AI transcript
0:00:07 Welcome to the Next Wave Podcast.
0:00:11 Today, we’re doing a live episode here at HubSpot Inbound in San Francisco,
0:00:13 and I have with me a repeat guest,
0:00:15 Matan Grinberg from Factory,
0:00:18 who the last time he was on here was our biggest episode ever,
0:00:21 where he did a live demo where he showed how
0:00:25 he could clone DocuSign in 20 minutes using Factory.
0:00:26 I think you’re going to love the conversation,
0:00:28 and so let’s just get right into it.
0:00:37 Using only 20% of your business data
0:00:40 is like dating someone who only texts emojis.
0:00:42 First of all, that’s annoying.
0:00:45 And second, you’re missing a lot of context.
0:00:47 But that’s how most businesses operate today,
0:00:50 using only 20% of their data.
0:00:52 Unless you have HubSpot, where all the emails,
0:00:56 call logs, and chat messages turn into insights to grow your business.
0:00:59 Because all that data makes all the difference.
0:01:05 Learn more at HubSpot dot com.
0:01:07 Matan, welcome back to the Next Wave.
0:01:08 Matan Grinberg: Thank you for having me.
0:01:09 It’s great to be here in person.
0:01:11 Matan Grinberg: Yeah, it’s awesome meeting you in person.
0:01:13 We’ve talked online so much, and you know,
0:01:15 the first episode that you did on the Next Wave
0:01:17 was actually our biggest episode ever.
0:01:19 I think we got like 67,000 views on it.
0:01:21 It was a slightly controversial episode,
0:01:23 because we were talking about cloning DocuSign.
0:01:25 And people were like, “Did you really clone all of DocuSign?”
0:01:26 There’s like one part of it.
0:01:27 It was like, “Yeah, it was one part of it.”
0:01:29 I still think it was a really cool demo.
0:01:31 And since then, you’ve raised a Series B, is that correct?
0:01:32 That’s right, that’s right.
0:01:34 Yeah, so we just raised our Series B.
0:01:39 we raised $50 million with some great folks like NEA, Sequoia,
0:01:41 JP Morgan, NVIDIA.
0:01:42 I’m very excited about it.
0:01:43 Matan Grinberg: That’s amazing.
0:01:46 So, NVIDIA, that seems like one of the best partners you could have,
0:01:50 and obviously NEA leading the round, and Sequoia.
0:01:52 What do you plan to do with the $50 million?
0:01:52 That’s a lot of money here.
0:01:56 Matan Grinberg: I think for us, the big thing is really building out a team
0:02:01 and growing the resources to reach more of the large enterprises
0:02:02 that we’ve been working with so far.
0:02:05 So, some of our large customers include Ernst & Young,
0:02:09 the large accounting firm, as well as MongoDB,
0:02:11 Bayer, and quite a few others.
0:02:17 And the reality is, to serve these very massive enterprises
0:02:19 and to help them with this behavior change
0:02:21 and this transformation of software development,
0:02:23 it actually requires a lot of resources.
0:02:27 Generally, if you have an org of 50,000 engineers,
0:02:30 maybe 5,000 of them are super excited to adopt
0:02:32 the coolest and latest AI technology.
0:02:36 And then maybe the 45,000 others are like, “You know what?
0:02:39 I’m actually happy in my VIM or in my Emacs and I don’t want to mess around
0:02:42 with any of this stuff or kind of change what I’ve been doing.”
0:02:45 And so, we want to invest in the resources to both build out our engineering
0:02:49 and product team so that we can make it that much easier to adopt.
0:02:54 And then also, our go-to-market team to reach developers and reach these organizations
0:03:00 to kind of raise awareness and make them understand this transformation that’s happening.
0:03:05 And, you know, going out in the field and helping these developers kind of adopt these new workflows.
0:03:08 Yeah, so I was wondering, like, how much do you have to help these companies?
0:03:11 Like, I would imagine that you show them Factory, you show them a demo,
0:03:14 but then, like, how do they actually use it and how do you get the whole team using it?
0:03:16 How much are you having to hold people’s hands?
0:03:19 Great question. It’s crazy how much it depends.
0:03:24 One thing that is for sure is just at the end of the day, developers are human.
0:03:27 And if you show them something that they used to do that was very painful,
0:03:32 that’s now going to be a lot less painful with Factory, generally, they’re going to start adopting it.
0:03:37 Now, really, the work is just what is the thing from their seat that’s really painful,
0:03:41 and how can we use Factory and the droids to accelerate that?
0:03:46 Something that often comes up is migrations or modernizations or refactors.
0:03:49 There are very few developers who actually enjoy doing things like that.
0:03:49 Right.
0:03:52 And so, kind of going in and getting a migration done with the droids
0:03:56 is kind of a quick way to get them to adopt to this new workflow.
0:03:59 And with your GA release, like, how do you feel like that went?
0:04:02 Like, I feel like you got a lot of attention.
0:04:04 Something I heard from people is like, they’re kind of confused.
0:04:07 Like, you know, is this for, like, vibe coders or is this for, like, large enterprises?
0:04:09 Or, like, what’s the focus?
0:04:11 Yeah, that’s a great question.
0:04:15 And I think, so, in May, we released Factory for GA.
0:04:16 It was a really great release.
0:04:21 We got tens of thousands of new users, but I think the important thing was we released Factory,
0:04:25 and basically the main way you accessed it at the time was via the web.
0:04:25 Right.
0:04:32 And something that we’ve learned is that, you know, in this transformation to what we’re calling agent-native development,
0:04:38 the reality is you need multiple surfaces where you can engage with these agents, with these droids.
0:04:41 And we only gave people one surface, which was the web.
0:04:47 That’s why, in this release, we’re really having a surface-agnostic or interface-agnostic approach.
0:04:53 In other words, if you’re going to be willing to change your behavior to become agent-native,
0:05:01 we need to meet you wherever you are, whether that’s in the web, or in your terminal, or in your IDE, or in Slack, or in Jira, or in Linear.
0:05:11 Wherever you might have an idea of, here’s a task I want to delegate to an agent, we need to meet you where you are, and make it very easy for you to do so.
0:05:20 Just like if we were co-workers, and you were like, oh man, here’s a perfect task for Matan, you’d know you can’t only tell me that via email or via Slack.
0:05:27 You could tag me in a ticket, you could, you know, send it to me on Slack, you could send me a text, like, kind of wherever is easiest for you personally, you’d be able to do that.
0:05:32 And we need to provide people kind of a similar ease of use for the droids.
0:05:38 And so that’s like, if you’re going to like a huge organization, enterprise, every single individual engineer gets to do that.
0:05:42 So like, they get to pick like, oh, I want to use it, I think you have a terminal thing coming out as well, right?
0:05:42 Yes, that’s right.
0:05:48 So like, whether I’m doing it web-based, or terminal, or I’m an executive in Slack, however you want to use it, you can.
0:05:49 Is that kind of correct?
0:05:53 Yes, pending, like, approval from IT to, you know, assuming they give access to these.
0:06:01 But basically, if you provide access, you can do it via Slack, via, honestly via email, like, you can do it obviously in the terminal.
0:06:07 But the point is to kind of mold to whatever, you know, the developer’s workflow looks like, to make it seamless.
0:06:16 Because the reality is we’re going from a world where developers wrote 100% of their code, to a world where developers will write 0% of their code.
0:06:20 And this new primitive in this world is going to be a delegation.
0:06:24 The primitive used to be writing a line of code, or writing a function, or writing a file.
0:06:25 It’s now going to be a delegation.
0:06:32 And so to kind of bring people into this new world with this new primitive, we need to make that delegation as easy and accessible as possible.
0:06:37 So people like the Primogen, people like that who love, like, almost like the art of coding.
0:06:42 You think that’s almost going to be like a niche thing in the future where it’s like, oh yeah, it’s kind of like have you have vinyl records or something like that?
0:06:43 You know, where it’s…
0:06:53 To a certain extent, maybe like the line by line might be, but there’s kind of a new art that emerges, which is, okay, great, you can delegate to an agent.
0:06:55 How much can you parallelize?
0:06:59 So similar, there’s kind of an analogy, you know, NVIDIA is one of our investors.
0:07:03 We love using this analogy between CPUs and GPUs.
0:07:03 Right.
0:07:06 Where CPUs, they’re all about processing in series.
0:07:11 You want to get as many computations as you can done in series in some amount of time.
0:07:13 Whereas the big unlock from GPUs was parallelization.
0:07:14 Right.
0:07:19 It turns out for certain tasks like matrix multiplication, you can actually parallelize a lot of the computation.
0:07:21 So that instead of like one, you know, operation at a time.
0:07:22 You can do them all at the same time.
0:07:29 Yeah, you can do four or ten or a hundred and the more you can parallelize, the faster the downstream computation.
0:07:39 It kind of emerges a similar thing with agents, which is if you can parallelize and have 20 droids working at the same time or 100 droids working at the same time, then you’re way faster.
0:07:42 There’s kind of an art of like, how do you actually parallelize to 20 agents at a time?
0:07:46 That’s kind of difficult, but I think that’s the new thing that will emerge.
0:07:52 And there was recently this study that came out from MIT, talking about like 95% of AI pilots have been failing.
0:07:52 Yes.
0:07:58 I was really, you know, I know you guys are having like a competition about this, like a man versus machine kind of competition upcoming.
0:07:59 Maybe it’s already live.
0:08:00 It’s actually this Saturday.
0:08:00 Okay.
0:08:01 It’ll be live by the time.
0:08:02 Yeah, so it’ll already happen.
0:08:05 So you can check in and see the results, see who won.
0:08:06 I’m kind of betting on machine.
0:08:07 Likewise.
0:08:08 Yeah.
0:08:09 So what are your thoughts on that study?
0:08:15 And like, you know, some people were saying like, oh, that’s, you know, AI has been completely overhyped and this is proof of it.
0:08:25 Yeah, I actually think it’s a really good wake up call that the way people have been measuring success in these deployments has been flawed.
0:08:28 So maybe stepping back, what did the study say?
0:08:37 I believe the specific phrasing was 95% of AI adoption efforts in the enterprise were not deemed successful.
0:08:41 And so that’s like across, not just software development, but AI across, you know, every vertical in the enterprise.
0:08:46 And I think the reason is because success criteria has been very ill defined.
0:08:55 For software development in particular, the metrics that people rely on are lines of code generated or self-reported developer productivity.
0:08:59 The problem is software development is fundamentally a pipeline.
0:09:02 And one step of that pipeline is implementing the code.
0:09:03 Right.
0:09:10 But the reality is before that there’s kind of understanding and planning, then there’s the coding itself, then there’s testing and reviewing and documenting.
0:09:20 And if you have like a literal pipe and you expand the like radius of one part of that pipe, but do nothing to the other parts, you actually don’t increase your throughput at all.
0:09:20 Right.
0:09:29 And so it’s kind of no surprise to me that if you’re just measuring things like lines of code generated, well, you give everyone lines of code generators, you’re going to have more lines of code.
0:09:33 But you’re not going to actually have success and like, you’re not going to start shipping software faster.
0:09:35 Yeah, what are those lines of code doing as well, right?
0:09:39 Like you could do 10,000 lines of code that are like, you know, unnecessary.
0:09:42 And someone has to actually go and figure out what are these lines of code doing.
0:09:47 And so now there are more people going and reviewing code or adding tests to that code or adding docs.
0:09:56 And it actually ends up not increasing the speed, which is why like this kind of problem that this MIT study pointed out is why when we do our deployments,
0:10:02 we’re really focused on really what are the business like ROI metrics that matter.
0:10:03 It’s not lines of code generated.
0:10:06 No business is succeeding or failing because of lines of code generated.
0:10:06 Right.
0:10:10 Like they’re succeeding or failing because we missed shipping this feature.
0:10:15 And then our competitor got it out and we had users churn and now they’re going to that tool instead of us.
0:10:19 And it’s whatever causing like those are the things at the business level that matter.
0:10:27 And that’s why we are focused when we do our deployments on let’s get this modernization done faster because we can’t ship this new feature until we modernize our database from whatever.
0:10:29 And you’re focused on like business outcomes as well, right?
0:10:30 Exactly right.
0:10:35 It’s like when I heard that study, actually, I thought about you guys because I was like, that’s actually something that’d be really great for factory.
0:10:40 Because like you guys are actually being hands on with, you know, enterprises, learning how they’re using these tools,
0:10:43 helping them like figure out the best way to actually get a great business outcome.
0:10:44 Yeah.
0:10:47 Whereas like cursor and the other one, there’s like, here you go.
0:10:49 Here’s this, you know, cool AI coding tool.
0:10:54 I think most of the tools in this space are focused on the individual and optimizing the individual.
0:10:58 Factory is focused on optimizing the team and the organization.
0:11:04 And surprise, software development is a team sport and it really matters to optimize for those team outcomes.
0:11:06 We have like a Michael Jordan quote in our office.
0:11:13 And I think he’s kind of a good example of someone who learned that you can optimize like the individual stats of a player.
0:11:17 But if you don’t end up winning the championship at the end of the season, no one cares, no one remembers.
0:11:19 All right. So it’s very similar to like writing more lines of code.
0:11:21 Exactly right. Exactly right.
0:11:29 Okay, let’s be honest. Your AI prompts aren’t giving you the results you deserve.
0:11:38 But with a little coaching, you can transform from basic prompts to engineering conversations that get you exactly what you want from ChatGPT.
0:11:47 That’s what this playbook delivers, not just random prompts, but a step-by-step system with the exact techniques top AI professionals use every day.
0:11:52 You’ll have your own personalized prompt engineering system that delivers consistent results.
0:11:57 Get it right now. Scan the QR code or click the link in the description below.
0:11:58 Now let’s get back to the show.
0:12:07 You know, when GPT-5 came out recently, you know, a lot of people were saying it’s kind of similar to the MIT study.
0:12:14 They were like, you know, AI was way overhyped and now like all, you know, we’ve hit a wall and there’s gonna be like an AI winter again and all this kind of stuff.
0:12:20 Like, what are your thoughts on it? Because I thought GPT-5 was pretty impressive, like not like revolutionary, but like a great step forward.
0:12:24 Things just kind of keep going forward and getting better and better. Like, what are your thoughts?
0:12:34 Yeah. So we work closely with the team at OpenAI. So we actually tested internally GPT-5 and it was sufficiently good that we changed it to our default model for like the self-service plan for factory.
0:12:41 But I also think basically there were really high expectations. They thought it was going to be the same jump as like three to four would be four to five.
0:12:49 Right. But I think the reality is a good way to think about these models is kind of like, I don’t know, imagine some subject that you’re not a deep expert in.
0:12:54 My go-to example here is like biology. I studied physics. I wasn’t that deep in biology.
0:13:00 How easy is it for you to tell the difference between someone who studied biology in high school versus college?
0:13:04 Pretty easy. Maybe on the order of a few seconds, you can tell the difference.
0:13:09 What about the difference between someone who studied it in undergrad versus PhD? Maybe it takes you a couple of minutes.
0:13:17 What about like a postdoc versus a professor? If you don’t know anything about biology, it’ll actually take you kind of a long time to figure out that difference.
0:13:20 I think it’s kind of similar as these models get better and better.
0:13:30 For most generic tasks, GPT-4 and 5 will perform just as well because like, you know, shocker to implement like, I don’t know, quick sort, GPT-4 and 5 will be able to do it.
0:13:39 Now, obviously, as the context gets bigger and like the problem gets more niche, or really, as you’re getting to like the tail ends of the distributions of these models and their performance,
0:13:41 that’s when you’ll notice the difference.
0:13:45 Only a few people are aware of those differences at those tail ends.
0:13:51 So like for the generic user or for someone kind of outside of domain, it’s really hard to tell the difference.
0:13:51 Yeah.
0:13:57 And I think having the expectations that there’s going to be some obvious like step function difference between these models,
0:14:05 it’s like assuming that as a non-expert in biology, you notice the difference between like a professor who’s been a professor for one year versus a 20-year one.
0:14:09 When it’s like, it’s outside of your domain, you might just not be able to notice quickly.
0:14:18 Yeah, it feels like ever since ’01 really, to be honest, like when ’01 first came out, I was actually doing a live episode with Matt Wolff, my co-host in Boston.
0:14:18 Yeah.
0:14:27 And it was right when ’01 came out, and I was like, this is huge, like there’s a new paradigm that’s been unlocked with reasoning, and now you can just throw more compute at the models and they’ll get smarter and smarter.
0:14:32 And a lot of people were just like, we’re not getting that, and they were like, oh, this is like, it takes longer and like the answer is slightly better.
0:14:37 I was like, it made me realize, like there’s already a thing where I think most people can’t recognize how much better the models are getting, right?
0:14:38 Yeah.
0:14:44 And just like what you were explaining, and it’s going to become more and more important for them to actually like see what the end result is.
0:14:57 Yeah. And actually, I think ’01’ is a good example where, in this analogy of like whatever high school, college, it’s kind of like the non-reasoning models are like a gut reaction from like whatever, professor, postdoc, college student.
0:15:01 And then the reasoning models are like, have them go and sit down and do something for a day.
0:15:04 And we end up being impressed more by that.
0:15:11 In the analogy, it’s like having a college student go and write you an essay on some topic in biology versus ask a professor for like a gut response on something.
0:15:12 Right.
0:15:25 And I think now, since we’re reaching the ends of our distributions for some of these things, so like we’re basically at the point where a lot of the users can’t tell the difference between a gut response from a college student versus a professor.
0:15:28 What we are now impressed by is like, oh, but look, they just wrote this essay.
0:15:28 So that’s impressive.
0:15:30 So that must be smart.
0:15:30 Right.
0:15:39 And I think now, as we see these models, it seems like the general consensus response gets more impressed by almost operational things.
0:15:43 Like write me an essay and put it in this file format and make it nice with pretty images.
0:15:43 Right.
0:15:49 Like that’s not necessarily like deep intelligence to create images, but it’s like in the process of building it, it’s like, oh, this is cool.
0:15:51 This is like convenient.
0:15:51 Right.
0:15:53 And that makes it feel like a best.
0:15:55 And most people, they saw that, they’re going to think it’s a better model.
0:15:55 Right.
0:15:55 Exactly.
0:15:56 It’s smarter.
0:15:56 It’s like, yeah.
0:15:59 But in this analogy, like that high school biology student could easily do that.
0:16:01 It doesn’t require more intelligence.
0:16:03 It’s more just like the mechanics of how it works.
0:16:03 Yeah.
0:16:04 Knowing how to put everything together.
0:16:04 Exactly right.
0:16:05 Yeah.
0:16:07 So are you still like really optimistic on like AI?
0:16:13 Do you think after GPT-5 came out, did it change anything in your mindset or like we’re still just thinking it’ll keep getting better and better?
0:16:21 I think it was a reality check that it’s not going to like be this sudden, you know, super intelligence knows everything can do all these like crazy things.
0:16:24 But it’s more like a general like kind of mind space.
0:16:30 It’s been a reminder that intelligence is not always like in your face obvious.
0:16:33 Because GPT-5 is significantly better than GPT-4 obviously.
0:16:33 Yeah.
0:16:45 But sometimes it shows itself in subtle ways in terms of how it deals with context of like enormous code bases or how when you give it like hundreds of files to deal with and it needs to synthesize two new files or make these modifications.
0:16:49 Can it handle that with all of those in context or is it going to forget some of them?
0:16:53 And I think those are the nuances where you really see the performance improvements emerge.
0:16:54 So I think that’s very interesting.
0:17:07 I think also there’s now it feels like we’re moving towards a little bit of a shift in the model development side where really RL is where a lot of that alpha is going to be found in these new models.
0:17:16 As well as some of the stuff that we just mentioned about like the convenience things or the presentation things or like the kind of little in-betweens that are not that like intellectually hard to do.
0:17:19 But make it more of a convenient setup for like output.
0:17:26 Yeah, it feels like the in-between things and the techniques that’s where OpenAI and Anthropic are really, you know, the leaders right now.
0:17:32 Like because XAI, they’re like catching up in computer, maybe like maybe they have the most computer, I’m not sure the exact numbers, I know it’s always changing.
0:17:35 But even though they have more computers, it still feels like GPT-5 is better.
0:17:39 Is that mainly just because like they have better techniques, you think? Or like, what do you think?
0:17:49 Yeah, I think it’s probably a mix of that. They probably spend more time on some of like the coding RL environments, but in terms of velocity and acceleration of model quality, Grok is ridiculous.
0:17:49 It is, yeah.
0:18:01 It’s spinning up the data center so quickly and like the Delta that they’ve had, like from like the Grok series of models in a vacuum, if the other model providers didn’t exist, this would be like the craziest thing ever.
0:18:07 It’s just because there are kind of so many models everywhere, it’s hard to realize the pace that they’ve had, so it’s…
0:18:12 Yeah, and Elon Musk is also kind of controversial right now, so like some people like love Grok and some people absolutely hate it.
0:18:31 I was actually telling Darren, our producer earlier, I was like, yeah, I think that might be like the surprise in the next three years is like, could be OpenAI and Anthropic, it also could be, oh, actually it’s XAI and Google, that could be the ones because they’re, especially because they’re all, they both also have more access to compute, but they also have more access to real data, which I think long term is going to be
0:18:34 really important for like robotics, and you start feeding that data back into the models, they get smarter.
0:18:40 What I think is interesting with XAI is they have a lot of access to data, but not coding data yet.
0:18:48 Now, a couple years ago, you could have said, great, they have smart people, but they don’t have access to data centers and GPUs, and you can see how quickly they solve that problem.
0:18:50 So, I think a looming thing is, well, they don’t have access to all the coding.
0:18:53 So, why do they not have access to it, but like Anthropic and OpenAI?
0:18:57 I mean, OpenAI probably through Microsoft, is that maybe GitHub?
0:19:04 So, Anthropic does, because they’re like, enterprise is huge for them, and they’re one of the most commonly used like APIs for coding.
0:19:04 Yeah.
0:19:07 OpenAI, as well, has a pretty sizable chunk.
0:19:10 On a relative basis, they’re more consumer, but they do have a very large enterprise chunk.
0:19:16 XAI, just because it’s a new player, there aren’t really enterprises who are like XAI shops, right?
0:19:18 Like, that just doesn’t really exist.
0:19:24 But again, a couple years ago, someone could have said that about them and GPUs and data centers, and you’ve seen how quickly they’ve addressed that.
0:19:30 So, I think what I’d be looking out on the XAI side is what they do to get that enterprise data.
0:19:32 Not even just coding, but just generally.
0:19:41 Because like, they’re not, as it stands today, you know, mid-2025, they’re not one of the most used APIs in the enterprise, but we’ll see how they solve it.
0:19:47 Okay, let me tell you about another podcast I know you’re going to love.
0:19:51 It’s called Billion Dollar Moves, hosted by Sarah Chin Spellings.
0:19:56 It’s brought to you by the HubSpot Podcast Network, the audio destination for business professionals.
0:20:06 Join venture capitalist and strategist Sarah Chin Spellings as she asks the hard questions and learns through the triumphs, failures, and hard lessons of the creme de la creme.
0:20:11 So, you too can make billion-dollar moves in venture, in business, and in life.
0:20:18 She just did a great episode called The Purpose-Driven Power Player Behind Silicon Valley’s Quiet Money with Mike Anders.
0:20:21 Listen to Billion Dollar Moves wherever you get your podcasts.
0:20:32 I do wonder if people are underestimating with XAI, like, the fact that Elon Musk has access to more power, like, in terms of, like, batteries and things like that.
0:20:37 Like, I’ve been looking at possibly building a data center with some partners.
0:20:37 Nice.
0:20:41 Like, looking at, like, should we build or acquire, or, like, what’s the strategy there?
0:20:48 And when you look at building, okay, not even talking about, like, building the building, but, like, all the equipment for, like, power and things like that, supplies, battery.
0:20:51 Some of this stuff, you’re, like, waiting, like, five years right now, like, the wait times.
0:20:51 Yeah.
0:20:59 And so, I do wonder if people are not realizing, like, he possibly could get around that with, like, all the different technologies, you know, he has access to his batteries from Tesla, etc.
0:21:00 Totally.
0:21:02 That seems like a gigantic advantage.
0:21:04 Maybe OpenAI will have partners that will help them get around that.
0:21:13 There’s one thing that time has shown is that Elon will use his advantages to whatever leverage he can to surpass competitors.
0:21:19 And he’s really good in the physical world, which actually starts to become a physical world thing, because how fast can you build the data center?
0:21:22 Because if it is coming down to data centers, that’s what’s better.
0:21:27 Especially given the exponential nature here, like, whatever is the bottleneck, you know, it’s kind of like whack-a-mole.
0:21:29 Or, okay, this thing, you know, GPUs was the bottleneck.
0:21:30 You buy it from someone else.
0:21:31 Then this becomes the bottleneck.
0:21:37 Eventually, that looks like, you know, potentially energy or, like, potentially owning the, like, physical hardware for these things.
0:21:41 The second it gets into that domain, XAI seems like they’re in a pretty good spot.
0:21:48 In the future, do you think we have, like, data centers that are just, like, 100 floors tall or, like, they’re out in space or, like, just kind of nerding out now?
0:21:48 Yeah, yeah.
0:21:49 I mean, listen, like, where’d this conversation go?
0:21:56 It seems like the incentive is becoming such that the ROI for building new data centers is very high and uncapped, in which case we’re going to keep building it.
0:22:04 Now, what that looks like is, you know, if they’re underground, above ground, stacked, flat, whatever, I think that’s outside of my domain of expertise.
0:22:05 Yeah.
0:22:08 But I don’t see when it, like, the exponential, like, tapers off.
0:22:09 It seems like it keeps going.
0:22:11 There’s going to be something, I’m sure.
0:22:15 Maybe that’s a new, like, algorithmic innovation where it actually, like, you need less or something.
0:22:17 But to me, I don’t…
0:22:19 Like, on device or something where you don’t need it, like, in the cloud?
0:22:26 Yeah, or maybe it’s something that we were talking about earlier where actually most people are satisfied with that, like, college-level, you know, intelligence.
0:22:29 And those tail ends aren’t as important and aren’t worth the marginal cost.
0:22:30 Right.
0:22:33 But for the enterprise, I think that’ll always be valuable because…
0:22:36 Yeah, it feels like the enterprise will always need the smartest model possible.
0:22:39 But the average person, like, yeah, it helps me create a document or whatever.
0:22:41 Which is actually, it’s funny that you mentioned that.
0:22:47 That reminds me of, like, some of the IDE tools out there or some of the, like, products that are almost…
0:22:48 Windsurf and…
0:22:49 You know, yeah.
0:22:50 …acquired by your competitor, Devin.
0:23:00 Yeah, but, like, so some of these, a kind of a key assumption has been for their, like, self-service plans to subsidize the tokens to bring in users.
0:23:03 And then eventually model costs go down and then now your margin becomes positive.
0:23:05 Okay, so you get the price there and the model goes down and your profit.
0:23:09 And a very interesting phenomenon that has occurred is that the best models are basically the same price.
0:23:14 You know, the best model today, in a year, will go down in price, but then the best model there will be back in the same…
0:23:16 Could also go up in the future as well.
0:23:17 That’s actually something I predicted last year.
0:23:20 I was like, okay, now that 01 is out, that changes everything.
0:23:20 Totally, yeah.
0:23:23 Because, like, yeah, the smartest model, like, why is that going to cost $100?
0:23:25 It could cost $100,000.
0:23:25 Right.
0:23:33 But it seems like regardless, like, even though model costs are going down, it seems like loosely flat is the cost of the best model.
0:23:38 But another phenomenon has also been self-service users want the best model.
0:23:42 So, in theory, it’s like, oh, hey, we’re subsidizing because the model cost is going to go down.
0:23:45 But the demand for the best model has remained the same.
0:23:54 So, it’s an interesting dynamic where you kind of are margin negative under the assumption that, okay, our margins will get better once the underlying LLM cost goes down.
0:23:56 But the users keep wanting the best model.
0:24:06 Yeah, it’s hard to switch, too, because, like, I noticed, you know, I’ve, like, signed up for the premium models of all these, like, you know, OpenAI, Anthropic, XAI, you know, all of them, Google’s.
0:24:11 I’ve tried all of them, like, the $200 to $250 a month plans.
0:24:13 And I’ve canceled them after a while.
0:24:14 I’m like, oh, a new model came out.
0:24:16 And now I’m kind of getting tired of doing it.
0:24:18 So, I’m like, maybe I’ll just stick with OpenAI.
0:24:19 I want to keep paying for it.
0:24:20 But it’s hard to go back.
0:24:24 Like, if you have any, like, really hard problems, like, Pro is so much better.
0:24:25 Yeah, yeah.
0:24:28 And there’s also probably something psychological, like, you want to use the best.
0:24:29 I want the Pro one.
0:24:31 It’s like getting a MacBook Pro versus looking at it.
0:24:31 Yeah, yeah.
0:24:33 So, it’s kind of an interesting dynamic there.
0:24:41 I think a big question is going to be, is there going to be a quality of model where this phenomenon of, like, okay, the best model gets cheaper, but now there’s a new best that’s more expensive.
0:24:45 Does that stop and people are like, actually, I don’t care about the best model?
0:24:46 So far, that hasn’t happened.
0:24:47 And I’m curious.
0:24:49 I think with professionals, it won’t.
0:24:49 Yeah.
0:24:51 Or, like, people at the highest levels, it won’t.
0:24:51 That’s what we’ve seen.
0:24:59 But maybe, like, average people, like, I would, yeah, like, I’m sure at some point, like, probably soon, probably, like, in two or three years, like, what else are they going to need?
0:25:01 Besides, like, it gets better at certain different things.
0:25:03 Like you said, creating a document or creating an email.
0:25:08 Okay, it got better at these different little use cases, but the intelligence of the model won’t have to get better.
0:25:10 I agree, but also human nature seems such that well.
0:25:12 We’ll just, like, keep filling it in.
0:25:20 Yeah, like, even now, like, some people will say, like, if you showed anyone from 10 years ago the models that we have now, they would, like, say this is AGI unequivocally.
0:25:20 Yeah.
0:25:23 Meanwhile, now the goalposts have shifted so much because it’s like, oh, but it keeps…
0:25:25 Now, if you say AGI, you’re like, oh, I’m like, you know…
0:25:26 What’s the definition?
0:25:31 Yeah, I’ve stopped tweeting as much because, like, I was, like, one of the guys, like, yeah, this is basically AGI.
0:25:32 And they’re like, you’re insane.
0:25:37 And I’m like, yeah, back five years ago, like, people in Silicon Valley would have said this is AGI or close.
0:25:37 Yeah.
0:25:39 I’m sure it’s not perfect at everything.
0:25:39 It’s not a human.
0:25:43 It doesn’t necessarily, you know, AGI doesn’t necessarily mean it’s, like, human level at everything.
0:25:45 But it’s better than humans at many things as of right now.
0:25:52 And I think also we as humans generally respond to, like, sudden change more so than slow change.
0:26:00 And so, like, if you were to show GPT-5 to someone in 2015, that’s a sudden change that’s like, oh, my God, this is the craziest thing ever.
0:26:01 They would be freaking out.
0:26:02 Like, I was here then.
0:26:04 I was in San Francisco then.
0:26:06 It would be like, this is the most insane thing ever.
0:26:11 Yeah, I remember, but then I started looking into AI, like, I bought NVIDIA stock, which I would have bought more.
0:26:17 But I bought NVIDIA stock and kind of like, oh, okay, so AI is going to be huge in the future and NVIDIA is going to be powering that.
0:26:23 We notice the, like, we notice the, like, large deltas of, like, whatever, 2015, no good models to, like, whatever, if we showed someone GPT-5.
0:26:28 Whereas GPT-4 to 5, it’s, like, really good to, like, also insanely good.
0:26:30 It feels like, oh, barely it changed, whatever.
0:26:35 And so I think it’s an interesting aspect of human nature where we like the big deltas.
0:26:39 And we can’t notice, like, the small ones that have been kind of adding up, I guess.
0:26:45 So one last thing I wanted to ask you about is, you know, obviously there was a lot of news around Devon acquiring Windsurf.
0:26:48 And obviously, you know, people compare you to Devon’s.
0:26:50 I would love to hear your thoughts on that deal.
0:26:52 Maybe, I don’t know how much you want to talk about it.
0:26:55 But, like, if someone was considering, like, you know, should they pick Devon or Factory?
0:26:58 Like, what’s the main differences as of right now?
0:27:03 Yeah, I mean, I think, first of all, a huge amount of respect for both the team at Cognition and the team at Windsurf.
0:27:10 I think that acquisition was great, especially for the folks that were left behind from the Google deal.
0:27:15 So I’m glad that there was kind of a setup there that hopefully worked for the people there.
0:27:23 I think the biggest difference is it shows that they’re kind of taking this approach of, like, really anchoring back to the IDE.
0:27:30 Whereas what we’re focused on is a little bit, like, kind of forward-looking of, okay, the IDE has been the optimal UI for the last 15 years.
0:27:39 Let’s now move to this future, which is agent-native, and meet people in the IDE, but kind of not have IDE as the fundamental primitive.
0:27:47 And so, from my perspective is, like, on the product side, it seems like it’s a little bit, like, almost anchoring to what was.
0:27:52 And our point is a little bit more, like, we want to make sure we meet developers where they are.
0:27:54 We are IDE agnostic.
0:28:00 You can delegate to droids in whatever IDE, whatever terminal, Emacs, Vim, whatever you want, you can delegate to droids.
0:28:09 But I think the end goal and what we want to help developers find is that the new optimal UI and the new interaction pattern is going to look very different.
0:28:14 You’re not going to have, like, this single-file viewer because you’re not going to be looking at lines of code.
0:28:18 I saw someone talking about, like, what does cursor 2.0 look like?
0:28:19 Or what’s the next cursor?
0:28:20 Like, it’s already here.
0:28:21 It’s, like, factory.
0:28:22 Yeah, and I think that’s right.
0:28:27 You know, we talked about going from that world of developers writing 100% of their code to developers writing none of it.
0:28:34 I think, from our perspective, it’s, I think, and I think I mentioned this last time, but it’s like the Henry Ford quote.
0:28:36 You ask people what they want, they’ll say faster horses.
0:28:39 We’re building the cars, and cars are a transformative technology.
0:28:44 But, importantly, you can’t just pump out the Model T and just throw it on any dirt road.
0:28:46 In fact, cities have to change a little bit.
0:28:55 Like, in the world where there were horses, and even San Francisco, by the way, there are some great YouTube videos, like, archive YouTube videos of, like, San Francisco in the early 1900s.
0:28:57 They’re, like, on Market Street, which is right next to us.
0:28:59 There are, like, horses going by.
0:29:00 There are, like, stables.
0:29:06 Like, the whole city dynamic is very different when you have these living beings that are, like, your means of transport.
0:29:11 And turning San Francisco into what it is now, you have to pave the roads.
0:29:13 You have to build gas stations.
0:29:25 Like, these are things that are slightly different, and we’re seeing this with the enterprises as well, where if you want to just throw in agents into your engineering org, and you don’t do any of this, like, road paving or creating of gas stations, they’re not going to succeed.
0:29:29 You’re going to have this problem of generating a ton of code, but then some human has to review it all.
0:29:45 Whereas if instead you do kind of the upfront work of making sure you have really good CICD and pre-commit hooks and monitoring and observability and, like, comprehensive testing, then you set up these agents for success such that they can actually, you know, go and drive and not, you know, break down every once in a while.
0:29:50 And so, you know, from our perspective, it’s kind of like, are you building for the world of horses or cars?
0:29:52 And that’s kind of a thing that we ask ourselves a lot.
0:29:59 And from my perspective, kind of going, like, anchoring back onto the IDE feels a little bit like building for the horses again.
0:30:03 And, you know, from our perspective, it’s more we need to pave the roads.
0:30:06 Let’s make sure that the roads, you know, don’t hurt a horse’s hooves, let’s say.
0:30:10 Like, you can still use it, but we’re still, we’re going in that direction towards the future.
0:30:11 I think that’s really important for us.
0:30:14 Yeah, I probably should disclose that I invested in factory.
0:30:18 And part of the reason, too, is, like, I feel like, you know, yeah, you guys are creating the cars.
0:30:22 Like, yeah, cursor’s great now, but where they’re going in the future, you’re already there.
0:30:25 So I think by the time they realize that, I think you’ll be way ahead.
0:30:26 It’s my personal thing.
0:30:31 So if someone’s listening to this episode right now, like, how can they, you know, get some value out of factory?
0:30:32 How can they try it out?
0:30:35 And what would be, like, the best thing that they could just try and kind of start to understand?
0:30:42 Yeah, I mean, I think for the non-technical folks, just go to factory.ai, you can sign up for a free trial and try out factory.
0:30:45 You know, you can build something zero to one, like what we did earlier.
0:30:53 For more technical folks, I think, you know, whatever environment I challenge, you know, whatever niche environment that you think maybe won’t support agents,
0:31:00 chances are, you know, factory will be able to support that, whether it’s through the terminal, through your IDE, whatever your development workflow is,
0:31:04 go in factory.ai, sign up, try out the droids.
0:31:13 And I think very importantly, it’s kind of on everyone right now to take a sledgehammer to every workflow that you find holy
0:31:19 and try to rebuild it with AI to see if you can come up with new efficiencies and new kind of speed-ups.
0:31:21 Yeah, the technology is amazing.
0:31:23 It feels like a lot of people are just trying to do small, little iterative changes.
0:31:25 Like, it’s capable of so much more, I feel like.
0:31:31 And as a very habit-oriented person, I constantly need to remind myself, like, everything is changing so quickly.
0:31:35 There are probably much more efficient ways to do the things that I’m currently doing.
0:31:40 And so it’s like, you need to be comfortable with the uncomfortableness of, I know I’ve done it this way for, you know, 10 years.
0:31:41 Right.
0:31:44 I’m going to break that down and, like, kind of re-derive what is the new way I want to be.
0:31:45 It’s really hard for humans, though, right?
0:31:46 Yes.
0:31:49 We have all these habits and it’s more efficient for our brain to have the habits.
0:31:53 I literally eat the same thing every morning for breakfast, for lunch, and for dinner.
0:31:57 I’m the most habit-oriented person, but I think even still, it’s really important to do this.
0:31:58 Matan, it’s been awesome.
0:32:00 It’s been great doing it in person versus just online.
0:32:00 And, yeah.
0:32:01 Thanks for having me.
0:00:11 Today, we’re doing a live episode here at HubSpot Inbound in San Francisco,
0:00:13 and I have with me a repeat guest,
0:00:15 Matan Grinberg from Factory,
0:00:18 who the last time he was on here was our biggest episode ever,
0:00:21 where he did a live demo where he showed how
0:00:25 he could clone DocuSign in 20 minutes using Factory.
0:00:26 I think you’re going to love the conversation,
0:00:28 and so let’s just get right into it.
0:00:37 Using only 20% of your business data
0:00:40 is like dating someone who only texts emojis.
0:00:42 First of all, that’s annoying.
0:00:45 And second, you’re missing a lot of context.
0:00:47 But that’s how most businesses operate today,
0:00:50 using only 20% of their data.
0:00:52 Unless you have HubSpot, where all the emails,
0:00:56 call logs, and chat messages turn into insights to grow your business.
0:00:59 Because all that data makes all the difference.
0:01:05 Learn more at HubSpot dot com.
0:01:07 Matan, welcome back to the Next Wave.
0:01:08 Matan Grinberg: Thank you for having me.
0:01:09 It’s great to be here in person.
0:01:11 Matan Grinberg: Yeah, it’s awesome meeting you in person.
0:01:13 We’ve talked online so much, and you know,
0:01:15 the first episode that you did on the Next Wave
0:01:17 was actually our biggest episode ever.
0:01:19 I think we got like 67,000 views on it.
0:01:21 It was a slightly controversial episode,
0:01:23 because we were talking about cloning DocuSign.
0:01:25 And people were like, “Did you really clone all of DocuSign?”
0:01:26 There’s like one part of it.
0:01:27 It was like, “Yeah, it was one part of it.”
0:01:29 I still think it was a really cool demo.
0:01:31 And since then, you’ve raised a Series B, is that correct?
0:01:32 That’s right, that’s right.
0:01:34 Yeah, so we just raised our Series B.
0:01:39 we raised $50 million with some great folks like NEA, Sequoia,
0:01:41 JP Morgan, NVIDIA.
0:01:42 I’m very excited about it.
0:01:43 Matan Grinberg: That’s amazing.
0:01:46 So, NVIDIA, that seems like one of the best partners you could have,
0:01:50 and obviously NEA leading the round, and Sequoia.
0:01:52 What do you plan to do with the $50 million?
0:01:52 That’s a lot of money here.
0:01:56 Matan Grinberg: I think for us, the big thing is really building out a team
0:02:01 and growing the resources to reach more of the large enterprises
0:02:02 that we’ve been working with so far.
0:02:05 So, some of our large customers include Ernst & Young,
0:02:09 the large accounting firm, as well as MongoDB,
0:02:11 Bayer, and quite a few others.
0:02:17 And the reality is, to serve these very massive enterprises
0:02:19 and to help them with this behavior change
0:02:21 and this transformation of software development,
0:02:23 it actually requires a lot of resources.
0:02:27 Generally, if you have an org of 50,000 engineers,
0:02:30 maybe 5,000 of them are super excited to adopt
0:02:32 the coolest and latest AI technology.
0:02:36 And then maybe the 45,000 others are like, “You know what?
0:02:39 I’m actually happy in my VIM or in my Emacs and I don’t want to mess around
0:02:42 with any of this stuff or kind of change what I’ve been doing.”
0:02:45 And so, we want to invest in the resources to both build out our engineering
0:02:49 and product team so that we can make it that much easier to adopt.
0:02:54 And then also, our go-to-market team to reach developers and reach these organizations
0:03:00 to kind of raise awareness and make them understand this transformation that’s happening.
0:03:05 And, you know, going out in the field and helping these developers kind of adopt these new workflows.
0:03:08 Yeah, so I was wondering, like, how much do you have to help these companies?
0:03:11 Like, I would imagine that you show them Factory, you show them a demo,
0:03:14 but then, like, how do they actually use it and how do you get the whole team using it?
0:03:16 How much are you having to hold people’s hands?
0:03:19 Great question. It’s crazy how much it depends.
0:03:24 One thing that is for sure is just at the end of the day, developers are human.
0:03:27 And if you show them something that they used to do that was very painful,
0:03:32 that’s now going to be a lot less painful with Factory, generally, they’re going to start adopting it.
0:03:37 Now, really, the work is just what is the thing from their seat that’s really painful,
0:03:41 and how can we use Factory and the droids to accelerate that?
0:03:46 Something that often comes up is migrations or modernizations or refactors.
0:03:49 There are very few developers who actually enjoy doing things like that.
0:03:49 Right.
0:03:52 And so, kind of going in and getting a migration done with the droids
0:03:56 is kind of a quick way to get them to adopt to this new workflow.
0:03:59 And with your GA release, like, how do you feel like that went?
0:04:02 Like, I feel like you got a lot of attention.
0:04:04 Something I heard from people is like, they’re kind of confused.
0:04:07 Like, you know, is this for, like, vibe coders or is this for, like, large enterprises?
0:04:09 Or, like, what’s the focus?
0:04:11 Yeah, that’s a great question.
0:04:15 And I think, so, in May, we released Factory for GA.
0:04:16 It was a really great release.
0:04:21 We got tens of thousands of new users, but I think the important thing was we released Factory,
0:04:25 and basically the main way you accessed it at the time was via the web.
0:04:25 Right.
0:04:32 And something that we’ve learned is that, you know, in this transformation to what we’re calling agent-native development,
0:04:38 the reality is you need multiple surfaces where you can engage with these agents, with these droids.
0:04:41 And we only gave people one surface, which was the web.
0:04:47 That’s why, in this release, we’re really having a surface-agnostic or interface-agnostic approach.
0:04:53 In other words, if you’re going to be willing to change your behavior to become agent-native,
0:05:01 we need to meet you wherever you are, whether that’s in the web, or in your terminal, or in your IDE, or in Slack, or in Jira, or in Linear.
0:05:11 Wherever you might have an idea of, here’s a task I want to delegate to an agent, we need to meet you where you are, and make it very easy for you to do so.
0:05:20 Just like if we were co-workers, and you were like, oh man, here’s a perfect task for Matan, you’d know you can’t only tell me that via email or via Slack.
0:05:27 You could tag me in a ticket, you could, you know, send it to me on Slack, you could send me a text, like, kind of wherever is easiest for you personally, you’d be able to do that.
0:05:32 And we need to provide people kind of a similar ease of use for the droids.
0:05:38 And so that’s like, if you’re going to like a huge organization, enterprise, every single individual engineer gets to do that.
0:05:42 So like, they get to pick like, oh, I want to use it, I think you have a terminal thing coming out as well, right?
0:05:42 Yes, that’s right.
0:05:48 So like, whether I’m doing it web-based, or terminal, or I’m an executive in Slack, however you want to use it, you can.
0:05:49 Is that kind of correct?
0:05:53 Yes, pending, like, approval from IT to, you know, assuming they give access to these.
0:06:01 But basically, if you provide access, you can do it via Slack, via, honestly via email, like, you can do it obviously in the terminal.
0:06:07 But the point is to kind of mold to whatever, you know, the developer’s workflow looks like, to make it seamless.
0:06:16 Because the reality is we’re going from a world where developers wrote 100% of their code, to a world where developers will write 0% of their code.
0:06:20 And this new primitive in this world is going to be a delegation.
0:06:24 The primitive used to be writing a line of code, or writing a function, or writing a file.
0:06:25 It’s now going to be a delegation.
0:06:32 And so to kind of bring people into this new world with this new primitive, we need to make that delegation as easy and accessible as possible.
0:06:37 So people like the Primogen, people like that who love, like, almost like the art of coding.
0:06:42 You think that’s almost going to be like a niche thing in the future where it’s like, oh yeah, it’s kind of like have you have vinyl records or something like that?
0:06:43 You know, where it’s…
0:06:53 To a certain extent, maybe like the line by line might be, but there’s kind of a new art that emerges, which is, okay, great, you can delegate to an agent.
0:06:55 How much can you parallelize?
0:06:59 So similar, there’s kind of an analogy, you know, NVIDIA is one of our investors.
0:07:03 We love using this analogy between CPUs and GPUs.
0:07:03 Right.
0:07:06 Where CPUs, they’re all about processing in series.
0:07:11 You want to get as many computations as you can done in series in some amount of time.
0:07:13 Whereas the big unlock from GPUs was parallelization.
0:07:14 Right.
0:07:19 It turns out for certain tasks like matrix multiplication, you can actually parallelize a lot of the computation.
0:07:21 So that instead of like one, you know, operation at a time.
0:07:22 You can do them all at the same time.
0:07:29 Yeah, you can do four or ten or a hundred and the more you can parallelize, the faster the downstream computation.
0:07:39 It kind of emerges a similar thing with agents, which is if you can parallelize and have 20 droids working at the same time or 100 droids working at the same time, then you’re way faster.
0:07:42 There’s kind of an art of like, how do you actually parallelize to 20 agents at a time?
0:07:46 That’s kind of difficult, but I think that’s the new thing that will emerge.
0:07:52 And there was recently this study that came out from MIT, talking about like 95% of AI pilots have been failing.
0:07:52 Yes.
0:07:58 I was really, you know, I know you guys are having like a competition about this, like a man versus machine kind of competition upcoming.
0:07:59 Maybe it’s already live.
0:08:00 It’s actually this Saturday.
0:08:00 Okay.
0:08:01 It’ll be live by the time.
0:08:02 Yeah, so it’ll already happen.
0:08:05 So you can check in and see the results, see who won.
0:08:06 I’m kind of betting on machine.
0:08:07 Likewise.
0:08:08 Yeah.
0:08:09 So what are your thoughts on that study?
0:08:15 And like, you know, some people were saying like, oh, that’s, you know, AI has been completely overhyped and this is proof of it.
0:08:25 Yeah, I actually think it’s a really good wake up call that the way people have been measuring success in these deployments has been flawed.
0:08:28 So maybe stepping back, what did the study say?
0:08:37 I believe the specific phrasing was 95% of AI adoption efforts in the enterprise were not deemed successful.
0:08:41 And so that’s like across, not just software development, but AI across, you know, every vertical in the enterprise.
0:08:46 And I think the reason is because success criteria has been very ill defined.
0:08:55 For software development in particular, the metrics that people rely on are lines of code generated or self-reported developer productivity.
0:08:59 The problem is software development is fundamentally a pipeline.
0:09:02 And one step of that pipeline is implementing the code.
0:09:03 Right.
0:09:10 But the reality is before that there’s kind of understanding and planning, then there’s the coding itself, then there’s testing and reviewing and documenting.
0:09:20 And if you have like a literal pipe and you expand the like radius of one part of that pipe, but do nothing to the other parts, you actually don’t increase your throughput at all.
0:09:20 Right.
0:09:29 And so it’s kind of no surprise to me that if you’re just measuring things like lines of code generated, well, you give everyone lines of code generators, you’re going to have more lines of code.
0:09:33 But you’re not going to actually have success and like, you’re not going to start shipping software faster.
0:09:35 Yeah, what are those lines of code doing as well, right?
0:09:39 Like you could do 10,000 lines of code that are like, you know, unnecessary.
0:09:42 And someone has to actually go and figure out what are these lines of code doing.
0:09:47 And so now there are more people going and reviewing code or adding tests to that code or adding docs.
0:09:56 And it actually ends up not increasing the speed, which is why like this kind of problem that this MIT study pointed out is why when we do our deployments,
0:10:02 we’re really focused on really what are the business like ROI metrics that matter.
0:10:03 It’s not lines of code generated.
0:10:06 No business is succeeding or failing because of lines of code generated.
0:10:06 Right.
0:10:10 Like they’re succeeding or failing because we missed shipping this feature.
0:10:15 And then our competitor got it out and we had users churn and now they’re going to that tool instead of us.
0:10:19 And it’s whatever causing like those are the things at the business level that matter.
0:10:27 And that’s why we are focused when we do our deployments on let’s get this modernization done faster because we can’t ship this new feature until we modernize our database from whatever.
0:10:29 And you’re focused on like business outcomes as well, right?
0:10:30 Exactly right.
0:10:35 It’s like when I heard that study, actually, I thought about you guys because I was like, that’s actually something that’d be really great for factory.
0:10:40 Because like you guys are actually being hands on with, you know, enterprises, learning how they’re using these tools,
0:10:43 helping them like figure out the best way to actually get a great business outcome.
0:10:44 Yeah.
0:10:47 Whereas like cursor and the other one, there’s like, here you go.
0:10:49 Here’s this, you know, cool AI coding tool.
0:10:54 I think most of the tools in this space are focused on the individual and optimizing the individual.
0:10:58 Factory is focused on optimizing the team and the organization.
0:11:04 And surprise, software development is a team sport and it really matters to optimize for those team outcomes.
0:11:06 We have like a Michael Jordan quote in our office.
0:11:13 And I think he’s kind of a good example of someone who learned that you can optimize like the individual stats of a player.
0:11:17 But if you don’t end up winning the championship at the end of the season, no one cares, no one remembers.
0:11:19 All right. So it’s very similar to like writing more lines of code.
0:11:21 Exactly right. Exactly right.
0:11:29 Okay, let’s be honest. Your AI prompts aren’t giving you the results you deserve.
0:11:38 But with a little coaching, you can transform from basic prompts to engineering conversations that get you exactly what you want from ChatGPT.
0:11:47 That’s what this playbook delivers, not just random prompts, but a step-by-step system with the exact techniques top AI professionals use every day.
0:11:52 You’ll have your own personalized prompt engineering system that delivers consistent results.
0:11:57 Get it right now. Scan the QR code or click the link in the description below.
0:11:58 Now let’s get back to the show.
0:12:07 You know, when GPT-5 came out recently, you know, a lot of people were saying it’s kind of similar to the MIT study.
0:12:14 They were like, you know, AI was way overhyped and now like all, you know, we’ve hit a wall and there’s gonna be like an AI winter again and all this kind of stuff.
0:12:20 Like, what are your thoughts on it? Because I thought GPT-5 was pretty impressive, like not like revolutionary, but like a great step forward.
0:12:24 Things just kind of keep going forward and getting better and better. Like, what are your thoughts?
0:12:34 Yeah. So we work closely with the team at OpenAI. So we actually tested internally GPT-5 and it was sufficiently good that we changed it to our default model for like the self-service plan for factory.
0:12:41 But I also think basically there were really high expectations. They thought it was going to be the same jump as like three to four would be four to five.
0:12:49 Right. But I think the reality is a good way to think about these models is kind of like, I don’t know, imagine some subject that you’re not a deep expert in.
0:12:54 My go-to example here is like biology. I studied physics. I wasn’t that deep in biology.
0:13:00 How easy is it for you to tell the difference between someone who studied biology in high school versus college?
0:13:04 Pretty easy. Maybe on the order of a few seconds, you can tell the difference.
0:13:09 What about the difference between someone who studied it in undergrad versus PhD? Maybe it takes you a couple of minutes.
0:13:17 What about like a postdoc versus a professor? If you don’t know anything about biology, it’ll actually take you kind of a long time to figure out that difference.
0:13:20 I think it’s kind of similar as these models get better and better.
0:13:30 For most generic tasks, GPT-4 and 5 will perform just as well because like, you know, shocker to implement like, I don’t know, quick sort, GPT-4 and 5 will be able to do it.
0:13:39 Now, obviously, as the context gets bigger and like the problem gets more niche, or really, as you’re getting to like the tail ends of the distributions of these models and their performance,
0:13:41 that’s when you’ll notice the difference.
0:13:45 Only a few people are aware of those differences at those tail ends.
0:13:51 So like for the generic user or for someone kind of outside of domain, it’s really hard to tell the difference.
0:13:51 Yeah.
0:13:57 And I think having the expectations that there’s going to be some obvious like step function difference between these models,
0:14:05 it’s like assuming that as a non-expert in biology, you notice the difference between like a professor who’s been a professor for one year versus a 20-year one.
0:14:09 When it’s like, it’s outside of your domain, you might just not be able to notice quickly.
0:14:18 Yeah, it feels like ever since ’01 really, to be honest, like when ’01 first came out, I was actually doing a live episode with Matt Wolff, my co-host in Boston.
0:14:18 Yeah.
0:14:27 And it was right when ’01 came out, and I was like, this is huge, like there’s a new paradigm that’s been unlocked with reasoning, and now you can just throw more compute at the models and they’ll get smarter and smarter.
0:14:32 And a lot of people were just like, we’re not getting that, and they were like, oh, this is like, it takes longer and like the answer is slightly better.
0:14:37 I was like, it made me realize, like there’s already a thing where I think most people can’t recognize how much better the models are getting, right?
0:14:38 Yeah.
0:14:44 And just like what you were explaining, and it’s going to become more and more important for them to actually like see what the end result is.
0:14:57 Yeah. And actually, I think ’01’ is a good example where, in this analogy of like whatever high school, college, it’s kind of like the non-reasoning models are like a gut reaction from like whatever, professor, postdoc, college student.
0:15:01 And then the reasoning models are like, have them go and sit down and do something for a day.
0:15:04 And we end up being impressed more by that.
0:15:11 In the analogy, it’s like having a college student go and write you an essay on some topic in biology versus ask a professor for like a gut response on something.
0:15:12 Right.
0:15:25 And I think now, since we’re reaching the ends of our distributions for some of these things, so like we’re basically at the point where a lot of the users can’t tell the difference between a gut response from a college student versus a professor.
0:15:28 What we are now impressed by is like, oh, but look, they just wrote this essay.
0:15:28 So that’s impressive.
0:15:30 So that must be smart.
0:15:30 Right.
0:15:39 And I think now, as we see these models, it seems like the general consensus response gets more impressed by almost operational things.
0:15:43 Like write me an essay and put it in this file format and make it nice with pretty images.
0:15:43 Right.
0:15:49 Like that’s not necessarily like deep intelligence to create images, but it’s like in the process of building it, it’s like, oh, this is cool.
0:15:51 This is like convenient.
0:15:51 Right.
0:15:53 And that makes it feel like a best.
0:15:55 And most people, they saw that, they’re going to think it’s a better model.
0:15:55 Right.
0:15:55 Exactly.
0:15:56 It’s smarter.
0:15:56 It’s like, yeah.
0:15:59 But in this analogy, like that high school biology student could easily do that.
0:16:01 It doesn’t require more intelligence.
0:16:03 It’s more just like the mechanics of how it works.
0:16:03 Yeah.
0:16:04 Knowing how to put everything together.
0:16:04 Exactly right.
0:16:05 Yeah.
0:16:07 So are you still like really optimistic on like AI?
0:16:13 Do you think after GPT-5 came out, did it change anything in your mindset or like we’re still just thinking it’ll keep getting better and better?
0:16:21 I think it was a reality check that it’s not going to like be this sudden, you know, super intelligence knows everything can do all these like crazy things.
0:16:24 But it’s more like a general like kind of mind space.
0:16:30 It’s been a reminder that intelligence is not always like in your face obvious.
0:16:33 Because GPT-5 is significantly better than GPT-4 obviously.
0:16:33 Yeah.
0:16:45 But sometimes it shows itself in subtle ways in terms of how it deals with context of like enormous code bases or how when you give it like hundreds of files to deal with and it needs to synthesize two new files or make these modifications.
0:16:49 Can it handle that with all of those in context or is it going to forget some of them?
0:16:53 And I think those are the nuances where you really see the performance improvements emerge.
0:16:54 So I think that’s very interesting.
0:17:07 I think also there’s now it feels like we’re moving towards a little bit of a shift in the model development side where really RL is where a lot of that alpha is going to be found in these new models.
0:17:16 As well as some of the stuff that we just mentioned about like the convenience things or the presentation things or like the kind of little in-betweens that are not that like intellectually hard to do.
0:17:19 But make it more of a convenient setup for like output.
0:17:26 Yeah, it feels like the in-between things and the techniques that’s where OpenAI and Anthropic are really, you know, the leaders right now.
0:17:32 Like because XAI, they’re like catching up in computer, maybe like maybe they have the most computer, I’m not sure the exact numbers, I know it’s always changing.
0:17:35 But even though they have more computers, it still feels like GPT-5 is better.
0:17:39 Is that mainly just because like they have better techniques, you think? Or like, what do you think?
0:17:49 Yeah, I think it’s probably a mix of that. They probably spend more time on some of like the coding RL environments, but in terms of velocity and acceleration of model quality, Grok is ridiculous.
0:17:49 It is, yeah.
0:18:01 It’s spinning up the data center so quickly and like the Delta that they’ve had, like from like the Grok series of models in a vacuum, if the other model providers didn’t exist, this would be like the craziest thing ever.
0:18:07 It’s just because there are kind of so many models everywhere, it’s hard to realize the pace that they’ve had, so it’s…
0:18:12 Yeah, and Elon Musk is also kind of controversial right now, so like some people like love Grok and some people absolutely hate it.
0:18:31 I was actually telling Darren, our producer earlier, I was like, yeah, I think that might be like the surprise in the next three years is like, could be OpenAI and Anthropic, it also could be, oh, actually it’s XAI and Google, that could be the ones because they’re, especially because they’re all, they both also have more access to compute, but they also have more access to real data, which I think long term is going to be
0:18:34 really important for like robotics, and you start feeding that data back into the models, they get smarter.
0:18:40 What I think is interesting with XAI is they have a lot of access to data, but not coding data yet.
0:18:48 Now, a couple years ago, you could have said, great, they have smart people, but they don’t have access to data centers and GPUs, and you can see how quickly they solve that problem.
0:18:50 So, I think a looming thing is, well, they don’t have access to all the coding.
0:18:53 So, why do they not have access to it, but like Anthropic and OpenAI?
0:18:57 I mean, OpenAI probably through Microsoft, is that maybe GitHub?
0:19:04 So, Anthropic does, because they’re like, enterprise is huge for them, and they’re one of the most commonly used like APIs for coding.
0:19:04 Yeah.
0:19:07 OpenAI, as well, has a pretty sizable chunk.
0:19:10 On a relative basis, they’re more consumer, but they do have a very large enterprise chunk.
0:19:16 XAI, just because it’s a new player, there aren’t really enterprises who are like XAI shops, right?
0:19:18 Like, that just doesn’t really exist.
0:19:24 But again, a couple years ago, someone could have said that about them and GPUs and data centers, and you’ve seen how quickly they’ve addressed that.
0:19:30 So, I think what I’d be looking out on the XAI side is what they do to get that enterprise data.
0:19:32 Not even just coding, but just generally.
0:19:41 Because like, they’re not, as it stands today, you know, mid-2025, they’re not one of the most used APIs in the enterprise, but we’ll see how they solve it.
0:19:47 Okay, let me tell you about another podcast I know you’re going to love.
0:19:51 It’s called Billion Dollar Moves, hosted by Sarah Chin Spellings.
0:19:56 It’s brought to you by the HubSpot Podcast Network, the audio destination for business professionals.
0:20:06 Join venture capitalist and strategist Sarah Chin Spellings as she asks the hard questions and learns through the triumphs, failures, and hard lessons of the creme de la creme.
0:20:11 So, you too can make billion-dollar moves in venture, in business, and in life.
0:20:18 She just did a great episode called The Purpose-Driven Power Player Behind Silicon Valley’s Quiet Money with Mike Anders.
0:20:21 Listen to Billion Dollar Moves wherever you get your podcasts.
0:20:32 I do wonder if people are underestimating with XAI, like, the fact that Elon Musk has access to more power, like, in terms of, like, batteries and things like that.
0:20:37 Like, I’ve been looking at possibly building a data center with some partners.
0:20:37 Nice.
0:20:41 Like, looking at, like, should we build or acquire, or, like, what’s the strategy there?
0:20:48 And when you look at building, okay, not even talking about, like, building the building, but, like, all the equipment for, like, power and things like that, supplies, battery.
0:20:51 Some of this stuff, you’re, like, waiting, like, five years right now, like, the wait times.
0:20:51 Yeah.
0:20:59 And so, I do wonder if people are not realizing, like, he possibly could get around that with, like, all the different technologies, you know, he has access to his batteries from Tesla, etc.
0:21:00 Totally.
0:21:02 That seems like a gigantic advantage.
0:21:04 Maybe OpenAI will have partners that will help them get around that.
0:21:13 There’s one thing that time has shown is that Elon will use his advantages to whatever leverage he can to surpass competitors.
0:21:19 And he’s really good in the physical world, which actually starts to become a physical world thing, because how fast can you build the data center?
0:21:22 Because if it is coming down to data centers, that’s what’s better.
0:21:27 Especially given the exponential nature here, like, whatever is the bottleneck, you know, it’s kind of like whack-a-mole.
0:21:29 Or, okay, this thing, you know, GPUs was the bottleneck.
0:21:30 You buy it from someone else.
0:21:31 Then this becomes the bottleneck.
0:21:37 Eventually, that looks like, you know, potentially energy or, like, potentially owning the, like, physical hardware for these things.
0:21:41 The second it gets into that domain, XAI seems like they’re in a pretty good spot.
0:21:48 In the future, do you think we have, like, data centers that are just, like, 100 floors tall or, like, they’re out in space or, like, just kind of nerding out now?
0:21:48 Yeah, yeah.
0:21:49 I mean, listen, like, where’d this conversation go?
0:21:56 It seems like the incentive is becoming such that the ROI for building new data centers is very high and uncapped, in which case we’re going to keep building it.
0:22:04 Now, what that looks like is, you know, if they’re underground, above ground, stacked, flat, whatever, I think that’s outside of my domain of expertise.
0:22:05 Yeah.
0:22:08 But I don’t see when it, like, the exponential, like, tapers off.
0:22:09 It seems like it keeps going.
0:22:11 There’s going to be something, I’m sure.
0:22:15 Maybe that’s a new, like, algorithmic innovation where it actually, like, you need less or something.
0:22:17 But to me, I don’t…
0:22:19 Like, on device or something where you don’t need it, like, in the cloud?
0:22:26 Yeah, or maybe it’s something that we were talking about earlier where actually most people are satisfied with that, like, college-level, you know, intelligence.
0:22:29 And those tail ends aren’t as important and aren’t worth the marginal cost.
0:22:30 Right.
0:22:33 But for the enterprise, I think that’ll always be valuable because…
0:22:36 Yeah, it feels like the enterprise will always need the smartest model possible.
0:22:39 But the average person, like, yeah, it helps me create a document or whatever.
0:22:41 Which is actually, it’s funny that you mentioned that.
0:22:47 That reminds me of, like, some of the IDE tools out there or some of the, like, products that are almost…
0:22:48 Windsurf and…
0:22:49 You know, yeah.
0:22:50 …acquired by your competitor, Devin.
0:23:00 Yeah, but, like, so some of these, a kind of a key assumption has been for their, like, self-service plans to subsidize the tokens to bring in users.
0:23:03 And then eventually model costs go down and then now your margin becomes positive.
0:23:05 Okay, so you get the price there and the model goes down and your profit.
0:23:09 And a very interesting phenomenon that has occurred is that the best models are basically the same price.
0:23:14 You know, the best model today, in a year, will go down in price, but then the best model there will be back in the same…
0:23:16 Could also go up in the future as well.
0:23:17 That’s actually something I predicted last year.
0:23:20 I was like, okay, now that 01 is out, that changes everything.
0:23:20 Totally, yeah.
0:23:23 Because, like, yeah, the smartest model, like, why is that going to cost $100?
0:23:25 It could cost $100,000.
0:23:25 Right.
0:23:33 But it seems like regardless, like, even though model costs are going down, it seems like loosely flat is the cost of the best model.
0:23:38 But another phenomenon has also been self-service users want the best model.
0:23:42 So, in theory, it’s like, oh, hey, we’re subsidizing because the model cost is going to go down.
0:23:45 But the demand for the best model has remained the same.
0:23:54 So, it’s an interesting dynamic where you kind of are margin negative under the assumption that, okay, our margins will get better once the underlying LLM cost goes down.
0:23:56 But the users keep wanting the best model.
0:24:06 Yeah, it’s hard to switch, too, because, like, I noticed, you know, I’ve, like, signed up for the premium models of all these, like, you know, OpenAI, Anthropic, XAI, you know, all of them, Google’s.
0:24:11 I’ve tried all of them, like, the $200 to $250 a month plans.
0:24:13 And I’ve canceled them after a while.
0:24:14 I’m like, oh, a new model came out.
0:24:16 And now I’m kind of getting tired of doing it.
0:24:18 So, I’m like, maybe I’ll just stick with OpenAI.
0:24:19 I want to keep paying for it.
0:24:20 But it’s hard to go back.
0:24:24 Like, if you have any, like, really hard problems, like, Pro is so much better.
0:24:25 Yeah, yeah.
0:24:28 And there’s also probably something psychological, like, you want to use the best.
0:24:29 I want the Pro one.
0:24:31 It’s like getting a MacBook Pro versus looking at it.
0:24:31 Yeah, yeah.
0:24:33 So, it’s kind of an interesting dynamic there.
0:24:41 I think a big question is going to be, is there going to be a quality of model where this phenomenon of, like, okay, the best model gets cheaper, but now there’s a new best that’s more expensive.
0:24:45 Does that stop and people are like, actually, I don’t care about the best model?
0:24:46 So far, that hasn’t happened.
0:24:47 And I’m curious.
0:24:49 I think with professionals, it won’t.
0:24:49 Yeah.
0:24:51 Or, like, people at the highest levels, it won’t.
0:24:51 That’s what we’ve seen.
0:24:59 But maybe, like, average people, like, I would, yeah, like, I’m sure at some point, like, probably soon, probably, like, in two or three years, like, what else are they going to need?
0:25:01 Besides, like, it gets better at certain different things.
0:25:03 Like you said, creating a document or creating an email.
0:25:08 Okay, it got better at these different little use cases, but the intelligence of the model won’t have to get better.
0:25:10 I agree, but also human nature seems such that well.
0:25:12 We’ll just, like, keep filling it in.
0:25:20 Yeah, like, even now, like, some people will say, like, if you showed anyone from 10 years ago the models that we have now, they would, like, say this is AGI unequivocally.
0:25:20 Yeah.
0:25:23 Meanwhile, now the goalposts have shifted so much because it’s like, oh, but it keeps…
0:25:25 Now, if you say AGI, you’re like, oh, I’m like, you know…
0:25:26 What’s the definition?
0:25:31 Yeah, I’ve stopped tweeting as much because, like, I was, like, one of the guys, like, yeah, this is basically AGI.
0:25:32 And they’re like, you’re insane.
0:25:37 And I’m like, yeah, back five years ago, like, people in Silicon Valley would have said this is AGI or close.
0:25:37 Yeah.
0:25:39 I’m sure it’s not perfect at everything.
0:25:39 It’s not a human.
0:25:43 It doesn’t necessarily, you know, AGI doesn’t necessarily mean it’s, like, human level at everything.
0:25:45 But it’s better than humans at many things as of right now.
0:25:52 And I think also we as humans generally respond to, like, sudden change more so than slow change.
0:26:00 And so, like, if you were to show GPT-5 to someone in 2015, that’s a sudden change that’s like, oh, my God, this is the craziest thing ever.
0:26:01 They would be freaking out.
0:26:02 Like, I was here then.
0:26:04 I was in San Francisco then.
0:26:06 It would be like, this is the most insane thing ever.
0:26:11 Yeah, I remember, but then I started looking into AI, like, I bought NVIDIA stock, which I would have bought more.
0:26:17 But I bought NVIDIA stock and kind of like, oh, okay, so AI is going to be huge in the future and NVIDIA is going to be powering that.
0:26:23 We notice the, like, we notice the, like, large deltas of, like, whatever, 2015, no good models to, like, whatever, if we showed someone GPT-5.
0:26:28 Whereas GPT-4 to 5, it’s, like, really good to, like, also insanely good.
0:26:30 It feels like, oh, barely it changed, whatever.
0:26:35 And so I think it’s an interesting aspect of human nature where we like the big deltas.
0:26:39 And we can’t notice, like, the small ones that have been kind of adding up, I guess.
0:26:45 So one last thing I wanted to ask you about is, you know, obviously there was a lot of news around Devon acquiring Windsurf.
0:26:48 And obviously, you know, people compare you to Devon’s.
0:26:50 I would love to hear your thoughts on that deal.
0:26:52 Maybe, I don’t know how much you want to talk about it.
0:26:55 But, like, if someone was considering, like, you know, should they pick Devon or Factory?
0:26:58 Like, what’s the main differences as of right now?
0:27:03 Yeah, I mean, I think, first of all, a huge amount of respect for both the team at Cognition and the team at Windsurf.
0:27:10 I think that acquisition was great, especially for the folks that were left behind from the Google deal.
0:27:15 So I’m glad that there was kind of a setup there that hopefully worked for the people there.
0:27:23 I think the biggest difference is it shows that they’re kind of taking this approach of, like, really anchoring back to the IDE.
0:27:30 Whereas what we’re focused on is a little bit, like, kind of forward-looking of, okay, the IDE has been the optimal UI for the last 15 years.
0:27:39 Let’s now move to this future, which is agent-native, and meet people in the IDE, but kind of not have IDE as the fundamental primitive.
0:27:47 And so, from my perspective is, like, on the product side, it seems like it’s a little bit, like, almost anchoring to what was.
0:27:52 And our point is a little bit more, like, we want to make sure we meet developers where they are.
0:27:54 We are IDE agnostic.
0:28:00 You can delegate to droids in whatever IDE, whatever terminal, Emacs, Vim, whatever you want, you can delegate to droids.
0:28:09 But I think the end goal and what we want to help developers find is that the new optimal UI and the new interaction pattern is going to look very different.
0:28:14 You’re not going to have, like, this single-file viewer because you’re not going to be looking at lines of code.
0:28:18 I saw someone talking about, like, what does cursor 2.0 look like?
0:28:19 Or what’s the next cursor?
0:28:20 Like, it’s already here.
0:28:21 It’s, like, factory.
0:28:22 Yeah, and I think that’s right.
0:28:27 You know, we talked about going from that world of developers writing 100% of their code to developers writing none of it.
0:28:34 I think, from our perspective, it’s, I think, and I think I mentioned this last time, but it’s like the Henry Ford quote.
0:28:36 You ask people what they want, they’ll say faster horses.
0:28:39 We’re building the cars, and cars are a transformative technology.
0:28:44 But, importantly, you can’t just pump out the Model T and just throw it on any dirt road.
0:28:46 In fact, cities have to change a little bit.
0:28:55 Like, in the world where there were horses, and even San Francisco, by the way, there are some great YouTube videos, like, archive YouTube videos of, like, San Francisco in the early 1900s.
0:28:57 They’re, like, on Market Street, which is right next to us.
0:28:59 There are, like, horses going by.
0:29:00 There are, like, stables.
0:29:06 Like, the whole city dynamic is very different when you have these living beings that are, like, your means of transport.
0:29:11 And turning San Francisco into what it is now, you have to pave the roads.
0:29:13 You have to build gas stations.
0:29:25 Like, these are things that are slightly different, and we’re seeing this with the enterprises as well, where if you want to just throw in agents into your engineering org, and you don’t do any of this, like, road paving or creating of gas stations, they’re not going to succeed.
0:29:29 You’re going to have this problem of generating a ton of code, but then some human has to review it all.
0:29:45 Whereas if instead you do kind of the upfront work of making sure you have really good CICD and pre-commit hooks and monitoring and observability and, like, comprehensive testing, then you set up these agents for success such that they can actually, you know, go and drive and not, you know, break down every once in a while.
0:29:50 And so, you know, from our perspective, it’s kind of like, are you building for the world of horses or cars?
0:29:52 And that’s kind of a thing that we ask ourselves a lot.
0:29:59 And from my perspective, kind of going, like, anchoring back onto the IDE feels a little bit like building for the horses again.
0:30:03 And, you know, from our perspective, it’s more we need to pave the roads.
0:30:06 Let’s make sure that the roads, you know, don’t hurt a horse’s hooves, let’s say.
0:30:10 Like, you can still use it, but we’re still, we’re going in that direction towards the future.
0:30:11 I think that’s really important for us.
0:30:14 Yeah, I probably should disclose that I invested in factory.
0:30:18 And part of the reason, too, is, like, I feel like, you know, yeah, you guys are creating the cars.
0:30:22 Like, yeah, cursor’s great now, but where they’re going in the future, you’re already there.
0:30:25 So I think by the time they realize that, I think you’ll be way ahead.
0:30:26 It’s my personal thing.
0:30:31 So if someone’s listening to this episode right now, like, how can they, you know, get some value out of factory?
0:30:32 How can they try it out?
0:30:35 And what would be, like, the best thing that they could just try and kind of start to understand?
0:30:42 Yeah, I mean, I think for the non-technical folks, just go to factory.ai, you can sign up for a free trial and try out factory.
0:30:45 You know, you can build something zero to one, like what we did earlier.
0:30:53 For more technical folks, I think, you know, whatever environment I challenge, you know, whatever niche environment that you think maybe won’t support agents,
0:31:00 chances are, you know, factory will be able to support that, whether it’s through the terminal, through your IDE, whatever your development workflow is,
0:31:04 go in factory.ai, sign up, try out the droids.
0:31:13 And I think very importantly, it’s kind of on everyone right now to take a sledgehammer to every workflow that you find holy
0:31:19 and try to rebuild it with AI to see if you can come up with new efficiencies and new kind of speed-ups.
0:31:21 Yeah, the technology is amazing.
0:31:23 It feels like a lot of people are just trying to do small, little iterative changes.
0:31:25 Like, it’s capable of so much more, I feel like.
0:31:31 And as a very habit-oriented person, I constantly need to remind myself, like, everything is changing so quickly.
0:31:35 There are probably much more efficient ways to do the things that I’m currently doing.
0:31:40 And so it’s like, you need to be comfortable with the uncomfortableness of, I know I’ve done it this way for, you know, 10 years.
0:31:41 Right.
0:31:44 I’m going to break that down and, like, kind of re-derive what is the new way I want to be.
0:31:45 It’s really hard for humans, though, right?
0:31:46 Yes.
0:31:49 We have all these habits and it’s more efficient for our brain to have the habits.
0:31:53 I literally eat the same thing every morning for breakfast, for lunch, and for dinner.
0:31:57 I’m the most habit-oriented person, but I think even still, it’s really important to do this.
0:31:58 Matan, it’s been awesome.
0:32:00 It’s been great doing it in person versus just online.
0:32:00 And, yeah.
0:32:01 Thanks for having me.
Want to implement AI agents like $50M startups do? Get our ultimate guide: https://clickhubspot.com/fcv
Episode 80: Are coders really being replaced by AI agents, or is this just the next tech hype cycle? Nathan Lands (https://x.com/NathanLands) is joined by repeat guest Matan Grinberg (https://x.com/matansf), co-founder of Factory—an agent-native software development platform backed by NEA, Sequoia, JP Morgan, and Nvidia.
This episode dives deep into Factory’s ambitious mission to transform software engineering by enabling developers—and entire organizations—to delegate painful, repetitive coding tasks to “droids,” Factory’s intelligent agents. Matan shares strategies for helping massive enterprises adopt new workflows, how Factory’s platform is built for surface/interface agnosticism (terminal, IDE, Slack, and more), and why optimization for teams—not individuals—will define the future of AI-powered development. Plus, debate about GPT-5’s impact, the myth of “AI winters,” and what the real business ROI of AI looks like in the enterprise.
Check out The Next Wave YouTube Channel if you want to see Matt and Nathan on screen: https://lnk.to/thenextwavepd
—
Show Notes:
-
(00:00) Scaling Teams to Empower Enterprises
-
(03:54) Agent Native, Surface Agnostic Approach
-
(09:07) Prioritizing Business ROI Over Code
-
(12:10) Assessing Expertise Levels Quickly
-
(16:01) AI Model Nuances and RL Shift
-
(18:26) AI Enterprise Market Dynamics
-
(22:41) Choosing AI Subscription Plans
-
(25:43) Future-Focused, IDE-Agnostic Development
-
(27:30) Adapting Cities and Enterprises
-
(30:11) Embracing Change and Growth
—
Mentions:
-
HubSpot Inbound: https://www.inbound.com/
-
Matan Grinberg: https://www.linkedin.com/in/matan-grinberg
-
Factory: https://factory.ai/
-
Docusign: https://www.docusign.com/
-
Nvidia: https://www.nvidia.com/
-
Anthropic: https://www.anthropic.com/
-
Cursor: https://cursor.com/
Get the guide to build your own Custom GPT: https://clickhubspot.com/tnw
—
Check Out Matt’s Stuff:
• Future Tools – https://futuretools.beehiiv.com/
• Blog – https://www.mattwolfe.com/
• YouTube- https://www.youtube.com/@mreflow
—
Check Out Nathan’s Stuff:
-
Newsletter: https://news.lore.com/
-
Blog – https://lore.com/
The Next Wave is a HubSpot Original Podcast // Brought to you by Hubspot Media // Production by Darren Clarke // Editing by Ezra Bakker Trupiano
Leave a Reply
You must be logged in to post a comment.