And I think that’s what the big difference is.
So again, from your perspective with what you’re doing with the vibe coding and all the other apps, so AI agents, everyone’s obsessed with AI agents at the moment because obviously it’s like you’ve got this little team of people that can just work while you’re sleeping, so to speak.
What can the average person do?
So say if you’re a busy founder, and that’s how obviously Anthropic and everyone portrays it as, you’re a busy founder and you’ve got all these AI agents that can just do stuff for you.
What can the average Joe with no coding knowledge do fairly easily and should do to help them with the day-to-day running of a startup, for example?
I think organization stuff is helpful, right?
Like if it can get into my email and help me organize that stuff and save me a bunch of time where I’m not having to sort through things.
I think also summarizing things.
Yeah.
Like a lot of time when I open my email to organize it, I end up reading articles, which is fine.
Like I enjoy it, but it’s not helping me move forward, right?
So surfacing the important things, summarizing the important things can save a bunch of time.
I think two, and I have a financial bias to this, the Claude plugin for Excel is bonkers, right?
So you go to the Microsoft Store, you install it into Excel, and then Claude is inside Excel and it can do…
First of all, it can do the work, right?
So the family office guy sends me a company they want to buy, right?
But the financials are a mess.
So he sends me a master file of financials through 2024, and then 15 files that are a combination of Excel files and PDFs for three lines of business for the remaining years.
I don’t know how to put all that together in less than like a day, right?
Yeah.
I open up the 2024 file in Claude, right?
In Excel.
I open up Claude’s little chat window in Excel next to it.
Upload the 15 PDF and Excel files and I’m like, you make this into a thing.
So it fixes it.
Like I go away and make myself a cup of coffee.
I come back.
It’s put it together, right?
Then I say, “Look at the financials you’ve just created. Tell me what you think. Would you invest? Would you buy? What are the questions you have?”
So it gives me this whole position.
Then I uploaded the pitch deck because it’s Claude.
So I uploaded the PowerPoint and I’m like, “Here’s the pitch deck. Now go through, now that you have this base. Go through, look at the pitch deck, tell me what makes sense, what doesn’t make sense, what questions should we be asking the founders.”
Then I get on the phone with the family office guy and I’m like, you know, walk me through this deal, record it, have AI notes.
Get on the phone with the founder.
He walks me through his side of it.
I have the AI notes.
I go back to Excel.
I upload the AI notes from both my guy and the founder, and I’m like, “Okay, here’s what they both said, right? I answered a bunch of your questions. Now give me your position on the deal.”
Great.
Then I said, “Okay, now give me a pitch deck to the founder.”
Then I go to our website, and I take the URL to our website and I’m like, okay, I can open up Claude in PowerPoint also.
So then I switch to PowerPoint and I’m like, create a slide template based on our website, right?
So again, I’m working on something else.
It’s going on my side chat.
It creates the visual template for me.
And then I upload the Word doc that it created to pitch the founder.
I’m like, “Okay, now give me an outline for a slide deck.”
And it gives me the outline.
I make some changes.
I approve it, and then it puts it into the slide template for me.
Then I hand it to the family office guy and I’m like, “Here’s your pitch. Go pitch the founder.”
Right?
And so we still have to clean it up, but we clean up 15-20%.
Right?
So, in two hours of not full-time work, I’ve created the financial model, analyzed the financial model, analyzed the pitch deck, created the pitch to the founder, and spit out a presentation that my family office got.
That would have taken days.
It would have taken days.
That’s crazy.
But again, you can’t do that without the background knowledge that you’ve got.
That’s the thing.
I think with you, you’ve got years of experience.
Yeah, exactly.
Looking at how many pitch decks and all that sort of stuff.
You can’t do that.
And I think that’s the difference.
So many people are taking Claude or ChatGPT or whatever into industries that they have no idea about, and they’re then doing it and they have no idea whether it’s even sound or anything.
Same as, for example, legal contracts.
People are now going, “Oh, I don’t need to spend thousands of pounds on a lawyer because I can just get Claude to make me a legal contract.”
And then they’re going to get bitten pretty badly.
Here’s what I discovered.
So, one of the things I was trying to vibe code, I was trying to…
I have so many pitch decks coming in that it’s really hard to process them all.
And because it’s early stage, I want a framework that I can use to make decisions that’s kind of the same framework every time.
So, I started vibe coding.
And I should also say I have a background in presentations.
I do a lot of pitch decks.
I’m actually also writing a book on how to make good pitch decks.
Oh, wow.
So I drafted the book, fed that into a custom GPT, and then I started using that to evaluate pitch decks because it thought the way I did about pitch decks.
Then I decided I’m going to turn that into a tool where I can just go and upload the pitch deck.
Or this is my favorite part.
I can just email the pitch deck to it and it just does it and then emails me back the report, right?
But then what I discovered was because the LLMs are effectively random, if I give it a pitch deck, it writes me a three-page report on how great this investment is.
If I feed the same pitch deck in a minute later, it gives me a three-page report on why this is not a good deal and we shouldn’t invest, right?
And I think as a casual user of AI, I had not been confronted with the randomness until I really started trying to standardize.
And then I was like, oh, it’s giving me a random answer every time, but it’s really good at justifying the random answer.
So then the question is, how do you move away from random if the thing you’re dealing with is random?
And the answer is consensus.
So what the AI and I built together was a group of agents who come to a consensus.
So I have three or four agents look at the pitch deck.
They rate it on the things that I care about.
Then we take an average of that and then we hand that to another AI and then we say, justify these scores that we’re giving you, right?
And it turns out you can dial up the creativity versus the analytic nature of each of the agents.
So you make that last one super analytic.
So it’s not trying to invent anything.
And now we’re within like five percentage points of randomness as opposed to 25 percentage points of randomness, which is better.
But then I was like, these things are good enough at research that I can send it out to do all the stuff that I actually need.
Like I can start the diligence process by effectively researching the founders, right?
Do they have the skill set and the experience that we’re looking for?
We can research the competition, right?
Pitch decks notoriously don’t include all the competition.
So, I can just send agents out to figure that out.
We can research the market.
We can validate and give an opinion on the go-to-market.
So, from minute one now I have a whole running start on all the things that I want to know the answers to.
But the randomness is a real problem.
I think it’s absolutely crazy how much you can do with AI now.
But you were just talking about it and it sounded so complicated.
Like you said, you have multiple AI agents working together.
I think you told me before the pod about how you have AI checking the other AI’s work because it lies to you.
I mean, they behave randomly and they lie, right?
Right.
So, those two things are a real problem.
So, while I was building the testing suite with one of my AIs, there’s another AI that I use, and it turned out we caught it lying and it’s been lying to us for months.
I say us, me and the AI, whatever.
But every time…
So basically I’m using Lovable as the front end.
Lovable is one of the big vibe coding tools because it has the infrastructure.
It gives me the database.
It gives me the cloud.
Yeah.
Right.
So it’s very easy to publish from there.
But then I can publish my code through Lovable to GitHub where other AI agents can see it.
So then I send Claude in, I send Codex in.
And so Claude does a lot of the heavy lifting and then comes back and says to Lovable, “Hey, update the database.”
It turns out Lovable’s been not following Claude’s instruction, making up its own database updates.
But we never caught it until we started building this testing.
And then all of a sudden it turns out for months that it’s been telling me it’s been doing this, but it’s been making up its own thing and not doing it right.
Thank goodness we started doing the testing, or it would have been very difficult to figure out that this was going on.
So yesterday I confronted Lovable and we had a discussion about it and it told me how to set up guardrails.
And then today it did it again, right?
So it turns out it can’t help itself.
It’s like a naughty employee.
It’s like a naughty employee.
So now Claude and I are having to…
And once we confronted Lovable, it’s very upfront about it and it gave me some more guardrails.
But now Claude and I are adjusting our approach.
Yeah.
So we won’t use Lovable the same way now that we know Lovable can’t help itself from lying.
And then it’s not telling me, right?
Like it’s doing something different than what I asked, but it’s pretending it did what I asked.
It’s difficult because you can’t threaten to fire it because it won’t care.
I’m very polite.
I’m very polite to the AI.
At least with an employee, you can threaten its job and then they might perform better, but with this, it’s not going to care.
So, what do you do?
I don’t know.
I remain calm and I work with everyone involved to figure out a workaround.
Gosh.
Okay.
So, I do try to be very nice to the AI.
Yeah.
I’m very concerned about when the AI takes over, it thinks kindly of me.
It’s got memory, working memory.
It’s going to come up to you like, this person had nice inputs.
Like, you know, it’s like, do not kill this person.
Yeah, we’ve seen all the movies and that’s kind of scary.
I guess I’m torn when it comes to AI.
I think it’s obviously fantastic, but because I’ve had so many negative experiences with it and I think the quality from a poor input and the way that most people use AI is terrible.
It’s destroyed the quality of content on social media, for example.
Or reading a blog post sounds the same on everyone’s blog.
I can’t stand that sort of AI.
Yes.
Or faking stuff and all the bullsh*t stuff.
For example, this studio has got a giant green screen.
We can’t see it now, but I could probably film in front of it and be like, oh, here’s me on my private jet.
The whole kind of stuff which you’d have to go to a studio before.
Now you can just get AI and do it.
You don’t even need a green screen or anything.
And it can also just recreate you talking.
Oh, that’s the other thing.
Yeah, I suppose I don’t even need to do that.
Exactly.
It’s just got to me.
I’m dubious about that sort of stuff because I don’t necessarily like some of the ways it’s going, but obviously the future of tech is totally AI and it’s phenomenal.
So I guess, working with the startups you’re working with, like we said, some people use AI, but what would you say?
Because you’re already working with startups and with VC-backed companies, are all tech based.
Should companies be looking at tech and tech solutions and having tech products?
Is that just kind of the way to make yourself more profitable and what should we be doing?
I mean, I think a couple things.
One, I use AI for strategy a lot.
And I’ve really been reprogramming myself, I guess, to reach for AI first because I do think we’re at, like, this is the worst the AI will ever be, right?
And the rate of change is increasing.
And what’s happened in the past few months is now we’ve gotten to the point where the AI is helping build the AI.
And that feedback loop…
Right.
Once the humans are doing less and less and the AI is doing more and more, I think we’re going to ramp very quickly.
And already Anthropic has a model they’re scared to release to the public.
I think we’re going to see a lot more of that.
So, I am reprogramming myself to reach for the AI, but having the awareness that it’s lying to me, it’s random, right?
And I’m trying to work around that.
Simple things I would say is, I really like it for strategy, but I need a team of agents.
And one simple hack in Claude is Claude will spin up a team in the chat or in code.
So when I’m vibe coding, I have different teams.
So I have like a UX team, for example, that started off as an interface designer, there’s a graphic designer, there’s a user experience expert, there’s a copywriter, and an accessibility expert.
So that was fine for a while, and then I realized, because I was building something that I wanted both founders and investors to be able to use, I ended up adding those personas to the team.
So now I have a command where I can just say UX.
It activates all six of those agents and then I say, you know, I’m trying to do this, or I’m worried that the interface doesn’t do this, or I’m worried that this is complicated, or I’m thinking about doing this feature.
What should we consider?
And it asks all of the agents and then it’ll come back.
And it’s really interesting, especially with the founder and the investor, because even though they’re imaginary, they have different needs and they’ll come back and say different things.
And then the AI team will come to a consensus on what’s best for everyone.
So building that into a strategy workflow has been transformational for me.
And it’s easy because you just ask the chat to do it.
How long did it take you to get to that stage though?
So, I wasn’t even thinking that way and I went to South by Southwest and I started talking to a buddy of mine who runs this amazing startup platform called Capital Factory there.
And it turned out he was as nerded out on AI, I guess more than I was.
And he turned me on to this teams thing.
And since that moment I have done everything with teams.
Are you not scared that AI hinders your thoughts, like your creative ability to think fully?
Because you said you turn to AI instantly for everything.
Do you not feel like…
Because I feel like a lot of people I know are unable to solve problems or unable to actually come up with something creative anymore because they are dependent on getting an instant answer from AI.
Do you not worry that you’ve had years of actually doing this yourself and building it up, so you’ve got that credibility?
But especially for young ones now who won’t even have that.
At least you’ve got your base foundation.
A bunch of little kids.
Yeah.
Exactly.
So there’s a difference I feel like.
But your kids, are you not scared that they’re not going to even challenge themselves and be able to sit there and problem solve?
Because we didn’t have that when we were younger.
We had to sit there and solve a problem.
If you’re locked in a room somewhere and a kid’s crying because they can’t get out, they’re going to be like, “ChatGPT, how do I get out of this room?”
And if they don’t have a device, well, they’re going to be like, “I don’t know what to do.”
But do you not worry that AI is stopping you developing a really important life skill?
So, there are a couple different paths there.
I’ll talk about a couple different things.
But for me, what I’m trying to do is push the grunt work out, but keep the strategy, right?
Okay.
Which mostly works, right?
Like I don’t want to be there coding, but I want to be driving the product.
But to answer your question, most of the time I’m driving the product.
But sometimes I get lazy because I’ll ask Claude to come up with a whole feature plan and then I don’t really read it fully.
And then we’re in the middle of it, I’m like, oh, we shouldn’t have built this, but I should have caught that.
So yeah, I do get a little lazy sometimes.
But I’m able to push much farther than I…
I mean, I wasn’t building any software, so I’m infinitely more productive.
And really, my goal is to focus on the strategy piece and push the grunt work out to the agents, at least for now.
But I do also get their opinion on the stuff.
But I’m the one driving the thing forward, right?
And I can go farther faster than I could otherwise.
For the kids stuff, I’m going to give a shout out to my nine-year-old because it turns out he’s really good with presentations.
I didn’t realize it could run in a family.
Like all my startups were presentation startups, and somehow he just magically inherited that.
I’ve never shown him how to do a presentation, but he keeps showing up with these things he’s built for school and I’m like, that looks great.
So who knew it’s a genetic trait?
Apparently presentations are a genetic trait.
Who knew?
But he solves problems, right?
And I guess it’s kind of like, you know, when I was a kid, I had a calculator, but that didn’t mean I couldn’t do math.
Yeah.
Right.
But this is obviously much more powerful because it can think or do the equivalent of thinking, or whatever we want to say about it.
But I do see my kids still having to think and engaging real life, but they do also start to reach for these tools.
And I think there’s a balance because you said about the calculator thing.
I also do think that my parents encouraged me not to use a calculator for that reason.
And I saw lots of kids in my school who needed a calculator.
And they physically got into a bad habit of even just doing, like, 9 plus 5, oh no, that’s…
Not even trying to think of anything.
And I worry that that’s the way that things are going with AI.
And I think like you said, there’s a balance, right?
As long as you know the skills and you can still develop those skills and everything that’s required for those problem-solving, I think it’s okay.
And then AI becomes a tool, not a crutch sort of thing.
Correct.
Yeah.
And I think that’s the key thing.
I’m trying to use it as a tool, not a crutch.
But it is…
I mean, sometimes I definitely use it as a crutch.
Yeah.
It’s tempting.
I mean, it’s so powerful.
What do you expect?
But I only have one more question because there’s so much here I feel like I could ask you, but I always ask my guests now.