Reflection Beam, Mistral Le Chonk & the Drone That Delivers Your Chipotle

0:00You know, they just kind of leaned into the internet meme and now it's true. Today's frontier tasks are tomorrow's open source workloads. From open models to open skies, we've got drone delivery from DoorDash. It's like you have a bunch of AI interns. It turns out for IPOs to happen, market conditions really do need to be perfect. Bot traffic has now exceeded human traffic. Is this our deep seek of the West?

0:32We're back with another big week. And this time we have Sharia Agrawal, a partner on Lightspeed's growth team, joining us as a guest co-host. Reflection and Mistral each unveiled a new open-weight model. DoorDash unveiled its own drone delivery and an AI agent that you can text. Meta and Sierra announced an open standard for how personal agents work with businesses. and Aura called off its IPO hours before pricing. Plus, two updates from across the Lightspeed universe. Let's get into it. Shaya, welcome to Lightwork. Thanks for having me on, Josh. It's great to have you in the studio. So we have so much to cover this week.

1:16There was a lot in AI and specifically for open models. So this week on Monday, Reflection introduced Beam. And then Tuesday, Mistral announced Large Four, also known as LeChonk. Have you seen some of the internet memes around like all of the history on LeChonk and why they wound up going with this like fat cat as the name of their newest model? I think it was their logo at some point in time. And it sounds like they've really leaned into it now. Yeah, it's kind of started as like a meme this summer where like people were saying that they were going to announce this like fat cat with all of these parameters.

1:54And it was this massive like model. And it turns out that, you know, they just kind of leaned into the Internet meme. And now it's true. Both of these models are very exciting. Beam is 501 billion parameters. Only 23 billion are active per token. It's a sparse mixture of experts model. It's built primarily for text, coding, reasoning, and agentic work. LeChunk, on the other hand, is a trillion parameters model, 49 billion active with a million context window. It's in 160 languages. The API is in preview now. Both of these models didn't publicly announce their weights, but both are coming later this month.

2:39What does this mean, like, that they didn't announce the weights right away? And also, like, what do these two open models mean overall for the AI industry and how they maybe relate to, you know, the frontier models? Yeah. Well, I think the weights will come. With regard to how this impacts the AI race, I think this is the first attempt from the West to really battle and compete with the Chinese open source models that are the majority of open source token flow today. I mean, they're selling the same thing. You can see inside model, you can adapt it and control it. And this has tons of utility for regulated industries, industries where you have private information, and also just for those that are cost conscious, because we know the gap between the cost per token for open source versus closed sources is massive.

3:33What does seeing inside do for you? Does it allow you to do more with the model? It does. It allows you to change certain actions, behaviors, steer the model in certain ways. And you can also then sort of RL and PostTrain on it and build your own version that's specific to your enterprise and your workflows and your knowledge. Sharia, explain to our audience, why do you pick one of these open-weight models instead of the frontier models? When do you use one of these? It comes down to two things, control, visibility, and then cost. So today's frontier tasks are tomorrow's open source workloads.

4:15And open weight models are maybe six to 12 months behind what the frontier model capabilities are. But that also means in six to 12 months, they'll catch up. And at a fraction of the cost, you pay a frontier model today. So even today, things that Frontier models released six to 12 months ago, they're able to do at a fraction of the cost. So running a thousand support tickets, it just makes more sense to run that through an open weight model versus something that's extremely complex, like a drug discovery task or a very long running agent that helps you complete a coding task. Those are things that are going towards the closed source models.

4:51And I'm sure it makes a lot of sense for control, because if you've got private customer information that you don't feel like you want to share with one of the large model firms, there's this great opportunity for you to control all of your data, have it hosted by you, and still be able to do the things that you'd like to do from a frontier perspective. That's sort of the pitch for sovereign AI. So Mistral is going after three to four very highly regulated industries, whether it's industrials, financial services. We know Reflections partner with the Department of Energy in the US. These are critical industries where this stuff matters a lot.

5:30One of the interesting things I saw in kind of both announcements were they were both pushing cost efficiency. It seems like both are quite a bit cheaper on a token basis. Why are they pushing cost? I think this is the value of open source models today. These companies are not really trying to compete with OpenAI and Anthropic on the frontier model race. What they're trying to do is push the Pareto frontier of intelligence per dollar. And what this allows enterprises to do is have some sort of equation that they're thinking about. For these tasks, the open-weight models are able to intelligently perform it and do it at a fraction of the cost.

6:15So why would I route it through Fable or some high-cost model? And you get to continue to control your data, like you were saying, right? So it's cheaper. And as we've been learning on the show, it feels like there are more and more tasks where you don't necessarily need the bleeding edge to accomplish the task you want to accomplish. Totally. Sharia, what does mixture of experts mean in regards to these kinds of models? So when you look at like the total parameters, that's the sort of initial memory bill that's hosted on servers. And think of the mixture of experts as a way to reduce your compute costs in runtime.

6:55So you have 501 billion parameters and 23 billion are active for Beam. and what that means is only 5 % or those 23 billion are actually leveraged and one analogy I can make is you know you have a group of a thousand scientists but you pick the the five percent that are best equipped to handle that specific task and instead of getting all thousand scientists to do that work let's choose the best and that's actually something we're seeing converge throughout the industry and certainly from the Chinese open models, but also Reflection and Mistral. Sharia, so Reflection has been cooking this model in stealth for a very long time, and people were wondering when this was coming, how good it would be.

7:41Is this our deep seek of the West? We certainly hope it is. Today, open source models are dominated by China, And Reflection Mistral's releases are giving the West a chance to compete on the open source front as well. From open models to open skies, we've got drone delivery from DoorDash, which is really exciting. DoorDash Air is a custom built in-house design six propeller aircraft that DoorDash says actually doesn't sound much louder than another car passing you on the highway. They're starting with restaurants in Northern California. They've got Chipotle and Popeyes as foundational partners.

8:28Talk to us about the state of drone delivery. Why is DoorDash entering this area? And why is this important to them in terms of their focus and last mile logistics like it's been for, you know, as long as DoorDash has been around? I think it's a continuation of the trend, you know. So Google Maps enabled Uber and DoorDash to actually begin, where you have drivers now able to go from point to point very easily. That was the mobile era. And now with increasing developments with drone delivery, it's natural for DoorDash to want to be able to increase access and get people access to the food that they want to deliver.

9:10And whether that's lower costs or simply being able to reach people that otherwise wouldn't be able to serve, that makes a ton of sense for DoorDash. I mean, yeah, it's also like a lot of the cost of their business is delivery, right? And so if you can have this drone fly across your yard and drop your food on a winch or your groceries or anything else that DoorDash has access to, it's pretty game changing for their business model, you'd have to imagine. Certainly impacts consumer appetite. What makes this different? There are so many drone delivery startups. What's DoorDash's edge in this area?

9:44I think DoorDash's edge is their network. You know, they're a true marketplace business. They have a very strong flywheel where more consumers bring on more restaurants that create sort of a compounding advantage. And it's very hard for a new drone delivery company to replicate that. Yeah. How are they going to get all of the businesses that DoorDash has spent years getting online so quickly? Yeah. DoorDash didn't just announce DoorDash Air. They also announced almost an instinct competitor. It's a chat bot that is in preview now and you can sign up if you'd like access. And basically now you can go and have a conversation with DoorDash around the food that you might order.

10:27And I think the interesting thing for me here was that DoorDash actually has a bot that is part of the app called AskDoordash. DoorDash says that nearly half of restaurant orders through AskDoordash from June to August went to places that customers had never tried before. I think this is cool because it's sort of showing that, you know, their version of a muse or an instinct is creating consumer behavior where they actually try new things. and I feel like for most people, and I don't know about you, but my lunch order from DoorDash is pretty much the same. I eat the same thing every day. I do too.

11:04So I wonder, like, does making it conversational make you try new things? You know, I haven't tried DoorDash's experience yet. I mean, it just came out. But I really want to. You know, I think this is something that I've actually always struggled with. I feel like the interface always presents the same options you're already used to seeing. So something that actually can search the entire breadth of their inventory and serve up things that might be interesting to me, I'd be curious. Yeah. And I do think that's a little bit by design. DoorDash says that more than 30 % of recent grocery orders are just routine restocks, right?

11:39They're just like, hey, this is the thing I want. Get it to me, which is you could imagine why the drone delivery would be even more helpful. But I think what's interesting here is just like that they could shift this consumer behavior a little bit. I mean, I certainly now with Muse and Instinct have this dialogue of a conversation that I might not have had with DoorDash. I would have just like picked the thing that's the easiest. But now maybe I'm going to say, hey, what's a Chinese food I haven't tried that I might like based on my past preferences? And DoorDash probably knows a lot about me.

12:16I'm sure it knows a lot about you. I think it could change a little bit of consumer behavior around this. I think they need to because they're facing a bit of a risk from the instincts and muses of the world. How do you view DoorDash's risk different from, we've talked a lot on the show about Amazon's risk, right? Like they've blocked Muse, they have this massive ads business that is built on the friction within its app that you kind of almost get frustrated and just take the one that they're advertising, right? And that is a big piece of their business model. Well, what's DoorDash's moat in this equation and how much of a risk are Muse and Instinct and all of these other agentic platforms?

13:00I think there's somewhat of a risk. And this is somewhat of an offensive, but also a defensive move from DoorDash. If they can deliver a better consumer experience and truly lead to the best order or the best meal experience, whatever the consumer is solving for, because they have more data, more knowledge, they're willing to expose more data to their own agent versus Muse or instinct, that could drive consumer behavior. But there is a world in which, you know, you ask user instinct about food choices and you don't go to DoorDash. You go to a restaurant's own delivery service, you go to Uber Eats, you go to some other service.

13:36So consumer attention and sort of being able to actually have the connection with the end user is so important. And I think that's what DoorDash is trying to do sort of end to end. Yeah, I think it's kind of the way we were saying that like, you know, maybe people aren't going to cut off Salesforce because everybody uses it as their CRM. And it's like, but for DoorDash, all of these restaurants have such slim margins. Like, are they really going to be like, oh man, and we love using the DoorDash platform. Like if these agents can kind of go around that and cut out some of the costs that they have with delivery, you could see that maybe their mode isn't as thick as it used to be.

14:20DoorDash also has been doing more on the restaurant side too. They've launched reservations. They launched reservations. They launched a whole platform for restaurants. It does seem like they are building the tools to stay as sticky as possible. 100%. In other personal agent news, This week, Instinct announced that it's allowing you to add it to a group chat. Sharia, how do you feel about adding your instinct that now probably knows so much about you to a group chat with your friends? And what happens if like your friend says, book me the hotel room? Does your instinct go and book a hotel on your behalf?

14:56You know, I think consumers have different levels of comfort with access to private information, with spending money on your behalf. I found that, you know, people in my cohort, younger people, a little bit more tech forward, we're more comfortable and we trust technology a little bit more. Obviously, there's a large part of the country and the world that may not feel that way. And I think instinct will have to traverse and figure out a way to get everyone on board. but I at least personally am I'm I'm a fan me too I mean the fun thing is the CEO Noah Shin said that 40 percent of users connect a personal credit card within three weeks and I know that I've experiences I right I think I did in seconds too but I feel like I'm I'm feeling what what they're actually talking about which is that people are spending a lot of money on agents and I know that when I got my first credit card bill after a month of Muse and Instinct, I was like, wait, I bought all of these unnecessary things because I was having fun having these conversations with these agents and buying things that I didn't necessarily need.

16:06But I think the average person who joins Instinct then gives us this credit card is spending$1 ,300 a month, which is a lot of money. It's a lot of consumer spend being captured by this new agentic workflow. What do you think this is going to do when it hits the masses? I think this is going to continue. I don't know what the average spend is for a person, but I think of more of the mix shift. You know, if you spent$1 ,000 a month clicking through different items on your cart and on Amazon or DoorDash or other platforms, now it's just being directed through some of these agents. Now, certainly there's an element of fun and discovery and you might be increasing your spend, but hopefully we're still being very thoughtful about how we're spending our money.

16:53Yeah, I think that is definitely the key here, but it certainly makes it a lot easier to shop. I feel like this has been such a game changer for how I get things done. And every person who's either tried Muse or Instinct has that same kind of an experience. But I do think the missing link has always been that it is still this bot going out and doing this stuff on your behalf. And I've had relationships with both platforms where they've almost gone a little too far and I've had to be like, no, no, don't buy this. Muse announced that it was partnering with Sierra to announce a personal agent protocol to make this a little bit more above board and actually create an authentication experience where it isn't just an anonymous bot.

17:42This is you. Then you confirm that you're signing in as you and you're authorizing this agent to act on your behalf so that it's not just booking you a flight without you having an option to cancel it. What do you think about these? There was both the Sierra announcement and then Instinct also made a partnership with Decagon, similar. What do you think about these sort of personal agent frameworks? I think we're going to see more of these. You think about Decagon, Sierra, they're serving the businesses. And Muse and Instinct and maybe even DoorDash's agent, they're serving the consumer. And how do they connect?

18:22How do they talk to each other? Today, it's a bit of a Wild West out there. And we've seen Amazon actually block Muse. I think in the fullness of time, this is how the majority of the internet traffic will continue to trend. In fact, actually, I saw, I think today or a few months ago, that bot traffic has now exceeded human traffic on the internet. So this is going to continue to proliferate and we need to figure out the rails with which agents for both consumers and businesses can interoperate. Yeah, it does feel like this is every other cycle, right? Like first the Wild West happens and then we build the structures around them to make them work.

19:00I mean, obviously, you're not using instinct and muse today for growth investing, but. I think we will. I mean, I think the future is, it's like you have a bunch of AI interns, you know? And right now, like, I feel like I have three or four cloud analysts working for me. And I'm able to do so much more than I used to a couple of years ago. And even then, I still have to be present in front of my desktop and prompt and iterate. And like, wouldn't it be so cool if I had a personal agent that was then interacting with a frontier model to do that? And I'm just checking off, signing in. And I think, yeah, over time as an investor, what makes a mode is personal relationships, sort of actually having instincts, let's say.

19:41But some of this analytical work, it's commoditizing. Right. I mean, venture has always been a thin business in terms of headcount, but it does feel like we all have gotten superpowers with AI and have these bench of interns. Sharia, this was a surprising one. So last week, Aura and Aura Ring, which many people know as the ring that monitors your sleep and your health, pulled its IPO last week. It was a pretty shocking thing because Bloomberg reported that the book was 4x oversubscribed. Fortune cited that it was 5x. And either way, that demand was really substantial. Help us understand, like, why didn't this company go public last week?

20:27You know, it's a bit of a mystery, Josh. I'm scratching my head a little bit myself, especially so close to an IPO, because I think this was only pulled a day or two before. Are the behemoths just sucking all of the air out of the room in the public markets to allow new entrants? I think that's what the rumor mill would suggest. But there's probably, it's not just one thing, it's many things. That's definitely part of it. I think there's been some debate about Aura and how much of their revenue is like one-time hardware revenue versus is recurring software. And it's still growing very well as a business.

20:58They're profitable. They have very strong metrics. But I think that is a concern and impacts the valuation of the business. So, you know, one can say four or five X are oversubscribed, but the question is at what price? How do you feel this changes your feeling about, you know, growth investing? Obviously, you know, as a growth investor, you care about the public markets and there being an opportunity for or IPOs, does this change the way you think about companies with how hard it's been to get out this year? I think it impacts late stage growth investing. If you're looking to do a crossover round and solving more for IRR, I think early investors in Aura and even early growth investors in Aura are still probably quite content.

21:39Very happy. Yeah. And I think that the reality is this is a great company with great numbers. It can get out when the time is right. It's just a matter of felt like maybe the market conditions weren't perfect. And I think that's an interesting thing to just digest is that it turns out for IPOs to happen, market conditions really do need to be perfect. It's almost like, you know, going out for a day on the water, you need to be sure that like all of the things are looking really perfect before you go out and sail.

22:17Avara, which builds AI avatars that train and coach sales teams, raised a$17 million Series A. Avara says annual recurring revenue grew more than 400 % year over year, with customers including Elastic, Proofpoint, and Netscope. Lightspeed is proud to have led Avara's seed round in 2024 and participated in this Series A. Robco, the industrial robotics company, is now valued at more than$1 billion, doubling its valuation in just nine months. Robco sells more than a dozen robots designed to automate logistics and manufacturing tasks. The company's flagship system is a two-armed mobile robot called Alfie.

22:58It's capable of automating tasks that can't be performed by a traditional robotic arm. For example, an automaker could configure Alfie to stabilize a car part with one arm and use the other to apply paint. Lightspeed led Robco Series B in 2024, co-led at Series C earlier this year, and we participated in this route. Sharia, thank you so much for joining us. While Claire was out, you make a terrific Claire's Owl. Thanks, Josh. Try my best.

23:39you