← This Week in AI
Loading summary
Transcript123 lines
- [00:00]
Richard Socher
I do think every time we've seen automation waves, there's short term disruption, there are waves of Luddites. Imagine all the economic outputs of all the car crashes before self driving cars, all the injury lawyers that are making money, the hospitals and the emergency rooms making money, the insurance making money. Everyone's making money on car crashes. And when you take car crashes out, indeed there will be less economic value in a weird way, but obviously it's an improvement. But ultimately we can answer the question like why are there no more rags to richer stories? Because no one has to wear rags anymore. No matter how poor you are, you can have clothing without holds in it.
- [00:43]
Tanay Kotari
Why?
- [00:43]
Richard Socher
Because automation made it cheaper and cheaper with machines.
- [00:48]
Jason Calacanis
This Week in AI is brought to you by Notion. Bring all your notes, docs and projects into one space that just works with AI built right in. Try notion with notion agent@notion.com twist and quadratic bringing the productivity boost of AI into your spreadsheets. Visit quadratic AI twist to sign up and use Code Twist to get one month free of their pro tier subscription. All right, everybody, welcome back to this Week in AI. You thought I was gonna say this week in Startups or Welcome back to all in? No. This is my new podcast called this Week in AI. This is episode two. What is this Week in AI? Kind of think of it as like a hybrid between this Week in Startups and all in roundtable format only. Experts, me plus two or three experts in the field. We're going to rotate for the first 20 weeks or so a bunch of different experts. We're going to talk and go deep into the issues around AI. Every week you can get our daily update on all things AI at this Week in AI AI ThisWeekinAI AI. Also, you can find us on YouTube, YouTube.com ThisWeekinAI and if you wanna subscribe on Spotify or Apple, there's QR codes on the screen right now if you're watching the video on YouTube. If not, just go to ThisWeekinAI AI Spotify or ThisWeekinAI AI Apple and you'll get taken there. We're on Twitter as well. This week Nai. This week the letter Nai. Follow us in all those places. All right. This week I am very, very excited because we have guests who have great products. The first is Richard Socher. Richard, you're the CEO and founder of U.com y o U.com, an incredible domain name and recursive AI. And you do search infrastructure for enterprise teams. We'll talk about that in A moment. And Tanay Kotari is here. I know you, Tanay, because I am a fan of your product Whisper. And Whisper is voice detection software. Now you're probably saying, wait, isn't that like built into every product or service? It is on your phone, it is on your Windows desktop, your Mac desktop, but it typically sucks. And Whisper actually works and it understands context. And congratulations, Tanay, you just launched on Android. I understand.
- [03:14]
Tanay Kotari
Yeah, yeah, that was yesterday.
- [03:16]
Jason Calacanis
Congrats on that. Company's doing great. Just ballpark how many users are using it now and why are they using it? Why are they using Whisper as opposed to the built in dictation that we have on various devices?
- [03:30]
Tanay Kotari
There's a few million people at this point.
- [03:33]
Jason Calacanis
Oh, wow.
- [03:33]
Tanay Kotari
We launched about 14 months ago. And the biggest reason is for people, it's just so much more natural to speak. Every voice dictation software so far hasn't been built the way people would like it to because you speak very differently than you write. And so what we aim to do with Whisper is solve an extremely simple problem. You just speak naturally with your rambles, with your ums and us, change your mind throughout it. And Whisper will produce something that's ready to send and it's contextual. So if you're in an email, it's formatted like an email. If you're in text message, it's casual with the names. Right. And the slang. Right. And it just. We spend a lot of time on getting that level of perfection. So people are like, hey, this looks fantastic, let's just send it. Unlike Siri or other products where you get a lot of mistakes that are
- [04:21]
Jason Calacanis
made and Yeah, I mean, Siri is so bad at doing dictation. I mean, it's almost like not worth using. I think a lot of people have given up on it and it costs like 10 bucks a month. What, 100 bucks a year? 200 bucks a year for Whisper these days?
- [04:35]
Tanay Kotari
Yeah, it's $12 a year and the dollar a month in the US and each country has its own pricing, so that's something we actually rolled out last year because Whisper had users in 140 different countries. So we just have local pricing.
- [04:52]
Jason Calacanis
Yeah. Now, one thing I wanted to ask you about. I see with the Open Claw and Revolution. We'll talk about that on today's show as well, that some people are getting frustrated. I know I am when I'm talking to my OpenClaw agent and we've been talking to all kinds of startups on this week. In startups, you can go Check it out for folks listening, you really need to talk to it and it understands what you're doing. So it's a pain in the neck sometimes to stop typing or whatever you're doing and initiate your microphone. There's a little bit of friction there. And I see people are using pedals. So I guess here's an image of a foot pedal hooked up to Whisper that we found on the Internet. Talk about this modality and if this is just three or four weirdos or you think this is going to become a persistent thing, or maybe you think the future is going to be devices like this Plodpin, which I've been testing. It's recording right now and I wore it when I was skiing last week and left it recording the whole time so I could give it a notes, random notes on things I needed to do and gave it some action items. It was pretty effective for that function. But tell us a little bit about foot pedals and how you think about that.
- [06:11]
Tanay Kotari
Whisper is at this point today where a lot of people have pretty much stopped typing on their computers completely. And so having a keyboard makes less sense. So what people are doing now, they're either getting foot pedals, they're getting rings with buttons on them, or they are getting these mics that have a button on them, like these DJI ones, and they just hold it up with their mouth and talk to it. And so you can get the $16 foot pedal from Amazon, right? Plug it into your computer and you can almost be like, lean back like this and just go to a whole app with your foot.
- [06:46]
Jason Calacanis
Amazing. And the ring, that's an interesting one. So the ring initiates the microphone or it is the microphone?
- [06:53]
Tanay Kotari
Both. I want to see played with both of them. And I'm extremely bullish about ring being one of the next big form factors for voice. Because if you think about it like the Apple watch, it's kind of weird, right? You have to hold your elbow weirdly up in the air to talk to it.
- [07:08]
Jason Calacanis
And even with ridiculous, you look ridiculous.
- [07:11]
Tanay Kotari
But a ring, you just do this and it's beautiful.
- [07:14]
Richard Socher
Maybe the pin will come back after all that. You main pin.
- [07:18]
Jason Calacanis
It's worth discussing. That product was way ahead of its time. What do you think it got right, Richard, and what do you think it got wrong?
- [07:26]
Richard Socher
I think the timing was tough. The eye wasn't quite there yet. The price point was not right. It had this cool sort of projection, but it wasn't quite high res enough. And ultimately you could just put your phone into a pocket like Here and then it felt very similar. So. But I think in the right hands, this kind of form factor could make sense. I've actually seen foot pedals first be used by some of the most sophisticated radiologists. Those radiologists use two mice foot pedals, voice recognition, tons of shortcuts to go quickly through radiology scans.
- [08:01]
Oliver
Ah.
- [08:02]
Jason Calacanis
And they would use two foot pedals. That's interesting. So they might have been using it to navigate or move around or go to the next slide. I bet that's right. And then one of them was probably to turn on the microphone and initiate it. Super fascinating. You're also an investor, Richard. I understand.
- [08:18]
Richard Socher
And yes, a very happy investor in today's company too.
- [08:21]
Jason Calacanis
Oh, okay. Tell us about AI ventures. And yeah, what you, what you do, it's obviously AI focused.
- [08:26]
Richard Socher
Yeah, that's right. AI plus X. You know, basically we look for seed stage AI companies. It started with some of my students and interns and friends and employees from my first startup, MetaMind, where I was lucky to invest in Hugging face at a $5 million valuation, which you don't see very much anymore. And then we started scaling it up and making it a proper venture fund with institutional investors and now happy investors in amazing companies. AIX Venture is pretty incredible. Excited to be in the seed. Rounds of Hugging Face, perplexity, weights and biases, Whisper Flow, Tollbit, Ambience, Windsurf, a bunch of other fun companies.
- [09:05]
Jason Calacanis
Nicely done. Nicely done. Well, I think the number one story everybody will agree, is a fictional piece this week. It was written by Citrini. Citrini, I think, is the name of the research firm. I'd never heard about them before today, but they did a farcical that they sort of pretended was written in June of 2028. So we're here in February of 2026. They put this out basically 18 months from now. And in this farcical post, this future post, they basically set it up that the S and p was near 8,000 by October of 2026. Unemployment was low, productivity was booming, and GDP was printing 5 to 9%. This is something I think we could all see and agree with. And they basically stated, quote in every was exceeding expectations. And the market was AI. The only problem, the economy was not. By June of 2028, unemployment in their future vision hit 10.2%. That's up from like the 4% we've been experiencing for the last couple years. And the S and p was down 38% from the highs. And that there was a ghost GDP output showed up in the national accounts, but it never circulated through the rest of the economy is the premise. And that the intelligent displacement displacement spiral and the quote intelligence displacement spiral became a feedback loop with no natural break. Basically AI improved so white collar people got laid off, Workers spent less on consumables. We live in a consumer driven economy and on enterprise software margins tightened, companies buy more API and this basically resulted in the economy collapsing. This farcical substack that came out on Sunday this week then caused the stock market to absolutely get crushed. And it basically in people's mind has created a doomerism loop. Anything that's gonna get touched by AI, the public markets, or some segment of retail in the public markets, I suspect thinks that this is gonna be a crazy headwind for everything from IBM to Salesforce. What was your take on it, Richard? If you wanna driven founder, and trust me, you do, you're going to need to spend some time in spreadsheets building models and doing projections. But so many of these spreadsheet programs are stuck in the 90s. Thankfully. Now there's Quadratic finally bringing the productivity boost of AI into your spreadsheets. But this isn't like some simple chatbot in the corner who can answer your questions. No, this is an AI native platform that handles all the number crunching and organization for you. You just describe what you want to do with your data and Quadrat Quadratic makes it happen right there in the spreadsheet. Now you can get insights about your business without fighting formulas and you can immediately share your results with your team and all of your collaborators. No setup or payments are required upfront. You can just start using Quadratic right now. It's going to blow your mind. Visit Quadratic AI Twist to sign up and use the code Twist to get a free month of their pro tier subscription. That's Q U A D R A T I C A I Twist. Quadratic AI twist.
- [12:34]
Richard Socher
I think the markets are very jumpy right now. I think they overestimate the speed of some of these changes. As much as we're in our bubble, we see that a lot now. There are cases where actually making something better with AI could result in less revenue in the overall economy. I'll give an example. Self driving cars. Imagine all the economic outputs of all the car crashes before self driving cars, all the injury lawyers that are making money, the hospitals and the emergency rooms making money, the insurance making money. Everyone's making money on car crashes. And when you take car crashes out, indeed there will be less economic value in a weird way. But obviously it's an improvement. I do think every time we've seen automation waves, there's short term disruption, there are waves of Luddites, and we'll see those many waves of Luddites coming up from AI. But ultimately we can answer the question, like, why are there no more rags to riches stories? Because no one has to wear rags anymore. No matter how poor you are, you can have clothing without holes in it.
- [13:47]
Jason Calacanis
Why?
- [13:47]
Richard Socher
Because automation made it cheaper and cheaper with machines. And so I think it's kind of a ridiculous prediction. I think we can predict actually which kinds of goods and services we will have access to. And those are the ones that only currently wealthy people have access to that are bottlenecked on intelligence. So we will all have access to personal healthcare teams, personal tutors for our kids, personal assistants, and that will make most people more productive.
- [14:15]
Jason Calacanis
Yeah, this is the abundant side of the argument. Tanay, when you saw this piece, what rang true about it? What rang false about it?
- [14:25]
Tanay Kotari
You know, when I first read it, the first thing that came into my mind was this notion of ghost gdp, which makes a lot of sense, right? So it's AI models paying other AI models. And then most of the money that's flowing through the system is between all of these AI tools. And after a couple of hours, when I was thinking about it, the thing that really hit me is you could have written this article about 100 years ago about the Industrial revolution, but people would have been terrified. Like, hey, money is just going to go from like one factory to another factory and like, it's not going to touch the people and that's how the world is going to be. But hey, that is true. But that's actually incredible because that significantly multiplies the GDP and it changes. But the thing that I believe about people is people adapt. Like humans adapt so well to new situations, environmental changes, how the world works so quickly. Like when Covid happened, within a couple of months, we were all living in a completely different world than we were before.
- [15:29]
Jason Calacanis
And society carried on. We figured out how to live with deliveries and wiping our groceries with Clorox wipes before we brought them into the house.
- [15:38]
Tanay Kotari
Yeah. And so AI is a magnitude or two faster than the Industrial Revolution. And it's going to cause a lot of change, a lot of chaos. People hate change. People hate chaos. But I think when all is said and done, in a few years from now, I think we'll figure it out.
- [15:55]
Richard Socher
It's part of the lump of labor fallacy, actually, what we're seeing often that people think labor is this fixed lump. And when you take 30% of labor away with some automation or AI, then 30% of people are unemployed. But that's just been proven wrong over and over again.
- [16:10]
Jason Calacanis
It creates what's called a cognitive surplus. People then have more bandwidth to do more things. So then the question becomes, as an average office worker, if you automate half your chores, which my team has been doing quite effectively, then you are faced with either doing nothing and leaving work at 1pm or saying, I'm going to make three more clips from this podcast or I'm going to write a blog post about the blog, about the podcast, or maybe I'll start another podcast. People will fill their time. And the cognitive surplus in history created a lot of very interesting things. Wikipedia used to talk about this. You had all these intelligent people in the world. They had a little bit of extra time and then collectively they built this incredible thing called Wikipedia. You now have a lot of cognitive surplus in the world with developers. Maybe they're getting their job done quicker and then on the weekend they work on open claw skills or they build out a side hustle. So if the human spirit continues as it has since the beginning, the cognitive surplus creates opportunities for other businesses. And one of those businesses was actually, actually DoorDash and Uber, which get mentioned in this piece, people had extra time to go do things. And that became an attractive opportunity to be able to make 20 or 30 bucks an hour. And, you know, if you have four extra hours a week and you can make a hundred bucks and then pay down your credit card or, I don't know, buy some extra comic books, whatever you use your surplus cash for, people were drawn to that. And this is where I think this doom post, this Dr. Doom post, kind of misses things specifically, kind of show some real lack of expertise in marketplaces. So that was the part where I was like, this doesn't make much sense. I'm not sure who wrote this or what their background is, but quote, the DoorDash moat was literally, you're hungry, you're lazy. This is the app on your home screen. An agent doesn't have a home screen. This is the first time in history the most productive asset in the economy has produced fewer, not more jobs. I thought this was completely naive. And I'll open it up to both of you because what's unique about DoorDash isn't exactly the software. It's a beautiful interface and you know, they match jobs and you know, it's very fast and great, awesome. And it's global now. But the relationship with the restaurant and the trust built with the user, that's actually what makes the network effect work. So your thoughts on the DoorDash section, either of you?
- [18:45]
Richard Socher
I think at a high level, DoorDash is a great example of the kinds of apps that you can cannot just vibe code. A lot of people say, oh, software is all going to go to hell and everything is going to vibe code. People are not going to vibe code. Things where permissions are mission critical so far. Where you need a two sided marketplace, where you need these complex network effects, those things are very hard. Sure, you can get a quick app for Instagram, like an Instagram clone, right? But if you don't have millions of users actually using Instagram, then it's not going to be a very important, interesting app to use. And So I think DoorDash is a perfect example of that as well.
- [19:19]
Jason Calacanis
How does it affect how you're running your company? Tanay and any thoughts on the DoorDash one? I didn't get to get your take on marketplaces, but also running a software company, people in this doomerism would say, oh, I could just create whisper, I could just create any software. I can create notion, I can create slack. Your thoughts on this sort of everything can just be vibe coded.
- [19:39]
Tanay Kotari
Yes, you can create a V0 of it. Like honestly, if you today wanted to go make a notion clone or whisper flow clone, it's going to take you two hours and you can just do it. But hey, here's the real thing. Once you do that, you will actually realize the key pieces that make these things really hard, that make these businesses really hard. Same for DoorDash, right? You make this thing and then you realize like, hey, I actually got to go talk to all of these restaurant owners and I got to build a brand that people trust and I got to figure all of the other things out. DoorDash is actually a three sided marketplace which is even harder to build than most other marketplaces that you have to. And so that's, I think one of the things that is all the grunt work that goes behind the scenes in building these businesses. That is really hard. Voice for flow, for example. Right. The hard thing isn't building this voice to text that works everywhere. The hard thing is getting about all the different edge cases is about how do you make it so seamless that my dad or my grandfather can easily use this product. It's about, hey, how do we make sure that a person doesn't get the feeling of loss. A person feels that delight like those Things are really hard. Same with openclaw, right? Openclaw is out and it caused a lot of forward. People are super excited about it. But hey, guess what? 99% of the human population is never going to set it up because they're not used to. They're not used to setting up systems like this. They're not used to installing the skills. Like, I know a lot of people in my family, like my dad, my grandparents, right? They're not going to do that. And so the really hard thing with OpenFlow, actually, and we get hundreds of requests a week, it's like, tanya, can you integrate with OpenCloud and build like a voice openflow thing? And I was like, yes, but that's not the hard problem. The hard problem is how do you actually make that something that is seamless to use by everybody? So it's just one tap and it works. And you remove all of the technical complexity that comes over it.
- [21:41]
Jason Calacanis
We're done hiring new humans at launch. Okay. Because Notion's new AI agent is like having twice as many. I'm not exaggerating here. It's like doubling your team size because the AI has been integrated into your Notion knowledge base right in your workspace. So things that used to take a researcher or an operations person, you know, 24 hours, 48 hours, 72 hour return time, maybe even a week, gets done in minutes. For me, Notion brings all your notes, your docs and projects into one connected space that just works. It's seamless, flexible, powerful and fun to use. With AI built right in, you spend less time switching between tools and more time creating great work. And with Notion Agent, your AI doesn't just help work, it finishes it. For example, we wanted to reorganize the Twist 500 list. These are the top 500 private companies. But to actually improve and refine the list, we had to remove all the companies that had exits. So we just asked Notion AI to do that for us. And it did it. And we build the docket every day for this week in startups, for this week in AI, and my notes for all in in Notion. And Notion's AI agent makes that completely searchable. So my producers or I can talk about guests, ask questions, make sure I have the ad reads incorrectly for each segment, and if I'm just looking for highlights, I can just ask a simple prompt. Notion is our system of record. It's where it all comes together for our organization. And then notion added these AI features that literally have made product 3, 4, 5 times more powerful for the same Price Try notion with notion agent@notion.com twist all lowercase letters notion.com twist it's in the show notes. Try your new AI teammate notion agent today.
- [23:24]
Tanay Kotari
And so the great thing about wipe coding that I love is what you thought of previously as the boiler parade. Setting up login, setting up a database, making sure things are syncing well together and all of that, that work can be taken care of. So you as engineers, when you are like problem solvers in general, you get to solve the really hard problems that matter that it would take years for you to get to otherwise because you're just spending time in boilerplate.
- [23:59]
Jason Calacanis
Yeah, this is, I think one of the things people are not recognizing which is smart people, entrepreneurs, they will when they get through their chores, the easy stuff like, oh, I need to have a help page. I need to have, you know, I need to write job descriptions for my next hire. I need to do accounting, I need to do my cap table. As this stuff gets easier and more abstracted away in terms of running a company, you release extra time. Again, cognitive surplus. So what do you do with that? Well, you refine the product in ways that are uniquely human. The design of it. Oh, the AI is going to make better design great. Give me your best designs and then I'll take them 20% better from there. And there'll be more features to add and there'll be ways to make it more reliable. So this concept that SaaS software is going to be static and it's not going to improve and therefore people are going to rip it out and make their own. Well, if the cost of the SaaS software is low enough, I do think people will be like, well why bother now if the software is too expensive? What I do think is going to happen. So this is where I think the market is getting some of the reaction. Correct. The cost per seat and your renewal is going to be very different. When you talk to. I don't know if I'll just take Slack as an instance. When we go back to Slack to do our yearly or two year contract, we're going to say, well, you know, our Open Claw AI replicants set up Matter post for us an open source version of Slack and it works pretty well. And you're charging us 30,000 a year. We can set this up and maintain it for 1000. Can we meet somewhere in the middle here in terms of cost? So we'll have some negotiating position that we'll be able to use to compress the cost, but that creates efficiency across the whole economy. And it forces Slack and Salesforce to make a better product that then counters us saying, well, we don't need all your features. Now, I could see us getting rid of a third of our SaaS spending if they don't change and they don't evolve. But I don't see a reason to rip out Slack right now. Unless the new version that we went to, the open source version, let's say, or the one we Vibe coded at our 20 person firm was materially better. Because the cost of this stuff, Richard, is so low already, right?
- [26:15]
Richard Socher
Yeah. So a few thoughts. I think defaults are really, really powerful here. Also, 80% of iPhone users, I heard this crazy statistic and are still working in consumer land. 80% of iPhone users never change any settings on their phone. It means whatever comes as a default is just going to be the default search engine and so on. Now, in terms of Vibe coding, you cannot just Vibe code a search engine either to actually implement an index and crawl the whole Internet, that costs tens of millions of dollars or more. So that's not going to happen. So we're seeing this reinvention of infrastructure that AI can use. Like GPUs first were used for gaming, now for AI browsers, search engines, they can all be reinvented for AI as the main user. Now, when you talk about cognitive surplus, I think the biggest thing that people don't often have is agency. They're so busy with their lives and just doing the robotically sort of the things that once they free up, there's going to be a new skill to actually get people excited about. Hey, now with this all, what are you going to do with all this in your free time? That kind of mindset will require a lot more agency from people. And then in terms of what you mentioned about the markets and what they're getting, right, with SaaS, it is true, for instance, that when you look at China, China has much less of a developed SaaS ecosystem and economy because labor was so cheap that a lot of times companies would just build their own tools instead of buying them. And now with AI and Vibe coding, some of those tools will get cheaper too. But the more complex they get, the more mission critical they are. The less you're going to vibe code it and the more repetitive the processes are. And then also in some cases in SaaS, you want your employees to follow a certain step by step procedure, right? Because you're in a regulated industry that requires you to go through those steps. You cannot just let them Vibe code it and then automate Some way.
- [28:10]
Jason Calacanis
This is a key piece. If you vibe code it and information leaks or gets hacked now you've got a whole other level of problem. There was an interesting quote in here I want you guys to react to. And this, this I think got a little bit panned. The marginal cost of an AI coding agent had collapsed too. This is in their Future World in 2028 had collapsed too. Essentially the cost of electricity. Where do you think you fall on this one? Either of you?
- [28:38]
Richard Socher
Yeah, high level. I'm very excited to bring electricity costs down. I think electricity and energy are at the very core of so many things a lot of people like, oh, we have a shortage of water. We don't have a shortage of water. If you had infinite or very, very cheap electricity, just desalinate ocean water. Right. So there are a lot of these things that are ultimately energy problems. And I'm personally excited to get the marginal cost of intelligence to be closer and closer to that of electricity. And I think that will then show that the bottleneck will be the agency and creativity of the people and less sort of the ability to execute on good ideas.
- [29:19]
Tanay Kotari
Coding agents, very similar to LLMs are going to be one of the things that a large fraction of people will start getting on and start using as going to be part of their day to day life. When something like that happens, when you have so much demand, there's going to be a race to the bottom with everybody trying to win on cost. Because right now it's very clearly one
- [29:41]
Oliver
of the
- [29:44]
Tanay Kotari
big blockers, honestly to building tools. You see a lot of white coating companies running on basically negative or very minimal margin. And so there's other people innovating on reducing that cost down. And so I think historically also you've seen that whenever you have extreme demand for certain commodities, you just basically go down to single digit margin that you're making on top of your base cost, which in this case would be electricity. So I wouldn't be surprised. I pretty much expect that to happen. For just if you're just talking about the pure model layer going forward,
- [30:28]
Jason Calacanis
it does feel like free or close to free is what tokens will cost. And then the question is, how complex is your actual work? We this week stood up Kimmy and it's obviously nowhere near as good as Opus 4.6. It's not comparable head to head, but for simple tasks, hey, summarize this or sort this or tag this. It does it just as well. So we are now in that process of hey, we, what should we send to Claude and pay for and then what should we just on our own? Kimi 2.5 for 10, 20 bucks a month, what should we send to that job? So that opens up the big open source discussion. Where do we see open source playing into this or accelerating it even? Because it does feel to me at least when you start using agents, you start seeing your Claude bill or OpenAI bill or whatever you're using Gemini, go past $100 a for the agent or $200 a day, you're like, wait a second, that's $70,000 a year. That's like an entry level salary. Is that doing the work of another human or should I get another human? Or it's at 150k, it's the same price as the developer. And then you start looking for other opportunities. How do we think open source plays into all this?
- [31:53]
Richard Socher
I mean, this week in AI, interesting thing that also happened was anthropic complaining about how some other labs were distilling its models by yes, like sourcing them, talking to them and sucking out information. Which then of course a lot of people on X sort of pointed out the irony because they also have done that to the original human knowledge. And I think the ability to distill knowledge out of closed models a reduces some of their moat. And it shows that open source will be better and better and catch up faster and faster as long as it has access in some way to these closed models as well. And overall open source here massively proliferates this technology and enables everyone to play at it. And just like this idea that you can predict the future based on the goods and services that wealthy people have access to, if you're wealthy enough, you can hire 10 programmers and do something for yourself, implement them an app idea that you have now we're going to give everyone those capabilities, just like a personal tutor and so on. So I'm all for it. I think open source is great. You just have to be careful about how you use it and exactly like you said, know when to use it. And that is, I think another sign of getting this new skill that is to the management of AI agents up in people's hierarchy of sort of required skills to do your job. I think in the next 10 years being able to say, oh, I'm not so good with this agent thing is like in the workplace saying, now I'm not so good with this computer thing or this Internet thing. You just don't have people anymore.
- [33:33]
Jason Calacanis
When PCs came out today, this is probably a little before your time. I was an IT Engineer. And we were putting them on people's desks and they were getting them for the first time. And like some of the lawyers, I was putting it on their desk. I was working for a firm that installed them for lawyers. The lawyer would be like, I don't want to use this. And I'd be like, I have to put it in your office. They're like, put it on that side desk over there. And they had a secretary. You know, now we call them executive assistants, but they were called secretaries. And they would type and they would take dictation, they would write a memo, they would come in with a legal pad. And those folks who didn't make the paradigm shift, they just retired, basically. But, Tanay, I'm wondering how you think about. And we'll circle back to the agent issue, but before we get there, how do you think about open source? What models is Whisper built on? You must have tried all of them. And so how do you think about your costs? Because you've got millions of users. Grammarly's got tens of millions of users. You have all these services at scale. The token usage is a big part of their expense. So how do you think about open source versus OpenAI, Claude and all the other paid proprietary LLMs?
- [34:39]
Tanay Kotari
I should have three different points, quick. So one is one of my favorite things that happened last year was when deepsea came out, it forced basically every single lab to have a deep research mode that they built in to release that extremely fast and to release that kind of cheap because they knew that other people had that alternative. The same thing that you were talking about, Jason, about kind of the slack alternative that you can just build in house. It gave people a reason to demand lower prices. And so I think that actually drove a lot of innovation. And we've always seen open source over the last 20, 25 years do that. It's even more accelerated today. And so I'm actually very excited for everybody who's building an open source because you're actually having a major impact in all of these companies that are driven by hundreds of billions of dollars to provide things to people for cheaper and make it much, much harder to create. Basically, oligopoly there it is great downward pressure, right?
- [35:41]
Jason Calacanis
Like, it keeps them honest. They can't hold you hostage because they have the best model, because people go, yeah, it's a great model, it's 20% better. But I don't need 20% better. I'm good with what I got. And that is materially, I think, keeping the prices going down. Whatever it is 5 or 10x a year per token.
- [36:00]
Tanay Kotari
Yep, exactly. Whisper Flow was initially built on open source models and that actually accelerated our development early on. The thing we realized, though, is every voice model today was solving a different problem. They're all transcription models, fantastic for getting subtitles of movies or getting summaries of YouTube videos, but actually terrible for human input. And so we spent basically the first six to nine months putting a lot of band aids on these open source models to make them not hallucinate. If you speak English in a Russian accent, it would actually transcribe Russian. It's a problem. So we had to fix all of these things. And then eventually we learned so much about what we needed to do with these models that we had the ability to go and build our own. And so Whisper today runs on our own models. As a business, it's fantastic. It's about 90% gross margin because you don't have to pay OpenAI, Claude or anybody else for that. And we get to innovate, ship new things to people. We're not beholden to anybody waiting for somebody else to do innovation so our product can be better. And generally, the way I view the world is, hey, here's where we are. We want to get to the point where we actually have Jarvis in the hands of people system that just gets you, understands you, you. It's with you 24 7. And there's a lot of problems that we need to solve at that point. If somebody else solves them well, fantastic. We'll integrate with that, right? I think openflaw is a fantastic resource and we'll probably have something built with that into Whisper that is accessible. But hey, there's a lot of other problems that still need solving. And so we spend all of our R and D energy at the company working on those instead.
- [37:43]
Jason Calacanis
All right, let's shift to building an AI first team. Richard, you mentioned, hey, if you don't know how to use agents or manage agents or train agents, what value are you going to be? You said 10 years. I'd say like 10 weeks or 10 months. How could you possibly operate at work if you don't know how to manage an agent? Because that's the new interface. So much. So I had everybody come into the company on Sunday. We did an optional course. It was optionally mandatory for folks. It was optional, but I mean, it's kind of mandatory if you want to survive in the workforce in the future. And Everybody started using OpenClaw and learning how it works, et cetera. I spent my whole morning Just educating my Open Claw agent on who I introduced them to each person in the company. I said study what they do, study their emails, study their notion, study their calendar and study their activity on Slack. And then on the weekend, I want you to act as an executive coach and I want you to send a DM to their manager. So if it's a salesperson, Ricky's their manager, if it's, you know, on the investment team, Jackie, whoever it happens to be, and myself. So you have the employee, the team member, their manager, myself, the CEO essentially, and then the replicant. And I said recap their week and then coach them on what they could have done better and what they got right and then ask them some probing questions about how they could do better at their job. And that will happen this weekend. But I gave them access and this was like a major breakthrough for me. I said, hey, let's give my agent system level access to Gmail. So they can see all the employees Gmails in and out. They can see all the notion edits in and out. They can see all the Zoom meetings and their calendar and they can see all their Slack messages. Now they'll have like a whole understanding of that person and what they did this way week. And then for the first time, the manager doesn't have to ask the team member, what did you do this week? And then have the team member selectively try to remember what they actually did. It will just be like, here's what you did, here's the trends I'm seeing in it. Now we start the discussion, which I think if you're a young person, that's what we all wanted, we always wanted, at least for me as a high performer, I wanted you to know what I did and just tell me, how do I get better? How do I get better? So how are you guys building AI first teams and how do you think about that? It's obvious with developers, ship more code, go faster, have less bugs, whatever. But for the rest of the team, how do you think about that, Richard? And then we'll go to you.
- [40:20]
Richard Socher
Yeah, you're right. Like, of course the majority in all my organizations are developers. So that is the biggest shift there. What we're seeing is just multitasking, keeping lots of things in your head and being able to delegate these tasks better and better. I think for everyone else, this management sounds kind of boring, but it is. One of the most important skills that people will have to have is the delegation. Knowing when and how to trust what context those people, but also those agents have access to so you ask them the right things. That is something that we're training and when we're rolling out actually agents with our customers inside, like for you.com too we're seeing basically that you have to do certification, you have to do training and then you can actually do with AI also role specific training. Whereas like, okay, you're a marketing manager. Here's a like, help us. We'll help you now create a marketing agent so you can think about like how to write a blog post for your company and you can personalize that those certification programs too. And we've seen massively more usage of our agent creation platform on you.com when we required or when our customers CEOs required those certification programs for their people. So I think that is a big part of that change management in people's heads, not just a technology question.
- [41:40]
Jason Calacanis
Tenay, how do you think about building an AI first team? Customer support, sales, marketing, hiring, operations, accounting. How are you thinking about that and getting the developers. You don't really need to convince them that a copilot works because they've been using them for years or watching other people use them. But this knowledge workers outside that group are gonna resist in some cases and some are gonna embrace. How do you think about it?
- [42:06]
Tanay Kotari
The way Whisper runs today is so different from Ran as a company four months ago. The first thing is, by the way, we didn't scale down hiring at all. We're still growing our team at about 50% every quarter. And the kinds of people we hire today are very different. And so on the engineering team, we basically are now hiring senior and staff level engineers because every person is running five to 10 cloud agents in Barlow to get work done. Which means like we just shipped for example our Android app yesterday. It is right now the number one voice app. It works across every single platform and all you have, it was all coded by one guy running 10 cloud agents simultaneously. Now he did work 20 hours a day for the last like three months to get it out there. But man, you don't need junior engineers anymore was the first big realization. So every engineer at the company is essentially an engineering manager now. Every engineering manager is basically a director of engineering who's managing this massive team because your output velocity is just so high, which means we can do so much more. But then that expands to all other areas of the company as well. So what we did was we actually hired this one guy, he'd built a $300 million GMV restaurant OS business before in his main job is to go Team by team, sit beside them, see what their day looks like, create an automation plan and then go and automate the entire system one by one and then ship those internal features. For example, let's take customer support. Now most people who think about AI and customer support, they think it responds to tickets. Yes. That is one of the 15 things a customer support team does. And so we basically broke down every single loop that happens inside customer support, creating automations for that. So hey, first one is, you know, customer complaints, you write them a response back, great. Then you realize, hey, we actually don't know what to say to this person. And so what we built another loop for is hey, for questions, we don't have answers, to get the answer from a human, make a doc. So we have that for the future. The next thing is, hey, if this is an engineering issue that's happening, let's actually get access to our code base, the logs, what the user complained about. Let's actually compile that into a specific ticket for the engineer and hey, you know what, let's also pull up a PR for it so we actually take a first stab at fixing that problem. Then there's another loop where people are like, hey, how do I do this thing? And we realize it's something they're confused about. So that usually goes to the product and product marketing teams because it's our fault that we haven't educated the users. And so that's a problem to fix. So we have all of these different loops built in and they're all just running now autonomously within the company. And so what you have is company that has millions of consumers and we want to give a seven star experience to all of them is you need a customer support team of 200 people to just manage that, to manage the tickets coming in. And we have four humans.
- [45:23]
Jason Calacanis
Wow.
- [45:23]
Tanay Kotari
In our customer support team, we have enough now where we can even get two of them to just give white glove service to our enterprise customers. And the rest of it is just managed by these two people who are taking care of tens of thousands of tickets every single week.
- [45:41]
Jason Calacanis
This is a concept I think Tanay, you've talked about before, which is what is your expertise as a human and then how quickly before what's in your skill stack? Shout out. Scott Adams, creator of Dilbert, he always talked about a skill stack. You spend a career adding things to your skill stack. One of them is identifying bugs in software. One of them is identifying when you have a confusing UX or confusing tool. Tips and product is not educating the consumer Properly or isn't designed properly. And then of course, you would just have regular old tickets, Hey, I lost my password. Or I don't understand. It just could be dumb customers. Sometimes customers act dumb and sometimes customers make mistakes. Right? That used to be people's jobs to navigate all that, and now it's abstracted away. So you went from what you think would have been 200 to 4. That's a 98% reduction. How does a human keep. How does a human process the fact that what they did for the past 10 years is now automated and it can be done better by an agent? And I need to learn something new. You must have had this happen. You must have some anecdotes, both of you, of somebody going, what am I supposed to do now?
- [47:02]
Richard Socher
Yeah, so I think the most interesting thing here that actually happened in a lot of people's head with claudebot is that they realize that when you're an entrepreneur or when you're someone who cares about the outputs of an industry or an organization, you love AI. When you're getting paid by the hour, you hate AI. And in this particular case is like, imagine you're an entrepreneur and you have a cloud bot or some kind of bot that looks at all the things you do and then as soon as it gets enough examples, it starts doing it for you. You'd love it, right, because you own it. But if you're getting paid by the hour and you're doing that work, then your company owns it and then they might just let you go in the end. And so that will, I think, create a massive positive, long term incentive to become an entrepreneur and not an hourly employee.
- [47:54]
Jason Calacanis
You will need to be entrepreneurial to make it through this. Because let's face it, Tanay, a lot of people, you know, they live, they work to live. Maybe they're not cut out of the same cloth as say, entrepreneurs who are like, I need to solve this mission, it's important to me and I'm going to burn the midnight oil until I do. So have you started to have this experience internally? And how have you managed the human response to wow, what is my job now?
- [48:25]
Tanay Kotari
I know this creates a lot of stress for people, right? Even the smartest people as well, because, yeah, a lot of them have imposter syndrome. So the way we've been dealing with this, separate from hiding just inside the company, is not just telling people like, hey, you need to figure out how to do better or use AI or like, your job will be cut. But more so just like, here, let's figure out together how we build the org, how we work on this. Like, and whenever somebody in the team learns something new, let's have a session where we teach that to everybody else and everybody's up leveling themselves. Because, you know, interestingly, I felt this personally. So I've been. I prided myself on being a great software engineer, and I've been writing code since I was 9 years old to today, where now whenever I build things, I read code, yes, that Claude code produces, but I've barely written a single line of code in the last one year, which is ridiculous to me. I'm sad because I spent 17 years honing my craft and now it's all kind of useless. And I felt that pain for a couple of months. But honestly, at this point now, I love it because the thing I realized was it wasn't writing code and the syntax that I learned. What I learned was how to think. How do you take an extremely complex problem and break it down into small pieces and actually go solve it? How do you have taste for what looks good, what feels good? And so now with cloud code, I was like, wow, I can just go way faster and I can 5x myself and produce things in a day that would have taken me a week otherwise. And so I'm getting a thrill of something different now. And I want to create this feeling and this realization for people, but also having the empathy that it's going to be hard to get used to this change. So just recognizing that humanity in people, too.
- [50:31]
Jason Calacanis
Yeah. So speaking of humanity, here's producer Oliver, works for me, producing podcasts and working on the investment team. And when he was on the podcast team, I said, let's take some of these repetitive tasks. And here we are AO30. It's a 30 day since we started covering OpenClaw. So after OpenClaw and we started to really refine skills and tasks. So I'll have Oliver come on and show just three of the tasks. Here is producer Oliver, who has been doing all these amazing demos. Producer Oliver, I asked you to take the work you've been doing over the last two or three months, and over the last two or three weeks, we've been trying to get OpenClaw to do this.
- [51:10]
Richard Socher
Them.
- [51:13]
Jason Calacanis
These are tasks that would take hours a day, I think, typically, which means they would take, you know, dozens of hours a month. Let's walk through each of the skills and then I want you to tell me, like, just broadly what it's been like to be able to offload this kind of work. What I'll call chores. So let's see the first scale that we gave, that Oliver created. Tell us about this first scale. And Tanay and Richard, you can feel free to grill producer Oliver, ask hard questions, et cetera. And you guys can ask each other questions as well, obviously on the pod.
- [51:47]
Oliver
This is one of the first skills that I actually automated using OpenClaw. And this is an attendance check. Every morning. Everyone that works at launch and this week in startups will explain what they're going to do that day. This is kind of, you know, housekeeping, make sure everyone stays on task. But also it's a way for everyone to know what everyone else at the company is doing. And this was.
- [52:08]
Jason Calacanis
It's accountability. And, you know, we created this sod EOD program during COVID because we weren't in the same room. Nobody knew what anybody was doing when it was just confusing. And so we just said, hey, five minutes in the morning when you're having your coffee, ten minutes at the end of the day when you're packing your bag. Just what did you. What are you trying to accomplish today and what did you get done? I had about three, four, had about four people in the company who resisted me doing this. Three of them are no longer at the company. One of them still at the company is a super high performer. And he actually does it now because he felt left out. But we had a challenge which people would forget them. So I had an Athena assistant actually doing this. Now I have the Athena assistant on better work. So explain what this does and how it's changed things for you.
- [52:53]
Oliver
Yeah. So every day everyone goes into this specific Slack channel, tells us what they're going to do that day. And as you mentioned, this was a tedious task that we had a human doing. And now this is done autonomously, tags people who hasn't done it yet flags. Jason says this person hasn't sent in their sod. That's their start of day. And basically this skill just automates that task. Saves The Athena assistant 30 minutes for the Athena assistant to go do other work. But this task specifically, it looks at everyone's messages in that Slack channel and then just basically reports back. So we created the attendance check skill. This happens every day at 12 for specifically the start of day in the skill, it has the format that it wants it to send, including tagging you, tagging the people who haven't sent it yet. So this is just one of those tasks that we.
- [53:39]
Jason Calacanis
Now, you wrote the skill or the agent wrote the scale and you would give it instructions on how to change it.
- [53:46]
Oliver
I rarely look at the markdown format, which is how the scale is made. I but when it does make mistakes, I will sometimes go in there and see what exactly it wrote, because sometimes that's what we're showing now. This is the markdown. And you know, sometimes this does it a little crowded. Sometimes it writes things it doesn't need. Sometimes your agent will actually write the task on the cron job and the skill, which basically will cloud the context. So you want to make sure that the way that you structure a skill or a task is very organized. And I think that's some things we've been learning as we're trying to create a ton of new tasks. You want to create a task one, one at a time. I think that's the right way to do it for Open Call and that's kind of what we're going to lean into for the rest of our team. As I onboarded some of our team over the weekend with their own Open Call agent. And as we continue to onboard them the way we're going to do it, we're going to ask them what is something that you believe an Open Call agent can help you with. They've learned, you know, through us, through watching the podcast, what it can do, but they're going to give us, you know, five to 10 things that they think could be automated. I'm going to talk through it with them how this would potentially work and then we're going to do it one at a time because if you don't, your open call context can get very clouded, your skills are going to be messy. And that's something that we've learned. We're doing the one at a time. You got to take it slow while you're taking everything else pretty fast.
- [55:08]
Jason Calacanis
Yeah. And we've now given this access to all the zoom meetings. And so instead of people having to say what Zoom meetings they're doing or having people link to notion pages for the related tasks they're doing that day. For example, if I said I'm meeting with this startup and here's my notes from the call, that could automatically be included or our replicants could understand the context of that and take it on a go forward basis. Any thoughts on this, Richard or Tenay?
- [55:36]
Richard Socher
I love it. This is exactly what I meant with like, you have to learn how to delegate this, like chop down the task into concrete chunks that you can verify that the AI doesn't get confused with too large of a context window. For now. And all of those Things. Yeah, thanks for sharing.
- [55:56]
Tanay Kotari
One thing I'm curious about is have you thought about experimenting with doing the EOD report automatically? I know Jason, you briefly mentioned like, you know, looking at their zoom meetings, notion docs and all, or is there some benefit that you still see from people doing it themselves?
- [56:12]
Jason Calacanis
That is a great question. Today I started asking my, essentially Ultron I've talked about as a concept, the God that understands the oracle of the entire organization. I've asked it to reply to people and do that weekend coaching based on not only what they self reported, but that it sees in the data. And so we do think this could, could ultimately be done just by the replicant. But there's something for the human to set their own priorities in the morning. So maybe the end of the day could be done by the replicant, or it could make a first pass at it and say, hey, it's 7:00am here's what I think you should be working on. You reply to it and say, yeah, those three things are low priority. I'll do those next week. And you missed this thing that Jason called me on the phone about last night from his car. That's actually top priority. Now, now, now, the Oracle is, you know, kind of like your coach. So I've been trying to frame it with the team that everybody gets an executive coach. Richard, I think you said at the beginning of the show, one way to think about it is what are things rich people have that everybody could have? Rich people had chauffeurs, then you got uber Rich people had summer homes or ski houses and now they, they have Airbnb. Rich people had executive assistants, now they have agents. Rich people had executive coaches. But executive coaches cost five to $25,000 a month. Boards are happy to pay for it. When a company raises its Series A. I'm sure you've seen this or you gentlemen have had a board member say you need an executive coach. And you get too many complaints from people saying you're hardcore and you're not really communicating well with your team. Your ratings are low. You need an executive executive coach to help you. Now imagine everybody has an executive coach. To me that's like a super inspiring thing for, you know, a person in their first job like Oliver2 with his first full time job at a school to have an executive coach. We can't. That doesn't exist in the world. An executive coach for a 23 year old, it's just not a thing. So that's how I look at it is executive coaching. It's a great Question. All right, Oliver, let's go to your next skill. Oliver's three skills and he's got many more, but we're only in month one of this.
- [58:29]
Oliver
Yeah, I mean these are just some that I'm excited about that have been, been working well. And I think these skills, particularly the open claw agent, you know, using Opus, using Kimmy, have specifically, you know, done a lot of the heavy lifting. So this is one that I'm really excited about. I talked about this, I believe last week. It's the twist archivist, which basically takes the job away from, you know, two hours. We, I think we discussed one hour last week of, you know, finding a good clip, downloading the clip and editing the clip. It can do all of this for you. So we're really excited about crazy. So I was able to make a, an agent that basically will go look through 15 years of this week in startup clips. Find a clip. You know, maybe there's a fun Travis clip, maybe there's a clip of Jason talking about what he thinks thinks the future of robotics will be. These are really fun clips. You know, we've seen Jason's tweet about everyone in crypto should move to AI. Those are fun. So things like that people like bringing in the past. So this is a skill twist Archivist. Really excited about that. I actually did see today someone had their open call agent build in visual editor so it can edit and basically it created a dashboard that looks like something like Premiere or Cap Cut. So there's really exciting things and Open Call. What I noticed here is what it can build for itself to be able to complete tasks. So it built the video downloader, it built the editor, it built the ability to put captions on and it built the ability to import things into Google Drive. So it's really cool. This is one of my favorites.
- [60:01]
Jason Calacanis
This is tremendous because I was gonna hire somebody to be the twist archivist, like a full time position. So I was like, okay, we'll hire somebody for 60, 70, 80 grand. It could be work from home, whatever. They're obsessed with startups and they just want to watch old episodes and clip it. The problem is we could never find somebody who wanted that job. It's like finding a librarian, like the people who want to be librarians. It's a really needle in a haystack type job. And if we did find somebody, they'd be like, well, I want to work on the new episodes and the new hotness. I don't want to work on that. So it'd be Tedious that's one of the things that I'm noticing is there are some tedious tasks that you would never, a human would never take that the AI is more than happy to do. And this is one of them. And so 2,000 episodes into this Week in Startups, 250 episodes into all, in two episodes into this Week in AI, you start to have this archive and it's sitting there and there's all these nuggets in it, but you need to have judgment and then you need to do seven tests. So I love this one. And then I think even having a layer where it tests and it learns, and that's going to be the key is can this one. Because we don't have this one doing any kind of learning yet. So what I'd like you to do is have it learn, or maybe it's a second skill where when it posts a clip, it says, that clip got a lot of engagement. Why did it get engagement? And then it goes and finds more with more, you know, that have more qualities like the ones that previously work, if that makes sense.
- [61:31]
Richard Socher
I love it. That pushes it also to just have longer time horizons, which is a thing a lot of people are thinking about for their agents. And if you can really close that loop now, it might eventually get better and better at finding the right agents that will create the most engagement. Now, of course, it will maybe do some reward hacking and just find out, well, if you show certain body parts, you get a lot of engagement and things like that. So you need to be careful about what you allow it to pull in in terms of content. But, yeah, that would be a really cool.
- [61:59]
Jason Calacanis
This one is just crazy. Okay, give us your last one there, Oliver.
- [62:04]
Oliver
Before I show the last one, I'd love to ask Richard a little bit more about recursive AI. I know that that's your new project, and I'd love to ask you, you know, you're focused on self improving. We're focused on self improving at a little bit of a lower level. But I'd love to ask you a little bit more about that project and what the future lies there.
- [62:22]
Jason Calacanis
There.
- [62:23]
Richard Socher
Yeah. You know, I think overall AI is in this dual state right now, where on the one hand it's electricity, like, and the other one it's still research. And on the electricity, we know what to do to automate certain tasks. We give it some inputs, X, we give it some outputs, Y. We can train these models. It gets better and better. And we just, like, with electricity, can infuse it into every different industry. But there's still this sort of elusive goal of building superintelligence, doing real research towards that. And what does that enable you to do? Well, you don't have to manually define the context. You don't need to do all this manual onboarding of an agent. You just give it very high level complex goals and rewards and inputs and it will itself onboard into a complex code base. It itself will complex like be able to deal with complex rewards that you give it. And it can eventually automate the scientific method itself, which is basically what allowed humanity to stop mocking around for hundreds of thousands of years and now create all these incredible technologies. And so what we're doing at Recursive is essentially automating the knowledge discovery piece and fully closing the loop on the ideation, implementation and validation of AI research ideas and eventually taking that whole machinery and applying it to all future digital jobs. But eventually also science. Most people don't say we want more scientists, but most people love good scientific breakthroughs of better batteries, better fusion, like functioning fusion reactors, better drug development and medications and cures for different diseases and eventually for aging. Most people love the outputs of research and so they're aligned similar to inheritance, healthcare on caring about outputs. And I think a recursive self improving superintelligence could eventually be the ultimate eureka machine for humanity.
- [64:18]
Jason Calacanis
Tanay, how do you think about recursive learning and making your product whisper better and better for the users? How do you do that today? I'm curious.
- [64:31]
Tanay Kotari
So one of the actually hardest and most important problems to solve overall is having that feedback loop back into the system that is relatively human free. Because what that lets you do is what you're saying, right? Like Oliver, you made this thing where you can, okay, now start to feed in the results of that and have it generate even better and better clips. But you would still need to have some level of discretion there. Like I saw it first posted in a Slack channel. You haven't let it run for fully free yet because you don't know yet if it's going to do the right thing or not. But imagine if it starts doing the right thing. And imagine if you can scale that up over millions of people, billions of people. Whenever a system does that, it creates so much delight because that is also one of the skills we respect so much in people. It's like an extremely smart intern versus a mid intern. Like, like what you realize the biggest difference is how quickly can you just learn from feedback and grow because you're kind of starting from like little to no experience at all. And so with Whisper, you see this one small thing we do where if Whisper makes a mistake and you manually go fix it with your keyboard, Whisper says, like, hey, I added that word to your dictionary. Sorry, I made the mistake. Not going to happen again. And whenever people see that, there's such a big smile on their face because that's something they've been dying to see from Siri for like the last 15 years.
- [66:01]
Jason Calacanis
Oh, my God. Siri doesn't know how to spell my fucking last name. Or like, I literally ask it to call my wife and it's like, who? And I'm like the number one in my favorites. The person I communicate with more than anybody on the planet. I just want to talk to her on the phone. I'm on a ski lift. Dial her number. It's so dumb. Siri is disgracia to the highest level. Show us your third skill here. You're doing great work, Oliver. Oliver's one of my favorites. Top three right now for me.
- [66:33]
Oliver
High praise.
- [66:35]
Jason Calacanis
It is High praise. It is. High praise.
- [66:37]
Oliver
Actually, I think once we. When we started openclaw, the task I was trying to automate was the. Was our Athena systems task because these are pretty low level tasks, repeatable. And you know, we have one of the top Athena assistants. He does a great job. He can spend his time elsewhere. But. So this, this task basically looks for podcasts and looks at who their sponsors are. And this would help our sales team. What this agent does is it'll basically go and look at what sponsors these other podcasts had, and then for the last five episodes and it cycles through a list of different podcasts and then it will send them into our sales channel so then they can potentially reach out to those potential partners. So this is a skill. Very tedious. Going through a lot of different YouTube channels, going through the transcriptions, but now the agent can do it in a minute. This is just another great example of a single task that's saving. You know, I would say we do this five times a week. It's. And we used to do it three times a week, I believe.
- [67:40]
Jason Calacanis
So now, you know, I was trying to do it five times a day, but just people don't want to do this task, so they do anything but this task because it's tedious.
- [67:48]
Richard Socher
I think this will be such a mainstay for the future. Like, people always look back at past jobs and be like, why would you want to work manually with your hands in a field? Why would you want to weave clothes manually?
- [67:59]
Jason Calacanis
Why would you Want to have SDR sales development rep, which is what this is.
- [68:02]
Richard Socher
Why would you want to drive a truck for hundreds of miles all by yourself? Getting diabetes, being alone and so on. People. People will never look back in 100 years and look at the jobs that most people do right now and be like, oh, I wish we had those back.
- [68:16]
Jason Calacanis
It's incredible. And then the next step it does is it looks in our pipedrive, which is a really cool CRM we use, and then says, are they in the CRM? What's the last date that they were contacted and who is their account rep at mention them. So this has become really interesting because that was a task I always had. Or now that we have a sales manager, Ricky, it was her task ask, which is, oh, when's the last time we talked to aws? And we have Microsoft Azure and we have Gemini. Why don't we have AWS involved in the podcast? And it's like, oh, okay, we never called them. We haven't called them in two years. Sometimes that happens. A lead will slip through the fingers or somebody was their sales rep and they left and then it didn't get reassigned. So all those things come out in the wash. 3. Great skills. The Athena assistants are really interesting because you can go to athenawild.com and get a couple of weeks for free. I'm an investor in the company and I love it. They're really good at doing business process outsourcing, BPO and then a bunch of the BPO companies. You can find this tweet, somebody on the research team while we're talking about it. They also got caught up in this Citron madness and the dooming because they. They got caught up in this because they thought, oh, well, all these agents are going to get rid of business process outsourcing and India is going to lose all these jobs where all these intelligent people. Actually, I think it's the opposite. I think those people are going to be the ones who actually manage the agents and you're going to just find more and more work for them to do down the long tail of jobs. But there was a tweet that came out just talking about business process outsourcing, if you can find it, or we'll throw it in in post.
- [70:05]
Richard Socher
Jason, what are some skills that you would love your AI agents to have, but they're still tripping over, like access to certain systems or capabilities.
- [70:16]
Jason Calacanis
So one of them is having editorial judgment. So this is something we are going to figure out with them, which is, can we get the assistant when it picks the title of the show or the best moment from a podcast. Will it actually pick the most interesting moment in the podcast to make a clip out of as an example? Or if it's looking up potential partners, will it be able to handicap and qualify the lead? This lead is, you know, 90% to close. They advertise on these four podcasts. Those have the same demographics. If they love those four, they're gonna love us. This one's a 10% chance and actually have that be more accurate than a human's assessment. So a lot of times our salespeople are like, oh, yeah, we shouldn't have, you know, this supplement company on the podcast because, you know, the supplement companies are going after pod. You know, they have products that cost $100 a year, not 10,000. So it's just better they are on Bill Simmons or Call Her Daddy than a niche podcast about technology or AI. It's just not aligned. Or this startup company that pitched us, how likely is it to raise a round of funding in the future? So will this clear market with venture capitalists. We get over 10,000 people applying for funding. I need to know which ones we should engage because they will pull through in the future. So those are the kind of, I call it judgment, artistic, knowing a joke's funny. And there's a company called Is it Human in the Loop. Oh, Rent a Human, who we had on the podcast a couple weeks ago. And Rent a Human is going to allow your agent when it gets to one of these moments and say, like, I don't know which thumbnail is the best thumbnail for this episode. You could just rent 100 humans. Say, here's, you know, four YouTube pages. Click on of these four thumbnails, which one you think is the most interesting? And just say, click on a thumbnail. And we won't even tell it which one's most interesting. Just say click on a thumbnail. Clicks on the thumbnail, and one of the four thumbnails is ours. And we just see if we beat, you know, Mr. Beast and MK. You know, Marques, can we beat those right in the human tag. So that's actually the next level for us. But we have so much more to do@rentahuman.AI. 560,000 humans have made themselves available for rent.
- [72:48]
Richard Socher
It sounds like Amazon Mechanical Turk. Yeah, exactly. But maybe those guys will go into physical space and catch something for you or something.
- [72:56]
Jason Calacanis
Yes. What he showed to Ney on our podcast when he was on last week was you could in Shibuya Station in Japan, you could have 10 humans say, use Whisper Flow, now available in Japanese and hold up a sign for $100 for two hours in Shibuya station when everybody's walking by. And you could say, okay, find me 10 Shibuya stations around the world and say Whisper is out for Android. In the markets that have the most Android users, like maybe there's one of the boroughs in Brooklyn has the most, maybe Queens has the most Android users. So you, you put it on the queen's train stations, like crazy stuff like that that an AI wants to get done. Or tanay. Another example might be you want to know how to pronounce calacanis or tanay and you just have humans say it and read clips for you or something. And you need humans to give you a feedback loop. I don't know. That's what I think it's good at.
- [73:57]
Richard Socher
I think the old open source models do give China a lot of soft power. A lot of countries that can't afford the high token costs that the closed models have right now are downloading the Chinese models. And so what used to be sort of Hollywood having a lot of soft power in the world and narrative and storytelling now become Chinese open source models.
- [74:22]
Jason Calacanis
That's an incredible point, Tanay. Why don't we have like a champion here in the US that's doing an open source model model? Do you think so strange, like, shouldn't we have like five of them by now?
- [74:32]
Tanay Kotari
You're talking about a big company trying to do it.
- [74:34]
Jason Calacanis
Yeah, like I know OpenAI released their, like, you know, 2.5 is now open source, whatever, things that don't matter in their mind, but you have these champions, like half dozen serious ones in China. And to Richard's point, thinking about soft power, if you're going into Africa, South America, the frontier markets, the emerging markets, and you say, hey, we want you to spend a billion dollars on tokens this year. They're gonna be like, we don't have a billion dollar spent on tokens. And then China comes in and says, yeah, we can just put up some servers for you. It's free for the first million tokens a day. This is like really on a soft power basis, a way to have. And if you focus on their languages and their data sets, this could be tremendous. Or do we have champions and nobody knows them?
- [75:18]
Tanay Kotari
It's just a capitalism thing, right? Because I think the way they, the economy companies work in, in China, at least from what I understand, it tends to be less capitalistic because of how tight to the government you have to be. And in the US you just are like, of course OpenAI is not going to put their best models on the market because guess what, they make way more money from people paying $20 to $100 to them than they would ever buy an API. And that is true for so many other businesses. And so that is the key thing, that if you just do a cost benefit analysis, you would cannibalize yourself if you try to open up APIs, because APIs actually don't make as much money in the long term as owning the customer, having a recurring pay that's coming from them and not having to just compete on price as a commodity in that whole LLM market.
- [76:22]
Jason Calacanis
All right, in breaking news, gentlemen, breaking news. Tim Cook has announced as part of their onshoring efforts that they are going to make the Mac MINI in Houston, Texas, just next door to me. Here's Tim Cook's tweet. As part of our $600 billion commitment, Mac Mini will be producing in the USA for the first time later this year. We're accelerating our progress even further, producing more AI servers and opening an all new Apple advanced manufacturing center for hands on training. In addition to this, they have also said in the past that they're going to purchase 100 million chips from TSMC's Arizona factory which is opening I think later this year. And as part of this, the New York Times ran a story about the tensions with Taiwan. It turns out Silicon Valley has been warned that the CCP, Chinese Communist Party, plans on taking military action in 2027 and that we should expect massive disruption and that these Silicon Valley companies have largely ignored this warning.
- [77:36]
Tanay Kotari
Warning.
- [77:37]
Jason Calacanis
I've heard from my sources that this 2027 action by the CCP has been pushed back to 2029 after Trump's out of office because China's a little bit concerned. Quote from the New York Times article, this invasion that could happen in this report could cut the supply of chips from Taiwan. According to this story, cutting the supply of chips from Taiwan would lead to the largest economic crisis since the Great Depression. Richard, your thoughts on the growing tensions around Taiwan and Apple? Onshoring. I mean this may be above all of our pay grade here, but it is the topic of the moment.
- [78:22]
Richard Socher
I do think trade is an incredible force for peace and I'm kind of saddened to see sort of it being used more and more against other countries. I do think when you trade with someone, you're not going to attack them, right? Because you downshoot yourself and you thought whether you're Supplying things to them or you're getting things from them, it usually doesn't make sense to attack deep trading partners. And so I hope that sort of international trade cannot be completely reduced. But I also understand why in terms of matters of national security, you do need to have certain crucial sectors to be more within your own borders today.
- [79:15]
Jason Calacanis
Any thoughts on the global chessboard and how this affects our business? Again, understand if it's something above all of our pay grades, but I think Richard makes a great point. Companies that trade together, companies that are in business together, they generally don't go to war. Companies that are not in business together and don't have that fabric, don't have that shared prosperity, they're more likely to go to war. So your take Tenay on Mac Minis, Taiwan fabs, all of it.
- [79:46]
Tanay Kotari
You know, if you look back over the last 200, 500 years or so, there's kind of this shift that keeps happening back and forth where you get more globalized and then you realize, oh wait, there's a lot of risks. Then you try to be a lot more independent as a country and try to build a lot of things in house. And then the world changes. You kind of forget the scars of the past and you start to get more globalized and more dependent on others. And I think that is what people are realizing today because so much of the US GDP now is built on top of technology. So much of the technology is built on top of chips, and those chips are in risk. So even if that is a 5% chance or less of that happening,
- [80:30]
Jason Calacanis
that
- [80:30]
Tanay Kotari
is a massive chance for a black swan event. And so I think just that threat itself presenting up and the fact that, okay, this is one of the biggest bottlenecks for technology all over the world, right? United States, it's big for, for India, it's even going to hurt China. And so you're seeing all of these countries trying to get more of these raw materials, more of the base on which everything else is built on top of, just in house. And as much as we would all love to live in a world that's peaceful and no countries attack each other and we all happily trade with each other forever after, that's just not the world we live in, unfortunately.
- [81:15]
Jason Calacanis
So
- [81:17]
Tanay Kotari
my reaction to this is not a surprise that this is happening. It's going to be incredibly sad if it does happen, because there's such incredible talent in Taiwan that's, you know, tsmc and in a number of regions of China as well, like China and US Trade so much among each other. And there's going to be the kind of hiccups that happen here because of this ongoing global power domination battle that all these countries are into. And you're just going to have a lot of casualties in different ways than previous wars.
- [81:55]
Jason Calacanis
Here's what's going to happen. I'll give you my prediction. Taiwan and China have a long history. China obviously wants to be respected as the owner in their mind of Taiwan. That's what they believe. They believe it's one country. Taiwan believes it's its own country. And the United States has had this amazing ambiguity, strategic ambiguity, saying, yeah, Taiwan's an amazing country, great leaders. And yeah, Taiwan and China have this amazing relationship that we cherish and we respect and you have had this great ambiguity. But China, I think, wants to enforce its sovereignty and test if that is actually true and if the west will fight for Taiwan. I can tell you the Taiwanese people are very proud. And the rest of the world that's dependent on this is going to take two or three steps over the coming years to avoid China getting control of those fabs. The first is they're building fabs here in the United States, as we talked about, Arizona, etc. That's going to take some time, but it's well underway. It should be moving faster. There should be a greater sense of urgency. The second is we're going to airlift every one of those engineers and everybody who works at those factories and give them United States citizenship or move them to another area outside, far enough outside of China's purview that they wouldn't invade. What countries would China not want to invade? They wouldn't want to invade India, Australia, Korea, Japan. So those talented individuals could be given citizenship and be underwritten to go to those four or five nations very quickly, just be given pure amnesty for them and their families to get the heck out of Dodge. And then there's the burn it down scenario. And I wouldn't be surprised after we airlift all of those fabs and dismantle them and send all the talent to the four corners of the earth to keep it away from China, if the Taiwanese people don't torture them. This sounds crazy and ludicrous, but if you were Taiwan and they said we're coming there because of those fabs, not just because of sovereignty, not just because we want to exert our power and we believe Taiwan's ours, if China's saying we want Taiwan for those fabs, what would the Taiwanese people do if they were actually facing innovation, if the ships were actually coming around they burned them to the ground before they would give them to the Chinese. This is the scenario. I know I sound like a loon, perhaps, but I think that these scenarios are being planned in that 5%. There's a 1% chance in each of those five, there's probably five different plans. The first is to defend Taiwan. The first is to make peace. The second is to defend Taiwan. The third is to replicate Taiwan. The fourth is to get everything moved out of there. And the fifth is burn it to the ground. That's like the craziest scenarios. You know, the CIA, the leadership of Taiwan and all the countries around it, Singapore, et cetera. And Bloomberg predicts a $10 trillion economic cost if the country, the sovereign country of Taiwan, gets invaded by China. This has been an amazing second episode of this week in AI. Thank you to my guests. Where can people find out more about your companies today? And who are you hiring for? If you're hiring at all? Maybe you're just firing people and you're going to be a solo shop of four people or something. I assume you're hiring. Who are you hiring for?
- [85:30]
Tanay Kotari
We're hiring a lot of people. I think we went from 20 to 45 in the last three months.
- [85:35]
Jason Calacanis
Wow, that's big. You raised some money. Yeah, yeah.
- [85:39]
Tanay Kotari
And we're hiring for talent across every single software engineering role. Great. Building our our sales team. Everybody in growth, customer support, marketing, and all the roles are up on our website, which is WhisperFlow, AI, that's WISP and you can find us there and follow me on LinkedIn or Twitter.
- [86:08]
Jason Calacanis
Perfect. Richard, people want to get your money as an investor. How do they reach you? And then tell us what you.com is offering.
- [86:18]
Richard Socher
So yeah, Aix Ventures. Yeah, you can just follow me on Twitter and X Richard Socher, ping me there for your company ideas. And then for you.com, we're hiring a ton of folks in sales businesses ramping.
- [86:36]
Jason Calacanis
Explain to people the business of you.com?
- [86:38]
Richard Socher
happy to. Yeah. We built AI search infrastructure. So whenever you want your LLM to be up to date, accurate and have citations, we are essentially the Google for LLMs, the search for your large language model. We have customers like Thoughtspot, OpenAI, Databricks, Salesforce, Windsurf.
- [86:58]
Jason Calacanis
So they can't get this through an API from Google. There's no search API anymore for Google. Yahoo Search Monkey back in the day. It's really just you and Brave are the only two people who offer anything like this.
- [87:09]
Richard Socher
Yeah, that's right. And we offer like all kinds of other complex content APIs. So you can because essentially the way search is very different to how people search, LLMs can read a ton more content before they summarize you these kinds of links that you would have gotten from Google, very short snippets. So we can do both very fast, very short, very accurate, slower but cheaper and just like around the Pareto optimal frontier. So we're hiring search engineers, salespeople, marketing folks. Marketing is something that I think is actually a big role where ultimately a lot of things aren't zero sum that people think are zero sum, but human attention is zero sum. And so we have to get, get better and better at marketing. And so I think that's something that we're looking forward to hiring, too.
- [87:58]
Jason Calacanis
All right, everybody, we will see you next time on this Week in AI coming out every Wednesday or Thursday. We're figuring it out this Week in AI AI.