just the basics of that, the cyber risk here, the data security risk here becomes huge. So we've heard all the problems that people have had with just sending out AI generated stuff, whether it's in legal where it's just making, or haven't heard about this in a couple years, but where it was making up cases to cite and those were being brought into court, or if it's consultants using it and not reviewing and just giving AI slop to their clients, that was one issue that hallucinations, I'm not saying it's solved, but people hopefully are a little bit more careful now And citations have helped this.
(30:01):
But to me, the next big danger is going to be around once people figure out how easy it is to vibe code something, how much trouble they're going to get into because they don't know the basics of cybersecurity, because it's never been part of what we've been asked to do. These apps will build everything and you can tell it to make it secure, but if you don't know what you're looking for, I don't know. I think maybe whether it's Cursor or OpenAI or Anthropic or any of the others out there, maybe they are building in the harness that it's going to have better guardrails automatically. But there is a big concern to me when people really start, because people haven't realised how easy and how much they can do with vibe coding yet. So there's going to be some exposure there. Unless the companies figure out if we want people to be doing more of this, we need to put the guardrails in place.
Host: Paul Barnhurst (30:48):
I think we're starting to see the software vendors are putting the guardrails in place within their AI. So remember we released Maria Nakic and she said, Hey, the goal is to have people be able to build their own platform. They're releasing language within. If you do it within a platform, they can add that security and a lot of those guardrails and allow you to build so you can have your own apps. I think K4 has done something similar. I know other tools that could do that in the finance space. And so I think early we often see the software vendors and then AI releases their own kind of version of it. And many of the vendors are like, oh, what do we do now? And so I do think AI will eventually get there, but you're right. And I'm sure we've already had some, we're going to have some major security issues with vibe coding.
(31:42):
I also think vibe coding regardless how easy it is. For most companies, you don't want your average employees vibe coding. You're just creating basically macros everywhere that nobody understands and you're going to have a nightmare. The guardrails and governance that's needed to really have vibe coding used at any large scale at a company I think is much bigger than any of us realise from a maintenance and a long term. And it's one thing if you're doing it on your own desk and you're running the process and it doesn't go out anywhere. Yeah, let people vibe code. That's much like the Excel guy, but when the next person inherits it, odds are they're going to start over because they don't like the way you designed it. So I'm mixed on this one. I see vibe coding. It's not going away. I think anyone who's trying to vibe code major tools is a mistake.
(32:38):
Just don't do it. It's not worth it with very few exceptions. Workflows, simple tasks, some manual things. Yes, you can get some great benefit, but it's like you said, with much power comes much responsibility. What else do we have on this page? I know we're already in about 30 minutes and I have one thing I want to cover. So what else do we want to run through here?
Co-Host: Glenn Hopper (33:00):
Yeah, I mean for me, really that tool level training, it becomes irrelevant so quickly. But if people take the training and they get in there, then that's a good starting point for them. But then they're going to have to keep up with it and know what's changed. Again, I think we've covered that one already is how work gets delegated and what we're doing.
Host: Paul Barnhurst (33:20):
I love your governance point. Who
Co-Host: Glenn Hopper (33:23):
Owns
Host: Paul Barnhurst (33:23):
Automated process? That gets back to what I'm talking about in vibe coding. Governance is going to be the issue, not that you can do it. Well, and security.
Co-Host: Glenn Hopper (33:33):
Yeah. And governance, I mean that's one of the most popular topics that I cover, whether it's in a keynote or a training or with my clients. And that's one thing with agent harnesses is you can build in much deeper, and we could do another episode where we dive into harnesses in particular. I think we
Host: Paul Barnhurst (33:49):
Do
Co-Host: Glenn Hopper (33:49):
Cover
Host: Paul Barnhurst (33:50):
That on a future episode for sure.
Co-Host: Glenn Hopper (33:52):
Yeah, because if you can build the harness, then the tracking becomes just built in part of what you're doing. So it's not like we're going to go save every prompt and response that we get from a chat as part of the audit trail. That's just not practical and individuals aren't going to do that. But if you haven't harnessed the audit trail and you can see and you have the guardrails and you tell the AI what it can and can't do and you have the human in the loop and you show with timestamp, this human signed off on this at this point, this is how everything was calculated. Oh, by the way, we can run this again so it's not just like asking a chatbot the same question and getting a different answer. This is all the things, it's kind of like when we were good at Excel, that's how we could stand out in the field.
(34:37):
And then after Excel, for me, it became Make and Zapier and InAden and all these workflow automation tools. It's another thing that people had to learn. But now with AI and with what these tools can do, I know there's still plenty of people using those, but really what we're doing with the off the shelf features from OpenAI and Anthropic and Microsoft as well, but is building our own agents and doing this. So yeah, like you said, we need to know what the ultimate outcome we're looking for, but the whole foundation of tools is shifting.
Host: Paul Barnhurst (35:11):
Yeah, agree. And anything else you want to share on here? Let's briefly touch on you at a cost, how that's changed. Let's take two minutes there and then I want to share something or maybe we'll wrap. We'll just see how time goes and I can share mine another day.
Co-Host: Glenn Hopper (35:24):
Yeah. The biggest thing on cost to me is if you look at on the left here on the screen, the cost per token, so per one million output tokens, when GPT-4 came out in 2023, it was $60 per million. Now if you go down to the cheapest ChatGPT model six Luna that was just released, 50 cents per million tokens. Now the difference is you're like, well, wait a minute, everybody's blowing through their budget. The difference is we're burning through so many more tokens. If you wanted to, you mentioned turning a light off before the broadcast started today, that light compared to what's happening, the energy consumption out of these data centres is nothing at this point.
Host: Paul Barnhurst (36:06):
My whole house's power is nothing. I think as I told you, the data centre they're building in Utah would power the entire state of Utah roughly twice. And that's one data centre just for military application.
Co-Host: Glenn Hopper (36:18):
Yeah. And I'm in Memphis where Elon Musk's the big XAI data centre is, and there's a lot of pushback locally on that. There's a lot
Host: Paul Barnhurst (36:25):
Everywhere. We won't get into the politics of it. We could spend hours on that and piss everybody off probably depending on how we decide to view it. But it's fascinating to say the least.
Co-Host: Glenn Hopper (36:39):
Yeah. My son told me this joke a year or so ago. It wasn't his joke, but I'm crediting him as the one who told me, but I think it kind of perfectly encapsulates where we are with AI. Everybody's really worried about AI taking their job and I didn't really take it very seriously, but recently a friend of mine just lost his job to AI and his job was to lie to people and drink 40 gallons of water an hour.
Host: Paul Barnhurst (37:08):
Lie to people.
Co-Host: Glenn Hopper (37:10):
Nice. Someone said this on a panel I was on or somewhere not long ago, they said that AI was as cheap as it's ever going to be. And I thought that is not a rule of economics. And you were completely mistaken because this is, you mentioned at the top of the show, it's becoming commoditized. I understand it's exorbitantly expensive to every query that we run and every model training run and all that, but for this to be sustainable, it has to get cheaper. And then whether it's the competitive forces or the technology itself that it's improving, it is going to keep getting cheaper, but we're going to be using more and more of it.
Host: Paul Barnhurst (37:54):
I think there's two competing things going on here. There's the Moore's Law. If we look at computers, how much did it used to cost for 512K? And now you get a terabyte for what that used to cost. That's going to apply here, but I think there's the flip side of people are going to use more and more. So the cost -
Co-Host: Glenn Hopper (38:17):
I just ordered a two-
Host: Paul Barnhurst (38:18):
Two terabyte.
Co-Host: Glenn Hopper (38:19):
I just ordered a two terabyte drive for under $200, which I think about -
Host: Paul Barnhurst (38:24):
Your solid state. Yeah. If you're doing the old, I mean you're a flash. If you're doing solid state, it'd be like nothing, the old brick instead of the newer technology. But what I was going to say is I think you have two competing things. You have the fact that yes, the cost gets cheaper and cheaper. You have the use going up. What consumers pay, now they may be getting a lot more value, has to increase for these companies to be profitable. So our $100 subscription, you could easily argue is probably worth 500 or your 200. How much value do you get out of your $200 a month subscription to Claude or ChatGPT, would you argue? If you had to pay for the value, would you pay
Co-Host: Glenn Hopper (39:10):
More
Host: Paul Barnhurst (39:10):
Than 200?
Co-Host: Glenn Hopper (39:12):
Yeah.
Host: Paul Barnhurst (39:12):
Would you pay 500?
Co-Host: Glenn Hopper (39:13):
Not across the board. I'd be more selective to pick one, but I had an admin last year. I have an AI admin this year and arguably better than my actual admin that I had last year. No, it doesn't run autonomously, but the end result with a little bit of nudging for me is so just on that alone, but I think about I'm running some really big client churn analysis type projects. So you're
Host: Paul Barnhurst (39:39):
Someone who uses it heavily and there's a lot like Goo. You're saying you're willing to pay two and a half times. There are a lot of others. I'm going to guess if they told you it was $800, you would make it work as a subscription. You wouldn't cancel it?
Co-Host: Glenn Hopper (39:53):
Probably,
Host: Paul Barnhurst (39:54):
Yeah. Yeah. You wouldn't like it. I'm going to guess even at a thousand you probably wouldn't, you'd be more selective, you might cancel some others, but you would still use one model because you're getting enough
Co-Host: Glenn Hopper (40:04):
Value. I don't know. Yeah, it would be very interesting to see where I would draw that
Host: Paul Barnhurst (40:08):
Line.Because I
Co-Host: Glenn Hopper (40:09):
Know
Host: Paul Barnhurst (40:09):
You enough know all things to automate that I'm guessing it has to be saving you at least 20 hours a month compared to the old way.
Co-Host: Glenn Hopper (40:18):
Oh, probably.
Host: Paul Barnhurst (40:20):
It's
Co-Host: Glenn Hopper (40:20):
Probably significantly more than that. Yeah.
Host: Paul Barnhurst (40:22):
Probably what? What would you guess? 60,
Co-Host: Glenn Hopper (40:25):
100? I would guess 40 to 60.
Host: Paul Barnhurst (40:28):
So let's just split the difference. Let's say it's 50.
(40:30):
Let's say an hour is worth, we'll use simple math. Someone in finance, we'll just use $100. You coul use $50, but let's say 100. If it's saving you 50 hours, 50 times 100 is $5,000. So anything less than that in theory is saving you. So if we get to the point where the average person is getting, let's say you're the extreme, the average person getting is 10 times and it's worth $100 an hour, some it's going to be worth more, some it's going to be a lot less, that's a thousand dollars. These AI to really make a profit have to increase the cost, but as the cost of economics goes down, we're getting more and more value. So you might be getting five times what you're getting for that 200 account because the token cost has come down so much, but you're paying more because when they go public, they're going to have to show profit.
(41:25):
So that's how I think about, I kind of mix the two together. I don't know if that makes sense if I'm off base, but I think there are two competing things taking place. So we as consumers are going to pay more, but the cost is going to come down in the sense of what it's costing to run it. But given that's been heavily subsidised, they have to offset that in charging us more.
Co-Host: Glenn Hopper (41:45):
Yep. Yep.
Host: Paul Barnhurst (41:46):
Well, I think we'll cover mine another day. I'll just say we did a survey of about 200 people and I'll share one little bit. I won't go into it deep. We'll cover it on another episode because I want to get your thoughts. But basically what we found is those who are using Agentic AI were much happier with their FP&A tools, their closed consolidation, et cetera. The number one thing people said that was a pain when we did the survey, two people said cost out of 120 people that provided a response. Guess how many said data?
Co-Host: Glenn Hopper (42:14):
Most
Host: Paul Barnhurst (42:15):
60 mentioned data. Only two mentioned cost. So that's something interesting to vendors. So I want to dig into that a little bit next time, but that's a primer. The other thing we found, if you're a big company, you're probably using Copilot. If you're a small company, you're using Claude because you're not locked in. Finance overwhelmingly is saying, "Hey, Claude's the tool to use right now." Now some of the things you mentioned, maybe that will shift again, but time will tell. So I think wrap up, what's the takeaway for our listeners from our ramble today, Glenn?
Co-Host: Glenn Hopper (42:49):
Yeah, the key takeaway to me is we're all going to stroke out if we keep trying to keep up.
Host: Paul Barnhurst (42:57):
I've never heard it said that way.
Co-Host: Glenn Hopper (42:59):
Maybe that might be a colloquial phrase if we try to keep up with the latest that's out there. I think pick a tool that you use that works for you. Learn how to maximise that, whether it's Copilot or Claude or ChatGPT or whatever. Get really good at using that. Understand that this right now is the worst AI we're ever going to have from here forward. So figure it out how to use it because I hate to say this, and I think our audience, I'm probably preaching to the choir, but if you're late to the game on AI, you're not too late. I mean, the tools are so good now, you can figure it out pretty quick, especially once you learn, oh, this is helping save me time. It seems like a fool's errand to me to keep trying to chase whatever the latest, greatest model is.
(43:45):
Just learn how to use the technology, learn how to change your thinking based on the existence of this technology and figure out what it is that you really add that provides value and lean into that and figure out, okay, this is me who I am in my role. Now I have a new tool to do my role. How can I get better at that role rather than how do I get better at using whatever the latest AI?
Host: Paul Barnhurst (44:06):
I think you nailed it and I think you gave me a couple ideas for future episodes. I think one episode would be really interesting to talk about how we need to change the way we lead, the way we manage, talking about the human side of all this. And then the second thing is, like you said, people learn how to orchestrate workflows, learn how to break apart your workflows, your processes, to write documentation that a system can use to be able to understand context. That's going to apply to different levels and different ways with all of the AI tools. So pick the tool you're going to learn with, yes, are some better than others, but it's changing so quickly and they're all good. You can get benefit from all of them. I don't care if it's Copilot, I don't care if it's Gronk, I don't care if it's Meta, I don't care which one it is.
(45:00):
If you can't get benefit from them, that's a you problem, not an AI problem.
Co-Host: Glenn Hopper (45:06):
Yeah. Well said.
Host: Paul Barnhurst (45:07):
I'm going to close with that. It's a you problem, people. No, I'm kidding.
Co-Host: Glenn Hopper (45:11):
Do better.
Host: Paul Barnhurst (45:12):
It's like the old Uncle Sam, we want you for the... Oh man, that's definitely a sign it might be time to wrap up. Well, enjoyed this episode. We'll be back with more. We'll sell a little bit on the survey. We have some other exciting things coming, but if you have questions, reach out to us. Let us know what you'd like us to cover. We'd love to hear from you. We haven't heard from you. What topics would be most helpful? Send us a list and we'll cover them. Thank you everybody. And until next time, we're signing off. Thanks for listening to the Future Finance Show and thanks to our sponsor, qflow.ai. If you enjoyed this episode, please leave a rating and review on your podcast platform of choice, and may your robot overlords be with you.
