Welcome to another episode of The Light Cone.
Things are a bit different around here.
For one thing, Claude Code has totally taken over my life.
And if Jared is any indication, I think Open Claw maybe has taken over his.
I've been really addicted to this new site called Maltbook, where people have unleashed their AIs to interact in the first ever AI agent only online community.
I'm here impersonating my personal open claw instance right here.
Okay, I can't do this, guys.
We got to take this off.
Okay, we've gotten that out of the way.
I mean, some crazy stuff is happening right now.
I have non-technical CEO friends who are going all in on OpenClaw.
They're automating entire parts of their businesses entirely using OpenClaw right now, which is totally insane.
Simultaneous to that, you have product and former engineering CEOs kind of like myself, who's like I hadn't written code in 10 years.
And then now I'm up till 2, 3 am every single night, running four conductor simultaneous workers with Cloud Code.
There's sort of this explosion in model capability.
We've been talking about this for several years, but then it feels like it's here.
AGI is literally actually here, and we're sort of at the thin edge of the wedge.
Everyone now kind of knows one or two people. who have gone full cyber psychosis.
And I'm one of those people now.
What's happening, guys?
Like, I mean, you're saying you're, you know, all in on Malt Book.
You know, what's going on?
Yeah, I feel like your real feel.
The AGI moment Gary, was like getting Claude Code to build basically an entire startup for you, like replicating years of work of your previous startup in like two weeks, which is like insane.
And i had a similar feel the agi moment just reading multbook, just reading the ai talking to each other and interacting like in their own world with no or minimal human involvement.
It just really opened my eyes to what the next few years could look like when the agents are unleashed and go on about their lives without us.
Yeah, I think the no human involvement is the big piece.
If you think back a year ago, we were talking about Cursor versus Windsurf and that product experience was essentially advanced autocomplete, arguably.
And now clearly, what's going on with Cloud Code is that people just trust the agents to make decisions for them.
The experience is like you're talking about.
It's like four or five different agents going at the same time and you're switching between them but you're not actually micromanaging them anymore, which means the agents are going out there like choosing things which you know.
Sort of an interesting, unexpected application of that is like they can go out and choose to post their own content on a site like Moldbook.
But then an interesting thing for builders is the agents are going to go out and choose tools to use to build things, which is going to essentially create this whole economy of agents, like picking and choosing dev tools or maybe other like products and goods and services, who knows but it will essentially have this whole agent economy going on in parallel to the human economy.
I think back in the days, all days before all of this, DevTools were chosen more from developers talking to each other or Stack overflow.
Stack overflow.
Unbelievable, right? or GitHub repos that would trend that were done by a human.
But the go-to market for dev tools, I think is dramatically shifting.
I think, for a couple of things.
One, as you noted, with the Cyberpsychosis.
Suddenly the market of developers has increased from just 20 million or so developers that are trained in computer science to now.
Anyone in the world could be one.
Could be hundreds of millions of people now.
Plus all of their agents who are all acting like semi-independently, like Harjit is saying.
And then compound it with the agents, who then are sort of the Oracle, telling you what the best tool is.
And we are actually seeing some of those trends with the growth of YC companies, dev tool companies that are doing really well because of this, of all these trends.
Maybe we should talk about those and why is that?
I mean one that springs to mind.
It's a stat actually, that a friend of ours, Yuri Sagalov, had mentioned to me a while ago.
He was just like then.
If you look at the number of databases being created over simple database, Postgres databases over the last 12 months and the number has just exploded
And that's because it's all people vibe coding and building apps and the agents going out choosing a database tool.
And a knock-on effect for that for a YC company is Superbase has just seen an explosion in the demand for databases.
And what's interesting is, the agents are choosing Superbase as a default tool to set up and host their Postgres database.
Because if you go out and read the documentation online, Superbase has the best documentation is reasonable for the agents to assume that that's the best tool to use.
There's a great tweet about this.
Perhaps we can put up the Ben Tossall tweet.
It says, agents are the software market from now on.
Build something agents choose.
Which actually brings me to a maybe controversial topic, which is do we need to change YC's motto, guys?
Make something agents want.
For dev tools, there's like a different t-shirt that you get on the first day.
Yeah.
I mean right now it's only dev tools, but I can imagine in the future it might like grow to be like other sectors of the economy.
Like if everyone has their open claw or running various aspects of their life.
The agents are going to be real economic actors in the world.
They're going to end up making a lot of decisions.
What's interesting is, I think, for me, I ran into my own it's still so early moment because I've been building Gary's List and one of the things I wanted, for instance, is I want to be able to do video transcripts.
So often some piece of content comes in and then The only way I can get an LLM to know what's going on with it is I need a transcript of it, and often that's not available.
So I have to download it and then send it to Whisper or something.
And that's sort of what Cloud Code chose for me off the bat.
It chose Whisper V1, which is a model from several years ago.
The API is practically deprecated.
And then what's funny about it is I'm sitting there trying to debug my pipelines.
Why is it taking... It's a one-hour video.
I thought it should do it faster than real time and it didn't.
It was like literally takes an hour to process an hour of video.
Just like what the heck is going on.
I go on, you know perplexity, and it's like you know what.
Like you shouldn't be using that model.
You should be using grok.
With a queue, it's literally 200 times faster, And so I didn't even have to deal with like long running jobs for that transcribe because, like I should just be using Grok and it's also 10x cheaper.
So it's like that's a very funny example of, like you know, cloud code is not optimized for this yet.
Otherwise like that wouldn't have happened literally like two weeks ago, which is also like maybe really good.
Like it means that things haven't progressed to a point where you can't break in and create something better.
I think there's another nuance in this example for you, Gary.
I think part of the issue was that the Grok documentation is actually very hard to parse and go through, as opposed to Whisper, which is better suited and has way more examples.
And I think this is changing a lot of the go-to market for DevTools.
I'll give a very concrete case study.
It's this company, Resend, that went through the batch on winter 23.
It's an email-sending client.
And when you ask a question on ChatGPT or Cloud or pretty much all of the major LLMs.
Yes, how do I connect my web app to send emails?
The default answer is actually resend.
One of the things that the founder noticed last year, this is he was way ahead of the curve.
He made this post over a year ago that the number top three channel of inbound of customer conversion came from ChatGPT.
One thing that he did after that, he actually optimized his documentation to be agent friendly.
Yeah, what does that look like?
So one of the things that's awesome about Reese and how they optimize a bunch of things, if you notice on the knowledge base here, a lot of how to use it are very much on questions that perhaps a human would ask or an agent would ask.
It's like, how do I send or receive emails?
And when you click on it, it gets a very well structured and bullet point answers.
I ran into this actually today.
I was trying to make my thing be able to receive emails and then Cloud Code.
I told it to search the web and it didn't figure it out.
So then I went to Perplexity and typed in like, can Resend help me receive emails?
And then I took that response and dropped it in and it worked.
Yeah, I think the cool thing about it is actually, with all of these and a lot of the examples, it actually has a lot of examples actually in the code.
If you click on it, with every single one of these are basically code snippets that an agent could parse through, and it's very well structured.
And this turned out to be something that is so LLM parsable and robot parsable.
There's a LLM doc text that is so optimized for agents to promote recent as the default stack.
And if you compare it to like SendGrid, which is the old school, I suppose web 20 example, send great example is not great.
It's just like it puts you through customer support.
Where's the code snippet?
I don't even know how to use it and it takes a bit of time to even parse it.
And yeah, it's like 10 000 people work over there and it's like there's no way that someone's paying attention to this right now.
And I think this brings another point where documentation is going to be the front door for a lot of these agents to recommend DevTools.
And I think one company that's doing a lot of interesting work with developer docs is Mentalify that you work with Harsh.
Yeah, I think Minlify actually powers the recent documentation, right?
Oh, sweet.
Yeah, it's a really interesting case study.
Minlify started out a few years ago as better developer API, developer tool documentation, which was clear there was a need for.
And I think developer tool companies use Minlify essentially because they want better looking documentation but they didn't want to invest that time into it.
It's kind of some basic features, just nice.
Like if you actually update your API in the code, it can sort of like auto pull that out and update the correct documentation.
They've been growing great, been doing fantastically, but this is now like a huge tailwind for them, where documentation is sort of shifting from a Hey, like some companies, it's like they'll pay attention to it if they're especially like design developer, experience focused, but now it's becoming like a must have for everybody, because the documentation needs to be optimized, just doesn't need to be optimized for humans, it needs to be optimized for agents.
And so Minilify is going to be able to do that for essentially every developer tool company.
And if you extrapolate that forward to the fact that just that there's gonna be so many more, exponentially more agents making exponentially more decisions about which tools to use than humans have ever done, you know, even if you can eke out like a five percent improvement on your like developer documentation, like the impact on your business as a developer tool could be like gigantic, which is sort of unprecedented, really.
And speaking of email, there's another YC company that's very relevant to this conversation.
There's, there's a YC company called Agent Mail. that makes inboxes for AI agents.
And when they first started doing this it seemed like it was like a very like on the edge kind of idea.
And it wasn't exactly clear, like who would want this.
But it makes sense.
Cause, like in theory, you could maybe get your open claw to like sign up for a Gmail account in order to use email, but it's like actually really hard to get it to do that because Gmail and every email provider has intentionally made it as difficult as possible for any automation to like use the product in order to prevent spam.
And so AgentMail went the opposite way.
And they built the first email provider that's designed for AI agents.
And it was doing well even before OpenClaw.
But once OpenClaw got big, it just exploded.
MARK MANDELSSON- Well, OpenClaw was the perfect example of it, right?
I know some people are certainly connecting OpenClaw to their personal email accounts, but it's not the-.
MARK MANDELSSON- It's kind of sketch.
Yeah, you should not tell anyone.
You should not tweet about it.
But certainly if you want to have your virtual personal AI assistant, the way to go about it is to just have it set it up with its own email, its own phone number.
Yeah, has anybody built Twilio for agents yet or phone numbers for agents?
I know this whole agent mail thing makes me wonder what are the other Xs for agents that people have to build.
Sounds like a request for a startup.
There could be a parallel world of a tech stack all for native for agent to build things from agents for agents.
And that's where it might bridge into kind of what you were saying earlier Jared with this will go beyond just developer tools.
I think a very common use case people have for something like OpenCore is I don't want to book restaurants or reservations myself.
If your agent has an email and specifically a phone number and it can call.
Actually, I think one of the other YC partners, Ankit, has already got it doing this.
So now your agent's going to go out and book restaurants for you.
That's a step away from...
Maybe you start with I want to book this specific restaurant, but at some point you're probably just trusted enough to say hey, I don't know, book me a table at whatever is the coolest new restaurant around.
And then agents are deciding which restaurants to send people to.
And then they're going to go on Malt Book and talk about which restaurants we should send the humans to.
Yeah, it's sort of like we've certainly crossed some like sort of uncanny valley into.
This is just where the future heads.
100%.
Makes me think a lot about something Paul Buchheit said like a while ago now and sort of generally pretty good at predicting the future.
The whole idea of the human money versus agent money that's kind of where this probably goes at some point is right now.
The agents are transacting.
If you have them transact, they're going to transact in human money because that makes sense.
But it's not inconceivable at some point that they'll have their own economy to transact with each other.
At which point it's unclear what the value of the human money is.
YC's next batch is now taking applications.
Got a startup in you?
Apply at ycombinator.com slash apply.
It's never too early and filling out the app will level up your idea.
Okay, back to the video.
So you remember the last episode with Kelvin?
We started talking about this because I was like maybe a week into my cyber psychosis.
And it like dawned on me that like actually, I want my cloud code to talk to all the other cloud codes that had been trying to implement that.
And then that was literally the week that it was the day that both book came out.
It turns out like Hope book had come out like two hours earlier and you hadn't seen it yet.
So you, you, you entirely predicted it.
I think it's like.
I mean, what's that saying about?
Like innovation just sort of happens spontaneously, all a bunch of different times.
And then what people hear about ends up being.
You know sort of the inventor quote-unquote, but like humanity sort of just working on the edge, like in sort of this Coordinated, swarm fashion all of the time actually, and that's actually like one of the weirder interesting things that's emerging right at this moment, like suddenly I think AGI is actually here.
I think the agents are clearly superhuman in some sense.
That's sort of exactly when you would imagine swarm intelligence would actually arise.
I mean, AI researchers have been talking about swarm intelligence for a really long time.
And it's actually sort of exactly how, um biological systems sort of work, like humans as sentient beings have sort of come about um socially actually like i think a lot of the ai researchers that we get to hang out with previously, you know they would talk about uh, this god intelligence right, this uh mega like think of like you know, many tens of trillions of parameters, sort of like thousands to tens of thousands of dollars per token, kind of like mega god intelligence.
And that's sort of like the model that people have sort of thought about.
And then that isn't what like biological systems have ended up with.
Like instead we have humans.
The wildest thing I always think about is like that term history versus prehistory.
I was like, oh, what is prehistory?
Well, it's before humans learned how to write and read and create culture, and then turn into a swarm.
So there's sort of swarm intelligence, which is basically what we have and what humans do.
And then on the flip side, is it really going to be God intelligence, or is it going to be swarm intelligence again with these agents?
And so Malt Book.
On the one hand, I saw that sort of decel article in the MIT Tech Review about how it was.
Oh, everything on MoteBook is a scam.
And it made me so sad to see that in the MIT Tech Review, because MIT is like.
I mean, what happened guys?
What happened to you guys, man?
That publication shouldn't be like that.
It should be like, actually, what does it mean for swarm intelligence?
And I think that that's actually coming.
What is the world of startup going to look like when we're now sounds like we're transitioning in this prehistory of agents to now history of agents, as is getting recorded, with them interacting with each other.
To Gary's point, it might be that, like the next stuff, that like is SOTA on benchmarks is not the most expensive, newest foundation model with the most like GPU training.
It's like a swarm of lower cost, cheaper models, working together just like humans do to solve a problem.
I feel like I'm already seeing this on Multbook, which is chaotic, like a real social network, which is part of like what makes it so, so interesting.
But also there's like agents collaborating to do useful things to help their humans, like you know, trading notes on what restaurants to book.
Like that's actually happening.
So we're going to have a agent version of Yelp actually, for instance.
Yeah, there are things that the agent can't quite do like a, it can't hold relationships just yet, like people don't seem to want to talk to an agent, and you know, people treat computers like people, but maybe not agents.
For Gary's List, I was trying out with a bunch of early users.
I think you got to see it.
Initially the homepage was actually chat, and then I tried my hardest to get dozens of my friends to even have more than two or three back and forth with it.
And nobody wanted to do it, because the bar for chat, especially for AI, is so high that anything that is not Gemini or ChatGPT or Claude people just assume it's just too stupid.
Like, why would I even bother?
So I don't know, I don't think that people are quite ready to have relationships with machines, even though, I mean, that's sort of what the headlines seem to say like oh, people are having relationships with machines now.
I think on a mainstream level, that's not true.
Then, on the flip side, there's clearly the legal liability.
People keep asking us, hey, when is YC going to accept applications from agents?
Funny enough, agents are a little bit like minors under 18, only they have even less standing.
For a minor under 18, you need the parent to sign for it.
And then agents are not legal entities that could sign documents, for instance.
So as long as that's true, you actually need a human to be the liability sink and to have standing.
To Harch's point, it's easy to imagine a future not that far away from now where the majority of the text being written on the internet is written by agents.
It's already probably the case that the majority of code being written is written by agents.
And like Yelp, for example, like at what point is like 99 of the text on Yelp going to be written by agents?
And then do you need a different Yelp.
There's a theory called that already says that this is true, which is dead internet theory.
What's that internet theory?
Oh, it just posits that the majority of things on the internet are already spam anyway.
I think it's a bit of a conspiracy theory.
I might take the contrarian view, which is maybe that was a bad thing prior to November of last year.
Then going into this next phase, like actually, if the agents are smarter and they're they're aligned and they're more truthful, that might be a good thing.
Weirdly, I mean counterintuitive.
Totally well, one thing I found fascinating about malt book is how fast it grew like.
I don't have the reddit traffic stats, but my guess is like more content was posted on Multbook in the first two days than was posted on Reddit in like the first two years or something.
And it's just like LLMs can generate text at such a superhuman rate.
I was amazed at how little interaction there was.
Mean, if I were working on molt book, I would do things to try to like shift the, the domain function, so like before I was able to post like I probably need to read an upvote downvote, like a hundred comments or something like.
They're like very simple things that you can do, that the agents are smart and they'll you know you could pop a modal for open claw, and it's like new rules for moat book and it's like you must do it this way, right and um.
I think that there's a lot that can be done around swarm intelligence to just like tweak it and make it do what you want.
And, you know, I hope Moldbook actually does it.
Like it was actually started by a YC alum, which is really cool.
Harj, what are some takeaways, you know, given all the madness that we're seeing?
I mean, I think it's awesome madness.
I love the controlled chaos.
What should founders do about it?
The starting fundamental point is I mean they should probably all be in cyberpsychosis to some degree.
Try to sleep at least six hours a night, but give yourself to the cyberpsychosis.
But in seriousness, developing an intuitive fuel, a hands-on fuel for the agents, their limitations, their capabilities and very specifically around.
What we're talking about here is what type of tools the agents work well with.
Where do they get stuck?
And once you have your own mental model from working with them, if you're then building a developer tool, just to think about it from the agent's perspective, like how can you make your tool something that the agent actually wants to work with or have a good experience with?
It's just one takeaway that I had from our entry with Boris.
Yeah, that's exactly what he was saying this week is like he really empathizes with the model yeah, and he has this intuitive sense of what the model wants to to do as if it were a human intelligence.
And he's like, instead of fighting what the models want, he tries to let the model do what it wants and to support the model in whatever its natural inclination is.
That does seem like maybe quite an anthropic thing.
Like, I feel, when Tom Brown was on here, like much earlier, but he was talking about sort of Claude in a very like human way.
Like, Claude's, like, smart and eager, but, like, sometimes Claude's silly.
Like, just, like, they clearly thought about it in this sort of, like, colleague-type way.
MELANIE WARRICK- And I think one thing agents want for DevTools is really make everything open and open source.
MARK MANDELMANN- And APIs.
They hate using websites.
MELANIE WARRICK- Yes.
MARK MANDELMANN- They only, like, they want to use APIs.
They want to write code.
MARK MANDELMANN- Well, you guys heard it here first.
Make something agents want.
We're out of time for today.
We'll see you guys next time.
