Jeff Bezos: "Never Hire Your Friends Under 40"

Does your experience with AI sound a little something like this? You've been prompting for 20 minutes. The output is polished, confident, and completely useless. That's what happens when AI doesn't know anything about your business.

HubSpot's AI works with your actual customer data, every interaction and every signal your team has generated over time. So instead of prompting and hoping, you ask once and ship it. Check out HubSpot.com, the agentic customer platform for growing businesses.

Let me give you a Jeff Bezos quote. I don't know if you saw this interview. Neil's about to fall asleep, but it's okay, guys. I got you. Which interview?

CNBC one? No, I forgot where it was. It was with Walter Isaacson who wrote the Elon Musk biography. So he said this. I thought this was really interesting.

So Jeff Bezos said, never hire your friends if you're under 40, but after 40, always hire your friends. And so- with blue origin right now he has an executive that was sitting next to him that worked with him for 20 years or so he's like i couldn't have chosen like a better partner to work with on this and neil let's think about this if we think about some of the people that we work with now um that we work then we're both over 40 now right um so do you think this rings true because i tend to think that after 40 you kind of know what you want and what you don't want and you know who you trust and who you don't trust so what do you think

so I would say for me particularly in my twenties, I made way too many mistakes with hiring as well. My thirties, I made less. Um, but now that I'm in my forties and I'm still young, I'm 41. I kind of know what's a good fit.

What's not a good fit, what works, um, who will work. And, um, I still make mistakes, but my chances of making mistakes are a lot less. And what I found is a lot of the people I've worked with over the years that I've learned to trust and I've enjoyed working with them, they're good at what they do and I would work again with them.

But they were first colleagues and now they're friends. But they started off as colleagues and I would work with a lot of them all over again. You know what?

I think about, for example, like, today you're sick, right? But we're supposed to record with Noah. So Noah, I've known for a while, right? And so, you know, it's like, it's tried and trusted.

Like, I know how he's going to do. I know what to predict. And I know he's reliable, like he's going to show up on time. And so... That's what it is. And I know sometimes you might forget a thing or two when it comes to recording, but for the most part, it's been good, right?

And so I think what Jeff Bezos here is true. I actually hadn't thought about it, but I do believe that, yeah, if you're under 40, you kind of don't know what you want or you kind of don't have enough pattern recognition. So you end up making the same stupid mistakes and then eventually you'll realize them. Yep.

Yeah, I agree with you on that. But yeah, I never really thought about it that way. In general right now, I don't really think about hiring friends or non-friends. I just try to hire people who are really good at specific things that I'm trying to get done.

I never really hired generalists, especially in marketing. People are like, oh, I'm really good at paid ads. I'd be like, no, no, no, which platform? I'm good at all the platforms.

It's okay. I don't want to work with you. But if someone's like, I'm really amazing at Facebook. Look what I can do on Facebook. Cool, sounds good. You just run Facebook.

But I tend to still hire very targeted specialists who don't just specialize in a form of marketing, but specialize in a channel, a very specific channel or platform. And I found that to work way better than hiring generalists.

Dude, speaking of which, yeah, I would just say that... The company we were talking about earlier, they have a lot of people that are looking for jobs.

Really? Yeah, we can talk about it later. But yeah, we can talk about it later. Okay, so Eric's talking about another popular marketer. They have a massive social following.

We won't go into their name. They do really well on the view count on social media. But why would you say there are a lot of people who are looking for a job? Is it culture?

Because I know the pay is pretty decent. The hours of work are hard, though. Culture, which leads to a lot of churn, especially within that team. Got it.

Yeah. So anyway, it's good to know. So anyway, I want to pull this up from Ramp. So Ramp does a really good job. You want to talk about data storytelling?

They have so much data around company spend in general. It's not all the data in the world. But I think there's some interesting stuff. So let me show this to you, Neil.

So they have this over here. They have a lot of startups and small businesses. So it's good to look at those kind of businesses through the lens of Ramp. So it says you're spending too much on AI.

You're also using too little, okay? So a big AI bill doesn't mean you're spending, you're using too much on AI. It means you're buying it wrong. So most companies I talk to have two contradictory problems.

They overspend on AI they use, but use far less AI than they should. So for example, like everyone just using the latest, like Fable 5 for, you know, like what's the...

what's the freezing temperature for water, right? Are you serious? That's one of the worst things to use it for, and it's expensive. Yeah, I'm saying people are using it in stupid ways, right?

So let's start with a simple question. How much does $100,000 buy an AI? So... For those of you that can't see this, if you're spending 100 grand on Fable 5 for 5 billion tokens, this is what you get, right?

And then Kimi 2.6, which is what you can call an open model or open-ish model, you get 210 billion tokens, so 21 times more, right? Opus 4.7 is a little higher than Fable 5. GPT 5.4 is a little more generous. Look at this.

Gemini 3 Flash. Neil, remember you were talking about how Google is just going to undercut everyone? They are undercutting quite a bit here. So 89 billion, right?

So you get... It's Kimi 2.6. Kimi 2.6, that's more of like an open source model. That's one of the Chinese models. Got it. Yeah. So for the same 100 grand... And you watch Gemini 3 and Microsoft and all of them, they'll get even cheaper and cheaper because they just print so much cash.

Yeah, Microsoft doesn't even register right now, but at least you just pay attention to Gemini. And like, honestly, Neil, I still continue to think that we're going to have to buy a lot more infrastructure. Your team's already buying computers. I'm looking at buying more computers because you just keep running all this stuff locally.

Like GLM 5.2 came out and that's like equivalent to 4.8 or pretty close and a lot of people are talking about it. So I think we're just going to buy more. And I think this too, Neil, I think clients, because clients aren't going to want us to just do this for them. So what we're going to do is we're going to sell them some of our computer and we'll probably mark it up a little bit too.

Dude, and what we've been starting to do is not buy computers and hardware. We've been starting to lease it. Because the big fear is you buy it and then it's not powerful enough for the future models or not enough RAM or whatever it may be. So it's just like, yeah, might as well just lease it.

Remember I told you I was leasing all my Apple devices? Uh-huh. Yeah. So anyway, so for $100,000, you can buy 5 billion tokens on the smartest model on the market or 210 billion tokens, 42x4.

Right. So anyway, you guys should read this one, but a hundred thousand dollars drawn as finished work. So Fable five, it can do like support tickets and sales calls analyzed. It'll cost you for $5,000 or sorry, 5,000 tasks, $20 each.

Okay. Now for Kimmy 2.6, you want 500,000 tasks, 20 cents per task. Okay. I like, I like how this is modeled out. This is pretty useful. Yeah, that is.

And on a side note, when it comes to leasing computers, you know not all computer companies do it, but there are third-party companies that will buy it for you and lease it to you. No, I didn't know that. Can you name one of them? I can find out from Tracy.

Okay. She was telling us in all hands. I was like, oh, I didn't know that. She's like, or all hands for leadership. And she was just like, yeah, it's just a more efficient way.

I was like, oh yeah, that's smart. And she's like, yeah. She's like, you buy it and then what? Because she's just like, some of the hardware that we bought a year ago is useless for some of this stuff.

It's like the last thing you want to do is make a big investment and then be like a year later, two years later, we all need new stuff. I told you, I've been, this device that I'm showing on my screen right now, this is a DGX Spark from NVIDIA. I'm like actively getting this repaired right now. They're sending me a replacement.

So some of this stuff, like it dies pretty quickly too. So, okay, I want to go a little more on this. So model type, frontier models for frontier problems. Okay, well, like, ideally, if you're doing a level, an S-tier problem, you're using the S-tier model on it, right?

So you're going to have a better time. It's going to succeed on a tougher problem versus if you run for like a tough problem, you're using a cheaper model. It's going to take longer to solve it and it might take like a few more turns to do it. So you're better off using PhD level intelligence there, right?

So, All that to say is it's just they have a conclusion here where you should actually spend more. OK, so for easy work, work that is well understood and scope gets the cheapest model that clears your benchmarks. Hard work.

So novel, ambiguous, high stake work gets the best model on maximum effort. Sample as many times as you like with much more leeway on spend. I think a lot of people just aren't thinking about that. So they think that and this is what I mentioned earlier, Neil, a lot of companies just leave the smartest one on by default and that ends up burning a lot of tokens.

You want to know why they do it? I know why. I think a lot of people just, one, I think there's a lot of reasons, but also I just think people don't know any better.

Real quick, if you want to acquire customers faster and more efficiently this year with the latest strategies and tactics, then check out singlegrain.com. That is my ad agency. Again, www.singlegrain.com. Check it out.

And if it seems like a fit, we'll get in touch and help you with a free marketing plan. I think that's part of it. But I think a bigger one is it's not their money.

Well, there's that too, yeah. Right, because if people had to be accountable for it and it affected their comp, you bet people would be really quick at figuring out ways to save money and figure out how to get more done with less money. You know what I noticed, Neil? So on our Claude usage, right, I don't know if this is correlation or causation, but, and I love, by the way, this is no knock.

I think if you speak Spanish, typically you're going to be a little more wordy. Would you agree or disagree with that? Like just their language in general involves the use of more words, yes?

I think so, but I don't know. Okay. So the answer is yes. You can research this. We're looking at our cloud team. So the native Spanish speakers, their usage is higher, significantly higher.

Yeah. And then like, you know, people have said, one of our mutual friends said that like, um, If you use Chinese words, it actually takes up less tokens, but I don't think that's true anymore because the number of letters counts as one character versus each Chinese character is just one, right? So it's more efficient to write in Chinese, but I don't know if that's necessarily true.

But I can see with my usage, at least, the different language, just how you do things as a native speaker is a little different than an English speaker.

The other thing that I'm seeing in marketing is people are using AI when they don't need to use AI. And I'll give you a great example of this. I'm seeing a lot of people go out there and creating things like social content and cranking out versions of it and then sending it to other people on the team like, here's just some random ideas that I got from AI. Hope this helps you come up with other ideas.

And I'm like, so you just spent all this money creating images and videos that won't be published. I'm like, you could have just shared a list of ideas without going there. And they're like, no, I want to show that I go above and beyond for a company. And I'm like, this is not going above and beyond.

This is just wasting money. And then I was talking to another friend at a conference. They work very closely with Salesforce. Some of the compensation supposedly, and I don't know this firsthand, but some of the compensation is supposedly based on token usage, depending on the department you are or how good you look.

So people are just burning up tokens and building stuff to just randomly use tokens and do a forum so that way they can get praise. And I'm like, this is just terrible ways to compensate people. People should be compensated on how efficient they are, the results that they're bringing in the most efficient way, not just on quote unquote token usage. Because then people just token maxed.

Let's talk about that more because I've had a thought. I'm like, hey, I'm realizing there's a problem here that I've caused in my own company, Neil. And so...

Yes, there's so much, not everyone, but a lot of people right now are trying to automate first before they even understand the problem. And I'm like, whoa, whoa, whoa, whoa. Yes, I talk about AI all the time, but you have to nail it before you scale it. Because if you don't understand the problem and you scale out using AI, it's going to be a bunch of garbage, right?

And for some reason, it's hard for me to get that point across. I'm like, guys, I don't want you to be a slot cannon, right? And the other thing is...

I still see pockets of AI theater, right? I'm like, guys, don't make dashboards just to make dashboards. And then yet, I still continue to see dashboards. I'm like, guys, we need to make things that lead to an outcome.

It needs to lead to an outcome. Ideally, it's a business outcome. Whether you're saving time or whether you're driving more leads or whatever it is exactly, it needs to lead to that. And not only that, we're big on making skills that other people can use.

We're like, guys, okay, It's really cool that you made this skill for yourself, but if not, at least one person is using it, it's useless, right? So we have a rule now where it's like, we're doing this training right now where it's like, okay, you have to at least upload two skills to our skills dojo and at least one other person has to be proven using it. So that's the verifiable proof.

Now, token maxing is one thing. I think we can look at token use as one thing. It shouldn't be the only thing, but we need to look at, that's the quantifiable piece. But the qualifiable piece is, hey, can you prove to me

that you're using in a smart way. And maybe that's your manager that's actually talking to you about it. But again, this goes back to nail it before you scale it. Because right now I'm starting to see documents, like artifacts and things like that, that we're showing to clients where it's clearly AI generated and we're not putting thought in front of it.

And people are just reading off the slides. And I'm just like, dude, enough. Dude, I was in a meeting with a financial firm. And when I was meeting with them...

we're talking about new ways to grow. So this company specifically targets rich people to market to, to try to convince them to park some of their money with them and they would manage their money. So think of like a wealth manager, wealth advisor, whatever you want to end up calling it, or financial advisor.

So I was like, what do you send reports? How do you get clients? What do you send them before? And they're showing me some of the workflows and systems that they do.

So one of them was showing me how they analyze clients' current portfolio. They'll tell them what their stock portfolio looks like, what numbers it can hit, do projections, and all this kind of stuff.

I was just like, show me an example of it. They were loading up how they're using ChatGPT and asking ChatGPT about a stock. It was a tech stock, so I know quite a bit. At least I track quite a few of the tech stocks out there.

They were wanting to know how likely a tech stock was to hit like 100 and something, like maybe like 120 or 150. I forgot what the exact number was. And it's only like 10, 15% away from that current number.

So they're showing me this analysis from ChadGBT that it shows that it's somewhere around 60 to 70% likelihood that it would hit that number for the next six plus months this year. And I'm like, this stock moves up and down 5% in days so many times you can't even count on two hands just in a matter of like a month or two.

And I was just like, do you really need to use ChatGPT to do some of this stuff? Or could you just use your brain and look at a chart? But the reason I'm using this example here is the probability that ChatGPT ended up giving was so off and so bad that I was just like,

using it for helping you do some of your work and creating some of your marketing collateral and analysis collateral to try to convince people to become more customers, in this case, is actually hurting you versus just using your own head and just using common sense.

And I'm just like, you don't need AI to analyze everything for you and have it do all the work. In many cases, not only will it do it inaccurately, it'll do a much crappier job than if you just use common sense. Yep. It's judgment.

People like to say this word taste. It still comes down to judgment and taste. If you know that, then you're going to know the right times to use this stuff. And to me, it's just another tool, right?

It's a very powerful tool. But I think right now, because people are talking about it so much, and then I'm talking about it so much, people just assume you should use AI for everything. But that means you're not using your best judgment, and that's not a good thing. And that's what people pay for.

People are paying for your taste or judgment, whatever you want to end up calling it. And I know a lot of companies that pay others to use AI forum because they know the other person has better taste in judgment. Yeah. I want to hear, I spoke to someone yesterday at a company agency.

Let's say this agency has a thousand people or so. And he was saying that, remember you had brought up agency, AI agency multiples being like 30X or something like that. So he said he's seeing something like around, call it 22 to 26X. Do you have an update on that?

Like how that's been looking? We're still seeing them go for up to 30-ish X or 30-something X. I don't think they all are, but they're definitely a lot of them getting well over 20. The problem with AI agencies is the sales process takes six months. Give it a year, year and a half, the multiples will come down.

Did you see Accenture's latest numbers? No. Except your stock has been taking because people not needing as much consulting in many different areas. But yeah, I know you and I have the same belief when it comes to some of the consulting that these types of firms do, especially on the management consulting end.

It's just like, I don't know who pays a management consultant. I think it's the biggest sham in the world. It's a bunch of, I've met some of these people and some of them are really smart, but I feel like it's a lot of paper pushing and a lot of posturing and acting like you're doing something and like trying to delay, like you're basically just trying to manage and string along the client for as long as possible. That's how I feel about it.

And I haven't seen anything to break my beliefs. Yes. It's so bad. And this is the thing. Hiring people does not necessarily fix a lot of your problems.

A lot of your problems will be fixed yourself. Yep. With good people. It's like the Finn thing, right? Finn got acquired by Salesforce for three point something billion dollars, which is the old intercom, right?

The reason it did really well is the founder went in and fixed it himself or herself. I don't know who the founder is. It's him. Yeah. But a lot of times it's like, you got to get your own hands dirty.

You can't expect someone new to come in and be like, oh, I hired this bank consultant. They're going to come and fix all our problems. Well, if they're able to fix all your problems and do this, they wouldn't be a bank consultant in the first place. That's it for today.

And we will see you tomorrow.