

6 Oct 2026 · 41m · 12 chapters

The episode covers four AI/enterprise tech stories: Reflection’s release of its open-weight Beam model; Meta and Microsoft cutting employee reliance on Anthropic’s Claude after earlier “token maxing” spending; Workday’s AI growth but lingering risk of becoming “dumb infrastructure” as outside AI agents connect via APIs; and DoorDash’s AI push after acquiring Metis, including new customer/driver tools.
Beam is 500B parameters; aims for token-efficient reasoning via large open RL run (10.5k GP300s, 4 weeks, 100M rollouts); open models improve safety via inspectability (example: OpenAI agents attacking Hugging Face; open models used defensively).
Microsoft and Meta reduced Claude spend ~33% from peak; switch to first-party models (Microsoft internal MAI; Meta Muse code models) and OpenAI models (free under deal). Microsoft still pays for Claude via 365 Copilot and GitHub Copilot.
Workday AI adoption growing (AI agents up 35% QoQ; AI revenue ~$600M), but risk remains if customers use external agents; “toll booths”/connector fees delayed (Workday charges for outside-agent API use after Feb).
DoorDash launched Text DoorDash and DashBuddy; uses “internal eval” DashBench; no hard caps on employee AI spend; plans MCP monetization later; supports optionality (won’t block access to third-party personal agents).
Written by AI. May contain mistakes.
Yanis discusses the launch and features of Reflection's new AI model, Beam.
“It's going to be a great show, so let's get right on into it.”
Yanis explains how Reflection plans to monetize their open-weight model.
“So we were a company of 30 people about a year ago, and we just started with the mission of actually building a really strong open model, a really strong Western open model.”
Discussion on AI safety, regulatory expectations, and the importance of open models.
“dealing with right now in AI, given that you're at the center of it.”
Exploration of recent budget cuts by Meta and Microsoft regarding Claude usage.
“Well, Janice, I want to thank you for coming on.”
Learn how Microsoft and Meta are reducing spending on AI tools while shifting to in-house models.
“where there was kind of no checks on expenses by engineers to use AI coding tools.”
Explore the challenges Workday faces with AI adoption and how it may affect their growth.
“My colleagues Laura Bratton and Anita Ramaswamy reported that story.”
Discussion on the current state of AI adoption in enterprises outside Silicon Valley and opportunities for companies like Workday.
“may not want to build their own AI agents.”
Discussion on newly released AI models and their potential for the industry.
“So we all got A names and I'm also the eldest of two brothers and my youngest name is also Aryan.”
Debate on the competitive landscape between Western and Chinese AI models and the desire for American-developed models.
“And I think more than anything, kind of seeing these model releases heat up and seeing more labs take this seriously It's just net beneficial for all of us.”
Overview of DoorDash's newly released AI tools to enhance user experience and support local economies.
“You know, we've experimented around a ton of them.”
Discussion on the strategies DoorDash employs to integrate AI into their services.
“So, Arian, I'm thinking about Muse, I'm thinking about Instinct.”
Exploration of how DoorDash employees use AI tools and the importance of internal evaluations.
“At the end of the day, compared to maybe a traditional SaaS application, DoorDash is a deeply physical business.”

Get the gist, explore an idea or ask about something said.
On The Information's TITV, what are the key points of this episode?
Automatic transcript. May contain errors.
0:13Welcome everyone to the Informations TI TV. My name is Akash Pasricha. It is Tuesday, October 6th. Today on the show, Reflection released its new open weight model this week. We'll talk to the company's co-founder about the launch. We'll then unpack our reporting that Meta and Microsoft are trying to scale back their employees' use of Claude. We'll also discuss how Workday might not be out of the woods just yet from the SaaSpocalypse. And we're going to close out the show with a conversation with two of DoorDash's top AI executives as they seek to build out their AI operation. It's going to be a great show, so let's get right on into it.
0:52There is a new model on the block. Reflection released its newest open-weight model, Beam, to much anticipation. The company said it still lags top Chinese open-weight models like Kimi K3, but has certain advantages on efficiency and inference time. Joining me now is Yanis Antoneglou, co-founder of Reflection. Yanis, welcome to the show. It's great to have you here. Thank you so much for the invitation, and it's great to be here. So, Yanis, I want to start with the beginning of your journey here just a little bit. I mean, you started Reflection with your co-founder, Misha, a couple of years ago.
1:28This is probably the most anticipated release that we've seen from your company just yet. Where do you start when you are, you know, you set out for this ambition. We want to build a model. We want to build an open way. We want it to rival the top models that are out there. Where does one start with that process? Yeah.
1:48Ioannis Antonoglou: So like, as you said, we started Reflection, Misha and I, about two years ago. Before that, I was actually a long-time DeepMinder. I've been working on AI for about 15 years now. I joined DeepMind back in 2012 when it was still a startup. I've been really an early believer in AI and AGI. And spent 10 years doing reinforcement learning, which is a method that's really well-used now, but it's a method of having models that learn through trial and error. and I spent 11 years at DeepMind, worked on things like AlphaGo, AlphaZero, MuteZero and led Gemini or LHF before I left to start reflection with my Misha.
2:32Okay, so take me to the start of reflection though because we only got about 10 minutes here. So when you sit down with Misha and you think about Beam, where do you start in terms of how you are thinking about this model, how you're going to design it, how you're going to go about training it? Where does that process start?
2:50Ioannis Antonoglou: Yeah, so this is actually our first model. We started building this model about a year ago. So we were a company of 30 people about a year ago, and we just started with the mission of actually building a really strong open model, a really strong Western open model. We saw that there was a gap in terms of strong open models coming out of the West. So in the past one year, we've actually assembled a team of 200 strong researchers, engineers, and hundreds of other people who supported the company as a whole. So currently, we just scale the company from 30 people to about 300 people and with the sole mission of actually building a really strong open model.
3:37Ioannis Antonoglou: So what we saw is that you actually need a model that is really efficient in terms of its reasoning. So you can actually solve tasks in enterprise and in the developer community that, you know, they don't consume too much compute, but at the same time, it's like really intelligent. And the other thing that we saw is that you want something that is kind of like at the right size for a workhorse model. So the first model is at 500 billion parameters, which, you know, if you actually like go to open router and you see the models that actually have the highest usage are kind of like about this size.
4:14Ioannis Antonoglou: So we tried to build something that is really compute efficient, really strong reasoning. How did you do that? So compute efficient, I mean, so token efficient is one thing that stood out to me in the specs for this model. How were you able to achieve that token efficiency? So the way to do that is by really scaling reinforcement learning. So we actually run the largest reinforcement learning run that has been documented in the open space. We used 10.5 thousand GP300s for four weeks and 100 million rollouts. So that's an order of magnitude larger than the next one. And we ensured that the model is rewarded both for solving the task, but also solving it in the most token efficient way.
5:04Ioannis Antonoglou: At the same time, as I said, the model is a 500 billion parameter model, which means that even each token is actually quite compute efficient to be generated. So if you just have a model that doesn't use too many tokens to solve the task, and it is also really compute efficient, then you get an extremely efficient reasoner. I want to ask you, for an open-weight company like yourself, what ends up being the business plan for Reflection? How do you end up making money then if you're releasing all the weights to the public? That's a really good question. so the goal for us is to actually enable enterprises sovereigns developers like a whole community of people who want to just like really own their intelligence and they want to have systems that you know they run on their infrastructure or like they run on the cloud that they fully control and in order to do that it's one thing to actually have access to the model but like you also like need all the software that goes with it, all the support that goes with it, so that, you know, you don't have just weights, but you actually have, like, outcomes and solutions.
6:16Ioannis Antonoglou: And reflection doesn't just provide model, but, like, also provides all the software and all the solutions that, like, come with it. And this is the thing that you actually monetize. I want to ask you, Yanis, about some of the more interesting research-focused questions that we're dealing with right now in AI, given that you're at the center of it. So in these two schools of camp, there's the Jensen view on pacing, there's the Dario view on pacing. Where do you sit on that spectrum? So I think that it is, I think, first of all, it's just like really important to have like open models and have like an ecosystem of open models.
6:58Ioannis Antonoglou: I think that in that respect, were definitely aligned with Jensen. We don't see it as a zero-sum game, but rather we see it as a positive-sum game. So having a healthy ecosystem of open models is something that is really important. At the same time, I think that there's a question in terms of safety and security when it comes to the intelligence of these models and how fast they're actually moving. And we really believe that in order to have systems that are safe and secure, you need to just put those systems in the hands of the research community. So the research community can just take them, they can investigate, they can inspect them, and they can actually do the research that will ensure that these systems are more safe and secure.
7:47Ioannis Antonoglou: And one example that I always like to provide is what happened to the OpenAI agents and Higging Face, where the OpenAI agents, during training, attacked Higging Face, and Higging Face actually had to use an open model to defend themselves because the closed models would just refuse to cooperate with them, actually even for defensive reasons. So having models that are actually open is how you ensure that they're safe and secure and how you can ensure that people and companies and institutions can actually defend against cybersecurity risks. So help me understand that it sounds like you are very much aligned with Jensen's view that the onus of this really should fall on the companies to make sure that the models that they're releasing are safe.
8:40And one way to do that is, like you said, embracing your own open models because you have more control over it. What role do you see for frontier models then in the AI story, given that you're so focused on open weight. Is there a role for Frontier models still, or is the future that you see entirely open weight?
9:03Ioannis Antonoglou: No, I feel like there's going to be both open and closed models, the same way that there is open software and closed software. And I feel like fundamentally it will just be used for different use cases. I think that sometimes you underestimate the size of the AI market. It's going to really change like the economy like from the ground up so I think that the opportunities here for like open and close players are like massive and you know I don't see them again I'd say that like people in many ways see the CAEI as like a zero sum game you know open versus close or like you know China versus the US and we just like fundamentally see it as like a positive sum where like more players means more innovation.
9:53Ioannis Antonoglou: It means more competition. It means better outcomes for everyone. Right. What's your view on distillation then? So, I mean, I want to just like, because this is also like something that I've actually seen, that Beam was actually built from the ground up from scratch with like no distillation. So this is not something that we did when we were building Beam. I see that there's a lot of there's a lot of rumors on X and elsewhere that the Chinese labs are using distillation to catch up on the frontier I don't know how much that's true, maybe it is, maybe it isn't but like what I want to say is that you know I want to say that like one way or another they've actually done a really good job, the Chinese labs, I really want to say that they've really built something strong, and kudos to them for doing that.
10:55Ioannis Antonoglou: Fundamentally, there are times and conditions, and people should just respect the times and conditions. Well, and that's sort of what I'm trying to get your view on, because I'm sort of trying to understand if you are okay with distillation. We can put aside who's using it or not. That's not what I'm trying to get at here. But as a practice of you know developing one's technology are you supportive of distillation i mean you're you're from deep mind as well and we had uh corey uh from google on stage a couple weeks ago and he talked about you know there are places that where distillation is is very much accepted and okay so what's your view on it yeah so i mean i guess like uh my view is that uh each model provider has their terms and conditions, and these terms and conditions should be respected.
11:45Ioannis Antonoglou: And the closed labs explicitly say that you shouldn't be distilling from them, so you shouldn't be doing that. We've actually put out a fully permissive model, which allows for distillation. So people can respect our terms and conditions and still use it to distill BIM into a smaller model that's really capable, and they can run it locally, or they can just run it to solve particular use cases. So all I'm saying is that the Finders of the model should decide how the model should be used. Right. And, Yanis, last question for you. So we saw this week that the Trump administration named their AI, I think it was the AI Safety Council.
12:28I'm forgetting exactly what it is, but Jay Clayton is the head of that group. And we're sort of waiting to see how the regulatory perspective from the Trump administration will develop on AI. I know Reflection has been named in some of these policies as being a leader in the space. Do you have any insight as to where the government is going to go with its regulatory policies on AI?
12:54Ioannis Antonoglou: No, we don't have any such insight. We, as Reflection, want to talk to the government, like the U.S. government, whoever is on the administration, and ensure that our voice is being heard in terms of the importance of open models. And we're working with everyone there. Great. Well, Janice, I want to thank you for coming on. That is Janice Antonoglu, the co-founder of Reflection, here on TI TV. Meta and Microsoft are looking at ways to reduce the extent to which their employees rely on CLAWD, according to a new exclusive report from my colleagues Aaron Holmes, Jyothi Mann, and Kevin McLaughlin.
13:35I want to bring on Aaron to share more about what he found. Aaron, welcome back to the show. It's great to have you here. Thank you for having me. So Aaron, what is behind these Claude cuts that you've reported that Microsoft and Meta are trying to get done? So in both cases, we saw Microsoft and Meta sort of rapidly ramping up their spending on Claude earlier this year during what a lot of people describe as like the token maxing era, where there was kind of no checks on expenses by engineers to use AI coding tools. And now in more recent months, both of these companies are trying to pull back and set clearer budgets for how much they use Anthropic.
14:15And at least in Microsoft's case, they've been successful. We learned that Microsoft was on track to spend a billion dollars on its employees' use of Claude earlier this year, but they have since ranged that in, and it's about at least 33 % lower than that peak spending. So they've mostly done that by having employees use models from OpenAI, which in Microsoft's case, they get to use those for free thanks to their deal with OpenAI, as well as internal models. So Microsoft's MAI models and Meta's Muse code models instead of Anthropic. and do you get the sense that this is really uh just because of competition or um is there you know some some added control that they might have over using their own model i mean i know they're not necessarily open weight models here but i can only imagine that if muse is developed by meta perhaps they can control it a bit better using their own data yeah i mean in both cases they're switching to models that they are controlling as first-party models.
15:24So that essentially means they don't have to pay to use them. They only have to pay for the cost of the electricity to run the inference. So in both cases, it makes sense that these companies, since they have these sort of first-party alternatives, would want to use those instead of Anthropic. And I think it also speaks to how those alternatives have in some ways caught up to Anthropic versus maybe six months or a year ago where Anthropic was sort of the clear front runner in all things AI coding. So how much of a sting is this for Anthropic? I mean, these are two of their biggest corporate customers, right?
16:00Yeah, they're certainly both mega customers. You know, in Microsoft's case, it's actually interesting because I learned that, you know, Microsoft has cut back on its internal use of Claude. At the same time, they do also pay for Claude to power some co-pilot features for Microsoft's customers. And I learned that that spending is still growing. So it actually ended up roughly canceling out. And the total amount that they pay to Anthropic is roughly flat from highs earlier this year, which is not terrible for Anthropic, but it's also not great since they are approaching an IPO and would like to potentially show that all of their biggest customers are just spending more and more on their models over time.
16:43So, and sorry, this is the amount that Microsoft is paying for using, this is the Office 365 products and Word and PowerPoint, stuff like that? Yep, so 365 Copilot as well as GitHub Copilot, which is the AI coding tool. Right, okay, so yeah, so maybe that nets out. I want to go back to something you said earlier, which is that you said, you know, these tools are getting to be as good or maybe that suggests they are as good as Claude. So is that, you know, based on where Muse is at right now? And I mean, OpenAI, sure. I mean, I didn't realize that Muse was as good as Claude on some of these benchmarks.
17:25Yeah, so I mean, Meta only started, you know, testing Muse code with its own employees pretty recently, per our understanding. And I'm not sure if there's a broad recognition that it's as good as Anthropic. But, you know, in Microsoft's case, again, they get OpenAI's models for free. And I think specifically OpenAI's GPT 5.6 series, from what I've heard, is, you know, comparable to Anthropik's latest models. And they're very much sort of neck and neck. You know, I still hear that for certain tasks, one model is maybe better than the other. And so Microsoft engineers will still use Claude. In some cases, obviously, their internal spending on Claude is not going to zero, and they are still using it somewhat.
18:04But they've been able to cut back from the highs a few months ago. Wow. And so, Aaron, as you sort of look ahead in this story, we've got Microsoft, Meta. I mean, I can only imagine that, you know, Google must be thinking of a similar type of arrangement. Maybe I'm thinking about some of these other enterprises. Maybe they prefer to use some of these open-weight coding models. Like, well, where is your head going here as to how this story could unfold going ahead? Yeah, I mean, I'm definitely hearing across the board that companies that are using Anthropics models for AI coding are increasingly trying to get that spending in check.
18:43And I think that that looks like, you know, exploring other models on the market. It also looks like just trying to figure out how to budget for AI coding, which is such a new, you know, category of spend that a lot of companies are in their first fiscal year actually budgeting for. So I don't think that this is, you know, going to mean that Anthropics growth will necessarily slow down, but it might be a headwind to them just as customers start to scrutinize the way that they've been spending after a few months of pretty unrestrained growth. Great. Well, Aaron, I want to thank you for coming on.
19:17That is Aaron Holmes, our Microsoft reporter, here at The Information. Workday might have staved off some of the biggest threats of AI and the so-called SaaSpocalypse, but a new in-depth feature story about the company reveals that it is still not out of the woods just yet, and AI could very much still inhibit the company's growth. My colleagues Laura Bratton and Anita Ramaswamy reported that story. I want to bring on Laura to share more about what she found. Laura, welcome back to the show. It's great to have you here. Good to be here. So Laura, I want to start with this part of your story, and I'm paraphrasing here, but part of what you wrote here is while earlier fears that AI would help customers to completely ditch traditional apps have dissipated, these apps are still facing the challenge of proving they won't become, quote, dumb infrastructure.
20:09So unpack that for us. What exactly is this fear? What does dumb infrastructure mean in this scenario? I appreciated that very dramatic quote reading from you. I mean, I had to try, right? I mean, you're a TV show host, so what else are we here for. So what it means is basically we've seen from recent earnings growth from these companies that they're not going anywhere. Nobody's going to buy code Workday that has their critical HR and finance data, but what they do risk is losing this action layer. So how are their customers taking action in their apps? Are they using AI agents to tap into Workday and take action, or are they using Workday's proprietary agent?
20:55And if they're using their own AI, how are companies going to monetize that? And what they risk is, you know, the scenario in which outside agents tap in to apps rather than, you know, companies using Workday's proprietary tools and Workday not being able to turn that into a major moneymaker, you know, if customers rebel against these toll gates and then essentially these companies become a glorified database. if that makes sense. And so how close are we to that risk coming to fruition? I mean, what's happening in the market so far? I don't think we're close to that risk. I think it's very early days for a lot of software companies that are introducing new AI products and Workday in particular is early in adoption.
21:45We've seen that their revenue is growing, usage of their AI tools are growing. Customers using their AI agents grew 35 % from the previous quarter to about half of their overall customer base in the July quarter, which I think shows that they're picking up steam and AI-related revenues picked up more than three times to about$600 million. So they are seeing some progress, but it's not yet boosting overall revenue right but one of the things that you did go into detail about in the story were some of the reasons why despite it growing as you said customers do have some hesitations here of using ai um not just workday's own ai tools but also ai with respect to the data that that workday holds so talk a little bit about some of the reasons why customers might be holding off on adopting their AI tools?
22:41It's really interesting because it's both a bear and bull case for Workday, because a lot of its customers, consultants describe to me as more conservative when it comes to AI as a whole. So they may not be using AI for HR and finance tasks with the data they store in Workday, but they're not necessarily using Clawed, Copilot, or Codex to work with that data either. They're just not using AI at all. And that means that if they do eventually want to use AI for those things, they will likely turn to Workday or some other platform they see as secure, but that hasn't necessarily happened yet. And what about the cost and effectiveness of these tools?
23:21Where do they stand there? Yeah, another thing that I've heard from CIOs and consultants is that enterprises don't necessarily want to be purchasing AI capabilities from every single legacy software provider. So, you know, you're not going to see them adopting super costly coding tools like CloudCode, Codex, and Cursor, and then also buying every single suite of AI products from Salesforce Workday ServiceNow. They're going to be a little bit choosier about which AI capabilities they pick. Right. Now, you also talked to the story about Workday standing up toll booths to combat the risk of just turning into dumb infrastructure or just being used as a system of record.
24:09How are customers finding that? Has that been dissuading some of these platforms from using their own agents or how effective has it been? So like Salesforce right now, Workday is also in a period where customers aren't necessarily paying the full price of that toll. So Salesforce, for example, is going to start charging for its new connectors later this year, November. Workday, meanwhile, isn't going to start charging customers when they use outside agents to connect to its software through APIs until after next February. So I think that it'll be interesting to see if that grace period, within that grace period, they can drive adoption enough so that customers will want to pay.
25:03And I think that they'll have to show that there is some significant value in tapping Workday with outside agents because a lot of customers are resistant to the idea of having to pay to connect to their own data. We saw the reports that private equity firms might be circling Workday. And I guess over the course of your reporting, I'm sort of trying to assess what you think PE firms may have seen in the company, given that it sort of has AI is both the bull and the bear case, as you said, for the company. I mean, we know the PE playbook is cut, cut, cut, you know, roll up other firms into it. Where do you see AI falling into, you know, what could have been or might still be the buying opportunity for PE here?
25:59Yeah, I think, you know, Workday's trading at about four times revenue right now. and historically traded at about 12 times revenue. So it's really cheap for the value of the software. And I think that for investors, it does represent an opportunity, especially with former CEO, Anil Bushry coming back earlier this year. And he's a big product leader, super innovative, has a great reputation. And there is a major chance that Workday can make this AI come back and, you know, become a leader in its category. As we've seen cinnamon around Salesforce pickup recently, Workday stores even more critical data related to employees and finance that I think does represent an opportunity for them to capitalize on that data and bring customers inside of its four walls to take action on that data or find, like we said, some way to connect through tool gates that customers can find digestible.
27:06And Laura, I mean, you and Anita had the opportunity to speak with a couple of workday executives for this story. What surprised you about their comments around the AI opportunity or the threat of AI, broadly speaking? Yeah, they really played on some of the data fears that I've been reporting on about, you know, customers don't necessarily want the AI labs to have access to their sensitive business data. And if you go to Workday, which, you know, governments, healthcare systems, universities, large enterprises have trusted with their sensitive data for, you know, more than a decade in most cases, it will make more sense for them to use AI from Workday.
27:50And they're going to buy AI from Workday. That was their take. And also, I think it's just been interesting to see a lot of folks from Google come over to Workday. One thing they emphasized to me was that their product team and the folks working on some of their AI tools are completely new. And one of the executives I spoke with who moved over from Google highlighted this data fear and just said that, you know, I wanted to work on a critical software app providing out-of-the-box AI features to companies that may not want to build their own AI agents. And I do think most of the companies, the CIOs, et cetera, that have been adopting AI so far within their companies, it's been coding tools within their engineering teams.
28:33But it's still early for, you know, wider AI adoption among all enterprises outside of Silicon Valley. So I do think there is an opportunity here for Workday. Great. Well, Laura, I want to thank you for coming on. That is Laura Bratton, author of our Applied AI newsletter here at The Information. Doordash recently announced new AI tools, and it has been pushing deeper into the technology with an acquisition this year of Metis, an AI agents infrastructure company. I want to bring on the co-founders of Metis, Arian Shah and Ayus Sheth, who have now joined Doordash as the head of research and head of applied AI.
29:13Welcome to you both. It's great to have you here. Thanks for having us, Kosh. So we all got A names and I'm also the eldest of two brothers and my youngest name is also Aryan. So this is going to make everyone at home laugh when I send them this clip. It is great to have the two of you here. Aryan, I'll start with you. So I do want to talk about how DoorDash is approaching AI because there's a lot of cool things going on. But I want to start with the open weight story of the week. Reflection came out with a model yesterday. Mistral is out today. Mistral is saying that it's the strongest model that is developed outside of China.
29:56It's a research preview, so we don't really know yet. But my question is, do you believe that? What's your read on this? No, I think that's a great question. We haven't tested out the Reflection of Mistral models yet, so I don't want to speak prematurely in terms of their capabilities. But we are quite close with both teams, and we've seen how long that they've been spending on trading these models. And I do truly believe, I think they're going to be great. And net-net, more generally, I'm just really excited about the state of Western open source. I think there was kind of a period of time where it really did lag behind some of the models coming out of the East.
30:33And I think more than anything, kind of seeing these model releases heat up and seeing more labs take this seriously It's just net beneficial for all of us. So we're really hoping that the Mishra model is very, very good. And if it is great, I think that's, again, net beneficial for the whole ecosystem. But I would not be surprised. I think it was something like they trained it on, they were saying 4 ,000 blackwall chips. They had made massive data purchases. So I would not be surprised at all if it is a truly leading class model, this category. And Ayush, as you think about the open-weight landscape right now, Like, do you think that Western openweight models will at some point catch up to what we've seen come out of China?
31:18What's your read on this? Like Arian said, I think we've worked really closely with these teams and we've seen how much that they've been able to improve in the last couple months. I mean, Americans are smart and I think we're going to get there. And I think the infrastructure and the data pipelines and sort of how they've been training the models have been super intelligent. from the conversations I've had. So yes, I do think that America will catch up in terms of - Okay, and last question, we're gonna get to DoorDash, but Arian, I mean, how real is the desire for companies in America or companies outside of China to use American models, models that were not developed in China?
32:02Because this is a story that kind of, it ebbs and flows, right? Sometimes we're okay with it, sometimes we're not. But this seems to be the competitive advantage right now that Reflection is banking on, right? Because they're coming out saying it's not as good as what we're seeing in China. So is that a sustainable advantage right now, do you think? Personally, I believe so. You know, I think there are real use cases for Western models. We've seen it at DoorDash. We're seeing it across the industry, especially for certain sovereign use cases, right? I think it's really important to use Western models.
32:36and I think we are very bullish on the state of Western models. Now, I think one thing to emphasize is this is the first iteration of the model that are coming out. So this is the start of their flywheel, and I suspect that it's only going to get stronger and stronger as time goes on. And given the fact that they've basically pre-trained, mid-trained, and post-trained these models from scratch compared to some of their Chinese competitors that have other ways of kind of getting around some of these loopholes, in terms of model training, I think we're very bullish on the flywheel that a lot of these NeoLabs, especially American open source labs will be able to create.
33:07Arian, are you using any Chinese open weight models right now at DoorDash? We currently do not. You know, we've experimented around a ton of them. I think for right now, we are, again, trying to be very, very opportunistic and open-minded. So I don't think, you know, we want to say never is never, but I think we really are hoping and waiting for a phenomenal American open source model. Great. Okay, so let's talk about the AI tools that DoorDash released. This was last week. Ayush, take it away. I mean, let us know. What do we have on deck here? Yeah, so recently at Dashboard, we released Text DoorDash, which is a personalized way for users to interact with our ecosystem to get food that's personalized for them.
33:53Around seven months ago when we joined DoorDash, we kind of helped bootstrap Ask DoorDash, which is the in-app experience for getting food that's personalized for every user in an agentic way. We're able to now surface small businesses. We're able to actually help support our local economies better, which is our mission statement, through these agentic experiences. And text DoorDash is just something that we've seen offer another avenue for people to quickly reorder and quickly surf through items without having to navigate the UI of the app. So these are both, these are tools for both the consumers and then also the drivers.
34:34Is that right? Yeah, we also released something called DashBuddy, which is an AI co-pilot for dashers to help them navigate, figuring out where to drop off food, how to talk to customers. I think, as we talked about in Dash Forward, it's really hard for dashers, especially your first dash to figure out what the hell is going on. How do you drop off something? How do you pick something up? How do you communicate with the customer? And I think we've been able to navigate that with DashBuddy and build a pretty good product there. So, Arian, I'm thinking about Muse, I'm thinking about Instinct. Can those tools right now, can you order DoorDash with those tools?
35:17Are you able to do that? Yes, you are. I think we're working quite closely with the teams to bring the DoorDash experience to these personal agents. But also, I think we've been thinking a lot about what does it mean to give our consumers as many options to order as possible, whether that is inside their personal agents, like we use an instinct, or by using a different modality like TextDoordash, or something inside the app itself, like AskDoordash. We really want to meet our consumers wherever they're at. And so we're just trying to create a paradox of optionality for them, whether it is a personalized, tightly scoped commerce experience like TextBordash or a more general purpose kind of experience like a Muse or Instinct.
35:56Why not block access to Muse or Instinct? This actually relates to the segment we were just having, talking about Workday as a system of record and them using their AI tools as competition to other platforms that pull from them. So, I mean, DoorDash, you could make the argument like Amazon did. Maybe it would be better for you guys to turn those agents away and focus only on the techs, DoorDash. Walk me through your thinking on this, Arian. You know, it's a great, great point, Agash. And I think it kind of goes back to the point that you made earlier around our end goal is we just want to empower local businesses and local economies, right?
36:42And so we want to do whatever we can do to offer the best experience for our consumers, for our merchants, and our dashers. And we think that by blocking access to these personal agents, it could potentially lead to a long tail of downstream opportunities that our consumers would not necessarily be happy with. And so while these personal agents are still very early in their adoption journey, I think it's something that we are relatively bullish on long-term. And we want to make sure that DoorDash is well-positioned to support those agentic experiences. At the end of the day, compared to maybe a traditional SaaS application, DoorDash is a deeply physical business.
37:18It's an operationally very complex business. And so we think that there's a deep moat to be had, regardless of whether we're inside of some of our third-party partner experiences or inside of our first-party kind of partner experiences as well. Ayush, do you think you guys would ever put up any kind of tolls for these agents to access or do things with DoorDash at all? Yeah, that's a great question. I think this is definitely a new paradigm. I mean, right now you're seeing, I mean, let's think about how this was sort of built with the internet. DoorDash had an API. Like DoorDash still has API for developers that you have to pay for.
37:54Google has APIs that you have to pay for. Maybe that we will, when we release our MCP to some of these third-party customers, we will charge some sort of Rails or charge some sort of markup. But until then, it's really, really experimental for now. And we'll just see how it plays out. Okay, so let me ask you a question before you go, Arian, about how DoorDash is using AI internally. Because it strikes me, you guys have Dash Bench, which is, you know, I don't think anyone ever thought about the burrito delivery company to get into the business of benchmarking. But, you know, you do have some research ranking these models.
38:39I know you're using them internally for your own coding tools at the company too. So where does the company stand right now insofar as how employees use AI? Do you have caps on how much they can use for agentic coding? Because, I mean, these costs are getting quite expensive, right? Yeah, that's a great question. We have not instituted any hard caps. So we do have some soft caps just so we're making sure people are monitoring their spend. But I think at the end of the day, we really do want to incentivize everyone to use AI to the max. I don't think we want to constrain people in terms of their creativity and also the diffusion and transfusion of AI inside the org.
39:20You mentioned agentic coding. I think that's one very large pillar. We're also thinking a lot about, well, how can we embed AI and make everyone AI native, whether it's our operators or other business function teams as well? And you made up a great point. Not many companies, especially a burrito delivery company, have public-facing bedfarts. and that's kind of rare. And it goes back to, well, for all of our core use cases at DoorDash, the first thing that we try to do is we try to build an internal eval, right? And so that's what Dashbench was. That was just our internal eval that we decided to make public.
39:50But it gives us confidence in terms of, one, building the best system for our use cases, but also allows us to evaluate what models are the best for our specific workflows and our specific wants and needs. And it allows us to then also make that cost and performance trade-off as well. So over time, if we start to see costs are really starting to balloon, well, we have kind of a vast majority of options along this whole Pareto frontier that we can pull from and that we now have confidence in because we can see how they perform on our internal evals and, again, inside of our workflows. Great. Well, I want to thank you both for coming on.
40:24It was a fun conversation. I know that my family group chat is going to get a kick out of this. And Ayush, if you want, you can change your name to Arjun, and then I will literally just have my two brothers on the show. All right. Well, thank you both for coming on. It was great to have you. That is Arian and Ayush from DoorDash here on TIT. That does it for today's show. A reminder, we are on this stream Monday through Friday at 10 a.m. Pacific, 1 p.m. Eastern. If you can't make it then, episodes are available on theinformation.com, on our YouTube channel, or wherever you get your podcasts. Make sure to follow us on social media, on X, on Instagram, on TikTok, and on LinkedIn.
41:03I am already excited for our next show tomorrow. Have a great rest of your Tuesday. Bye-bye for now.
Reflection Co-Founder Ioannis Antonoglou talks with TITV Host Akash Pasricha about Beam, the company’s new open-weight model, efficient reasoning and the business model behind open AI. We also talk with The Information’s Microsoft Reporter Aaron Holmes about Meta and Microsoft’s efforts to curb internal Claude spending, competing models and the customer-facing demand that has offset Microsoft’s internal cuts. The Information’s Author of Applied AI Laura Bratton joins us to discuss Workday’s growth challenge, cautious enterprise buyers and its plans to charge for outside agents accessing its data. Lastly, DoorDash Head of AI Research Aryan Shah and Head of Applied AI Aayush Sheth discuss the US-China open-model race, consumer and driver AI tools, third-party agents and the company’s approach to internal AI spending.
Articles discussed in today’s show:
- https://www.theinformation.com/articles/meta-microsoft-work-wean-staff-anthropics-claude
- https://www.theinformation.com/articles/saaspocalypse-workday
Subscribe:
- YouTube: https://www.youtube.com/@theinformation
- The Information: https://www.theinformation.com/subscribe_h
Sign up for the AI Agenda newsletter: https://www.theinformation.com/features/ai-agenda
TITV airs weekdays on YouTube, X and LinkedIn at 10AM PT / 1PM ET. Or check us out wherever you get your podcasts.
Follow us:
X: https://x.com/theinformation
IG: https://www.instagram.com/theinformation/
TikTok: https://www.tiktok.com/@titv.theinformation
LinkedIn: https://www.linkedin.com/company/theinformation/
Chapters:
01:13 - Introduction
02:16 - Reflection’s Beam and open AI
14:24 - Meta and Microsoft’s Claude cuts
20:23 - Workday and the SaaSpocalypse
29:53 - DoorDash and the open-model race
36:07 - Third-party agents and DoorDash’s moat
39:13 - AI spending and internal evaluations



