WEBVTT
00:00:00.000 --> 00:00:09.439
It allows you to turn these SOPs, these standard operating procedures that you might have in your company, and put that into kind of an automated flow.
00:00:16.399 --> 00:00:23.280
Welcome to Partner Path, a podcast that unpacks the venture capital and growth equity ecosystem from a junior perspective.
00:00:23.519 --> 00:00:29.839
Young entrepreneurs and investors have already had a massive impact on the industry, having started unicorns and launched billion-dollar funds.
00:00:30.239 --> 00:00:36.560
We discuss these success stories and more by sharing perspectives and advice from some of the industry's most prominent role models.
00:00:36.719 --> 00:00:43.119
This podcast expresses our views as of the date published and does not represent the views of and is not endorsed by any company for which we work.
00:00:45.119 --> 00:00:50.159
Before we get to the episode, Laura and I are excited to announce a partnership with Overlap.
00:00:50.640 --> 00:00:59.920
Fresh out of Y Combinator's Summer Batch, Overlap is an AI-driven app that uses large language models to curate the best moments from podcast episodes.
00:01:00.479 --> 00:01:12.400
Imagine having a smart assistant who reads through every podcast transcript, finds the best parts or parts most relevant to your search, and strings them together to form a new curated stream of content.
00:01:12.640 --> 00:01:14.560
That is what Overlap does.
00:01:14.959 --> 00:01:22.879
I personally use Overlap to explore content on future podcast guests and highly recommend it as another diligence tool for investors.
00:01:23.280 --> 00:01:25.599
Check out Overlap on the App Store today.
00:01:25.920 --> 00:01:30.239
Okay, today we are chatting with Robert Chandler, co-founder and CTO at Wordware.
00:01:30.400 --> 00:01:35.200
Robert started his career as a software engineer at 5AI before founding Hey Daily.
00:01:35.519 --> 00:01:40.480
In early 2023, Robert founded Wordware, a toolkit for building AI agents.
00:01:40.719 --> 00:01:41.519
Welcome, Robert.
00:01:41.840 --> 00:01:43.040
Great to be here, Laura.
00:01:43.359 --> 00:01:53.519
I don't have a deep understanding of your background, so I'd love if you could go back to the early days of entrepreneurship and maybe when your itch to start something began, since this is not your first rodeo.
00:01:53.840 --> 00:01:54.480
Wow, yeah.
00:01:54.640 --> 00:02:02.079
I mean, truly, I think the start of my entrepreneurial journey was back at school and I used to like bake and sell brownies on the school bus.
00:02:02.319 --> 00:02:03.200
You probably didn't even know that.
00:02:03.280 --> 00:02:04.719
That's uh this is like going way back.
00:02:04.879 --> 00:02:06.319
Not a venture scale business.
00:02:06.560 --> 00:02:12.719
Then God, I I got onto the show called The Young Apprentice, or nearly got onto the show called The Young Apprentice.
00:02:12.800 --> 00:02:15.680
So you know Donald Trump was the the apprentice.
00:02:15.840 --> 00:02:17.919
In the UK, I think it probably started there.
00:02:18.080 --> 00:02:23.759
We had a guy called Lord Sugar, and he also did a young apprentice, which was people like 16 to 18.
00:02:23.840 --> 00:02:28.639
And I got to the final 20 of 10,000 people to be on that show, which was it's kind of a rogue anecdote.
00:02:28.719 --> 00:02:29.039
That's true.
00:02:29.199 --> 00:02:30.960
Did that motivate you to do something else?
00:02:31.199 --> 00:02:32.639
It made me even more hungry.
00:02:32.719 --> 00:02:38.479
I think uh yeah, I wanted like always always been an entrepreneur, always wanted to like make something, sell something.
00:02:38.560 --> 00:02:45.199
And I guess I didn't go straight into founding a company because I got to join a really interesting startup straight out of university working on self-driving cars.
00:02:45.439 --> 00:02:50.080
One of the founders was the former CEO of ARM, and between them they'd sold like a billion dollars of businesses.
00:02:50.319 --> 00:02:54.639
And I was like, these are the people to learn from to learn how to build like an incredible company.
00:02:54.800 --> 00:02:59.759
There were 10 people when I joined, crewed to 200 people, got acquired by Bosch for like 150 million.
00:02:59.919 --> 00:03:04.879
So really awesome, kind of working on the cutting edge of machine learning and kind of learning from these incredible founders.
00:03:05.199 --> 00:03:07.280
Was where were the self-driving cars at?
00:03:07.840 --> 00:03:10.400
Yeah, they were in mostly like London.
00:03:10.479 --> 00:03:25.439
So we we probably probably still the largest like public trial of self-driving vehicles in the UK, where we actually like ran like a test route between like two places in London.
00:03:25.599 --> 00:03:32.319
And we still had a safety driver, but people from the public got to ride in the self-driving cars and kind of you know, the safety driver would take over occasionally.
00:03:32.479 --> 00:03:33.840
But interesting.
00:03:34.000 --> 00:03:37.599
So, how does the self-driving car scene compare to San Francisco?
00:03:37.840 --> 00:03:41.120
Oh, well, like there are actually self-driving cars in San Francisco.
00:03:41.199 --> 00:03:42.080
You've got like Waymo.
00:03:42.479 --> 00:03:45.360
It's interesting because the tech is basically the same.
00:03:45.599 --> 00:03:50.719
It's just Google has spent$20 billion getting to the point where it like worked as well as it does.
00:03:51.039 --> 00:03:54.560
We could talk about self-driving cars for like the entire entire podcast.
00:03:54.719 --> 00:03:57.599
I think it's super cool that you can ride in a self-driving car in San Francisco.
00:03:57.759 --> 00:04:06.319
Do I think it's gonna scale to be better than Uber in terms of like cost per mile?
00:04:06.960 --> 00:04:08.080
Not soon.
00:04:08.319 --> 00:04:09.919
Um yes, I agree.
00:04:10.000 --> 00:04:11.759
I think one, that's for a different episode.
00:04:11.840 --> 00:04:21.199
And two, I think the infrastructure of San Francisco, as well as how the road, how everything is laid out, and I think the population density makes sense here, but you it would be hard to replicate in New York.
00:04:21.519 --> 00:04:22.319
Yeah, for sure.
00:04:22.480 --> 00:04:29.120
And and London, especially where it's like rainy and like dark, and people like cross the tiny, narrow, winding roads.
00:04:29.199 --> 00:04:29.839
It's yeah.
00:04:30.160 --> 00:04:30.720
Yeah.
00:04:31.040 --> 00:04:37.279
So then fast forward, you wanted to start something and you did, and you started a company called Hay Daily.
00:04:37.759 --> 00:04:38.000
Yeah.
00:04:38.160 --> 00:04:42.720
So actually, I met my co-founder 10 years ago at Cambridge where we were starting studying machine learning.
00:04:42.959 --> 00:04:53.680
We, you know, built the drone society, we broke into parties, we uh like started a dating app as you do whilst you're in your university, solve your own problems.
00:04:54.000 --> 00:04:59.360
And I think every every every founder has to start a dating app at some point.
00:04:59.600 --> 00:05:01.920
And uh, so we've been friends for a very long time.
00:05:02.079 --> 00:05:07.040
He actually left university, started his first company in San Francisco using GPT2 and BERT.
00:05:07.279 --> 00:05:13.360
I kind of joined this self-driving car startup, but we'd always like kind of knew that at some point we might work together.
00:05:13.519 --> 00:05:18.639
And so when he he kind of he'd so he built a company using GPT 2, the tech wasn't quite ready for what he was doing.
00:05:18.720 --> 00:05:22.639
It's like augmenting human memory, a lot like these pendant devices that we're seeing all over the place nowadays.
00:05:22.959 --> 00:05:23.519
Rabbit.
00:05:24.000 --> 00:05:35.279
Not like Rabbit, like rewind or limitless, like friend, like the other one that was called friend and yeah, no, tab that turned into friend or like no longer your friend, an unfriend blocked.
00:05:35.360 --> 00:05:39.279
Uh and uh yeah, so that the GPT 2 wasn't quite ready for that.
00:05:39.439 --> 00:05:44.399
So he pivoted into like tech remote teams, and that was just before COVID hit, so had much more success there.
00:05:44.560 --> 00:05:48.879
Took a partial exit and actually kind of explored for a little while.
00:05:49.120 --> 00:05:52.959
Gave me a call three years ago, said, Hey, come be my co-founder, I've got an idea.
00:05:53.199 --> 00:05:56.639
I was like, hell yeah, this is like someone that I've wanted to work with for a long time.
00:05:56.720 --> 00:05:59.680
We were just about to get acquired by Bosch, so the timing was ideal.
00:05:59.839 --> 00:06:09.600
I did have to leave behind a bunch of equity, but I knew like missing the opportunity to work with like a very close friend and someone who I trust immensely would have been like too big an opportunity cost to miss.
00:06:09.759 --> 00:06:12.480
So we started, we were both a little bit burned by deep tech.
00:06:12.720 --> 00:06:20.800
Like, you know, there were in netter in San Francisco, there's no self-driving cars driving around at that point in was it 2021, late end of 2021.
00:06:21.040 --> 00:06:23.839
GPT-3 hadn't really done much beyond GPT 2.
00:06:24.079 --> 00:06:28.240
It was still like, okay, like let's do something where we can generate revenue and get fast feedback.
00:06:28.319 --> 00:06:38.399
So we created or started a creator economy business, a subscription service for influencers where you would a lot like Substack or Patreon or OnlyFans deliver exclusive content in return for like a subscription fee.
00:06:38.480 --> 00:06:40.079
But it was through like iMessage and WhatsApp.
00:06:40.240 --> 00:06:41.920
So it was fine.
00:06:42.240 --> 00:06:43.920
It's hard to scale those, right?
00:06:44.240 --> 00:06:45.439
That was one of the big things.
00:06:45.600 --> 00:06:48.000
I think we disproved some hypotheses.
00:06:48.079 --> 00:06:54.079
So we grew it to like 30k MRR, but we disproved some hypotheses that led to basically realizing it wasn't gonna be a billion dollar business.
00:06:54.240 --> 00:07:04.160
There was you you actually the hypotheses meant that you didn't need that big a scale in order to get to like big like revenue.
00:07:04.480 --> 00:07:09.839
You just have to increase your RPU, like your average revenue per customer for each like interaction, probably.
00:07:10.079 --> 00:07:11.199
Yeah, so you see it like sustain.
00:07:11.519 --> 00:07:23.199
I think we worked out that we needed basically the kind of simple maths was if you had a hundred thousand followers and you could convince 1% of them at$10 a month, or to pay you$10 a month, then you could be making$10,000 a month.
00:07:23.360 --> 00:07:24.639
We'd take 25% of that.
00:07:24.800 --> 00:07:28.800
We needed about 3,000 influences at the revenue multiples at the time in order to be a billion dollar business.
00:07:28.879 --> 00:07:31.759
So it's like, okay, this could be like a a sales-led thing.
00:07:31.920 --> 00:07:40.959
And you'd actually like 3,000 influencers, given there's 20 million or so creators with 50,000 followers or more on Instagram alone, was like, it seemed pretty easy.
00:07:41.199 --> 00:07:46.879
Turned out like actually very few creators have a following that are willing to pay the money.
00:07:47.519 --> 00:07:50.160
That's the like the big thing, isn't it?
00:07:50.560 --> 00:07:51.120
Interesting, yeah.
00:07:51.279 --> 00:08:00.560
Nowadays, uh how I view the influencer market is the ones I think with what you alluded to with like 50,000, 100,000 followers, it's almost like there's too many, the degrees of a closeness are too close.
00:08:00.720 --> 00:08:04.160
Yeah, and then you have the influencers who now are almost hitting celebrity level.
00:08:04.240 --> 00:08:04.399
Yeah.
00:08:04.480 --> 00:08:10.879
And so their income streams are probably fine without some of this additional passive income because they already have additional passive income.
00:08:11.199 --> 00:08:18.079
Yeah, and like if you can do a brand deal for like$50,000, why would you create exclusive content and like make less money?
00:08:18.319 --> 00:08:18.639
Yeah.
00:08:19.360 --> 00:08:22.720
And yeah, it's like there's all sorts of different ways to monetize.
00:08:22.879 --> 00:08:28.079
Also, to sign up for a subscription requires like a lot tighter relationship than to sign up for like a one-time purchase.
00:08:28.319 --> 00:08:30.160
So basically, we just proved that idea.
00:08:30.480 --> 00:08:35.840
Realized like there was there was a path to something that could have been more sustainable.
00:08:35.919 --> 00:08:41.279
It was like microcourses delivered through like influencer branded, like a bit like Skillshare, but for like microcourses.
00:08:41.919 --> 00:08:44.000
But we weren't the right people to be building that.
00:08:44.240 --> 00:08:44.480
Okay.
00:08:44.720 --> 00:08:46.320
And AI got super exciting.
00:08:46.480 --> 00:08:50.720
We were using like stable diffusion to reach out to these creators to like make fan art.
00:08:50.879 --> 00:08:53.440
So we were like already exploring with AI.
00:08:53.679 --> 00:08:59.440
And we were like, we should be doing something in AI, not just kind of working with influencers and doing subscription content.
00:08:59.600 --> 00:09:00.559
So that's fair.
00:09:00.720 --> 00:09:08.320
And you also want something that hopefully you can build a billion-dollar business or build something that can like scale to the level that you and your co-founder are hoping to achieve.
00:09:08.639 --> 00:09:08.879
Exactly.
00:09:08.960 --> 00:09:09.279
Yeah, yeah.
00:09:09.440 --> 00:09:12.000
So there was that's perfectly good lifestyle business to be built there.
00:09:12.320 --> 00:09:13.279
That wasn't what we're interested in.
00:09:13.360 --> 00:09:23.200
We were like, we we want to build something huge and something that's like gonna dramatically change the world rather than just be yeah, make a few people a little bit richer.
00:09:23.519 --> 00:09:24.240
That's fair.
00:09:24.559 --> 00:09:27.840
Okay, but then AI in and of itself is quite a broad term.
00:09:27.919 --> 00:09:29.039
And even your backgrounds.
00:09:29.120 --> 00:09:37.200
I mean, your backgrounds are, as you mentioned, more conducive for hardware type build, but that probably didn't make sense for what you were looking to do next.
00:09:37.279 --> 00:09:40.399
So, how did you decide on agenteg use case specifically?
00:09:40.720 --> 00:09:48.320
Yeah, I think the period before we pivoted, so you know, you've got to like make a strong like pivoting company's heart.
00:09:48.399 --> 00:09:51.279
You've got to decide that, okay, what you're working on is not good.
00:09:51.360 --> 00:09:54.879
You've got to decide that, like, are you gonna do something different?
00:09:55.200 --> 00:09:56.559
Are you gonna keep some of the team?
00:09:56.639 --> 00:09:58.080
Are you gonna keep none of the team?
00:09:58.480 --> 00:10:00.000
What are you even gonna work on?
00:10:00.320 --> 00:10:16.720
So there was kind of a phase where we were exploring everything from like working in like foundational image models to, I mean, like we explored like neural radiance fields and synthetic data generation, and even like, can we do video generation and these kind of things?
00:10:16.879 --> 00:10:25.919
And we realized what a lot of people think about LMs, where they think about them as knowledge like engines, like search engines, and like how can you make a better Google?
00:10:26.159 --> 00:10:38.799
So basically the point where we pivoted, which ended up being March 2023, everyone was like doing chat with your docs, or like, can we take in your all your knowledge of your like company's data and then like put it into a chat bot?
00:10:38.960 --> 00:10:44.240
And we were like, okay, that's cool, but what's way more exciting is this is the first thing that we've ever seen that can think.
00:10:44.320 --> 00:10:46.000
So the first thing that came can reason.
00:10:46.159 --> 00:10:51.120
And you know, it's like it's simulated reasoning, it's it's kind of fuzzy and and not amazing, but it's pretty cool.
00:10:51.279 --> 00:10:52.399
And like there's something there.
00:10:52.480 --> 00:10:55.120
So we were looking at this from a reasoning perspective.
00:10:55.279 --> 00:10:56.639
We started like just building.
00:10:56.720 --> 00:11:03.840
So we were like, let's read all the papers, let's read the React paper, let's read the Tor Former paper, let's I guess we had to discover these papers even existed.
00:11:03.919 --> 00:11:11.840
So like let's just like look at the literature, catch up with everything that's happened with LLM since you know attention is all you need all the way up to like the state of the art of the day.
00:11:12.159 --> 00:11:15.759
And we used frameworks like Langchain, we used niche ones like LMQL.
00:11:15.919 --> 00:11:26.320
We started building these things from scratch and really learned what it meant to build an agent and kind of felt like there's gonna be a huge amount of value created here.
00:11:26.399 --> 00:11:28.559
It was like an LLM plus a tool.
00:11:28.799 --> 00:11:30.080
We were like, this is AGI.
00:11:30.159 --> 00:11:41.440
You know, it's weak AGI, it's inefficient AGI, but it's it's the kind of an LLM is the unit of reasoning required to deliver a weak AGI.
00:11:41.679 --> 00:11:45.120
So we were all about like how can we turn that weak AGI into a narrow expert AI?
00:11:45.279 --> 00:11:47.600
And you do that through basically prompting.
00:11:47.840 --> 00:11:56.320
And so um, I think what was most exciting, or like one of the reasons it was really exciting to us is kind of for the last 30 years, software has been about automation.
00:11:56.559 --> 00:11:59.120
But you've only been able to do if-else and structured data.
00:11:59.279 --> 00:12:02.720
For about the last 10 years, you've been able to do machine learning.
00:12:02.960 --> 00:12:07.440
So you can do predictive like analysis, or you can do like classification type tasks.
00:12:07.919 --> 00:12:14.720
And that's really cool and really useful, but it's only works if you can gather enough data.
00:12:14.799 --> 00:12:30.240
So you have to gather like millions of examples of data, you have to annotate them, you have to train one model to do one thing, and at the end of that, you've kind of spent a few million dollars, you've hired some machine learning PhDs, and you've waited six to twelve months to like get some results, um, and you've automated one thing.
00:12:30.399 --> 00:12:34.559
With LLMs, that that activation entry drop dropped through the floor to basically zero.
00:12:34.720 --> 00:12:48.080
And so there was this kind of new wave of software that through prompting alone, through this very low-hanging fruit, where you didn't need to be insanely technical, you could build incredibly powerful automations that previously would have just been impossible or intractable.
00:12:48.399 --> 00:12:51.039
Okay, so then that brings us to agents.
00:12:51.360 --> 00:12:51.759
Agents.
00:12:51.919 --> 00:12:52.240
Yep.
00:12:52.559 --> 00:12:54.960
Now a quick word from our sponsor.
00:12:55.279 --> 00:13:01.600
Laura and I are thrilled to announce a partnership with Grata, an intuitive deal sourcing intelligence platform.
00:13:01.919 --> 00:13:05.919
Keep missing out on private market deals, ditch the Rolodex and get Grata.
00:13:06.399 --> 00:13:14.559
Grata's AI-powered platform provides accurate data, deep insights, and time-saving workflows that help get more deals done.
00:13:14.799 --> 00:13:20.000
And it's trusted by over 500 firms in 3,000 deal makers, including myself.
00:13:20.240 --> 00:13:29.919
I've personally used Grata in both my deal sourcing and diligence workflows and would recommend their tool to any investor across private equity, growth equity, corpdev, or investment banking.
00:13:30.159 --> 00:13:32.960
We've negotiated a special offer for our audience.
00:13:33.200 --> 00:13:42.480
Mention Partner Path to the Grata team during sign-up, and you'll get free executive seats for your partners and MDs included in your subscription.
00:13:42.799 --> 00:13:44.080
Now back to the episode.
00:13:44.399 --> 00:13:52.720
Can you just break down generally agents and how kind of wordware fits into this revolutionizing human productivity?
00:13:53.039 --> 00:13:53.360
Yeah.
00:13:53.600 --> 00:13:59.039
So I think whenever you think about agents, you can basically think about them as a collection of prompts.
00:13:59.200 --> 00:14:01.120
So there's kind of a little bit more to it than that.
00:14:01.279 --> 00:14:06.960
They are kind of an orchestration of prompts, and those prompts can use tools so they can pull in extra information.
00:14:07.200 --> 00:14:16.720
They're also able to, you know, do inter-agent communication, which basically means having the outputs of one prompt feed into the inputs to the next prompt.
00:14:16.879 --> 00:14:20.240
And together you get something that kind of forms this reasoning system.
00:14:20.320 --> 00:14:26.000
Some of that orchestration like gets it to self-reflect, gets it to like think, and it can like look at its own output.
00:14:26.320 --> 00:14:28.080
And really, you can think about it a little bit like this.
00:14:28.240 --> 00:14:34.960
When you use Chat GPT, you're kind of a wizard and you're casting a spell and you're saying, I want you to write me a blog post.
00:14:35.120 --> 00:14:38.320
And if you just put that into ChatGPT, you get a pretty mediocre output.
00:14:38.399 --> 00:14:40.879
And maybe you say, or write me an essay on the Roman Empire.
00:14:40.960 --> 00:14:43.679
You'll get like an okay essay, but it's not gonna be very good.
00:14:43.840 --> 00:14:49.279
But if you're a bit more skilled with ChatGPT, you can be like, okay, I'm gonna write an essay about the Roman Empire.
00:14:49.440 --> 00:15:07.360
First, I'm gonna get paste in some research from Wikipedia about like the some some articles, and then I'm gonna tell ChatGPT to create an outline, and then I'm gonna get it to refine that outline, and then I'm gonna get it to write the introduction, then the conclusion, and then fill in the paragraphs, and maybe I'll get it to proofread and critique, and then I'll get it to rewrite.
00:15:07.440 --> 00:15:09.759
And that's like what a skilled prompt engineer would do.
00:15:09.919 --> 00:15:20.399
And what you can do with Wordware is you can kind of distill that that whole process into a flow, into something that would be like almost like an SOP in a business.
00:15:20.639 --> 00:15:25.279
So it's I would like you you can basically orchestrate this agent as a potion.
00:15:25.360 --> 00:15:34.480
So you've gone from being a wizard casting a spell to a person making a potion, and then you can share that potion with your friends or your colleagues, and suddenly everyone becomes as powerful a wizard as you are.
00:15:34.559 --> 00:15:37.759
So that's a very high-level kind of version of what Wordware is.
00:15:37.919 --> 00:15:48.399
But I guess concretely, it allows you to turn these SOPs, these standard operating procedures that you might have in your company, and put that into kind of an automated flow.
00:15:48.559 --> 00:15:52.960
I think another way that we quite like to think about it is we have this process, this thing that we call the intern test.
00:15:53.120 --> 00:16:00.320
And basically with our customers, we ask them, what is, if you're trying to explain this to an intern, what would what would they do?
00:16:00.480 --> 00:16:05.679
Where would they look for data to answer the kind of question you want them to answer or do the thing you want them to do?
00:16:05.840 --> 00:16:07.200
What are the steps they would take?
00:16:07.279 --> 00:16:11.840
And if you can do that, if you can explain that, then you can probably get an AI to do it for you.
00:16:12.000 --> 00:16:19.679
So you break down the problem into this set of substeps, you then kind of narrow the scope of what this agent can do, but you get it to be much more reliable.
00:16:19.759 --> 00:16:33.200
And that's really where WordWare helps like you craft these chains of prompts, this orchestrate these LLMs and interact with the tools and pull in the right data to build automations that would otherwise previously not have been possible.
00:16:33.519 --> 00:16:35.120
That's a very graceful explanation.
00:16:35.279 --> 00:16:35.919
Thank you.
00:16:36.159 --> 00:16:37.679
And let's put this into practice.
00:16:37.759 --> 00:16:40.480
What are some of the initial use cases you're seeing on Wordware?
00:16:40.799 --> 00:16:42.799
Yeah, I mean, it's it's really, really broad.
00:16:42.879 --> 00:16:52.559
So we have customers that are in finance, we have customers that are in logistics, we have customers that are in a lot of like AI sales-based products.
00:16:52.799 --> 00:16:55.759
But I think it's it really is anything that you could give to an intern.
00:16:55.840 --> 00:17:01.919
It's like if you could give it to an intern, suddenly you can have a thousand interns, and then what use cases are valuable for that?
00:17:02.080 --> 00:17:09.039
And so, like in sales, it's how can I write much better cold email outreach copy that's like really personalized to the end user.
00:17:09.200 --> 00:17:19.119
If you had an intern, you could teach them how to like look at their LinkedIn and pull in the right data and craft a hook and then like write a good email, and that would be intractable to train enough interns that that would be worthwhile.
00:17:19.279 --> 00:17:24.000
But with AI, you train one intern and you have a thousand interns or a million interns and then zero again.
00:17:24.160 --> 00:17:28.480
With like the legal use cases, it's how can I create a trademark application?
00:17:28.559 --> 00:17:32.160
Because that's the kind of thing that actually is pretty much run off the mill.
00:17:32.400 --> 00:17:34.559
And so this is one where it's actually not fully fully automated.
00:17:34.640 --> 00:17:36.559
You you still have the lawyer review it.
00:17:36.720 --> 00:17:39.519
But you know, 8% of the work is kind of grunt work.
00:17:39.599 --> 00:17:55.119
It's okay, I need to describe, take in this company's like landing page and description and turn it into a kind of slightly more legalese description that that covers a broader as much part of their trademark without infringing on other people's trademarks.
00:17:55.200 --> 00:18:13.279
I also need to do some like research on who could they possibly infringing, be infringing, and then compare their description to their other their kind of potential competitors or like infringement descriptions and you know it's basically what are the reasoning processes that would that make up knowledge work?
00:18:13.440 --> 00:18:23.680
I think we're in in a decade where that's kind of like automate or be automated, and a lot of people are building companies that are looking to automate the things that they have domain expertise in.
00:18:23.920 --> 00:18:24.559
I like that.
00:18:24.799 --> 00:18:26.079
Automate or be automated.
00:18:26.480 --> 00:18:33.119
How would wordware differ from like traditional no-code tools or maybe more opinionated customers that actually want to go in?
00:18:33.279 --> 00:18:36.400
They're coding, they're engineers, and we want to make more of the sausage.
00:18:36.720 --> 00:18:41.119
Yeah, so I I guess say traditional code tools or traditional no-code tools.
00:18:41.279 --> 00:18:42.000
Sorry, no code tools.
00:18:42.160 --> 00:18:44.960
Because wordware is, I mean, it's like a like you said, it's a wide audience.
00:18:45.119 --> 00:18:46.480
People like me, people like you.
00:18:46.720 --> 00:18:51.680
We have very different backgrounds, but everyone can use wordware, which is awesome because it's a horizontal AI tool.
00:18:51.920 --> 00:18:52.319
I think so.
00:18:52.400 --> 00:18:54.319
I think you can think of it a little bit like Excel.
00:18:54.480 --> 00:19:02.880
And you know, there's a long tail of use cases in which end users built, like basically programmed using Excel.
00:19:02.960 --> 00:19:04.240
They don't really think about it as programming.
00:19:04.319 --> 00:19:10.000
You're like, sometimes you're just using the sum function, you're adding some data, and that's like a useful tool that people have created for their own use.
00:19:10.160 --> 00:19:13.599
There's the similar kind of thing where, you know, like there's the v lookups and then there's the macros.
00:19:13.759 --> 00:19:19.359
With Wordware, you really have like progressive ability to get more and more complex with like a very high ceiling.
00:19:19.519 --> 00:19:23.039
So I think kind of wordware sits between two ends of the spectrum.
00:19:23.200 --> 00:19:33.920
One end is like fully code-based, so things like Langchain, doing it from scratch, just hitting the APIs directly, where you have to be a developer, you have to have experience writing software, you have to be able to deploy this yourself.
00:19:34.079 --> 00:19:38.720
That's the the big blocker is this very slow feedback cycle between the engineers and the people who know what good looks like.
00:19:38.799 --> 00:19:40.640
And so that's a very limited use case.
00:19:40.799 --> 00:19:48.559
On the other end, you have a lot of these low code tools, so things like Zapier, other tools that are like more AI-first Zapias, or the kind of like flowchart-based things.
00:19:48.720 --> 00:19:50.720
And the problem there is you hit a ceiling very quickly.
00:19:50.799 --> 00:19:53.359
If you want to add any kind of logic, you struggle.
00:19:53.519 --> 00:19:59.519
You can basically do the quite okay linear flows, but as soon as it gets quite complicated, then it's it's kind of game over.
00:19:59.759 --> 00:20:08.240
Wordware sits between the two and it gives you the kind of flexibility of all this, uh, or like the low floor of the no-cools with the high ceiling of software.
00:20:08.319 --> 00:20:19.119
So you can combine the flexibility of natural language with the structured programming and build these like genuinely complex, genuinely useful agents, whether you're technical or non-technical, or you're collaborating between technical and non-technical teams.
00:20:19.200 --> 00:20:22.559
And then you deploy, you can deploy it through an API, or you can trigger it through Zapier.
00:20:22.720 --> 00:20:29.039
And that way you've kind of got this broad base, whether you're trying to build an AI-powered product or you're trying to automate parts of your workflow.
00:20:29.359 --> 00:20:29.680
Okay.
00:20:30.000 --> 00:20:31.279
That makes a ton of sense.
00:20:31.519 --> 00:20:34.880
So right now we are, I don't think we've hit AGI yet.
00:20:34.960 --> 00:20:36.880
And I think that's universally agreed upon.
00:20:37.039 --> 00:20:43.039
And I think, well, depending who you ask, but I think we're really making steps towards automating more and more of the workflow.
00:20:43.119 --> 00:20:48.319
And by using tools like Wordware and other domain-specific tools, where do you see the future of Wordware going?
00:20:48.480 --> 00:20:53.200
You kind of described where it's at right now, which is like amazing, just the progress you've made in under two years.
00:20:53.359 --> 00:20:55.599
But like, what does the future of Wordware look like?
00:20:55.920 --> 00:20:56.319
Yeah.
00:20:56.640 --> 00:21:02.640
I would argue that if you asked someone two years ago, that is like what you're building today AGI?
00:21:02.720 --> 00:21:03.920
Probably the answer would be yes.
00:21:04.160 --> 00:21:04.400
Yeah.
00:21:05.039 --> 00:21:05.680
It's relative.
00:21:05.920 --> 00:21:08.160
I think just like our expectations have got higher.
00:21:08.559 --> 00:21:09.279
Definitions have changed.
00:21:09.440 --> 00:21:10.720
Definitions didn't exist.
00:21:10.960 --> 00:21:12.000
Yeah, yeah, well, exactly.
00:21:12.079 --> 00:21:14.160
It was like, oh we'll we'll kind of know it when we see it.
00:21:14.240 --> 00:21:17.519
It's AGI, it's uh human-level intelligence, whatever that means.
00:21:17.759 --> 00:21:18.079
I like that.
00:21:18.160 --> 00:21:19.039
The bar keeps getting higher.
00:21:19.359 --> 00:21:20.319
The bar gets higher, yeah.
00:21:20.400 --> 00:21:25.200
It's uh yeah, I think it's like as yeah, yeah.
00:21:25.519 --> 00:21:29.440
Anyway, well, where is where we're going is probably the uh the slightly easier question.
00:21:29.519 --> 00:21:32.799
Like the we could also spend an entire podcast just discussing AGI.
00:21:33.039 --> 00:21:33.839
Where is where we're going?
00:21:33.920 --> 00:21:36.559
I think the goal is to be the AI operating system.
00:21:36.960 --> 00:21:44.720
We want to empower teams to augment their themselves.
00:21:44.880 --> 00:22:02.480
We want you know the people to be focusing on the really differentiated work, the stuff that's like uses their special skills of being a human, their experience, their all of that, and then offload the kind of grunt work to AI and that's Actually, really tailored to their taste.
00:22:02.640 --> 00:22:08.720
So I think one of the problems with like a lot of off-the-shelf AI systems is you lose that taste.
00:22:08.880 --> 00:22:17.759
You don't get much control over how good the outputs are and like whether they align exactly with what you want versus what just like the generic optimal thing is.
00:22:17.920 --> 00:22:32.640
And I think as the kind of more and more AI happens, more like there's more and more content being put out there by AI or like work being done by AI, the big differentiator will be like, have you imbued your own companies or your own personal taste on that output?
00:22:32.720 --> 00:22:34.640
It's like we see a lot of content that says delve.
00:22:34.799 --> 00:22:39.440
You might not want your content to say delve and you might want it to be sort of more fun and more upbeat.
00:22:39.599 --> 00:22:52.720
So I think WordWeb becomes this AI operating system by starting with narrow agents that are like highly specialized, highly like reliable, repeatable, very good at doing one thing, and then expanding into being orchestrated by more general agents.
00:22:52.880 --> 00:22:55.440
So for now it's very much like event-based triggering.
00:22:55.519 --> 00:23:00.799
So you know, you receive an email, you run a WordWare agent, that creates a draft response, or maybe even just sends the response.
00:23:00.880 --> 00:23:14.799
You'll start seeing maybe like a slightly higher level agent that will act as your like EA, and then that'll have the ability to like craft, like delegate to a narrow agent to respond to maybe in your case, like a an inbound fundraising email.
00:23:14.880 --> 00:23:19.599
And you'll be like, okay, let's you know do some research on this company, and like what are the things I need to know?
00:23:19.759 --> 00:23:28.640
I need to like find out when it was founded, like what the previous fundraising uh history was, who's backed them, like what are the background of the founders?
00:23:28.799 --> 00:23:29.599
What's the product?
00:23:29.680 --> 00:23:30.960
Like, do I care about this market?
00:23:31.119 --> 00:23:37.039
And like, you know, and then there'll be a bunch of different responses that you'll give based on that, and maybe you'll ask follow-up questions, or maybe you'll be like, let's schedule a meeting.
00:23:37.200 --> 00:23:40.880
I don't like it's all down to you and like what do you do day to day?
00:23:41.119 --> 00:23:41.519
Yeah.
00:23:41.759 --> 00:23:44.480
Yeah, I have a lot of use cases, so that makes sense.
00:23:44.720 --> 00:23:46.319
One, the growth has been phenomenal.
00:23:46.480 --> 00:23:56.160
I actually read that your launch on August 2nd, you grew from 10,000 users on the main product to over 300,000 users today, and I'm sure that number is outdated.
00:23:56.559 --> 00:23:58.160
It's yeah, we're we're growing a lot.
00:23:58.400 --> 00:23:59.279
Yeah, which is amazing.
00:23:59.359 --> 00:24:02.079
And I think it's worth tapping in honestly just to the growth strategy.
00:24:02.160 --> 00:24:21.759
And I know I know you're more on the tech side as I introduced you, your CTO and co-founder, but how important is scaling like quality of users versus like quantity of users, especially given that the use cases are quite dispersed and to burn calories in all these different areas of like helping onboard and get people up to running, you have to know kind of where to burn them.
00:24:22.079 --> 00:24:23.519
No, it's it's it's a massive challenge.
00:24:23.599 --> 00:24:31.039
And you know, actually, as the CTO, I'm the one that ends up like we've got all these requests for different features across like a very broad spectrum.
00:24:31.200 --> 00:24:35.519
It's like, how do we prioritize to make sure enough people that we care about are happy?
00:24:35.599 --> 00:24:36.720
And that's that's really hard.
00:24:36.880 --> 00:24:41.839
I think really and I guess do founders listen to this podcast as well.
00:24:42.000 --> 00:24:42.160
Yes.
00:24:42.720 --> 00:24:43.039
Okay, yeah.
00:24:43.119 --> 00:24:49.119
So this is like a it's one of those things, as a founder, you have to like work out where do you place your bets.
00:24:49.279 --> 00:24:52.319
And actually, when we launched, we weren't expecting to get quite as much traction as we did.
00:24:52.400 --> 00:25:03.440
We thought, you know, we're building something really useful, but we got a lot more prosumer traction than we'd expected, even though we hadn't done anything to like build the onboarding, to build like the kind of B2C type product that you'd expect.
00:25:03.599 --> 00:25:07.839
And so we are now prioritizing more of those features than we were before.
00:25:07.920 --> 00:25:12.720
We were focused largely on B2B, on these scale-ups and startups who are building AI-powered backends.
00:25:12.880 --> 00:25:22.880
We're still, they are still our most important customer, but we are starting to add features and like onboarding and like integrations with Zapia that help the slightly more prosumer users, at least the ones that are really like passionate.
00:25:23.039 --> 00:25:25.279
And over time, we're going to be adding more and more features.
00:25:25.359 --> 00:25:28.559
But I think there's all sorts of things you need to do building a company.
00:25:28.640 --> 00:25:30.960
One is get your name out there so people know you exist.
00:25:31.519 --> 00:25:37.440
And so a lot of what we've been doing has always been well, and actually, we're building a horizontal tool.
00:25:37.599 --> 00:25:42.079
So not only do people need to know we exist, they need to know like what can we do with this tool.
00:25:42.319 --> 00:25:44.160
And so we've always been dog fooding wordware.
00:25:44.240 --> 00:25:58.400
So we've always been using wordware to build fun or useful tools to demonstrate, like, oh, you can build, we have one called Audio Scribe, and it's like you can build a tool, uh a thing with wordware where it takes in a voice note and turns that into like a summary.
00:25:58.480 --> 00:25:59.839
And it'll first classify that voice note.
00:26:00.000 --> 00:26:05.839
So it could be you're making a shopping list or it could be you're making a journal entry, and it has different prompts for those different kind of aspects.
00:26:06.079 --> 00:26:07.759
We've kind of been using Webware.
00:26:08.000 --> 00:26:11.599
We did one that went insanely viral, which was the kind of Twitter roast.
00:26:11.680 --> 00:26:14.160
I think we got like eight million people run that, probably more now.
00:26:14.319 --> 00:26:19.359
Haven't looked at the numbers recently, but it's that was obviously great for getting attention.
00:26:19.599 --> 00:26:25.680
Most of those users are like most of those kind of people who play used the twit Twitter roast are not going to be our ICP.
00:26:25.839 --> 00:26:32.960
Like a small fraction of them then signed up for wordware, and like some of those are our ICP, and some of those are the more prosumer ICP.
00:26:33.039 --> 00:26:37.359
And there's benefits to having like people using webware and playing with webware and giving us feedback.
00:26:37.519 --> 00:26:39.440
There's also a lot of noise to cut through when you have that.
00:26:39.519 --> 00:26:40.880
So it's it's a challenge.
00:26:40.960 --> 00:26:42.079
I don't think there's uh a good answer.
00:26:42.559 --> 00:26:47.119
And hopefully the goal of ProSumer is a lot of it can be self-served, honestly, and people can get up and running.
00:26:47.519 --> 00:26:51.119
Yeah, and what's what's awesome is like it helps test our infrastructure.
00:26:51.200 --> 00:27:04.000
So, you know, having run, I think we ran like six or seven million eight like executions of an agent within a 24-hour period or like a two-week, uh, few day period, and that really tests our infrastructure, and that's good.
00:27:04.079 --> 00:27:13.839
That's good for people putting their trust that actually if we can deal with that level of traffic, then we'll be fine because it's pretty rare to have that scale of just like usage of a product, especially overnight.
00:27:14.000 --> 00:27:17.279
It went from like zero to millions in.
00:27:17.680 --> 00:27:19.599
You have to share your tech stack after this.
00:27:19.839 --> 00:27:21.279
We'll put it in the show notes.
00:27:21.759 --> 00:27:28.799
Well, uh, yeah, we uh we'll plug wordware and the or yeah providers you're using probably more.
00:27:29.039 --> 00:27:33.200
But actually, I want to dig into something you mentioned with users, finding your ICP, cutting through the noise.
00:27:33.359 --> 00:27:35.920
I think it's relevant in the same vein of being in San Francisco.
00:27:36.160 --> 00:27:39.039
And there's so many startups, especially in agent infrastructure right now.
00:27:39.119 --> 00:27:44.400
And honestly, even frameworks or orchestration tools that try and be adjacent but friendly competitors to you guys.
00:27:44.640 --> 00:27:47.279
How do you think about honestly standing out in this space right now?
00:27:47.680 --> 00:27:52.480
Yeah, I think there's a lot of people who have gone prompting is the new programming.
00:27:52.720 --> 00:27:56.319
Let's build the same tools we have for programming, but for prompting.
00:27:56.400 --> 00:27:59.359
Or but and like you know, prompting agents, they're all basic the same thing.
00:27:59.440 --> 00:28:03.119
I think people use the word agent a lot when they really just mean a collection of prompts.
00:28:03.279 --> 00:28:05.759
And it's because it sounds more fancy, it sounds more exotic.
00:28:06.079 --> 00:28:06.880
Prompting for GitHub.
00:28:07.119 --> 00:28:07.599
Prompting, yeah.
00:28:08.000 --> 00:28:09.359
I read that on your YC website.
00:28:09.599 --> 00:28:13.359
That's it's yeah, I mean, yeah, GitHub for agents, I think we probably say somewhere.
00:28:13.759 --> 00:28:14.319
Yeah, that's better.
00:28:14.559 --> 00:28:16.960
And so a lot of people have gone, okay, let's build the same tools.
00:28:17.200 --> 00:28:27.759
They've built the observability layer, or they've gone via the evaluation layer, or they've gone for like the user feedback, or like one of these like niche parts of the stack.
00:28:27.920 --> 00:28:35.440
And the problem there is that they're kind of obvious and they're kind of not useful until you've got something that works well enough.
00:28:35.599 --> 00:28:37.440
And they're also very easily replaceable.
00:28:37.599 --> 00:28:42.160
So you can very easily replace your like your observability layer with another observability layer.
00:28:42.319 --> 00:28:53.759
If it's easy for you, if you've built your tool, your observability layer, such that it's easy to plug in, which you kind of have to for Git adoption, then everyone's gonna build something that's easy to plug in, and therefore you've got no moat, like not a big moat.
00:28:53.920 --> 00:28:57.519
So it's there's a lot of people trying to compete for the same, it's like small subset of users.
00:28:57.599 --> 00:29:03.359
I think the way we took it differently was to say, okay, prompting's the new programming, but from first principles, what should the tools look like?
00:29:03.519 --> 00:29:06.720
And to us that really meant starting with the orchestration layer.
00:29:06.799 --> 00:29:09.440
That is like the most important part of all these agents.
00:29:09.519 --> 00:29:19.200
It's how do you combine the prompts and the tools and the domain experts and the user input into something that can build like genuinely useful AI agents.
00:29:19.359 --> 00:29:22.079
And so we have had to take quite a lot of that programming stack.
00:29:22.160 --> 00:29:38.400
And we so we started with the orchestration, we also did the deployment, so we're a little bit like Vercell in the sense that we we own, we have the layer that is the Next.js that is the wordware like programming language, and then we have the deployment where you like Vercel, where you deploy for Vercell, but you deploy your agent through WordWare.
00:29:38.559 --> 00:29:44.480
We're working on the GitHub for agents as well, so where people can fork and share and we get that network effects.
00:29:44.720 --> 00:29:54.640
I think really think about it from a perspective of like this is something new, let's treat it as something new, rather than like this is something new, let's do the obvious things that will be needed.
00:29:54.799 --> 00:30:02.480
Because if we own the orchestration layer, you can then add in the monitoring and observability and feedback and fine-tuning and all this other stuff.
00:30:02.640 --> 00:30:05.039
But you can't rip out the orchestration layer very easily.
00:30:05.200 --> 00:30:12.799
Like once you built using Langchain or once you built using wordware, it's very difficult to replace your orchestration layer.
00:30:12.960 --> 00:30:18.720
And I think the orchestration layer is honestly the nucleus that feeds into customers' opinions when they say, we just want to work.
00:30:18.799 --> 00:30:19.519
We wanted this to work.
00:30:19.599 --> 00:30:21.759
And yeah, it's that sounds it's easier said than done.
00:30:21.839 --> 00:30:25.279
But it's like once the orchestration layer is working, then it really can only get better.
00:30:25.599 --> 00:30:29.440
Yeah, and I think what matters is that how tight can you make that feedback cycle?
00:30:29.519 --> 00:30:31.119
And we saw this in self-driving cars.
00:30:31.200 --> 00:30:37.359
We saw this like if you can get more information from every drive you take, like, where did I go wrong, even if I didn't have a driver disengagement?
00:30:37.440 --> 00:30:45.119
And if I did have a driving disengagement, how quickly can I work out why the driver had to disengage and like improve that part of my stack?
00:30:45.200 --> 00:30:51.839
And if I do improve part of my stack, how quickly can I get that deployed such that when we can actually tell if the fix has improved it?
00:30:51.920 --> 00:30:55.759
It's it's all about how can you make responses faster.
00:30:55.920 --> 00:31:12.880
And we did that by like putting the domain expert in the driving seat so you can like edit and iterate instantly rather than having this kind of back and forth between the engineer writing the prompts in the code base, showing the results to the PM or the CEO, them going, hmm, it's good on X, it's bad on Y, the vibes are off, an engineer going, I don't like vibes, what am I doing?
00:31:13.039 --> 00:31:15.759
And like trying to optimize towards this fuzzy direction.
00:31:15.920 --> 00:31:18.799
It's like, no, just put the person who knows what good looks like in the driving seat.
00:31:18.880 --> 00:31:24.160
They can iterate like a hundred times in an afternoon and build something genuinely useful much, much faster.
00:31:24.559 --> 00:31:25.680
That makes a ton of sense.
00:31:25.839 --> 00:31:31.680
You probably see yourself partnership opportunities with the other agent infrastructure companies.
00:31:32.559 --> 00:31:34.079
Yes, maybe.
00:31:34.400 --> 00:31:36.480
Um so that's another thing.
00:31:36.559 --> 00:31:39.759
I'm seeing more and more companies go to other parts of the stack.
00:31:40.079 --> 00:31:41.440
Yeah, I mean, they are.
00:31:42.000 --> 00:31:46.079
Like when I look at my market map now, the market map started here and we're slowly converging.
00:31:46.400 --> 00:31:47.359
You you kind of have to.
00:31:47.440 --> 00:31:55.839
If you're not yeah, once you've done monitoring, you might as well do evaluation and then you add a playground.
00:31:56.079 --> 00:31:59.759
And then yeah, like it's no, I agree with you.
00:31:59.839 --> 00:32:03.839
Yeah, if you're monitoring logs and watching that, you might as well set evals, benchmarks as well.
00:32:04.000 --> 00:32:06.160
And like kind of when you're feedback, yeah.
00:32:06.480 --> 00:32:09.599
Yeah, I think that there's there'll be consolidation over the next few years.
00:32:09.680 --> 00:32:14.720
I'm not yet sure who I think will win, but then I'm also biased because I think we're gonna win.
00:32:14.799 --> 00:32:16.079
So uh fair enough.
00:32:16.240 --> 00:32:17.440
I don't know if this will be on the podcast.
00:32:18.240 --> 00:32:19.920
Yeah, what's next for you all right now?
00:32:20.000 --> 00:32:21.839
You just finished, I should have given context.
00:32:21.920 --> 00:32:26.079
You just finished a very impressive YC batch, just raised a very impressive round.
00:32:26.160 --> 00:32:28.720
Feel free to touch on that, or you can just touch on what's next.
00:32:29.039 --> 00:32:31.519
I think by the time this podcast is out, we will have announced.
00:32:31.759 --> 00:32:37.599
But we just I think we raised the largest round ever for a company in a YC batch.
00:32:37.759 --> 00:32:42.799
So we raised 27 million at a well, a big valuation.
00:32:42.960 --> 00:32:43.119
Yeah.
00:32:43.440 --> 00:32:45.839
And yeah, we are now, I guess, expanding the team.
00:32:45.920 --> 00:32:50.000
So we're looking for amazing kind of everything across the entire team.
00:32:50.079 --> 00:32:53.839
But like I'm I'm especially interested, invested interested in the technical team.
00:32:53.920 --> 00:32:59.200
So full stack product engineers, designers, we're like looking for great people.
00:32:59.279 --> 00:33:09.200
A lot of our team are ex-founders and love to really build in what is the new paradigm of like it's we're basically this is the new internet.
00:33:09.359 --> 00:33:16.880
There's a Microsoft scale opportunity to be built, and we are trying to take that opportunity and build something like a generational company.
00:33:16.960 --> 00:33:22.960
So if you like working on heart problems, you want to invent things that have never been done before, then come join us.
00:33:23.119 --> 00:33:25.519
And I guess what are we building next?
00:33:25.599 --> 00:33:28.640
It's kind of building that air operating system.
00:33:28.799 --> 00:33:33.119
So I guess I can I can go concretely into our roadmap, but that'd be a bit boring for the uh for the podcast.
00:33:33.200 --> 00:33:35.359
So uh show notes type of thing.
00:33:35.519 --> 00:33:36.000
Yeah.
00:33:37.119 --> 00:33:40.880
So lots of interesting and hard problems and not enough people to work on them at the moment.
00:33:40.960 --> 00:33:42.160
So in San Francisco.
00:33:42.319 --> 00:33:43.759
In San Francisco, in person.
00:33:44.079 --> 00:33:45.039
It's the best kind.
00:33:45.200 --> 00:33:51.279
Yeah, fully, fully in person, and we're I won't tell you where our office is gonna be because it's not yet 100% confirmed.
00:33:51.359 --> 00:33:53.279
But fully in person in San Francisco.
00:33:53.680 --> 00:33:55.920
Why don't we end with a quick fire round if that works?
00:33:56.240 --> 00:33:56.640
Let's do it.
00:33:56.799 --> 00:33:58.960
Okay, what's an agent trend that you're most bullish on?
00:33:59.200 --> 00:34:04.400
I'm most bullish on narrow AI agents, which I think has been under discussed.
00:34:04.559 --> 00:34:08.800
It's quickfire, so I'll I might stop there, but I think just quickly define narrow AI agent.
00:34:09.039 --> 00:34:19.679
Yeah, so so an uh like an agent that does one task, and these are somewhere between software and AI.
00:34:19.760 --> 00:34:30.400
They might have self-reflective reasoning loops, but they are only trying to do one thing, like write a blog post or a code email, or kind of write do one part of the code execution process.
00:34:30.480 --> 00:34:39.760
And I think narrow agents, you're able to get more reliability, you're able to use smaller, faster, cheaper models, or get squeeze more value out of the smarter, bigger models.
00:34:39.920 --> 00:34:43.679
They will be quicker because they don't have to work out how to solve the problem every time.
00:34:43.840 --> 00:34:56.239
I think one of the problems you see with a lot of these inter-agent, like auto-gen style agents is that they pontificate a lot and they think a lot and they like, you know, you get less reliable results, and there's a reasonably high chance they go off in some random direction and start self-prompting.
00:34:56.320 --> 00:35:03.760
And I think narrow agents are the way to go, at least with today's capabilities, but I still think that will be true as the models get better.
00:35:03.920 --> 00:35:07.119
You will always get better results from a narrow agent than you do from a general agent.
00:35:07.199 --> 00:35:16.320
It's just like you constrain the best AI out there to do something like smaller in scope, it will be better at that task.
00:35:16.800 --> 00:35:20.639
And what's an industry agent's our poised to disrupt that people aren't talking enough about?
00:35:21.119 --> 00:35:22.159
Oh god, everything.
00:35:22.320 --> 00:35:25.920
It's the entirety of knowledge work is about to be disrupted.
00:35:26.320 --> 00:35:29.119
What's your favorite agent built on wordware so far?
00:35:30.800 --> 00:35:33.599
Uh most of my favorite agents are fun.
00:35:33.760 --> 00:35:37.599
So we we have one that we use a lot in our Slack channel called Shrekify.
00:35:37.760 --> 00:35:43.840
And it takes in takes in some description and turns it into a Shrek cartoon.
00:35:44.000 --> 00:35:50.000
So uh yeah, that's I should definitely say something that's like one of our customers, but yeah.
00:35:50.320 --> 00:35:50.880
That's fun.
00:35:51.119 --> 00:35:54.960
Prediction what percent of enterprise companies will be using agents in 2025?
00:35:55.360 --> 00:35:58.000
A hundred percent that still exist in twenty thirty-five.
00:35:58.639 --> 00:35:59.840
That's a great note to end on.
00:36:00.079 --> 00:36:00.800
Thank you, Robert.
00:36:01.039 --> 00:36:01.519
This was fun.
00:36:01.920 --> 00:36:03.199
Awesome, that was a great idea.
00:00:00.000 --> 00:00:09.439
It allows you to turn these SOPs, these standard operating procedures that you might have in your company, and put that into kind of an automated flow.
00:00:16.399 --> 00:00:23.280
Welcome to Partner Path, a podcast that unpacks the venture capital and growth equity ecosystem from a junior perspective.
00:00:23.519 --> 00:00:29.839
Young entrepreneurs and investors have already had a massive impact on the industry, having started unicorns and launched billion-dollar funds.
00:00:30.239 --> 00:00:36.560
We discuss these success stories and more by sharing perspectives and advice from some of the industry's most prominent role models.
00:00:36.719 --> 00:00:43.119
This podcast expresses our views as of the date published and does not represent the views of and is not endorsed by any company for which we work.
00:00:45.119 --> 00:00:50.159
Before we get to the episode, Laura and I are excited to announce a partnership with Overlap.
00:00:50.640 --> 00:00:59.920
Fresh out of Y Combinator's Summer Batch, Overlap is an AI-driven app that uses large language models to curate the best moments from podcast episodes.
00:01:00.479 --> 00:01:12.400
Imagine having a smart assistant who reads through every podcast transcript, finds the best parts or parts most relevant to your search, and strings them together to form a new curated stream of content.
00:01:12.640 --> 00:01:14.560
That is what Overlap does.
00:01:14.959 --> 00:01:22.879
I personally use Overlap to explore content on future podcast guests and highly recommend it as another diligence tool for investors.
00:01:23.280 --> 00:01:25.599
Check out Overlap on the App Store today.
00:01:25.920 --> 00:01:30.239
Okay, today we are chatting with Robert Chandler, co-founder and CTO at Wordware.
00:01:30.400 --> 00:01:35.200
Robert started his career as a software engineer at 5AI before founding Hey Daily.
00:01:35.519 --> 00:01:40.480
In early 2023, Robert founded Wordware, a toolkit for building AI agents.
00:01:40.719 --> 00:01:41.519
Welcome, Robert.
00:01:41.840 --> 00:01:43.040
Great to be here, Laura.
00:01:43.359 --> 00:01:53.519
I don't have a deep understanding of your background, so I'd love if you could go back to the early days of entrepreneurship and maybe when your itch to start something began, since this is not your first rodeo.
00:01:53.840 --> 00:01:54.480
Wow, yeah.
00:01:54.640 --> 00:02:02.079
I mean, truly, I think the start of my entrepreneurial journey was back at school and I used to like bake and sell brownies on the school bus.
00:02:02.319 --> 00:02:03.200
You probably didn't even know that.
00:02:03.280 --> 00:02:04.719
That's uh this is like going way back.
00:02:04.879 --> 00:02:06.319
Not a venture scale business.
00:02:06.560 --> 00:02:12.719
Then God, I I got onto the show called The Young Apprentice, or nearly got onto the show called The Young Apprentice.
00:02:12.800 --> 00:02:15.680
So you know Donald Trump was the the apprentice.
00:02:15.840 --> 00:02:17.919
In the UK, I think it probably started there.
00:02:18.080 --> 00:02:23.759
We had a guy called Lord Sugar, and he also did a young apprentice, which was people like 16 to 18.
00:02:23.840 --> 00:02:28.639
And I got to the final 20 of 10,000 people to be on that show, which was it's kind of a rogue anecdote.
00:02:28.719 --> 00:02:29.039
That's true.
00:02:29.199 --> 00:02:30.960
Did that motivate you to do something else?
00:02:31.199 --> 00:02:32.639
It made me even more hungry.
00:02:32.719 --> 00:02:38.479
I think uh yeah, I wanted like always always been an entrepreneur, always wanted to like make something, sell something.
00:02:38.560 --> 00:02:45.199
And I guess I didn't go straight into founding a company because I got to join a really interesting startup straight out of university working on self-driving cars.
00:02:45.439 --> 00:02:50.080
One of the founders was the former CEO of ARM, and between them they'd sold like a billion dollars of businesses.
00:02:50.319 --> 00:02:54.639
And I was like, these are the people to learn from to learn how to build like an incredible company.
00:02:54.800 --> 00:02:59.759
There were 10 people when I joined, crewed to 200 people, got acquired by Bosch for like 150 million.
00:02:59.919 --> 00:03:04.879
So really awesome, kind of working on the cutting edge of machine learning and kind of learning from these incredible founders.
00:03:05.199 --> 00:03:07.280
Was where were the self-driving cars at?
00:03:07.840 --> 00:03:10.400
Yeah, they were in mostly like London.
00:03:10.479 --> 00:03:25.439
So we we probably probably still the largest like public trial of self-driving vehicles in the UK, where we actually like ran like a test route between like two places in London.
00:03:25.599 --> 00:03:32.319
And we still had a safety driver, but people from the public got to ride in the self-driving cars and kind of you know, the safety driver would take over occasionally.
00:03:32.479 --> 00:03:33.840
But interesting.
00:03:34.000 --> 00:03:37.599
So, how does the self-driving car scene compare to San Francisco?
00:03:37.840 --> 00:03:41.120
Oh, well, like there are actually self-driving cars in San Francisco.
00:03:41.199 --> 00:03:42.080
You've got like Waymo.
00:03:42.479 --> 00:03:45.360
It's interesting because the tech is basically the same.
00:03:45.599 --> 00:03:50.719
It's just Google has spent$20 billion getting to the point where it like worked as well as it does.
00:03:51.039 --> 00:03:54.560
We could talk about self-driving cars for like the entire entire podcast.
00:03:54.719 --> 00:03:57.599
I think it's super cool that you can ride in a self-driving car in San Francisco.
00:03:57.759 --> 00:04:06.319
Do I think it's gonna scale to be better than Uber in terms of like cost per mile?
00:04:06.960 --> 00:04:08.080
Not soon.
00:04:08.319 --> 00:04:09.919
Um yes, I agree.
00:04:10.000 --> 00:04:11.759
I think one, that's for a different episode.
00:04:11.840 --> 00:04:21.199
And two, I think the infrastructure of San Francisco, as well as how the road, how everything is laid out, and I think the population density makes sense here, but you it would be hard to replicate in New York.
00:04:21.519 --> 00:04:22.319
Yeah, for sure.
00:04:22.480 --> 00:04:29.120
And and London, especially where it's like rainy and like dark, and people like cross the tiny, narrow, winding roads.
00:04:29.199 --> 00:04:29.839
It's yeah.
00:04:30.160 --> 00:04:30.720
Yeah.
00:04:31.040 --> 00:04:37.279
So then fast forward, you wanted to start something and you did, and you started a company called Hay Daily.
00:04:37.759 --> 00:04:38.000
Yeah.
00:04:38.160 --> 00:04:42.720
So actually, I met my co-founder 10 years ago at Cambridge where we were starting studying machine learning.
00:04:42.959 --> 00:04:53.680
We, you know, built the drone society, we broke into parties, we uh like started a dating app as you do whilst you're in your university, solve your own problems.
00:04:54.000 --> 00:04:59.360
And I think every every every founder has to start a dating app at some point.
00:04:59.600 --> 00:05:01.920
And uh, so we've been friends for a very long time.
00:05:02.079 --> 00:05:07.040
He actually left university, started his first company in San Francisco using GPT2 and BERT.
00:05:07.279 --> 00:05:13.360
I kind of joined this self-driving car startup, but we'd always like kind of knew that at some point we might work together.
00:05:13.519 --> 00:05:18.639
And so when he he kind of he'd so he built a company using GPT 2, the tech wasn't quite ready for what he was doing.
00:05:18.720 --> 00:05:22.639
It's like augmenting human memory, a lot like these pendant devices that we're seeing all over the place nowadays.
00:05:22.959 --> 00:05:23.519
Rabbit.
00:05:24.000 --> 00:05:35.279
Not like Rabbit, like rewind or limitless, like friend, like the other one that was called friend and yeah, no, tab that turned into friend or like no longer your friend, an unfriend blocked.
00:05:35.360 --> 00:05:39.279
Uh and uh yeah, so that the GPT 2 wasn't quite ready for that.
00:05:39.439 --> 00:05:44.399
So he pivoted into like tech remote teams, and that was just before COVID hit, so had much more success there.
00:05:44.560 --> 00:05:48.879
Took a partial exit and actually kind of explored for a little while.
00:05:49.120 --> 00:05:52.959
Gave me a call three years ago, said, Hey, come be my co-founder, I've got an idea.
00:05:53.199 --> 00:05:56.639
I was like, hell yeah, this is like someone that I've wanted to work with for a long time.
00:05:56.720 --> 00:05:59.680
We were just about to get acquired by Bosch, so the timing was ideal.
00:05:59.839 --> 00:06:09.600
I did have to leave behind a bunch of equity, but I knew like missing the opportunity to work with like a very close friend and someone who I trust immensely would have been like too big an opportunity cost to miss.
00:06:09.759 --> 00:06:12.480
So we started, we were both a little bit burned by deep tech.
00:06:12.720 --> 00:06:20.800
Like, you know, there were in netter in San Francisco, there's no self-driving cars driving around at that point in was it 2021, late end of 2021.
00:06:21.040 --> 00:06:23.839
GPT-3 hadn't really done much beyond GPT 2.
00:06:24.079 --> 00:06:28.240
It was still like, okay, like let's do something where we can generate revenue and get fast feedback.
00:06:28.319 --> 00:06:38.399
So we created or started a creator economy business, a subscription service for influencers where you would a lot like Substack or Patreon or OnlyFans deliver exclusive content in return for like a subscription fee.
00:06:38.480 --> 00:06:40.079
But it was through like iMessage and WhatsApp.
00:06:40.240 --> 00:06:41.920
So it was fine.
00:06:42.240 --> 00:06:43.920
It's hard to scale those, right?
00:06:44.240 --> 00:06:45.439
That was one of the big things.
00:06:45.600 --> 00:06:48.000
I think we disproved some hypotheses.
00:06:48.079 --> 00:06:54.079
So we grew it to like 30k MRR, but we disproved some hypotheses that led to basically realizing it wasn't gonna be a billion dollar business.
00:06:54.240 --> 00:07:04.160
There was you you actually the hypotheses meant that you didn't need that big a scale in order to get to like big like revenue.
00:07:04.480 --> 00:07:09.839
You just have to increase your RPU, like your average revenue per customer for each like interaction, probably.
00:07:10.079 --> 00:07:11.199
Yeah, so you see it like sustain.
00:07:11.519 --> 00:07:23.199
I think we worked out that we needed basically the kind of simple maths was if you had a hundred thousand followers and you could convince 1% of them at$10 a month, or to pay you$10 a month, then you could be making$10,000 a month.
00:07:23.360 --> 00:07:24.639
We'd take 25% of that.
00:07:24.800 --> 00:07:28.800
We needed about 3,000 influences at the revenue multiples at the time in order to be a billion dollar business.
00:07:28.879 --> 00:07:31.759
So it's like, okay, this could be like a a sales-led thing.
00:07:31.920 --> 00:07:40.959
And you'd actually like 3,000 influencers, given there's 20 million or so creators with 50,000 followers or more on Instagram alone, was like, it seemed pretty easy.
00:07:41.199 --> 00:07:46.879
Turned out like actually very few creators have a following that are willing to pay the money.
00:07:47.519 --> 00:07:50.160
That's the like the big thing, isn't it?
00:07:50.560 --> 00:07:51.120
Interesting, yeah.
00:07:51.279 --> 00:08:00.560
Nowadays, uh how I view the influencer market is the ones I think with what you alluded to with like 50,000, 100,000 followers, it's almost like there's too many, the degrees of a closeness are too close.
00:08:00.720 --> 00:08:04.160
Yeah, and then you have the influencers who now are almost hitting celebrity level.
00:08:04.240 --> 00:08:04.399
Yeah.
00:08:04.480 --> 00:08:10.879
And so their income streams are probably fine without some of this additional passive income because they already have additional passive income.
00:08:11.199 --> 00:08:18.079
Yeah, and like if you can do a brand deal for like$50,000, why would you create exclusive content and like make less money?
00:08:18.319 --> 00:08:18.639
Yeah.
00:08:19.360 --> 00:08:22.720
And yeah, it's like there's all sorts of different ways to monetize.
00:08:22.879 --> 00:08:28.079
Also, to sign up for a subscription requires like a lot tighter relationship than to sign up for like a one-time purchase.
00:08:28.319 --> 00:08:30.160
So basically, we just proved that idea.
00:08:30.480 --> 00:08:35.840
Realized like there was there was a path to something that could have been more sustainable.
00:08:35.919 --> 00:08:41.279
It was like microcourses delivered through like influencer branded, like a bit like Skillshare, but for like microcourses.
00:08:41.919 --> 00:08:44.000
But we weren't the right people to be building that.
00:08:44.240 --> 00:08:44.480
Okay.
00:08:44.720 --> 00:08:46.320
And AI got super exciting.
00:08:46.480 --> 00:08:50.720
We were using like stable diffusion to reach out to these creators to like make fan art.
00:08:50.879 --> 00:08:53.440
So we were like already exploring with AI.
00:08:53.679 --> 00:08:59.440
And we were like, we should be doing something in AI, not just kind of working with influencers and doing subscription content.
00:08:59.600 --> 00:09:00.559
So that's fair.
00:09:00.720 --> 00:09:08.320
And you also want something that hopefully you can build a billion-dollar business or build something that can like scale to the level that you and your co-founder are hoping to achieve.
00:09:08.639 --> 00:09:08.879
Exactly.
00:09:08.960 --> 00:09:09.279
Yeah, yeah.
00:09:09.440 --> 00:09:12.000
So there was that's perfectly good lifestyle business to be built there.
00:09:12.320 --> 00:09:13.279
That wasn't what we're interested in.
00:09:13.360 --> 00:09:23.200
We were like, we we want to build something huge and something that's like gonna dramatically change the world rather than just be yeah, make a few people a little bit richer.
00:09:23.519 --> 00:09:24.240
That's fair.
00:09:24.559 --> 00:09:27.840
Okay, but then AI in and of itself is quite a broad term.
00:09:27.919 --> 00:09:29.039
And even your backgrounds.
00:09:29.120 --> 00:09:37.200
I mean, your backgrounds are, as you mentioned, more conducive for hardware type build, but that probably didn't make sense for what you were looking to do next.
00:09:37.279 --> 00:09:40.399
So, how did you decide on agenteg use case specifically?
00:09:40.720 --> 00:09:48.320
Yeah, I think the period before we pivoted, so you know, you've got to like make a strong like pivoting company's heart.
00:09:48.399 --> 00:09:51.279
You've got to decide that, okay, what you're working on is not good.
00:09:51.360 --> 00:09:54.879
You've got to decide that, like, are you gonna do something different?
00:09:55.200 --> 00:09:56.559
Are you gonna keep some of the team?
00:09:56.639 --> 00:09:58.080
Are you gonna keep none of the team?
00:09:58.480 --> 00:10:00.000
What are you even gonna work on?
00:10:00.320 --> 00:10:16.720
So there was kind of a phase where we were exploring everything from like working in like foundational image models to, I mean, like we explored like neural radiance fields and synthetic data generation, and even like, can we do video generation and these kind of things?
00:10:16.879 --> 00:10:25.919
And we realized what a lot of people think about LMs, where they think about them as knowledge like engines, like search engines, and like how can you make a better Google?
00:10:26.159 --> 00:10:38.799
So basically the point where we pivoted, which ended up being March 2023, everyone was like doing chat with your docs, or like, can we take in your all your knowledge of your like company's data and then like put it into a chat bot?
00:10:38.960 --> 00:10:44.240
And we were like, okay, that's cool, but what's way more exciting is this is the first thing that we've ever seen that can think.
00:10:44.320 --> 00:10:46.000
So the first thing that came can reason.
00:10:46.159 --> 00:10:51.120
And you know, it's like it's simulated reasoning, it's it's kind of fuzzy and and not amazing, but it's pretty cool.
00:10:51.279 --> 00:10:52.399
And like there's something there.
00:10:52.480 --> 00:10:55.120
So we were looking at this from a reasoning perspective.
00:10:55.279 --> 00:10:56.639
We started like just building.
00:10:56.720 --> 00:11:03.840
So we were like, let's read all the papers, let's read the React paper, let's read the Tor Former paper, let's I guess we had to discover these papers even existed.
00:11:03.919 --> 00:11:11.840
So like let's just like look at the literature, catch up with everything that's happened with LLM since you know attention is all you need all the way up to like the state of the art of the day.
00:11:12.159 --> 00:11:15.759
And we used frameworks like Langchain, we used niche ones like LMQL.
00:11:15.919 --> 00:11:26.320
We started building these things from scratch and really learned what it meant to build an agent and kind of felt like there's gonna be a huge amount of value created here.
00:11:26.399 --> 00:11:28.559
It was like an LLM plus a tool.
00:11:28.799 --> 00:11:30.080
We were like, this is AGI.
00:11:30.159 --> 00:11:41.440
You know, it's weak AGI, it's inefficient AGI, but it's it's the kind of an LLM is the unit of reasoning required to deliver a weak AGI.
00:11:41.679 --> 00:11:45.120
So we were all about like how can we turn that weak AGI into a narrow expert AI?
00:11:45.279 --> 00:11:47.600
And you do that through basically prompting.
00:11:47.840 --> 00:11:56.320
And so um, I think what was most exciting, or like one of the reasons it was really exciting to us is kind of for the last 30 years, software has been about automation.
00:11:56.559 --> 00:11:59.120
But you've only been able to do if-else and structured data.
00:11:59.279 --> 00:12:02.720
For about the last 10 years, you've been able to do machine learning.
00:12:02.960 --> 00:12:07.440
So you can do predictive like analysis, or you can do like classification type tasks.
00:12:07.919 --> 00:12:14.720
And that's really cool and really useful, but it's only works if you can gather enough data.
00:12:14.799 --> 00:12:30.240
So you have to gather like millions of examples of data, you have to annotate them, you have to train one model to do one thing, and at the end of that, you've kind of spent a few million dollars, you've hired some machine learning PhDs, and you've waited six to twelve months to like get some results, um, and you've automated one thing.
00:12:30.399 --> 00:12:34.559
With LLMs, that that activation entry drop dropped through the floor to basically zero.
00:12:34.720 --> 00:12:48.080
And so there was this kind of new wave of software that through prompting alone, through this very low-hanging fruit, where you didn't need to be insanely technical, you could build incredibly powerful automations that previously would have just been impossible or intractable.
00:12:48.399 --> 00:12:51.039
Okay, so then that brings us to agents.
00:12:51.360 --> 00:12:51.759
Agents.
00:12:51.919 --> 00:12:52.240
Yep.
00:12:52.559 --> 00:12:54.960
Now a quick word from our sponsor.
00:12:55.279 --> 00:13:01.600
Laura and I are thrilled to announce a partnership with Grata, an intuitive deal sourcing intelligence platform.
00:13:01.919 --> 00:13:05.919
Keep missing out on private market deals, ditch the Rolodex and get Grata.
00:13:06.399 --> 00:13:14.559
Grata's AI-powered platform provides accurate data, deep insights, and time-saving workflows that help get more deals done.
00:13:14.799 --> 00:13:20.000
And it's trusted by over 500 firms in 3,000 deal makers, including myself.
00:13:20.240 --> 00:13:29.919
I've personally used Grata in both my deal sourcing and diligence workflows and would recommend their tool to any investor across private equity, growth equity, corpdev, or investment banking.
00:13:30.159 --> 00:13:32.960
We've negotiated a special offer for our audience.
00:13:33.200 --> 00:13:42.480
Mention Partner Path to the Grata team during sign-up, and you'll get free executive seats for your partners and MDs included in your subscription.
00:13:42.799 --> 00:13:44.080
Now back to the episode.
00:13:44.399 --> 00:13:52.720
Can you just break down generally agents and how kind of wordware fits into this revolutionizing human productivity?
00:13:53.039 --> 00:13:53.360
Yeah.
00:13:53.600 --> 00:13:59.039
So I think whenever you think about agents, you can basically think about them as a collection of prompts.
00:13:59.200 --> 00:14:01.120
So there's kind of a little bit more to it than that.
00:14:01.279 --> 00:14:06.960
They are kind of an orchestration of prompts, and those prompts can use tools so they can pull in extra information.
00:14:07.200 --> 00:14:16.720
They're also able to, you know, do inter-agent communication, which basically means having the outputs of one prompt feed into the inputs to the next prompt.
00:14:16.879 --> 00:14:20.240
And together you get something that kind of forms this reasoning system.
00:14:20.320 --> 00:14:26.000
Some of that orchestration like gets it to self-reflect, gets it to like think, and it can like look at its own output.
00:14:26.320 --> 00:14:28.080
And really, you can think about it a little bit like this.
00:14:28.240 --> 00:14:34.960
When you use Chat GPT, you're kind of a wizard and you're casting a spell and you're saying, I want you to write me a blog post.
00:14:35.120 --> 00:14:38.320
And if you just put that into ChatGPT, you get a pretty mediocre output.
00:14:38.399 --> 00:14:40.879
And maybe you say, or write me an essay on the Roman Empire.
00:14:40.960 --> 00:14:43.679
You'll get like an okay essay, but it's not gonna be very good.
00:14:43.840 --> 00:14:49.279
But if you're a bit more skilled with ChatGPT, you can be like, okay, I'm gonna write an essay about the Roman Empire.
00:14:49.440 --> 00:15:07.360
First, I'm gonna get paste in some research from Wikipedia about like the some some articles, and then I'm gonna tell ChatGPT to create an outline, and then I'm gonna get it to refine that outline, and then I'm gonna get it to write the introduction, then the conclusion, and then fill in the paragraphs, and maybe I'll get it to proofread and critique, and then I'll get it to rewrite.
00:15:07.440 --> 00:15:09.759
And that's like what a skilled prompt engineer would do.
00:15:09.919 --> 00:15:20.399
And what you can do with Wordware is you can kind of distill that that whole process into a flow, into something that would be like almost like an SOP in a business.
00:15:20.639 --> 00:15:25.279
So it's I would like you you can basically orchestrate this agent as a potion.
00:15:25.360 --> 00:15:34.480
So you've gone from being a wizard casting a spell to a person making a potion, and then you can share that potion with your friends or your colleagues, and suddenly everyone becomes as powerful a wizard as you are.
00:15:34.559 --> 00:15:37.759
So that's a very high-level kind of version of what Wordware is.
00:15:37.919 --> 00:15:48.399
But I guess concretely, it allows you to turn these SOPs, these standard operating procedures that you might have in your company, and put that into kind of an automated flow.
00:15:48.559 --> 00:15:52.960
I think another way that we quite like to think about it is we have this process, this thing that we call the intern test.
00:15:53.120 --> 00:16:00.320
And basically with our customers, we ask them, what is, if you're trying to explain this to an intern, what would what would they do?
00:16:00.480 --> 00:16:05.679
Where would they look for data to answer the kind of question you want them to answer or do the thing you want them to do?
00:16:05.840 --> 00:16:07.200
What are the steps they would take?
00:16:07.279 --> 00:16:11.840
And if you can do that, if you can explain that, then you can probably get an AI to do it for you.
00:16:12.000 --> 00:16:19.679
So you break down the problem into this set of substeps, you then kind of narrow the scope of what this agent can do, but you get it to be much more reliable.
00:16:19.759 --> 00:16:33.200
And that's really where WordWare helps like you craft these chains of prompts, this orchestrate these LLMs and interact with the tools and pull in the right data to build automations that would otherwise previously not have been possible.
00:16:33.519 --> 00:16:35.120
That's a very graceful explanation.
00:16:35.279 --> 00:16:35.919
Thank you.
00:16:36.159 --> 00:16:37.679
And let's put this into practice.
00:16:37.759 --> 00:16:40.480
What are some of the initial use cases you're seeing on Wordware?
00:16:40.799 --> 00:16:42.799
Yeah, I mean, it's it's really, really broad.
00:16:42.879 --> 00:16:52.559
So we have customers that are in finance, we have customers that are in logistics, we have customers that are in a lot of like AI sales-based products.
00:16:52.799 --> 00:16:55.759
But I think it's it really is anything that you could give to an intern.
00:16:55.840 --> 00:17:01.919
It's like if you could give it to an intern, suddenly you can have a thousand interns, and then what use cases are valuable for that?
00:17:02.080 --> 00:17:09.039
And so, like in sales, it's how can I write much better cold email outreach copy that's like really personalized to the end user.
00:17:09.200 --> 00:17:19.119
If you had an intern, you could teach them how to like look at their LinkedIn and pull in the right data and craft a hook and then like write a good email, and that would be intractable to train enough interns that that would be worthwhile.
00:17:19.279 --> 00:17:24.000
But with AI, you train one intern and you have a thousand interns or a million interns and then zero again.
00:17:24.160 --> 00:17:28.480
With like the legal use cases, it's how can I create a trademark application?
00:17:28.559 --> 00:17:32.160
Because that's the kind of thing that actually is pretty much run off the mill.
00:17:32.400 --> 00:17:34.559
And so this is one where it's actually not fully fully automated.
00:17:34.640 --> 00:17:36.559
You you still have the lawyer review it.
00:17:36.720 --> 00:17:39.519
But you know, 8% of the work is kind of grunt work.
00:17:39.599 --> 00:17:55.119
It's okay, I need to describe, take in this company's like landing page and description and turn it into a kind of slightly more legalese description that that covers a broader as much part of their trademark without infringing on other people's trademarks.
00:17:55.200 --> 00:18:13.279
I also need to do some like research on who could they possibly infringing, be infringing, and then compare their description to their other their kind of potential competitors or like infringement descriptions and you know it's basically what are the reasoning processes that would that make up knowledge work?
00:18:13.440 --> 00:18:23.680
I think we're in in a decade where that's kind of like automate or be automated, and a lot of people are building companies that are looking to automate the things that they have domain expertise in.
00:18:23.920 --> 00:18:24.559
I like that.
00:18:24.799 --> 00:18:26.079
Automate or be automated.
00:18:26.480 --> 00:18:33.119
How would wordware differ from like traditional no-code tools or maybe more opinionated customers that actually want to go in?
00:18:33.279 --> 00:18:36.400
They're coding, they're engineers, and we want to make more of the sausage.
00:18:36.720 --> 00:18:41.119
Yeah, so I I guess say traditional code tools or traditional no-code tools.
00:18:41.279 --> 00:18:42.000
Sorry, no code tools.
00:18:42.160 --> 00:18:44.960
Because wordware is, I mean, it's like a like you said, it's a wide audience.
00:18:45.119 --> 00:18:46.480
People like me, people like you.
00:18:46.720 --> 00:18:51.680
We have very different backgrounds, but everyone can use wordware, which is awesome because it's a horizontal AI tool.
00:18:51.920 --> 00:18:52.319
I think so.
00:18:52.400 --> 00:18:54.319
I think you can think of it a little bit like Excel.
00:18:54.480 --> 00:19:02.880
And you know, there's a long tail of use cases in which end users built, like basically programmed using Excel.
00:19:02.960 --> 00:19:04.240
They don't really think about it as programming.
00:19:04.319 --> 00:19:10.000
You're like, sometimes you're just using the sum function, you're adding some data, and that's like a useful tool that people have created for their own use.
00:19:10.160 --> 00:19:13.599
There's the similar kind of thing where, you know, like there's the v lookups and then there's the macros.
00:19:13.759 --> 00:19:19.359
With Wordware, you really have like progressive ability to get more and more complex with like a very high ceiling.
00:19:19.519 --> 00:19:23.039
So I think kind of wordware sits between two ends of the spectrum.
00:19:23.200 --> 00:19:33.920
One end is like fully code-based, so things like Langchain, doing it from scratch, just hitting the APIs directly, where you have to be a developer, you have to have experience writing software, you have to be able to deploy this yourself.
00:19:34.079 --> 00:19:38.720
That's the the big blocker is this very slow feedback cycle between the engineers and the people who know what good looks like.
00:19:38.799 --> 00:19:40.640
And so that's a very limited use case.
00:19:40.799 --> 00:19:48.559
On the other end, you have a lot of these low code tools, so things like Zapier, other tools that are like more AI-first Zapias, or the kind of like flowchart-based things.
00:19:48.720 --> 00:19:50.720
And the problem there is you hit a ceiling very quickly.
00:19:50.799 --> 00:19:53.359
If you want to add any kind of logic, you struggle.
00:19:53.519 --> 00:19:59.519
You can basically do the quite okay linear flows, but as soon as it gets quite complicated, then it's it's kind of game over.
00:19:59.759 --> 00:20:08.240
Wordware sits between the two and it gives you the kind of flexibility of all this, uh, or like the low floor of the no-cools with the high ceiling of software.
00:20:08.319 --> 00:20:19.119
So you can combine the flexibility of natural language with the structured programming and build these like genuinely complex, genuinely useful agents, whether you're technical or non-technical, or you're collaborating between technical and non-technical teams.
00:20:19.200 --> 00:20:22.559
And then you deploy, you can deploy it through an API, or you can trigger it through Zapier.
00:20:22.720 --> 00:20:29.039
And that way you've kind of got this broad base, whether you're trying to build an AI-powered product or you're trying to automate parts of your workflow.
00:20:29.359 --> 00:20:29.680
Okay.
00:20:30.000 --> 00:20:31.279
That makes a ton of sense.
00:20:31.519 --> 00:20:34.880
So right now we are, I don't think we've hit AGI yet.
00:20:34.960 --> 00:20:36.880
And I think that's universally agreed upon.
00:20:37.039 --> 00:20:43.039
And I think, well, depending who you ask, but I think we're really making steps towards automating more and more of the workflow.
00:20:43.119 --> 00:20:48.319
And by using tools like Wordware and other domain-specific tools, where do you see the future of Wordware going?
00:20:48.480 --> 00:20:53.200
You kind of described where it's at right now, which is like amazing, just the progress you've made in under two years.
00:20:53.359 --> 00:20:55.599
But like, what does the future of Wordware look like?
00:20:55.920 --> 00:20:56.319
Yeah.
00:20:56.640 --> 00:21:02.640
I would argue that if you asked someone two years ago, that is like what you're building today AGI?
00:21:02.720 --> 00:21:03.920
Probably the answer would be yes.
00:21:04.160 --> 00:21:04.400
Yeah.
00:21:05.039 --> 00:21:05.680
It's relative.
00:21:05.920 --> 00:21:08.160
I think just like our expectations have got higher.
00:21:08.559 --> 00:21:09.279
Definitions have changed.
00:21:09.440 --> 00:21:10.720
Definitions didn't exist.
00:21:10.960 --> 00:21:12.000
Yeah, yeah, well, exactly.
00:21:12.079 --> 00:21:14.160
It was like, oh we'll we'll kind of know it when we see it.
00:21:14.240 --> 00:21:17.519
It's AGI, it's uh human-level intelligence, whatever that means.
00:21:17.759 --> 00:21:18.079
I like that.
00:21:18.160 --> 00:21:19.039
The bar keeps getting higher.
00:21:19.359 --> 00:21:20.319
The bar gets higher, yeah.
00:21:20.400 --> 00:21:25.200
It's uh yeah, I think it's like as yeah, yeah.
00:21:25.519 --> 00:21:29.440
Anyway, well, where is where we're going is probably the uh the slightly easier question.
00:21:29.519 --> 00:21:32.799
Like the we could also spend an entire podcast just discussing AGI.
00:21:33.039 --> 00:21:33.839
Where is where we're going?
00:21:33.920 --> 00:21:36.559
I think the goal is to be the AI operating system.
00:21:36.960 --> 00:21:44.720
We want to empower teams to augment their themselves.
00:21:44.880 --> 00:22:02.480
We want you know the people to be focusing on the really differentiated work, the stuff that's like uses their special skills of being a human, their experience, their all of that, and then offload the kind of grunt work to AI and that's Actually, really tailored to their taste.
00:22:02.640 --> 00:22:08.720
So I think one of the problems with like a lot of off-the-shelf AI systems is you lose that taste.
00:22:08.880 --> 00:22:17.759
You don't get much control over how good the outputs are and like whether they align exactly with what you want versus what just like the generic optimal thing is.
00:22:17.920 --> 00:22:32.640
And I think as the kind of more and more AI happens, more like there's more and more content being put out there by AI or like work being done by AI, the big differentiator will be like, have you imbued your own companies or your own personal taste on that output?
00:22:32.720 --> 00:22:34.640
It's like we see a lot of content that says delve.
00:22:34.799 --> 00:22:39.440
You might not want your content to say delve and you might want it to be sort of more fun and more upbeat.
00:22:39.599 --> 00:22:52.720
So I think WordWeb becomes this AI operating system by starting with narrow agents that are like highly specialized, highly like reliable, repeatable, very good at doing one thing, and then expanding into being orchestrated by more general agents.
00:22:52.880 --> 00:22:55.440
So for now it's very much like event-based triggering.
00:22:55.519 --> 00:23:00.799
So you know, you receive an email, you run a WordWare agent, that creates a draft response, or maybe even just sends the response.
00:23:00.880 --> 00:23:14.799
You'll start seeing maybe like a slightly higher level agent that will act as your like EA, and then that'll have the ability to like craft, like delegate to a narrow agent to respond to maybe in your case, like a an inbound fundraising email.
00:23:14.880 --> 00:23:19.599
And you'll be like, okay, let's you know do some research on this company, and like what are the things I need to know?
00:23:19.759 --> 00:23:28.640
I need to like find out when it was founded, like what the previous fundraising uh history was, who's backed them, like what are the background of the founders?
00:23:28.799 --> 00:23:29.599
What's the product?
00:23:29.680 --> 00:23:30.960
Like, do I care about this market?
00:23:31.119 --> 00:23:37.039
And like, you know, and then there'll be a bunch of different responses that you'll give based on that, and maybe you'll ask follow-up questions, or maybe you'll be like, let's schedule a meeting.
00:23:37.200 --> 00:23:40.880
I don't like it's all down to you and like what do you do day to day?
00:23:41.119 --> 00:23:41.519
Yeah.
00:23:41.759 --> 00:23:44.480
Yeah, I have a lot of use cases, so that makes sense.
00:23:44.720 --> 00:23:46.319
One, the growth has been phenomenal.
00:23:46.480 --> 00:23:56.160
I actually read that your launch on August 2nd, you grew from 10,000 users on the main product to over 300,000 users today, and I'm sure that number is outdated.
00:23:56.559 --> 00:23:58.160
It's yeah, we're we're growing a lot.
00:23:58.400 --> 00:23:59.279
Yeah, which is amazing.
00:23:59.359 --> 00:24:02.079
And I think it's worth tapping in honestly just to the growth strategy.
00:24:02.160 --> 00:24:21.759
And I know I know you're more on the tech side as I introduced you, your CTO and co-founder, but how important is scaling like quality of users versus like quantity of users, especially given that the use cases are quite dispersed and to burn calories in all these different areas of like helping onboard and get people up to running, you have to know kind of where to burn them.
00:24:22.079 --> 00:24:23.519
No, it's it's it's a massive challenge.
00:24:23.599 --> 00:24:31.039
And you know, actually, as the CTO, I'm the one that ends up like we've got all these requests for different features across like a very broad spectrum.
00:24:31.200 --> 00:24:35.519
It's like, how do we prioritize to make sure enough people that we care about are happy?
00:24:35.599 --> 00:24:36.720
And that's that's really hard.
00:24:36.880 --> 00:24:41.839
I think really and I guess do founders listen to this podcast as well.
00:24:42.000 --> 00:24:42.160
Yes.
00:24:42.720 --> 00:24:43.039
Okay, yeah.
00:24:43.119 --> 00:24:49.119
So this is like a it's one of those things, as a founder, you have to like work out where do you place your bets.
00:24:49.279 --> 00:24:52.319
And actually, when we launched, we weren't expecting to get quite as much traction as we did.
00:24:52.400 --> 00:25:03.440
We thought, you know, we're building something really useful, but we got a lot more prosumer traction than we'd expected, even though we hadn't done anything to like build the onboarding, to build like the kind of B2C type product that you'd expect.
00:25:03.599 --> 00:25:07.839
And so we are now prioritizing more of those features than we were before.
00:25:07.920 --> 00:25:12.720
We were focused largely on B2B, on these scale-ups and startups who are building AI-powered backends.
00:25:12.880 --> 00:25:22.880
We're still, they are still our most important customer, but we are starting to add features and like onboarding and like integrations with Zapia that help the slightly more prosumer users, at least the ones that are really like passionate.
00:25:23.039 --> 00:25:25.279
And over time, we're going to be adding more and more features.
00:25:25.359 --> 00:25:28.559
But I think there's all sorts of things you need to do building a company.
00:25:28.640 --> 00:25:30.960
One is get your name out there so people know you exist.
00:25:31.519 --> 00:25:37.440
And so a lot of what we've been doing has always been well, and actually, we're building a horizontal tool.
00:25:37.599 --> 00:25:42.079
So not only do people need to know we exist, they need to know like what can we do with this tool.
00:25:42.319 --> 00:25:44.160
And so we've always been dog fooding wordware.
00:25:44.240 --> 00:25:58.400
So we've always been using wordware to build fun or useful tools to demonstrate, like, oh, you can build, we have one called Audio Scribe, and it's like you can build a tool, uh a thing with wordware where it takes in a voice note and turns that into like a summary.
00:25:58.480 --> 00:25:59.839
And it'll first classify that voice note.
00:26:00.000 --> 00:26:05.839
So it could be you're making a shopping list or it could be you're making a journal entry, and it has different prompts for those different kind of aspects.
00:26:06.079 --> 00:26:07.759
We've kind of been using Webware.
00:26:08.000 --> 00:26:11.599
We did one that went insanely viral, which was the kind of Twitter roast.
00:26:11.680 --> 00:26:14.160
I think we got like eight million people run that, probably more now.
00:26:14.319 --> 00:26:19.359
Haven't looked at the numbers recently, but it's that was obviously great for getting attention.
00:26:19.599 --> 00:26:25.680
Most of those users are like most of those kind of people who play used the twit Twitter roast are not going to be our ICP.
00:26:25.839 --> 00:26:32.960
Like a small fraction of them then signed up for wordware, and like some of those are our ICP, and some of those are the more prosumer ICP.
00:26:33.039 --> 00:26:37.359
And there's benefits to having like people using webware and playing with webware and giving us feedback.
00:26:37.519 --> 00:26:39.440
There's also a lot of noise to cut through when you have that.
00:26:39.519 --> 00:26:40.880
So it's it's a challenge.
00:26:40.960 --> 00:26:42.079
I don't think there's uh a good answer.
00:26:42.559 --> 00:26:47.119
And hopefully the goal of ProSumer is a lot of it can be self-served, honestly, and people can get up and running.
00:26:47.519 --> 00:26:51.119
Yeah, and what's what's awesome is like it helps test our infrastructure.
00:26:51.200 --> 00:27:04.000
So, you know, having run, I think we ran like six or seven million eight like executions of an agent within a 24-hour period or like a two-week, uh, few day period, and that really tests our infrastructure, and that's good.
00:27:04.079 --> 00:27:13.839
That's good for people putting their trust that actually if we can deal with that level of traffic, then we'll be fine because it's pretty rare to have that scale of just like usage of a product, especially overnight.
00:27:14.000 --> 00:27:17.279
It went from like zero to millions in.
00:27:17.680 --> 00:27:19.599
You have to share your tech stack after this.
00:27:19.839 --> 00:27:21.279
We'll put it in the show notes.
00:27:21.759 --> 00:27:28.799
Well, uh, yeah, we uh we'll plug wordware and the or yeah providers you're using probably more.
00:27:29.039 --> 00:27:33.200
But actually, I want to dig into something you mentioned with users, finding your ICP, cutting through the noise.
00:27:33.359 --> 00:27:35.920
I think it's relevant in the same vein of being in San Francisco.
00:27:36.160 --> 00:27:39.039
And there's so many startups, especially in agent infrastructure right now.
00:27:39.119 --> 00:27:44.400
And honestly, even frameworks or orchestration tools that try and be adjacent but friendly competitors to you guys.
00:27:44.640 --> 00:27:47.279
How do you think about honestly standing out in this space right now?
00:27:47.680 --> 00:27:52.480
Yeah, I think there's a lot of people who have gone prompting is the new programming.
00:27:52.720 --> 00:27:56.319
Let's build the same tools we have for programming, but for prompting.
00:27:56.400 --> 00:27:59.359
Or but and like you know, prompting agents, they're all basic the same thing.
00:27:59.440 --> 00:28:03.119
I think people use the word agent a lot when they really just mean a collection of prompts.
00:28:03.279 --> 00:28:05.759
And it's because it sounds more fancy, it sounds more exotic.
00:28:06.079 --> 00:28:06.880
Prompting for GitHub.
00:28:07.119 --> 00:28:07.599
Prompting, yeah.
00:28:08.000 --> 00:28:09.359
I read that on your YC website.
00:28:09.599 --> 00:28:13.359
That's it's yeah, I mean, yeah, GitHub for agents, I think we probably say somewhere.
00:28:13.759 --> 00:28:14.319
Yeah, that's better.
00:28:14.559 --> 00:28:16.960
And so a lot of people have gone, okay, let's build the same tools.
00:28:17.200 --> 00:28:27.759
They've built the observability layer, or they've gone via the evaluation layer, or they've gone for like the user feedback, or like one of these like niche parts of the stack.
00:28:27.920 --> 00:28:35.440
And the problem there is that they're kind of obvious and they're kind of not useful until you've got something that works well enough.
00:28:35.599 --> 00:28:37.440
And they're also very easily replaceable.
00:28:37.599 --> 00:28:42.160
So you can very easily replace your like your observability layer with another observability layer.
00:28:42.319 --> 00:28:53.759
If it's easy for you, if you've built your tool, your observability layer, such that it's easy to plug in, which you kind of have to for Git adoption, then everyone's gonna build something that's easy to plug in, and therefore you've got no moat, like not a big moat.
00:28:53.920 --> 00:28:57.519
So it's there's a lot of people trying to compete for the same, it's like small subset of users.
00:28:57.599 --> 00:29:03.359
I think the way we took it differently was to say, okay, prompting's the new programming, but from first principles, what should the tools look like?
00:29:03.519 --> 00:29:06.720
And to us that really meant starting with the orchestration layer.
00:29:06.799 --> 00:29:09.440
That is like the most important part of all these agents.
00:29:09.519 --> 00:29:19.200
It's how do you combine the prompts and the tools and the domain experts and the user input into something that can build like genuinely useful AI agents.
00:29:19.359 --> 00:29:22.079
And so we have had to take quite a lot of that programming stack.
00:29:22.160 --> 00:29:38.400
And we so we started with the orchestration, we also did the deployment, so we're a little bit like Vercell in the sense that we we own, we have the layer that is the Next.js that is the wordware like programming language, and then we have the deployment where you like Vercel, where you deploy for Vercell, but you deploy your agent through WordWare.
00:29:38.559 --> 00:29:44.480
We're working on the GitHub for agents as well, so where people can fork and share and we get that network effects.
00:29:44.720 --> 00:29:54.640
I think really think about it from a perspective of like this is something new, let's treat it as something new, rather than like this is something new, let's do the obvious things that will be needed.
00:29:54.799 --> 00:30:02.480
Because if we own the orchestration layer, you can then add in the monitoring and observability and feedback and fine-tuning and all this other stuff.
00:30:02.640 --> 00:30:05.039
But you can't rip out the orchestration layer very easily.
00:30:05.200 --> 00:30:12.799
Like once you built using Langchain or once you built using wordware, it's very difficult to replace your orchestration layer.
00:30:12.960 --> 00:30:18.720
And I think the orchestration layer is honestly the nucleus that feeds into customers' opinions when they say, we just want to work.
00:30:18.799 --> 00:30:19.519
We wanted this to work.
00:30:19.599 --> 00:30:21.759
And yeah, it's that sounds it's easier said than done.
00:30:21.839 --> 00:30:25.279
But it's like once the orchestration layer is working, then it really can only get better.
00:30:25.599 --> 00:30:29.440
Yeah, and I think what matters is that how tight can you make that feedback cycle?
00:30:29.519 --> 00:30:31.119
And we saw this in self-driving cars.
00:30:31.200 --> 00:30:37.359
We saw this like if you can get more information from every drive you take, like, where did I go wrong, even if I didn't have a driver disengagement?
00:30:37.440 --> 00:30:45.119
And if I did have a driving disengagement, how quickly can I work out why the driver had to disengage and like improve that part of my stack?
00:30:45.200 --> 00:30:51.839
And if I do improve part of my stack, how quickly can I get that deployed such that when we can actually tell if the fix has improved it?
00:30:51.920 --> 00:30:55.759
It's it's all about how can you make responses faster.
00:30:55.920 --> 00:31:12.880
And we did that by like putting the domain expert in the driving seat so you can like edit and iterate instantly rather than having this kind of back and forth between the engineer writing the prompts in the code base, showing the results to the PM or the CEO, them going, hmm, it's good on X, it's bad on Y, the vibes are off, an engineer going, I don't like vibes, what am I doing?
00:31:13.039 --> 00:31:15.759
And like trying to optimize towards this fuzzy direction.
00:31:15.920 --> 00:31:18.799
It's like, no, just put the person who knows what good looks like in the driving seat.
00:31:18.880 --> 00:31:24.160
They can iterate like a hundred times in an afternoon and build something genuinely useful much, much faster.
00:31:24.559 --> 00:31:25.680
That makes a ton of sense.
00:31:25.839 --> 00:31:31.680
You probably see yourself partnership opportunities with the other agent infrastructure companies.
00:31:32.559 --> 00:31:34.079
Yes, maybe.
00:31:34.400 --> 00:31:36.480
Um so that's another thing.
00:31:36.559 --> 00:31:39.759
I'm seeing more and more companies go to other parts of the stack.
00:31:40.079 --> 00:31:41.440
Yeah, I mean, they are.
00:31:42.000 --> 00:31:46.079
Like when I look at my market map now, the market map started here and we're slowly converging.
00:31:46.400 --> 00:31:47.359
You you kind of have to.
00:31:47.440 --> 00:31:55.839
If you're not yeah, once you've done monitoring, you might as well do evaluation and then you add a playground.
00:31:56.079 --> 00:31:59.759
And then yeah, like it's no, I agree with you.
00:31:59.839 --> 00:32:03.839
Yeah, if you're monitoring logs and watching that, you might as well set evals, benchmarks as well.
00:32:04.000 --> 00:32:06.160
And like kind of when you're feedback, yeah.
00:32:06.480 --> 00:32:09.599
Yeah, I think that there's there'll be consolidation over the next few years.
00:32:09.680 --> 00:32:14.720
I'm not yet sure who I think will win, but then I'm also biased because I think we're gonna win.
00:32:14.799 --> 00:32:16.079
So uh fair enough.
00:32:16.240 --> 00:32:17.440
I don't know if this will be on the podcast.
00:32:18.240 --> 00:32:19.920
Yeah, what's next for you all right now?
00:32:20.000 --> 00:32:21.839
You just finished, I should have given context.
00:32:21.920 --> 00:32:26.079
You just finished a very impressive YC batch, just raised a very impressive round.
00:32:26.160 --> 00:32:28.720
Feel free to touch on that, or you can just touch on what's next.
00:32:29.039 --> 00:32:31.519
I think by the time this podcast is out, we will have announced.
00:32:31.759 --> 00:32:37.599
But we just I think we raised the largest round ever for a company in a YC batch.
00:32:37.759 --> 00:32:42.799
So we raised 27 million at a well, a big valuation.
00:32:42.960 --> 00:32:43.119
Yeah.
00:32:43.440 --> 00:32:45.839
And yeah, we are now, I guess, expanding the team.
00:32:45.920 --> 00:32:50.000
So we're looking for amazing kind of everything across the entire team.
00:32:50.079 --> 00:32:53.839
But like I'm I'm especially interested, invested interested in the technical team.
00:32:53.920 --> 00:32:59.200
So full stack product engineers, designers, we're like looking for great people.
00:32:59.279 --> 00:33:09.200
A lot of our team are ex-founders and love to really build in what is the new paradigm of like it's we're basically this is the new internet.
00:33:09.359 --> 00:33:16.880
There's a Microsoft scale opportunity to be built, and we are trying to take that opportunity and build something like a generational company.
00:33:16.960 --> 00:33:22.960
So if you like working on heart problems, you want to invent things that have never been done before, then come join us.
00:33:23.119 --> 00:33:25.519
And I guess what are we building next?
00:33:25.599 --> 00:33:28.640
It's kind of building that air operating system.
00:33:28.799 --> 00:33:33.119
So I guess I can I can go concretely into our roadmap, but that'd be a bit boring for the uh for the podcast.
00:33:33.200 --> 00:33:35.359
So uh show notes type of thing.
00:33:35.519 --> 00:33:36.000
Yeah.
00:33:37.119 --> 00:33:40.880
So lots of interesting and hard problems and not enough people to work on them at the moment.
00:33:40.960 --> 00:33:42.160
So in San Francisco.
00:33:42.319 --> 00:33:43.759
In San Francisco, in person.
00:33:44.079 --> 00:33:45.039
It's the best kind.
00:33:45.200 --> 00:33:51.279
Yeah, fully, fully in person, and we're I won't tell you where our office is gonna be because it's not yet 100% confirmed.
00:33:51.359 --> 00:33:53.279
But fully in person in San Francisco.
00:33:53.680 --> 00:33:55.920
Why don't we end with a quick fire round if that works?
00:33:56.240 --> 00:33:56.640
Let's do it.
00:33:56.799 --> 00:33:58.960
Okay, what's an agent trend that you're most bullish on?
00:33:59.200 --> 00:34:04.400
I'm most bullish on narrow AI agents, which I think has been under discussed.
00:34:04.559 --> 00:34:08.800
It's quickfire, so I'll I might stop there, but I think just quickly define narrow AI agent.
00:34:09.039 --> 00:34:19.679
Yeah, so so an uh like an agent that does one task, and these are somewhere between software and AI.
00:34:19.760 --> 00:34:30.400
They might have self-reflective reasoning loops, but they are only trying to do one thing, like write a blog post or a code email, or kind of write do one part of the code execution process.
00:34:30.480 --> 00:34:39.760
And I think narrow agents, you're able to get more reliability, you're able to use smaller, faster, cheaper models, or get squeeze more value out of the smarter, bigger models.
00:34:39.920 --> 00:34:43.679
They will be quicker because they don't have to work out how to solve the problem every time.
00:34:43.840 --> 00:34:56.239
I think one of the problems you see with a lot of these inter-agent, like auto-gen style agents is that they pontificate a lot and they think a lot and they like, you know, you get less reliable results, and there's a reasonably high chance they go off in some random direction and start self-prompting.
00:34:56.320 --> 00:35:03.760
And I think narrow agents are the way to go, at least with today's capabilities, but I still think that will be true as the models get better.
00:35:03.920 --> 00:35:07.119
You will always get better results from a narrow agent than you do from a general agent.
00:35:07.199 --> 00:35:16.320
It's just like you constrain the best AI out there to do something like smaller in scope, it will be better at that task.
00:35:16.800 --> 00:35:20.639
And what's an industry agent's our poised to disrupt that people aren't talking enough about?
00:35:21.119 --> 00:35:22.159
Oh god, everything.
00:35:22.320 --> 00:35:25.920
It's the entirety of knowledge work is about to be disrupted.
00:35:26.320 --> 00:35:29.119
What's your favorite agent built on wordware so far?
00:35:30.800 --> 00:35:33.599
Uh most of my favorite agents are fun.
00:35:33.760 --> 00:35:37.599
So we we have one that we use a lot in our Slack channel called Shrekify.
00:35:37.760 --> 00:35:43.840
And it takes in takes in some description and turns it into a Shrek cartoon.
00:35:44.000 --> 00:35:50.000
So uh yeah, that's I should definitely say something that's like one of our customers, but yeah.
00:35:50.320 --> 00:35:50.880
That's fun.
00:35:51.119 --> 00:35:54.960
Prediction what percent of enterprise companies will be using agents in 2025?
00:35:55.360 --> 00:35:58.000
A hundred percent that still exist in twenty thirty-five.
00:35:58.639 --> 00:35:59.840
That's a great note to end on.
00:36:00.079 --> 00:36:00.800
Thank you, Robert.
00:36:01.039 --> 00:36:01.519
This was fun.
00:36:01.920 --> 00:36:03.199
Awesome, that was a great idea.
