The Week AI Officially Became SI | Lightwork

0:00Do you want the smartest intelligence or do you want the one that like makes you want to give it a hug?

0:05Josh Machiz: Now I'm sort of really attached to my muse because I have it wearing little outfits. Being able to file a passport online and have it be done through a chatbot just feels like the future. I think they were debating the slide for like 20 minutes and then finally some of the engineers ran the numbers, did a bunch of analysis and they were like, let's go for it. Their version of building in public. It's a building in public masterclass.

0:36Josh Machiz: We're back with another big week. Tech leaders gather for lunch in D.C. for the White House Accord on superintelligence. The federal government launched America.gov, an AI chatbot that helps you find your way around government services. OpenAI's Dove Day debuted always on AI agents called Dots. and a new model that costs a fifth as much as its flagship. SpaceX got Starship to orbit for the first time, carrying 26 Starling satellites, plus two updates in the Lightspeed universe from Eleven Labs and MODIS. Let's get into it. Lots of important figures in DC this week. So many figures in DC this week, including us.

1:18Including us. Kidding, not us. We did get stuck in the traffic. We definitely got stuck in that DC lunch, AI spectacular traffic.

1:29Josh Machiz: And we actually had no clue it was happening until we were in line and saw the seating plan. And we were like, no wonder. No wonder. It's incredible that they were able to keep it under wraps that this many people from the AI world were going to show up in DC this week. Yeah. Who was at the lunch? It was a pretty spectacular crowd. It was everyone from Dario to Sundar. Who made up the lunch attendees? Yeah, about 20 execs made up of the usual White House crowd like Trump, Vance, Besant, Lutnik. But from the tech crowd, Jensen Huang, Elon Musk, Satya Nadella, Jeff Bezos, Sundar Pichai, Zuckerberg, Greg Brockman.

2:11A really who's who, but no Apple.

2:13Josh Machiz: Yes, no Apple. So I would say you had a mix of frontier labs like the OpenAI's Anthropics of the World. You had the chip providers like NVIDIA, AMD present, Micron, and then a bunch of sort of this weird mix of like cloud and enterprise players like Microsoft, Amazon, ServiceNow, even Palantir on the defense side. And then a couple of investors as well, but no Apple notably. But I think really that was mostly because one, Apple really hasn't been doing as much on the frontier, if you will. I think they're very much so focused on the diffusion of all this stuff. I think that makes sense. So it's understandable why they weren't there.

2:51Josh Machiz: Yeah. I also feel like they've tried to keep John Ternus out of the spotlight politically. Tim Cook, I feel like, has always been. He was very much in that zeitgeist. Exactly. Versus I feel like the move right now is for John Ternus having just become CEO to sort of be heads down, actually focus on company and product as opposed to, you know, eventually what I imagine he will engage in some of the more political representations and all that. But for now, I feel like it's actually a good thing that he's not necessarily beckoned to D.C. to be representative of Apple. So they weren't just at a lunch.

3:25What was the outcome from this lunch? And what does it mean for all of us who are trying to understand AI?

3:32Josh Machiz: This lunch was really one of the first gatherings in response to all of the safety incidents that we've been hearing about this summer. Notably, OpenAI hugging face. In the last few days, you've heard of a couple more incidents, including OpenAI hacking the Australian government, a bunch of other government websites, the DOE, the German DSC wiki incident. So you've had a whole slew of these events. And in response to that, as we've discussed extensively, there's been a lot of discourse around whether or not to pace the frontier. Are there going to be government regulations around this? And so this was really Trump and the administration's effort to bring everyone in a room and sort of hash it out.

4:18Josh Machiz: But I think as of many of these things, it is a launch. And I don't think oftentimes you get very tangible regulation or actionable next steps out of it. What they did end up coming out with was something called a White House accord on superintelligence. We know Donald Trump loves a rebranding. So So we decided after he did this Twitter poll, I think a couple of weeks ago with different options that he's landed on superintelligence as his chosen word for this. Yes, he has decided to officially rename artificial intelligence as superintelligence. And that's actually part of his executive order where all federal agencies must now say SI instead of AI in all official documents.

5:02Josh Machiz: He's also asked the science advisor's office to produce a working definition of SI as well. My sense is that this really stems from wanting to reframe artificial intelligence away from something that's a hazard, artificial to, you know, something that can demonstrate America's superiority. And so really framing it more around super intelligence as opposed to artificial intelligence. Well, on the timeline, someone was saying this is a big moment for Slovenia because it means that all of those SI domains are now going to be used by startups because it's the parlance of the U.S. government. Yeah, I think it was Anguilla who had been benefiting.

5:43And Anguilla had been making most of its money from tourism before AI became a thing. And now they make it all from the domain names.

5:52Josh Machiz: Yeah, we'll see for Slovenia. They're.SI. A dot SI. And isn't Melania from Slovenia? Is this a coincidence? No. I didn't actually put two and two together. You might be the first person I've heard actually connect those two dots. I mean, it's Donald Trump and everything is on the table and it makes perfect sense that he chose SI. On specifically the accord on superintelligence, the one pager that came out at lunch, though, on top of the executive order around renaming things, there were a couple also layers to that document, which was basically this morally binding, almost sort of like a constitution over actual law.

6:31Josh Machiz: But they basically had a one-pager signed by Trump, Dario Amadei, Sundar Pichai, Mark Zuckerberg, Greg Brockman, Jensen Huang, and Elon Musk. And it laid out four layers. One, that they would all embed more internal controls around monitoring these safety capabilities, especially in bio, chem, warfare, and unintended system access. Two, actually building some internal teams to verify how these controls work. Three, an independent external auditor and evaluator. We don't know what the shape of that looks like, but it is interesting that that was baked in. And then four, an independent board committee receiving reports from both.

7:11Josh Machiz: They said that these companies, the ones that signed this, will meet regularly to set standards. And over time, they may codify these steps into actual laws and regulations. Well, that sounds like a good first step. I don't think anyone was expecting regulation. There's a lunch. A lunch. So it feels like at least there's, and we've talked about this on the show, at least now we're getting steps towards having some sort of regulation around AI, or as we now have to call it, SI. from a lunch that didn't have Apple to something that felt very Apple-esque. We had the U.S. government basically do a product launch.

7:53It felt like Marco Rubio was giving me Tim Apple classicness as he unveiled America.gov on a big stage in D.C. What is America.gov? What do people need to know about this? And what does it mean about maybe design in Silicon Valley coming to a little bit of the District of Columbia.

8:15Josh Machiz: Yeah. Interesting question on the design part. And I would be so curious for your thoughts on all of that. But yeah, America.gov basically had its America Dev Day, if you will. America Dev Day. We'll have another OpenAI Dev Day to discuss. But there basically was the government's Dev Day where they pitched a chatbot as a singular front door where you can access every single question you have for the federal government. And so if you think about your existing workflow with a lot of government sites, there's about 29 ,000 of them. It is very hard to navigate information. You have to do a lot of Google searching.

8:51Josh Machiz: There are a lot of broken links. Now this is supposed to be the search and answers layer over all these government sites. Interestingly, in my research, I realized that about 40 million people visit government websites daily. So I actually can see how this is a problem for 40 million people on a daily basis and why something like this might be needed, even if some critics are calling this a LLM wrapper. I think you have to give the Trump administration credit for this. Before this administration, all of these websites were fairly arcane. And especially under the chief design officer, Joe Jebbia, who was co-founder at Airbnb, he's really put that stamp of design polish that maybe you're used to seeing in a travel app or how we shop into the U.S.

9:42government. And I think it's a really interesting thing to see all of these websites be branded in this fashion where it almost feels like Airbnb light, but honestly, in a very elegant way. Have you used the chat bot?

9:55Josh Machiz: I have not. I've seen clips from the presentation and different use cases, but I've not personally tried it yet. Yeah, Yeah, I played around with it a little bit, did my best to try and, you know, exploit it in the way you would some of the other bots. And it was pretty good. It managed to redirect itself back to doing like the ordinary functions. It'll be really interesting when they add some of the things that Marco Rubio spoke about on stage, like being able to file a passport online and have it be done through a chat bot just feels like the future. Yeah. It also is interesting that their website still says AI and doesn't say SSI.

10:37But I assume that's coming.

10:38Josh Machiz: Yeah. And so that's where I also have questions is, you know, I think it is unprecedented to see this deployed by a government. But I agree. I think it is important to recognize that there is a massive change needed in how we navigate a lot of websites. For example, to your point, filing a passport or running a name change, that has historically required four to five separate transactions. I have a lot of friends who have not ended up doing it because you end up having to go through your social security, the IRS, the VA, Medicare, yeah, exactly. And all of them have different applications just for a name change.

11:19Josh Machiz: And that's just one workflow. And so you can see how in the world where this scales and works, this saves people a lot of time. And I do think that is the value of AI is it's able to take a lot of information and make it navigable and searchable. They are launching also federal job applications, Medicare enrollment, but it'll likely be after December 2026 through 2027. So all those are still coming for now. I would say it's still very much so sort of like a glean, but for all things government. Yeah, it's still a little bit on the surface, but it seems like all of these things are coming. And it's exciting.

11:59I think it's nice that one of the stats I think they shared was something like 10 billion hours a year that are spent on these government websites and probably not accomplishing very much, right? And so it's nice to think that there is this path forward that's elegant. And I think it's really cool to see us finding a way to make navigating the U.S. government a little bit of a more seamless thing. Claire, we had a very exciting OpenAI Dev Day. And as usual, all startups were scared with what OpenAI would announce. But it turns out they introduced Dots, which is an always-on agent that has its own computer.

12:40This is similar to what we've talked about on the show with Instinct and Muse. What did you think of Dev Day and the major announcements?

12:49Josh Machiz: Yeah, they have officially entered the AI assistant agents chat. and I would probably categorize Grokbot in here, Instinct, Muse, and now you have Dots. This is going to be a very busy space as we've predicted and I'm sure there will be other products that come from the Apples and the Googles of the world. I imagine they're also working on similar products but for those who haven't been able to try a Muse or an Instinct, pretty similar in that because it has its own computer and connectors and computer use, it can help you complete actions the way that a human would. doesn't require an API or an integration.

13:26Josh Machiz: It can literally look at what's on screen. Some of the new changes that Dots has versus some of the others is that you can voice call your Dot, which I thought was interesting. That is cool. Right now, you can't really verbally talk to your Muse or talk to you. I think Muse said that's coming, but it's not available yet. Not available yet, but on the roadmap. And we already know how good OpenAI's voice chat is, right? Yes. And I think another thing, heavy again on this emphasis on safety as a moat is that they really were very clear on what boundaries you could set. So you could state what me or sorry, what I keep mistaking muse and dots, which I know a couple of execs also did.

14:05Right. I feel like everyone in the timeline is being like Grokbot plus muse equals dots. Yeah. Yeah.

14:12Josh Machiz: I'm like muse plus instinct dots. They're like all sort of sometimes interchangeable, but you can tell dots what you want it to do alone, what would mean your approval, and what it must never do. And so that's pretty cool that you can sort of set these guardrails up front. And you can optionally let it drive your laptop, which is another breakthrough. I think that's only available if you have the app downloaded on your desktop for Muse right now. I think for me, the interesting thing about this announcement was that everything that's old is just new again. Like I spent 10 years doing Chinese IPOs at NASDAQ and everyone came in with a cute, furry, like adorable mascot that like was there to represent their brand and was how you connected with the company.

14:59And it's interesting to see that from Muse to GrokBot, we're all trying to put these like friendly, furry faces on intelligence.

15:10Josh Machiz: The Tamagotchification of tech. The Tamagotchification of tech. I know it almost begs the question, like, do you want the smartest intelligence or do you want the one that like makes you want to give it a hug? Well, that's what I thought Meta did a really good job with at MetaConnect last week was that their upfront framing around Muse was not, oh, I'm putting super intelligence. Obviously, Meta super intelligence is the team behind it. But at no point during the entire presentation. Did they even mention intelligence? Yeah, I don't think they really mentioned. I mean, they mentioned AI, obviously, but that wasn't the story that they were really centering around.

15:46Josh Machiz: They were really focused on Muse will save you time and money versus I think sometimes when you put a super intelligent product and you're like, this achieves whatever benchmarks. I think your everyday consumer doesn't really care about some of these benchmarks. They don't care about the frontier, right? Like the average consumer wants to know, can it help get the thing I wanted done? Can you save me money? Can you look through my inbox of a thousand emails? Totally. It's about saving all those other things. I also think like, you know, it's amusing that I didn't mean to say that. And then I said it.

16:20Josh Machiz: Given all the releases and how busy this space is, do you have thoughts on who might sort of emerge as the winner? Is it a world where it's sort of multi-product and I have a bunch of these? I find myself today using both Instinct and Muse. I'm in the same bucket. I actually think I'm going to use all of them kind of the same way I use Claude for work and I use some of chat's things for personal or when it comes to design and I need the best image model, I'm using chat. Like, I feel like it's similar to that. It's also we had Instinct raise a ton of money this week. Right. So it's interesting to have this dark horse and then a very well-established player that gave us Instagram and Facebook.

17:07Josh Machiz: Massive distribution. With massive distribution going up against one of the frontier labs with the most, I think, current technology in its product, it would seem. And so it'll be interesting to see how the three of these, plus GrockBot, all kind of play out. And I believe what Apple and Google launch in this space. Right. And I think that's what we're learning from how Meta's succeeding is like there are all these other players that we know and love that can do product very well and don't necessarily need frontier intelligence. They've just needed it for it to be as good as it's gotten today, which is already, I think, smarter than your average person.

17:49Josh Machiz: Yeah. The thing I struggle with is I am starting to get frustrated by setting up a lot more like re-signing into things. How many of them do you want to give your email? I think that there are things that you only kind of give away once. And I don't think you give it to everybody. The other thing is like I started finding myself getting tired going through the process of setting up my cute little avatars. Like now I'm sort of really attached to my muse because I have it wearing little outfits and I'm like, I can get pretty much the same output. why would I restart? And maybe yes, if Google inevitably launches something better, I might move over there.

18:28Josh Machiz: But it just feels like not necessarily that there is high switching costs, but I do think there is consumer fatigue and constantly needing to set up a lot of workflows and your preferences and tastes. So I'd be curious to see. I also think, is there a world where you have one horizontal agent and a bunch of more vertical ones? Is it, you know, I have one for personal. I have one for work. I just I it still feels like an open question for me. Yeah, I think the interesting thing for me is attachment. Like you already heard when people when they used to shut down open AI models, like how upset people would be when their model was discontinued.

19:03Josh Machiz: Yeah. Now with the mascotification of all of these products, like, are you going to be sad if something about your mascot from your your like little muse is lost between versions? Like, I do think they're doing a really good job of just thinking about attachment to these things rather than caring so much about, is it the best? As long as it's got my data and can do the thing, I think I'm going to be satisfied. Yeah. They did launch a couple other things as well at Dev Day that I want to point out. Actually, about like 20 plus products. There were so many. Yeah. There were a couple notable ones, and we can speed through these because there's a lot.

19:45Josh Machiz: But I think a couple that caught my eye were, one, they launched something called GPT 6.1 Sol. So this one is near Astra intelligence, you know, for agentic coding, computer use, any professional work, but at one-fifth the token price. So it's... A cheaper Astra. Yes. It's basically frontier intelligence, but becoming the sort of like commodity input. So they didn't ship GPT 6.1. This sort of made the headlines a couple days ago where OpenAI said it was withholding given security issues, flagging potential deception during testing, and the model had misrepresented what it had done or not done.

20:29Josh Machiz: And so they decided to launch basically this lighter version at a cheaper rate. And what you see here is that safety is really a key consideration now. OpenAI didn't launch their flagship 6.1. Rather, they launched a cheaper model instead. Another model I thought was interesting was, or sorry, not necessarily a model, but something called UltraFast. Yes. Plus Pro. Yeah. Which is basically a paid speed tier. So really here, the positioning is, yes, we can create a product that's better around costs, like, you know, like GPT 6.1 Astra, but we can also give you speed as a luxury. And this model is a page viewed tier where, or it's not a model, sorry, it's a tier that allows you to have your models generate tokens up to eight times faster in codecs.

21:21Josh Machiz: So at about 300 tokens per second and six times faster in API. die. And the reason you might need this, you know, for your everyday consumer, not really a big difference. But if you are a developer, this is a huge difference. It's the difference between, you know, waiting three steps. You know, we probably have a highway for development. Exactly. And I would pay for this fee because instead of waiting for my code to generate in three minutes, I can get it in 0.4 seconds. And so you're actually seeing also some level of price discrimination on speed as well, which is interesting as intelligence gets cheaper too.

21:55Josh Machiz: And then a couple other things. I thought Decisions API was interesting. That was sort of their jab at Jev. We talked about Jev last week, which is basically this decision model, sort of this instant decision model that can classify things very quickly. It picks from a preset answers, and in 100 milliseconds, it can classify and route accordingly. And it's great for this high-volume work like customer service, decision-making. And what's interesting here is that OpenAI is focused on having this optionality. Yes, they want to own the frontier. They want to own the cheap models. They want to own the speed plan.

22:36Josh Machiz: And they also want to start diversifying to decision-making models, too, which I think is interesting given Jeb raising a pretty sizable round this week as well. Yes, but we don't cut out the upstarts here on Lightwork. Of course. And there's so much, I think, you know, whether it's across dots, across things happening at the model layer, I think it's all happening in real time. And my takeaway is not that like, oh, Jeb or Instinct are dead, but rather it's just such an interesting time and that i mean it's interesting that they're not just leaving that one area unturned yeah like i think the labs are recognizing they also have to play defense too it's like these early upstarts like instinct and jev are are i think it's probably learnings from instinct right it's like oh there's this there was there's this massively inflecting consumer tool and we could have been there first.

23:34Josh Machiz: Yeah. And I think for consumers too, there's actually sometimes a desire to use something new. Obviously these big labs and tech companies have so much surface area in their distribution and product and to their advantage sometimes when it means getting in front of users and automatically having integrations. But I also think a lot of people love instinct because you have a small sort of underdog team that is going after this space that is shipping at a velocity as quickly, if not faster than some of these labs. They still have no X account. They still have no website. They're just growing virally through word of mouth and, you know, I think for referral.

24:16Josh Machiz: And a good product. And a good product. That is conducive to that. So SpaceX had a very big launch this week. They managed to get Starship into orbit and release 26 of these V3 Starlinks. Yeah, it was pretty phenomenal. Have you seen some of the photos that came out of it? I watched the video. I mean, those videos are so captivating. It's insane how good they've gotten at making each one of these launches as its own kind of theater. I just think that whole team has really mastered it. It's really enjoyable to watch because you feel like you're on the team. And I think there's something special about feeling like when something goes wrong, like on this one, I think they lost an engine, right?

25:06Like that you almost feel like gutted and you're like, oh, no, like you do feel that attachment to their success in a way that if you weren't a part of the process, you might not. So in a way, it's building in public. It's a building in public masterclass.

25:23Josh Machiz: I actually think there are a lot of elements of that. You know, we've seen them also struggle. And there's also been a lot of these test flights that haven't worked out as successfully. And I do think this has been the closest thing to, you know, when I think about in my parents' age, when they were sort of seeing a lot of the advancements in space, there is a sort of like humanity coming together moment. And, you know, I often see when there are critiques of tech right now, a lot of people are like, why are we wasting our time on, you know, flimsy tech and AI boyfriends and girlfriends when we used to put people on the moon?

26:01Josh Machiz: And so I think this is almost like a return to, wow, we as humans are pretty awesome and we can build amazing things and go to space. And, you know, all of these are very much so collective efforts. And it's just, I think that we are not as important, you know, in a humbling way that like we are a little speck in the universe. And the fact that we've been able to accomplish breaking into orbit out of that is, I think, pretty inspiring. yeah how far the SpaceX team has come from their very first launch to this one yeah is pretty awe-inspiring yeah this one was extremely notable though because I'm sure a lot of folks have seen other test flights right and they're like oh this is just another test flight I've seen a lot of you know for those a little less space-pilled as maybe we are it just seems like another rocket launch what is different here is that the last 13 times were really much more like test flights So all of those flights were suborbital.

26:59Josh Machiz: None of them actually made it all the way into orbit. And so what would happen is they would have these test flights, they'd fly very high, very quickly, and then come back down on their own. And the main reason for that is because of safety. If you have things actually go into orbit, it's going to start spinning around the Earth and it can fall anywhere, which is probably not something you want to risk because there are a lot of people on the ground that might be subject to that. So the first 13 were really more so test flights where they were testing things like, can you actually have, you know, safe reentry?

27:33Josh Machiz: Can you can this thing actually launch? And now this was the first time they actually made it to orbit. What happened to your point is, you know, this is Starship. It's a humongous. It's enormous. Yes, it is 400 feet tall. Wow. Yeah. And it's made of two parts. So there is a booster, which has 33 engines. that's the thing that's doing a lot of the heavy lifting and then on top of that that's the actual ship itself called starship and the big idea of spacex entirely is reusability of these rockets and so the big idea is that you have these rockets that can fly back and then get caught by a giant arm on this launch tower and you can refuel and fly again more like an airplane as opposed to these unusable one-time use rockets something did go wrong though to your point on the live stream where they had one engine start failing.

28:24Josh Machiz: And it was actually interesting to your point of like feeling like part of the journey. I think during the live stream, a spokesperson was like, we may have to, you know, rely or we have to. Yeah, they were thinking they might not be able to go into orbit for a second there. Right. Like, but that is also the tension of the building public that makes it so exciting is, is like, you realize that there are these tradeoffs where like, hey, we might actually not do this, and then we're going to have to bring it down, and it'll still be a success, but not a full success. Yeah, and I think they were debating this live for like 20 minutes.

29:00Josh Machiz: And then finally, some of the engineers ran the numbers, did a bunch of analysis, and they were like, let's go for it. And actually, what happened is they were able to recognize that with the fuel that they had and the remaining engines, they could actually compensate for it to get that very precise push into orbit and it worked. The only thing that changed was this was originally supposed to be, you know, the deployment of 26 Starlink satellites plus a, I want to make sure I'm getting right numbers, but around six orbits around the Earth in 10 hours. I think it was only in orbit for like three hours.

29:36Josh Machiz: Yes, exactly. So they still deployed the Starlink satellites, which makes this the first commercial Starship flight ever, which is pretty huge. And ended up cutting up the rest of the orbits to, you know, I think they only did two orbits to three hours. But still a major accomplishment. I think my other takeaway from this is if you haven't had a chance to really dig into SpaceX business, you would know that the SpaceX business is heavily reliant on Starlink as its most profitable and revenue generating business arm. And so to be able to actually have a vehicle that can scale that is really impressive.

30:18Josh Machiz: And I think that's why you had SpaceX's team call this the first commercial flight. It's very exciting. It was really cool to watch. And what is it just curious for our audience? Like, what is the V3 of Starlink actually capable of? It's quite large. So that's why ships like Falcon 9 can't transport them. Each V3 is really impressive because that adds about one terabit per second of capacity. So with 26 of them, you have about 26 terabits per second, which is about 10 times what a Falcon 9 is able to deliver. So, yeah, pretty great for the Starlink business to be able to actually serve customers from this revenue-generating Starship flight.

31:05Josh Machiz: 11 Labs, the voice AI company, launched 11V4, its most expressive voice model yet, plus a faster turbo version built for real-time voice agents. 11 Labs says V4 now covers more than 90 languages and can clone a voice from just 10 seconds of audio. Lightspeed joined Eleven Labs'$500 million Series D earlier this year. Modus, the AI-native accounting firm, released Financial Audit Bench, which is an open benchmark for how well AI models handle audit work. Modus says it tested 11 frontier models across 90 audit tasks, and it published a paper and the code so anyone can run it. Today, audit is one of the highest-stakes places you could put AI, And until now, there was no public way to check whether these models are actually good enough to do the work.

31:56Josh Machiz: Modus raised an$85 million seed in Series A, led by Lightspeed.