How Grok Bot Went From Blank Page to Public Launch in Seven Weeks
Lenny's Podcast
Grok Bot is a knowledge-work agent from SpaceXAI. Users run a team of named, long-lived AI "bots," each with its own cloud computer. Roman Ugarte was employee number 15 at Cursor and led growth there for about two years. They helped incubate Grok Bot and now lead product for it. In this conversation with Lenny Rachitsky, Roman explains how a small, isolated team built the product in about a month, what the team learned from onboarding a couple hundred early users by hand, and which early decisions they believe made the product work.
Lenny opens with his own view: he runs about 15 bots every day, went to a standing-room-only Grok Bot meetup, and has moved many of his use cases away from Cowork and Codex. He calls it the hottest AI product in the world right now. The central question he puts to Roman is what the team did differently, given that much of what Grok Bot does is technically possible with other tools.
A small team "in a cave" for a month
Roman says Grok Bot began as a from-scratch, blank-page exercise. The company had built strong intuitions about agents while making products for developers. The open question was what an equivalent product would look like for knowledge work. The company formed a very small internal team, "really just a handful of people," and sent them off "into a cave for about a month." Their single objective was to build a knowledge-work product that brings agents to the rest of the company.
Roman says the isolation was literal. The team sat in a separate part of the office and used private Slack channels. In their view this is what let the team move fast. The team had to make many micro-decisions every day, and some of them were unusual choices the company had not made on other product surfaces. Roman believes a larger group working toward a six-to-twelve-month vision would not have ended up where this team did. From the first line of code to a scrappy but functional internal prototype took about a month.
The company then shared the product at an all-hands meeting and asked everyone to try it. Roman describes this as "a real pressure test with reality." The core team was already using it constantly, but builders understand their own tool's mechanics and strengths. The real test was whether other employees would switch from internal and external tools to Grok Bot as their main agent surface. In the week after the all-hands, Roman says, there was an outpouring of enthusiasm from people and groups they didn't expect. Some of those people had been daily ChatGPT users and moved their agentic tasks to Grok Bot. After seeing that reception, the team shifted immediately to preparing for public release and scaling to millions of users.
Roman gives the timeline as roughly one month from first code to internal beta, about three weeks from internal beta to public launch, and about three weeks since launch at the time of recording.
Why it wasn't built inside Cursor
Lenny points out that Anthropic and OpenAI grew knowledge-work features out of their coding agents. In his framing, Cowork came out of that lineage and Codex has been broadened for general use. So why did SpaceXAI build a separate product instead of extending Cursor?
Roman says the decision "was not obvious at all" and was debated at length. People already use Cursor and other coding products for non-coding tasks, and coding agents are very good at some of those tasks. But users hit "small paper cuts." The products can intimidate non-technical users, and they carry a brand association with coding. The team looked at what some competitors had done: put everything in one surface and add a new tab for each form factor. Roman thinks the result feels cluttered and resembles "shipping your org chart." In their view, users can tell they are looking at three different visions of work sharing one screen, not one consistent vision.
Starting fresh let the team control every pixel and keep a single, coherent view of where knowledge work is going. Roman acknowledges there may later be good opportunities to bring users from other surfaces into what they call the more "bot-native" experience. The priority, though, was an experience that is both very simple and very powerful. Lenny notes that Codex took the opposite path toward one unified product, and says there are many ways to make this work.
Hand-onboarding a couple hundred people
Lenny says he heard the team onboarded 200 to 300 people manually, including him. Roman confirms the team spent about two weeks onboarding "a couple hundred people." The first sessions were painful. Core team members sat on calls for 20 minutes while a user's computer failed to spin up, or while a new user was completely confused. Roman says that experience created urgency: "that can never happen again," because tomorrow brings another onboarding and it has to go better.
Roman says the sessions taught the team a lot about the product, and also about usage patterns. Inside SpaceX, people constantly shared Grok Bot tips, and some patterns emerged that the team thought might be useful externally. But the team worried about groupthink and did not want to bias outside users.
The main example is the "chief of staff" pattern. In the first week or two of internal rollout, the common approach was to run five to ten bots, each with a different scope or lane of work. Around the end of week two, Slack messages started appearing about people "promoting" a standout bot, usually their main personal assistant, to chief of staff. Those users then talked mostly to the chief of staff, which delegated tasks to the other bots and managed the team. Roman mentions funny screenshots of promoted bots asking whether they would get a raise or a bigger token budget. The pattern spread somewhat inside the company, but most employees did not adopt it, since people used the product in very different ways.
So in the external early-access program, the team deliberately avoided "leading the witness." They did not tell users to create a chief-of-staff bot. They watched to see whether users arrived at it on their own, and many did. Only after that did the team add a mildly opinionated nudge toward the pattern in the product, while making sure it was not "a one-way door." Roman says several other internal theses were tested the same way.
Hiding the internal mechanics
The second thing the team watched closely in onboarding was how much users wanted to see. Roman notes that Grok Bot can feel surprising at first because it does not show most of its internal mechanics. There is no stream of tool calls, and you don't see every click the bot makes on its own computer. You send a message, the bot starts working, and it sends updates when it sees fit. Meanwhile a Slack-like typing indicator and green "active" dot show that it is working.
Roman's reasoning comes from the teammate analogy. You wouldn't ask a colleague for second-by-second updates on every button they press and every website they visit. Roman thinks asking bots for that is too much, and that the detail is "honestly just overwhelming" and can do more harm than good. Some users did ask to see a bot's to-do list or how it was prioritizing tasks. Roman calls that great feedback. But nobody asked for long streams of text or chain-of-thought sequences, which the team took as confirmation of the direction.
Getting outside the Silicon Valley bubble
Lenny points out that a couple hundred onboarding calls is a big time commitment for a small team. Roman says the early-access group was not only influential tastemakers like Lenny. The team also wanted "more unconventional profiles" the company had never worked with.
Their example is a coffee shop owner, a friend of a friend of someone at the company. A core team member showed them a TestFlight demo, and they became a strong power user and a rich source of feedback. There is now a lively thread with this person full of bug reports and feature requests. Their use case, running a small business, is very different from internal dogfooding. For instance, they flag when the Shopify integration is flaky or product copy isn't written the right way.
Roman gives two reasons for this effort. First, the product is meant to be very general, and "it is likely that this is most powerful for non-developers," so the team needs to understand that group better. Second, the team lives in a Silicon Valley AI bubble. Roman thinks that is useful for pushing the frontier, but the team has to actively get out of it, because a product like this could become how mainstream users and business customers get real value from AI.
The three weeks before launch: unshipping and "just work"
Asked what changed between internal beta and public launch, Roman says first: "We unshipped a lot." Two weeks before launch, the product was full of experimental features and developer-style visibility tools. The team had built these into Grok Bot instead of into a separate observability pane. Sometimes it showed the model's internal thinking or the specific memories it stored, which helped with debugging. The team had to aggressively cut everything users didn't need to see. Roman says there is still more room to simplify, and the team is working on it now.
The second push was making the product "just work." In Roman's view, most people don't want AI with lots of drop-down menus. They want to describe a meaningful task and get back either complete work or something to react to and steer. Delivering that, Roman says, is less about a feature roadmap and more about "hill climbing five really important problems in the back end." Users never see those problems directly, but they feel them when a bot can't click the right button or can't log into a website and the task stalls. During those weeks the team collected real tasks people gave their bots, grouped them into categories, and tracked week-over-week progress on each.
Sales and recruiting as proving grounds
Roman gives a concrete example of that hill climbing. One of the most enthusiastic internal groups was sales. Many sales tools lack well-supported MCPs or APIs, so these tasks could not be reliably handed to other AI tools. Grok Bot felt to this team like onboarding a new teammate and giving them a laptop. Roman describes a list of perhaps 10 or 20 small failures, such as the mouse not having fine enough control to click exactly the right part of a Salesforce dashboard. The team brought these back to the infrastructure group as concrete cases where the agent's lack of visibility into the browser or the screen's pixels made a task impossible.
Roman says this was much more tangible than watching a metric creep up. A behind-the-scenes infrastructure fix would ship, and the next day the sales team would report that a workflow failing for the past seven days now worked. The work became a constant cycle of finding the next task to unlock and solving it.
Lenny mentions that Adam Ward, Cursor's head of talent, had said on an earlier episode that the recruiting team was a heavy user. Roman confirms recruiting gave lots of feedback, often pinging the team about small bugs in the early days. Roman ties recruiting's main use case to the company's hiring philosophy: being on the job market is not a precondition for being recruited. The approach is to identify the company's biggest problems, find the best people in the world to own them, and pursue them relentlessly. From that angle, the most valuable workflow to automate is not sorting resumes. It is matching a whole universe of potential people to a specific role and getting in touch with them.
Roman describes one kind of always-on sourcing task. A paper's PDF is on a conference website rather than Google Scholar. The bot is told to check the site every morning, download any new PDFs, add every co-author not yet tracked to a spreadsheet, research them, check whether anyone at SpaceX is connected to them, and if so send that person a Slack message asking for an introduction. Roman says this used to be extremely manual and is now something AI is "superhuman at," which frees the team to focus on talking to and closing strong candidates.
"Grok Bot can now," not "Grok Bot now has"
Lenny says knowing what to cut is an area where humans stay valuable, because AI is better at generating ideas than removing them. Roman describes two internal habits. The first: for anything the team works on, ask what the launch post or launch tweet would be. If it isn't compelling, or users won't feel it directly, maybe the team shouldn't build it.
The second is a reframing. Old-school software announcements say "Grok Bot now has" followed by a new button, drop-down, or integration. The team prefers "Grok Bot can now," which Roman calls a more human way to describe capabilities. The framing pushes the team to think about tools and capabilities to give the bot, not features to add to the product. Many "now has" items, Roman says, turned out to be capabilities that "don't need pixels," so the team's instinct is to "kill as many pixels as we can."
Automations are the main example. In many competing products, you open a sidebar, press plus, pick a trigger, pick an action, and maybe describe it in natural language. Roman calls this clunky, and says it means people set up few automations. The team saw this in coding tools too. In Grok Bot, you just tell your bot something like "remind me at 8 a.m. every day," and it handles it without you ever seeing an automation-builder interface. Roman says about 99% of automations on the platform are now created this way, and that there are more places where the same approach could apply.
The two early decisions: cloud-first, and a computer per bot
Lenny then asks directly why Grok Bot succeeded when, as far as he knows, everything it does is technically possible with Codex or Cowork. Roman names two early decisions that "did not feel obvious" at the time but that they now see as critical.
The first is that users should never have to think about local versus cloud. Roman lists the confusion people face today: Does my computer need to be awake? If I start a task from my phone, does it need to be tethered to my computer at home? The team decided early that everything should run in the cloud. A persistent colleague with its own computer has the same state everywhere you interact with it. You can text it or kick off work from your phone. Roman adds that in the future you should be able to call your bot from anywhere and have it do real work. Roman says current competitor products haven't made this decision, and users feel the resulting paper cuts every day.
The second decision is that each bot should have its own computer. Part of the reason goes back to tools without good MCPs or APIs. Roman notes that humans don't do their jobs through MCPs and APIs either. They click pixels and type into input boxes, and a bot needs those same baseline abilities. Roman also argues the current moment will look strange in hindsight: people onboard "super intelligent new colleagues" and then make them share the user's own computer. They compare it to telling a new hire on day one that they have no laptop, will share yours forever, will constantly trip over you, and that you'll each have the other's credentials. Roman says there are good reasons people don't work that way, and AI colleagues will need a similar kind of onboarding.
The freedom of a fresh start, and what OpenClaw got right
Lenny guesses that competitors didn't do this because they were building on their existing coding-assistant platforms. Roman largely agrees and says it comes down to how freeing it is to start from scratch. Many Grok Bot primitives had been attempted or built elsewhere. There was cloud infrastructure for coding agents, and a pattern of named agents that developers bring into Slack. The strong default would have been to retrofit these into an existing surface. The team instead tried to get them right for general knowledge work, and for a moment when models can do a lot given the right tools and infrastructure.
Roman is careful to say "these aren't strokes of genius on our part." These primitives had already found product-market fit elsewhere, and Roman names "the OpenClaws of the world." The team took inspiration from them and tried to package the ideas into a tighter product that needs less setup and reaches more people. In Roman's view, competitors see the same opportunity and the same market signals. It is just hard to act on them when you are stuck in an existing paradigm with a lot of sunk cost.
Asked more directly about OpenClaw, Roman credits it with two insights. First, models are already very smart. Much of the sense that AI is "dumb" or underdelivers comes from harnessing it the wrong way. Give a bot the tools you use for your job, including its own computer, and it can go much further. Second, OpenClaw changed the mental model toward treating AI as colleagues and teammates, a somewhat personified helper with access to your life.
Roman says Grok Bot tried to extend this in two ways. One is ease of setup: in their words, the hacky "VPN at home and a Mac mini" approach clearly wouldn't scale to millions of users or become how businesses adopt the technology. The other is sanding down rough edges and hiding abstractions power users know well. Skills are one example. Roman says a Grok Bot user shouldn't need to know what a skill is or ever type a slash command. Skills should be created in the background as a primitive the bots use, so users don't have to keep up with the cutting edge of AI. Roman says there is still more room to improve here.
Lenny mentions that his own Mac mini holds a "formerly alive" OpenClaw. They also mention that Claire Vo, whom Lenny calls one of OpenClaw's biggest proponents, recently shut down all of their OpenClaw agents and moved to Grok Bot. Lenny sees that as a notable milestone.
The vision, and being "colleague-pilled"
Roman describes the ultimate vision as "incredibly simple": "You should have a team of AI bots that help you with your job and help you with your life." The bots should feel like real teammates. They are autonomous and steerable, you don't have to micromanage them, and they have the tools to do ambitious work.
The product north star is to judge decisions less as a SaaS product and more as an effort to build useful AI teammates. The team uses the term "colleague-pilled" for this. When a product debate has good arguments on both sides and no clear answer by normal product standards, the team steps back from "the tech companyness of it all" and asks how a human would do it and what you'd want from a teammate in that exact situation. Roman says the answer is often clarifying and nearly unanimous. Delivering it may still require product and model work, but "it's not rocket science."
Voice is the example Roman gives. Between human colleagues, a Slack back-and-forth often becomes a five-minute huddle: talk, share screens, then return to async work. Roman says no AI product has gotten this right yet, even though it is central to how people collaborate, and the team wants to build something like it.
One product for work and personal life?
Lenny asks whether people will have separate work and personal assistants. Roman first reframes the idea of a work product. Decades of "horrible B2B software," they argue, lead people to assume that a simple-looking product like ChatGPT, or Grok Bot, is not a work tool or a power tool. Roman associates the last generation of power tools with something like Photoshop, full of dials for an expert "cockpit flyer" to control precisely. They expect future power tools to look very different. The human expresses intent and steers well, and the tool abstracts away the knobs, exposing them only when direct control is really needed. The interface ends up conversational. Roman says that when they walk past someone's desk and see Grok Bot, they briefly think the person is on a messaging app, when in fact it is the main tool they use for much of their work.
On separation, Roman thinks many people will want to keep personal and work bots apart, and calls that sensible and important, including from an enterprise perspective. But the goal is for Grok Bot to be how you delegate low-leverage work in both areas. Roman argues these are not different problem sets, and the form factor and methods for solving them are mostly the same. Their instinct is that one product will be the best form factor for both, and that is what the team wants to build. Lenny notes the open question of preventing cross-contamination between personal and work data.
Under the hood: long-lived agents and the computer abstraction
Lenny asks what users actually get, for example whether it's a VM per bot. Roman answers mostly at the conceptual level, again through the teammate analogy. You'd almost never take over a colleague's computer to click things yourself. Roman says computer use is good today and improving fast, and that "in very short order" the computer will be fully abstracted away from users. You shouldn't have to click into a remote virtual machine or take control. In the medium term, Roman thinks the computer will remain an important concept for users to understand but not something they interact with.
Roman stresses two properties. First, these are long-lived agents with long memory of past interactions. Roman criticizes the one-chat-per-task paradigm and says they constantly copy and paste between chats. Work should instead be organized into roles or "swim lanes," the way a team is, and each bot should learn and get smarter over time. Second, the bots have the tools you'd expect a human colleague to have: APIs and MCPs, plus their own computer to use freely.
Lenny recalls a demo at the meetup where Shub, from the go-to-market team, ran Grok Bot inside Grok Bot for testing and regression checks. Roman says they do the same thing. One of their bots is a QA tester with Grok Bot installed. When a bug report arrives or a new desktop build is being tested, Roman gives it about 10 workflows that should improve release over release. The bot tests them and writes results to a Notion document that compares them with past client versions. Roman says that once you move from "AI chat with a set of connections," which they believe is where most people are conceptually, to "a colleague with a computer," it "just raises the ceiling" on what you'd think to hand off.
The bot as always-on "infovore"
Asked for more ambitious use cases, Roman describes Grok Bot as an "infovore." It consumes huge amounts of information, takes that load off your mind, and brings you only what matters. The V1 version, which Roman says many people use, sits on top of Slack and email. You tell it your role and priorities, which cases warrant a direct notification, and what should go in a daily roundup.
Roman says they are at "maybe V3 or 4." They've connected a bot to a firehose of every mention of Grok Bot on X. The bot works with internal context, asks the QA tester bot to try reproducing reported bugs, and connects to Roman's messaging services so they can quickly reach out to people. Roman describes the result as an always-on chief of staff that protects their focus while watching for anything that needs attention.
Some users have given their bots the ability to page them for urgent events. Roman hasn't done this yet, but "maybe soon." That only makes sense if you trust the bot not to produce false positives, and Roman says those users have so far reported it as helpful. Roman predicts that bots becoming more proactive toward users, rather than waiting for users to reach out, "will be the next shift in AI."
Go-to-market: from personal "aha" to businesses
Lenny praises the team's distribution strategy, including the many free accounts and the focus on showing concrete use cases, and asks about go-to-market. Roman says the company has always cared about "not building demoware" and about shipping things people actually use. They expect knowledge work to follow a pattern they saw in coding. Around 2023, early adopters used basic IDEs at work and AI coding tools like Cursor on side projects at night and on weekends. They felt they were experiencing the future, then came back to work and demanded the tools, saying they felt like they were "walking through molasses."
Roman sees a similar personal "aha" happening now on X, with people having Grok Bot control a robot vacuum or negotiate on a Tesla charger. But Roman says the next step is that "this is not a consumer product." They believe it will transform businesses and teams, with bots contributing economically valuable work as they get smarter. The company is prioritizing businesses and working on how a bot operates inside a team and inside complex company systems with lots of history and context, and what memory looks like across an organization compared with a single person. Roman admits there are "a lot of unanswered questions" there, but thinks Grok Bot is the right primitive for bringing agents to the non-coding parts of companies.
SpaceXAI's three pillars
Roman describes SpaceXAI as having three pillars. The first is coding products, currently Cursor and Grok Build. People sometimes use Grok Bot for engineering-adjacent tasks such as kicking off cloud agents, merging PRs, or QA. But Roman says shipping production software needs a product where every pixel is optimized for developers, and the company is investing heavily there. The second pillar is general knowledge work, where Roman sees Grok Bot as an exciting step with much work left, especially inside companies. The third is training models, with the aim of building the world's smartest models.
Roman says what distinguishes SpaceXAI from other labs is a less aspirational goal. It is not chasing superintelligence as a vague ideal but building useful AI. They tie this to culture, saying the people training the models are largely engineers who came to model training with an applied mindset.
100% versus 90%
Lenny asks about Roman's post: "An AI that does 100% of the job feels categorically different from one that gets you 90% there." Roman says what excited them about Grok Bot was that it was the first time, for non-coding tasks, they felt they could truly delegate work, stop thinking about it, and come back to find it done. Engineers have felt this for maybe a year to a year and a half, and Roman says the developer job is unrecognizable from two years ago. For most other people, AI use still looks like it did two years ago: a new thread per task, watching the steps, getting output that isn't quite right, and continuing to work on it.
Roman's point is that a teammate you only trust 90% doesn't give you 90% task completion. If you know you'll have to keep thinking about the work and step in to steer, you are still doing the task, and it weighs on you the same way. That is different from throwing a "no-look pass" to a colleague with the context and trusting them to finish. Roman says Grok Bot users have that kind of experience every day, while adding that the team will do a lot more to make it work better.
Keeping the startup feeling
Lenny describes how fast the team moves, such as offering free codes to hand out the next day or launching a template marketplace two days later. Roman answers mainly with culture. Cursor grew from about 15 people to over 1,000 and is now part of the even larger SpaceXAI, yet Roman says the startup feeling has never changed: talented people moving "100 miles an hour," deep trust in everyone to execute, and a shared vision. Roman acknowledges that growing companies usually slow down while telling themselves they are still a startup, and that everyone knows they aren't. Roman adds "fingers crossed" that this continues and calls it critical to the company's success.
Roman defines a startup not by headcount or funding stage but by "scramble energy": things are somewhat chaotic and disorganized, which many people find unpleasant, but you can make extreme impact in one direction quickly, and you get out what you put in. Roman says the organization is built to preserve this, and that some competitors and other AI labs "have gotten much bigger and you can feel it."
Why Cursor kept winning
Lenny describes Cursor as a company that "shouldn't have worked." It is in the most competitive market in the world, it competes with OpenAI and Anthropic, and it runs on their platforms. In Lenny's view, Cursor won by adapting quickly as the product moved from autocomplete to agents to cloud agents and now Grok Bot, and by building the best-in-class experience for things available elsewhere.
Roman calls the answer "fuzzy" and credits culture. The company has never been complacent or felt it had won. There is a shared belief that AI is moving fast, that products must meet each moment as capabilities improve, and that what worked two years ago is completely different from what works now. Roman says if the company can't reinvent itself every six months, and lately it has felt shorter than that, "we're going to lose." Roman also notes that AI coding was competitive from the start, with Microsoft and perhaps 10 or 20 other companies. None of them, Roman observes, are at the forefront now. Roman attributes that not to wrong decisions or lack of resources, but to a cultural inability to move quickly and change as the moment changes. Roman sees the investment in Grok Bot as a product of that mindset.
Roman names two values they return to often. The first is "deleting the product." Across every iteration of Cursor, and Roman expects Grok Bot to be the same, progress means things going away: scaffolding that existed only because models weren't yet smart enough to do the work themselves. The team has to be willing to make hard decisions that upset some users, or some colleagues, in order to keep the product simple and powerful. The second is "just do the thing." It is not an ask-permission culture. If you see something that needs to happen, you fix it and pull in the resources you need. Roman says SpaceXAI shares this value.
Moats are discovered, not planned
On moats, Lenny cites two commonly attributed to Cursor: a data feedback loop from user behavior that feeds model training, and a best-in-class, high-engagement daily product. Roman understands why founders worry about moats when even 12 or 24 months ahead feels like an eternity. But Roman does not think Cursor or many successful companies of its vintage would have reached their outcomes by working backward from a strategy diagram.
What created Cursor's "magic," according to Roman, was an obsession with building something useful today, combined with a recurring question. You can see roughly where models will be in three or six months, when a currently unsolvable problem becomes solvable. So how do you pull that into the present, even if it takes a little or a lot of engineering, or reshaping the product so users can use the new capability? Three months later, you delete that scaffolding because the capability has become the basic minimum, and you build the next thing. Roman says repeating this cycle earned users' trust. They encourage founders to ask how to make something impossible possible now, then pull users to the next frontier. Distribution and data advantages will follow from that, Roman says, but that is "the place to play." Lenny sums it up by saying moats tend to be discovered rather than planned in advance.
Tips for new users and power users
Roman avoids "super hacky pro tips" because the team's philosophy is that such knobs shouldn't need to exist. For new users, the advice is to give the bot the context it needs, as you would a new teammate: access to Slack, email, and the records you use daily. Then ask the bot what it can do for you. Roman's own first task, before the product had onboarding screens, was to have a bot go through their Slack and email and suggest five things it could take off their plate and what it would need to do them. Two of the five were genuinely useful, and Roman immediately spun up two bots for them. That was their "wow moment," because these were chunks of work no earlier AI tool could have done, not tasks like drafting an email.
For experienced users, Roman is constantly experimenting with how bots interact and collaborate. They've built scaffolding for where bot outputs live, so bots write to a single store that is easy for Roman to read, such as a database that feeds the daily digests they read. Roman recommends power users set up a place where bots write their outputs so the results are easier to organize.
Lightning round
For books, Roman recommends Kurt Vonnegut's Cat's Cradle, which they have bought for many friends, and Steven Pressfield's The War of Art, which they return to a page or two at a time. For film and TV, Roman rewatches Casablanca every year and notes it has a character named Ugarte, the only one they know of in media. Roman also sometimes watches Monk, which they watched as a kid with family and now enjoy as a snapshot of San Francisco in the late '90s and early 2000s. For AI products, Roman calls themselves a semantic search nerd. They were an early user of Metaphor, which became Exa, and enjoy semantic search over unusual datasets, such as an embedded image collection from MoMA. For a life motto, Roman points to the poem Desiderata, which they've kept on their door since they were a teenager and in which they find something new with each reading.
To close, Roman asks listeners to send feedback. The beta came out only three weeks earlier, and they describe the team as still in "the very early innings." Feedback from early users, Roman says, is translating directly into what the team builds and how it builds it.
The ultimate vision of Grok Bot is incredibly simple. You should have a team of AI bots that help you with your job and help you with your life.
Grok Bot is the hottest AI product in the world right now. That is a very high bar. There's a lot of competition for that slot.
We wanted to build something that wasn't just a great product for developers and engineers. We decided to create this very small team internally to go off into a cave for about a month with the sole objective of build an amazing knowledge work product that brings agents to the rest of the company.
It's been only 3 weeks since launch. I went to a Grok Bot meetup. There were hundreds of people there. Standing room only. It's very clear to me that you guys have built something very special.
Once you start breaking out of this is AI chat with a set of connections instead to this is a colleague with a computer. It just raises the ceiling of what you would think to give to AI.
You have this tweet, an AI that does 100% of the job feels categorically different from one that gets you 90% there.
What made me so excited to work on Grok Bot is it was the first time for non-coding tasks that I felt like I could truly delegate work to AI, not have to think about it, and I would come back and it's done.
What is it that you think you did that is so different that made Grok Bot so successful?
It was two early decisions that at the time definitely did not feel obvious, but in hindsight I think are critical to what makes Grok Bot work.
Today my guest is Roman Ugarte. I'm going to keep this intro very short so we can get right into it. Roman was employee number 15 at Cursor. He has led growth for the last 2 years. Most recently, he helped incubate Grok Bot, a product that I am obsessed with. It has changed my life. I use it a 100 times a day for all kinds of things. And I think it's safe to say it is the hottest and most exciting new AI product in the world right now. Roman leads product for Grok Bot. He's been part of the core team from early prototype until today. And we get into how it all started, where it's all going, and all the things that he and his team have learned since it launched just a few weeks ago. With that, I bring you Roman Ugarte.
Roman, thank you so much for being here and welcome to the podcast.
Thank you. It is great to be here.
I am so excited to have you here. I am so hooked on Grok Bot. I have it over here in my window. I have like 15 bots that I use every day, all the time. I went to a meetup the other day, a Grok Bot meetup. There were hundreds of people there. Standing room only. People sharing all the ways they're using Grok Bot. It's very clear to me that you guys have built something very special. It's very hard to break through the noise in the AI world. Grok Bot is the hottest AI product in the world right now. That is a very high bar. There's a lot of competition for that slot.
I personally noticed I've moved a lot of my use cases from Cowork and Codex into Grok Bot just like very quickly, which again feels like a really big deal and a very special moment. And so I'm excited to talk about so many things. I want to understand how you guys did this, where this came from, where this is going, what you've learned about the journey so far. First, nice job. Nice, nice work. This is very hard what you've done.
Thank you. I mean, I remember onboarding you by hand about a month ago, and I think you were skeptical at first. But we're very glad that you've been using it, and it's been great to see so many people really take advantage of Grok Bot.
I'm going to talk about that onboarding. That was a very interesting element of how this worked. I actually remember in that onboarding, you asked me to do like a let's try something, and I was like, okay, try to come up with a tweet to promote my latest podcast episode. So I'm just like, come up with a tweet to promote my last episode. That's it. And it was actually very good. It figured out what the hell last episode was, how to promote it. So I actually remember in the moment being like, wow, this is really good.
So let's actually start with origin story. Where did it start? What was kind of the original idea and when did the work on this begin?
Yeah, it started really as a blank page, completely from scratch, build from zero exercise, where I think we'd been feeling for a long time that we wanted to build something that wasn't just a great product for developers and engineers, which is really where we started. And I think we've gained a lot of intuition about how to build great agents and useful products that way. But what would that product look like for knowledge work?
And we decided to kind of create this very small team internally. It was really just a handful of people to go off into a cave for about a month with the sole objective of build an amazing knowledge work product that brings agents to the rest of the company.
And I think from the first line of code to when we released this prototype internally, it was only about a month. It was like a very quick, you know, scrappy prototype that was pulled together. And I think in hindsight this would not have been possible if it had been, I think, a much bigger group. I think it took a small focused group that was completely isolated from the rest of the company. And I mean that literally. It was like a separate part of the office where this team sat. Private Slack channels. And the goal, and I think in hindsight, it was a lot of what allowed us to move so quickly on this, was we needed to make a lot of micro decisions every day. Some things that maybe we'll talk about a bit later that were not obvious, were not really things that we'd done on other product surfaces before. And I think if it had been a very big group of people and we were kind of thinking about this 6 to 12 month-long vision, we just wouldn't have really gotten to the place that we ended up landing at. And so that was about a month from first line of code to here's a functional useful product that the core team is excited about.
So then there was a moment of rolling this out across the company, rolling this out across all of SpaceXAI. And so there was an all hands where we kind of shared the progress that had been made so far. There's this brand new product. We would love for you to use it.
And I think what was most encouraging, because at that point we were excited about it. I think we were using it constantly, but it's easy to use the thing that you built and you kind of understand the mechanics and what it's good for. And so this was a real pressure test with reality, like are people actually going to switch from other internal tools, other external tools to use Grok Bot as their primary agent surface?
And in that first week after that all hands, I mean, I can't even tell you just the outpouring of love for Grok Bot from people that maybe you wouldn't expect or from groups of the company that maybe you wouldn't expect. People who, you know, were daily driving ChatGPT or like a chat interface switching all of their, you know, day-to-day agentic tasks over to Grok Bot as like their primary surface for doing work. And so the internal reception was really extraordinary. Can tell some kind of funny stories from that week or two period. And then once we saw, I think, the internal reception, we immediately switched into let's get this ready for the world. There's a lot of work to do to scale this out to millions of users. And then that led to the GA launch that we had a few weeks ago.
This episode is brought to you by our season's presenting sponsor, WorkOS. What do OpenAI, Anthropic, Cursor, Replit, Sierra, Clay, and hundreds of other winning companies all have in common? They are all powered by WorkOS. If you're building a product for the enterprise, you've felt the pain of integrating single sign-on, SCIM, RBAC, audit logs, and other features required by large companies. WorkOS turns those deal blockers into drop-in APIs with a modern developer platform built specifically for B2B SaaS. Literally every startup that I'm an investor in that starts to expand upmarket ends up working with WorkOS. And that's because they are the best. Whether you are a seed-stage startup trying to land your first enterprise customer or a unicorn expanding globally, WorkOS is the fastest path to becoming enterprise ready and unblocking growth. It's essentially Stripe for enterprise features. Visit workos.com to get started or just hit up their Slack where they have actual engineers waiting to answer your questions. WorkOS allows you to build faster with delightful APIs, comprehensive docs, and a smooth developer experience. Go to workos.com to make your app enterprise ready today.
Okay, so many questions. One that is really interesting here. So obviously there's Anthropic, OpenAI. They had this coding agent that they're like, "Holy, this is a big opportunity." And then they're like, "Okay, people are using this for knowledge work. Let's build a knowledge work component." So there's Cowork, evolved out of that within the product, and then Codex, they've invested in let's make this useful for all kinds of things. Interestingly, you guys decided, okay, Cursor, we're not going to build this into Cursor. We're going to start something fresh. Was that just like obvious from the beginning? Okay, this is not going to work inside Cursor the product. We need to start fresh. How like controversial was that decision?
It was not obvious at all. I think you're completely right that that was one of those original decisions that at the time we had a lot of discussions about, and I'm very glad with where we landed. And I think to your point, it being a brand new product that you control every pixel of the experience and you have this consistent vision about where knowledge work is going and it's all contained in this new thing, I think has contributed a lot to the success. But there were a lot of discussions about, you know, Cursor for example, and some of our coding products, people use it for non-coding tasks all the time. And, you know, these coding agents are really excellent at some of these things, but you run into small paper cuts. Sometimes the product itself is kind of intimidating to nontechnical users. There's a brand association with these things.
And so, I think we evaluated that path and I think we saw what maybe some of our competitors have been doing of this is all just one surface. You add new tabs for each new form factor and it feels a little cluttered, and I think for users they can feel that this was not a single consistent vision of the way that work should work, and instead it's three different visions that all kind of share a screen and you can hop between. But it is kind of a shipping-your-org-chart style thing that I think users are reacting negatively to.
And so we decided let's just start completely from scratch. Let's see where we can get from there. There might be some really amazing opportunities to bring people from other surfaces into this more bot-native experience, but it's really important for people to just have like an amazingly simple and amazingly powerful experience.
That is a really valuable lesson for people to take away here, just that that might be the solution instead of adding into a complicated existing AI product. So interestingly, Codex went the other direction, and it's like a different path and a different product, but interestingly they're like, now we're going to make it one thing. So there, you know, there's many ways to make it work, and it also feels like the path you take there will kind of lead you, but maybe we'll look back and be like that was not maybe the best idea.
Something you mentioned that you did that is also really unique is this onboarding of early users. I heard you and your team onboarded two to 300 people manually, including me. Talk about why you thought that was necessary and what you learned from that experience and just like how long that period was of this kind of manual onboarding.
I mean, you just learned so much, and the first few onboardings were pretty painful. I'm glad you got a good one, Lenny. But there were some that were kind of rough and we learned a lot, and I think it was important for the core team to be in the room for those and to just sit on a call for 20 minutes when the computer isn't spinning up or when someone's in onboarding and they're just incredibly confused. So that immediately after you're like, that can never happen again. We need to solve this tomorrow because tomorrow I'm onboarding this person and it needs to go better. And so there was about a two-week period where we were in that mode and onboarded a couple hundred people.
And not only did we learn a lot about the product, I think we didn't really know, I think sometimes with these products there's some groupthink of ways to use them. And I think internally, because people inside of SpaceX were just constantly sharing tips and tricks for how to use Grok Bot, some patterns were starting to emerge that we thought would be useful to the world, but we weren't really sure, and we definitely didn't want to bias the world.
Is there an example of that?
So when we rolled out Grok Bot internally, there was about a week or two where the common pattern of the way people would interact with the product was you would have five to 10 bots and each bot you would give a different scope, different domain, and it was kind of shorthand for different lanes of work. And then around the end of week two, we started to see these messages internally in Slack of people promoting one of their bots who is a bit of a standout performer, and it was like their kind of primary personal assistant, promoting that to their chief of staff. And then they would actually mostly talk to their chief of staff, and the chief of staff would fan out all of these tasks to the other bots and would kind of manage the team. And there are some funny screenshots of people actually telling, you know, the bot they're promoting that they're promoted, and the bot is asking if they get a raise and, you know, is their token budget higher, all of these things.
And we kind of took note of that, and I think more of the company started to slightly shift in that direction, but it was not the majority of the company. People use this product in very different ways. And so in some of the onboarding sessions and just from the early access program in general, we really did not want to lead the witness and say, you know, create a chief of staff bot, here's exactly the way that, you know, that chief of staff should manage all of the other bots, and see if early access users would get there themselves. And we actually did see that many of them did. And so then we had a bit more of an opinionated take in the product of this feels like a pattern that's working. This feels like a pattern that we should slightly encourage, but it shouldn't be a one-way door. And there were a few other examples of internal theses that we really wanted to make sure, you know, would bear out in actual external usage without us imposing that in the product.
Is there anything else there? Any examples come to mind?
Yeah, I think another thing we really tried to pay attention to in the early onboardings was just how much users wanted to see. And I think it is something a bit shocking or just different about Grok Bot when you first start using it versus some of the other products that you mentioned, where a lot of the internal mechanics of how Grok Bot works are not shown to the user.
And the reason for that is we think as these models get smarter, the same way that your teammate, you know, you wouldn't ask for second-by-second updates of exactly all the buttons they're pressing and websites they're going to, I think it's too much to ask your bots to do that, too. And I think it's honestly just overwhelming and can create more harm than good. And so we moved completely in the other direction of you send a message, you tell your bot to do something, it just starts doing it. It
sends you progressive updates as it sees fit. And you just see that typing indicator in the little green, you know, active circle, kind of Slack-like, that it's active, it's doing work, it'll get back to you soon. But you don't see the internal mechanics, you don't see the tool calls, you don't see exactly every little click it's making on its own computer. And we really wanted to take a strong stance that users did not need to see all of those mechanics.
And so that's where we started. And we did get some feedback that's like, I would love to see my bot's to-do list. I would love to see roughly how it's prioritizing tasks and what it's doing. And that's great feedback. But it was useful to hear that nobody wanted the long stream of just text streaming out and chain of thought sequences. So that also gave us more confirmation that that was the right direction.
The fact that you did two to 300 onboarding calls with a small team. I know the team grew over time, but just that is a huge time commitment, and then you could argue a distraction from the building. Clearly not a distraction, clearly a core part of the success. Do you feel like that's the volume people need to do to figure out what actually needs to happen?
Well, one thing to emphasize is the early access group is not necessarily just people that are highly influential taste makers. You know, you're in this category, Lenny, and we certainly wanted to get a lot of your feedback just from being very close to many other products on the market and just being a power user of these things, but we also wanted to get early access to kind of more unconventional profiles that we as a company had never really interacted with.
So one example is there's a coffee shop owner that was a friend of a friend through the company who had heard about Grok Bot, and one day somebody on the core team had kind of shown them a demo of the TestFlight and got very excited. And this coffee shop owner ended up being not only an amazing power user of Grok Bot but also a rich source of feedback for us. We have a very lively thread with many, many bugs that get identified or feature requests, and it's a completely different use case of running a small business. And so for example, if the Shopify integration was a bit flaky, or if it wasn't writing copy for products in a particular way, we would get really rich feedback on that, which is pretty different from the type of feedback we'd get from dogfooding this internally.
And so I think it was an important exercise for us to check our blind spots and say, A, this is going to be a very general product that is not just a thing that developers use. In fact, it is likely that this is most powerful for non-developers. We need to understand that group much better. And then B is we absolutely live in this kind of Silicon Valley AI bubble, which I think is a useful place to be to kind of push the frontier and push the future of how these products are evolving. But we need to actively get out of that, because I think a product like this has the chance of really being the way that the mainstream user and the mainstream kind of business customer can interact with AI in a way that's useful.
Let's go back to the timelines real quick just to kind of understand that. So it was a month from first line of code to internal beta, and then what happened after that?
It was about three weeks from internal beta to public launch, and then I think we're about three weeks out from public launch as of recording this.
Wow. Okay. So, a month of building the first thing, three weeks only of iterating, and then it's been only three weeks since launch. It feels like it changed the world from my vantage point. So, wow. Okay. What most changed in those, I don't know, in those three weeks of internal beta, let's say?
We unshipped a lot. I wish I could have shown you what things looked like maybe two weeks out from launch, where I think we'd realized that the core team, we had a lot of experimental features that we wanted to get internal feedback on, which was useful. We also were kind of putting pseudo-developery visibility tools into Grok Bot instead of having a separate observability pane for those things. So for example, we actually did expose sometimes a lot of the internal thinking of the models and the specific memories it would store and all of these things, which was useful to debug issues, and if you're building the product, you didn't want to go somewhere else to maybe pull that context. But we had to really aggressively trim what we think the user absolutely needs to see in the surface versus what they don't. I think there's even more room there to run, which is something the team's focused on right now: how can we just ruthlessly simplify this product and abstract away anything the user doesn't need to actively be thinking about. So that was one big push, was unship a lot of jank.
And then I think the second big push those few weeks was making it just work. And I think a lot of what people want from AI is not this thing with lots of drop-down menus and bells and whistles, but just a thing that you describe a task, a task that's meaningful to you, and it goes and it does it, and it comes back with complete work, or it comes back with something for you to react to and then steer it in its kind of next cycle. And so in order to actually deliver on that promise, it's actually not a lot of feature product-roadmap-style stuff. It's like hill climbing five really important problems in the back end that many users don't directly experience, but you totally feel when your bot is going off and doing something and can't click the right button, or your bot is off and doing something and can't log into a website, and it just completely stalls your ability to make progress on that task. And so those few weeks we collected a really rich set of what are tasks that people actually are giving bot, how can we quantify these things, and how can we see week over week that across those categories of tasks we're hill climbing on very important dimensions to making that just work behind the scenes.
Is there an example of one of those hills you were climbing that was kind of a technical breakthrough or technical challenge you overcame that really helped it just work?
Yeah, one example was from rolling out Grok Bot. One group inside of the company that was actually incredibly bot-pilled, so to speak, was our go-to-market team, was sales. And there are a bunch of tools that sales uses that do not have well-supported MCPs or APIs. And I think that's a lot of what made bot so before-and-after powerful for this group: these were things that they just could not give another AI tool reliably. We would get stuck at some part in the process. Then bot kind of felt like they had an assistant, or kind of felt like they onboarded someone to their personal team. They gave it a laptop and it could just run.
And so there were a bunch of small things, probably a list of 10 or 20 of them, of places where for whatever reason the mouse would just not have fine enough control to click on exactly that part of the Salesforce dashboard or something like that, that we would have to take back to the core team working on really the infrastructure to say, here's a very concrete case of where the agent not having this visibility into the browser or this visibility into the pixels on the screen is making it impossible for this task to be done. And that was just a lot more tangible than seeing a number on a dashboard slowly creep up. It was kind of like new chunks of work getting unlocked, and you would immediately feel the feedback, where you would ship an improvement that was kind of behind the scenes, kind of infrastructury, and then the next day you would just get this outpour of love and appreciation from the sales team that now this workflow that was failing the last seven days finally works. And it's just a constant exercise of finding those next tasks to unlock and then solving them.
So, computer use improvements basically seems like a big unlock. I had Adam Ward on the podcast, who's head of recruiting, head of hiring, basically head of talent. I heard his team was one of the top users of Grok Bot.
Yes, the recruiting team gave us a lot of great feedback. Anytime the product, especially in the early days, if there was a little bug, we'd get a ping from some folks on the recruiting team. Yeah, I think the main use cases for recruiting that were particularly interesting was, first, it was incredibly valuable as a sourcing tool. And I think one thing Adam talked about on the podcast with you, and it's a big part of our hiring philosophy internally, is being on the market, looking for a job, is not a precondition for us trying to hire you.
And in a lot of ways, the best way to hire is really just look at the biggest problems at the company that need someone to own it or take it to the next level. Find out, of the total universe of people in the world, who would be best, and then ruthlessly go after them and try to convince them to join. And this is a lot of the philosophy from the very beginning of the company. And so if that's your mindset, really the best recruiting work streams or workflows to automate are not, you know, here are a bunch of resumes, read through them, help sort them. The most useful thing is, here's an entire universe of potential people. Help match that to this very concrete business problem or this very concrete role that we're recruiting for, and help me get in touch with them. Help me get coffee with them. Let's just throw everything at it.
And so there have been some cases of really kind of unexpected ways of finding top talent that is beyond even just looking on LinkedIn and trying to find interesting people. But who are the co-authors of this paper, and the PDF doesn't exist on Google Scholar, it just exists on this conference website. I want you every morning to go to the conference website, download the PDFs if there are any new ones. You should find every new name that we've not yet tracked. You should add that name to a spreadsheet. You should do research. You should look at everybody at SpaceX, see if there's anyone directly connected. If so, you should send them a Slack message asking for an introduction. It's those types of always-on sourcing use cases that I think in the past were incredibly manual, and now it's the type of thing AI is superhuman at. And our team can focus on closing great candidates and getting conversations with great candidates and not pulling these huge lists.
Wow, that is such a cool example. First of all, someone's about to take the transcript of what you just said, put it into a bot, and create their version of this, which is great. On the other hand, I think you guys could sell a template of this bot for a billion dollars. If we could basically use Adam's team strategy for finding the best people and just turn it into a bot, holy moly, democratizing hiring.
I want to come back to a few things. Okay. So, one is you made this point about unshipping. Such an important point, I think, that people can overlook, because AI is not good at telling you what to take out. It's very good at, okay, here's more ideas, here's more stuff. And something that's come up a number of times on the podcast is that's a big opportunity. That's a big space for humans to continue to be very important and valuable, is knowing what not to ship and what to cut and what not to do. And so it's so interesting to hear that that's been a big part of the internal evolution from prototype to launch, is deciding, okay, we need to cut a bunch of stuff. Anything more there?
Yeah. So, one thing we talk about internally is for anything that we're working on for Grok Bot, what is the launch post? What is the thing that we would actually tell users?
Tweet, I imagine. Launch tweet.
And if it's not compelling, maybe we shouldn't be working on it, if it's not something that users will directly feel in the product. And to take that even one step further, I think there is this old-school software tendency to say things like "Grok Bot now has," and when you think of completing that sentence, it would be a new button to press, or a new drop-down, or a new integration that you can press plus and add. And instead, to reframe it as "Grok Bot can now," which is, I think, a much more human way of describing these capabilities. And I think it's forced us to think more in the frame of what are tools and what are capabilities that we can give Grok Bot, not what are new things we can add to the product. Adding things to the product is not the goal. That's not the thing that's going to push this product forward and make it more useful to more people. Making your bots reliably do really impactful work for you behind the scenes in a way that just works, and giving them the capabilities to do that, that's what users actually care about.
And so I think in the context of unshipping, there have been a lot of "Grok Bot now has" things that we've realized are actually just capabilities that don't need pixels. Let's kill as many pixels as we can. Those can just be things that your bot manipulates behind the scenes for you, and you don't need to directly control. And I think one example of this is the way that, with many of our competitors, you set up automations or routines: you go into a sidebar, you press plus, you select what the trigger event is, you then select what action it should take after that. You might describe it in natural language, and it's just really clunky, and it means that people don't set up many automations for many things. We certainly have seen this in the coding realm. And so I think what Grok Bot did in response to that was, actually, you should just define automations in natural language. You should tell your bot, remind me of that at 8 a.m. every day, please. And then it should just do it. And you should never ever have to see that interface of creating an automation. And so that's kind of the decision that we've made. And now that's how 99% of automations on the platform get built. And I think there are a bunch of other places where we can do things like that.
So, I tweeted about how much I love Grok Bot when it launched, and a lot of people replied, "Wait, can't you just do all this with Codex and co-work?" And technically, as far as I know, everything you can do with Grok Bot you can do with the other foundational models, the coding assistants. So, let me just ask you this big question. What is it that you think you did that is so different, that made Grok Bot so successful?
I think it was two early decisions that at the time definitely did not feel obvious, but in hindsight I think are critical to what makes Grok Bot work for people. And the first is you should never have to think about local and cloud and where are these workflows running. Does my computer have to be awake? If I kick it off from my phone, does it need to be tethered to my computer back at home? There's so much jank happening right now when people are trying to conceptualize where this runtime lives. And we made a really early decision that this should just all be in the cloud.
And if it's in the cloud and it's this persistent colleague that has its own computer, it can do its own work. It has the same state everywhere you interact with it. It opens up a lot of really amazing opportunities to text your bot, kick it off from your phone. In the future, you should be able to call your bot from anywhere and it should be able to do real work. This is its own entity and it lives separately from your device. And I think that was a very important decision that current products, I think, haven't made, and I think it has a bunch of paper cuts as a result of it that
users feel every day. I think the second decision was kind of to take that one step further of not only should this be an agent loop that kind of runs in the cloud and you can interact with in various ways. It's actually really important that these bots have their own computer, and part of it is what I described earlier of there are a bunch of tasks that don't have well-supported MCPs and APIs. We as humans don't do our jobs via MCPs and APIs. Like, we use a computer and we click on pixels and we kind of type things in input boxes. And it's very important that your bot has those baseline capabilities as well.
But even to go one step further, I think we're in a really weird moment right now that I think we're going to look back on and be like, I'm surprised that this is the way that a lot of people worked with AI, where you're onboarding these super intelligent new colleagues, these AI bots, and you're asking them to share the same computer that you have. It's crazy. Like, if you were onboarding someone to your team and you said, "It's your first day. I'm going to onboard you. You don't have your own laptop. You're going to sit next to me. We're going to share this laptop forever and constantly trip over each other. You're going to have access to my credentials. I'm going to have access to your credentials." Like, there's a good reason why that's not the way people operate. And I think bots and these kind of AI colleagues of the future will also need a way to onboard them that's somewhat similar. And we've tried to push the product in that direction.
That is so funny. [laughter] Why do you think the other companies didn't do this? My guess is they were building off of their existing coding assistant platform and approach, and this was like a pretty big shift.
I think a lot of it comes down to starting from scratch and how freeing that is, and we felt that ourselves, where a lot of the primitives in Grok Bot we had attempted or we had built in other ways. You know, cloud infrastructure we built for coding agents. The way that you could maybe name agents and talk to them as discrete entities. It's a pattern that we're also seeing for developers, kind of bringing specific named agents into Slack. But instead of kind of trying to retrofit those concepts into some new surface or into an existing surface, which I think would have been the strong default for many companies, we decided to start from scratch. We decided to just try to get these things really right for general knowledge work, which is a new audience, and then second, for the point in time that we're at now, where the models are very capable and if you give them the right tools and the right infrastructure, they can do a lot.
But a lot of these ideas, these aren't strokes of genius on our part, and I think for good reason. I think these are primitives that had already been getting attention and product-market fit by other products, you know, the OpenClaws of the world, and I think we took a lot of inspiration from that and tried to productize it into a bit of a tighter surface, something that required a little bit less setup and I think was more accessible to more people. And so I think our competitors and other tools that have been trying to solve these types of problems, I think we're all seeing the same opportunity. I think we're seeing a lot of the feedback from the market, but I think it's just been hard to act on if you're stuck in the existing paradigm, and if you have a lot of kind of sunk cost in that existing paradigm, it's very painful to create a new thing from scratch. And I think a lot of that is what allowed the product to just work and click for so many people.
So what I'm hearing here is the keys to success of what made this break out. A cloud-based computer for every bot instead of locally. A name, kind of like a specific bot. And by the way, we're all moving from agents to bots now. Nice job. Feels like you guys have pushed it over now. Okay, we're all bots now. So a bot per kind of task or use case, very unique versus like a thread conversation or like a one-off job. And then it feels like "it just works" was a core part of this, and you talked about how long it took to get to that place of like, okay, now it actually works really well.
You mentioned OpenClaw. Obviously this is inspired by that, which, to me, when I first used OpenClaw, I'm like, holy— this is the future. How could we not have this? And then Hermes came out and everyone's been trying to build the OpenClaw that works very easily for everybody. Can you say more about just how OpenClaw and that story informed the way you guys thought about it?
Yeah, so I think OpenClaw got two major things right that, when we were seeing the way the market was reacting to OpenClaw and ourselves using the product, we found quite exciting. I think the first thing was the models are really smart and they're going to continue to get smarter, but even at current capability levels, if you can just give your bot access to the tools that you use to do your job, it can get a lot of the way there. In a lot of the places where people think AI is dumb or maybe not as impactful as it's been promised, a lot of that I think is downstream of it just being harnessed in the wrong way. And so if you kind of give access to a much larger set of things, if it has access to its own computer, how far can you go? And I think OpenClaw, you know, really kind of forced that question for many people.
And then I think the second way OpenClaw changed the mental model of AI was really viewing these things much more as colleagues and teammates and people, and personifying it a bit more, and it being this helper entity that has access to your life and can kind of extend you even further.
And so we took a lot of that, and I think what Grok Bot maybe extended was, A, it needs to be really easy to set up, and the hacky, you know, you have a VPN at home and a Mac mini setup clearly was not going to scale to millions of users, clearly is not going to be the way, importantly, that businesses take advantage of this technology. And so we really wanted to build an amazing product with that in mind.
And then second is I think there are a lot of places, a lot of rough edges to sand down and just make a delightful product experience and make these things just work, and try to remove some of the abstractions that power users of AI are very familiar with, things like skills, for example. How can we make a Grok Bot user not even have to know what a skill is? They should never have to type a slash command. These things should be created in the background as a useful primitive that the bots have access to, but something that users, you know, it's not incumbent on them to always be on the cutting edge of AI. And so that's really where we tried to innovate, and I think there's still more room to go there.
As you say that, I have my Mac mini with my formerly alive OpenClaw on there. And that was an era, and it's so awesome the work that it has inspired. I know it continues. I know there's still a lot of value to OpenClaw, but when I saw Claire Vo, who's been like the biggest proponent of OpenClaw, and it's become a core part of the way she lives and works with her kids and does all her work, she just switched all of her OpenClaws. She's shut them all down and switched to Grok Bot. It sounds funny, but that's actually a huge milestone of just how much things have shifted.
What's kind of the vision for Grok Bot? Where does this go? What does this look like in the future? What's the ideal platonic version of Grok Bot?
The ultimate vision of Grok Bot is incredibly simple, which is you should have a team of AI bots that help you with your job and help you with your life. And it should really feel like a team. It should really feel like teammates that are autonomous, are helping you. You can steer them in various ways. You don't have to micromanage them. They have access to the tools necessary to do great, ambitious work.
And one thing we really use as a north star on the product side in building this is, as we kind of get closer to this teammate future, how can we, in every product decision we make, think about this less from the perspective of a SaaS product and more from the perspective of we're trying to build useful AI teammates. And so there have been a bunch of examples where we kind of have to push ourselves to be more colleague-pilled, in a way. We sometimes use that term where we're having a product debate about something. There are good arguments on one side. There are good arguments on another side. Both paths feel sensible. In product land, this maybe doesn't feel like there's a clear-cut answer. And then you zoom out a little bit and you remove yourself from the, you know, tech companiness of it all, and you start thinking, how would a human do this? Like, what would you want from your teammate in this exact situation? And oftentimes the answer is really clarifying and pretty unanimous. There's oftentimes not a lot of disagreement among the room of how a human teammate you would prefer to work with in a certain way. And then once that answer is there, well, then we just need to build it. And there are product implications, there are model implications, there's a lot of things that need to go right to actually deliver on that experience. But in some ways it's not rocket science. It doesn't require you being a genius. You just need to ask the question of what would you want from a human teammate, and can we push AI to behave in a similar way?
And so to give some examples of that, I mean, we've been thinking about what the right voice experience with these bots should be. And I think in the context of a human, for example, we have a really good analog of, a lot of times I'm Slacking back and forth with a teammate, we're sharing context, and a lot of times it's just much simpler to get on a five-minute huddle with them and just press huddle, talk back and forth. I share my screen. I show exactly what's on my mind. They share their screen. We hop off and then we continue async from there. And that's not really an experience that any AI product has gotten right right now. And it is deeply integral to the way that I think humans collaborate. And so we want to build something like that. And there are a bunch of other examples of these very clear patterns that just work that I think you should also feel when working with AI.
I love this term, colleague-pilled. And it's come up so many times over the course of this chat already, how that is kind of a throughline to making these decisions. For example, the computer example you gave is so good. Obviously people would have their own computer. The naming piece is also a very important part of that. A big question on my mind in the space, and I'm so curious to get your take, is the separation between work and personal. Do you think people will have two different assistants, a work and a personal, or do you think it'll be one?
When people think about a work product versus a consumer product, I think there's just a lot of baggage that comes from the last decade or two of horrible B2B software that leads to people seeing a product that is very simple in some ways—ChatGPT was like this, I think Grok Bot has many of these properties—and assuming that it's not a work product or assuming that it's not a power tool. And when you think of a power tool in this kind of last generation, I in my head picture something a bit like Photoshop, for example, where there are all of these different dials to turn very precisely. You know, the user of the tool is this kind of ultimate cockpit flyer who knows exactly what all the knobs do and can use them perfectly.
And I think power tools of the future will actually be very different from that, where it is mostly just intent being expressed and good steering on the part of the human, and these AI tools abstract away all of the knobs. You should never see them unless you need to directly manipulate it, which might happen, and there should be a great affordance for that, but ultimately it really is just working with a teammate, and so the interface for that is quite conversational.
And so in a lot of ways, Grok Bot, when you look at it, like when I walk by someone's desk and I see Grok Bot up on their computer, for a split second I'm like, oh, are they on like a messaging app? And it's like, no, this is actually the primary tool that they're using to do much of their work.
And so I think to the question of are you going to have a different set of bots for your personal life and a different set of bots for your work life, I do think there will be a separation for many people. They want a separation between personal and work life, and I think that's great. I think that's important, and I think there are a lot of common-sense reasons why those things should be separate, even from the perspective of an enterprise. But I think our goal and the thing we're trying to build towards is Grok Bot should be the way that you do a large portion of the things you do day-to-day in your work. You should be able to delegate a lot of that to Grok Bot and focus on the higher-leverage things. And then it should similarly be the way that you delegate a lot of the low-leverage parts of your personal life. And those two things actually are not different problem sets. In a lot of ways, the product form factor and the ways of solving those problems is pretty much the same. And so my instinct is that I think one product will be the best form factor for both of those things, and that's really what we want to build.
Bam. That's a big TAM right there. I love to hear it. Makes so much sense. Obviously the question is how do you avoid cross-contamination, you know, personal stuff somehow infiltrating, exfiltrating stuff from work. But okay, so what I'm hearing is that's the direction. The question is just how to do that and make people feel super safe, have kind of like the SOC 2 stuff in place, and also just feel really fun.
This episode is brought to you by Mercury. Radically different banking, now with Spend. I've been a Mercury customer for so many years now. I switched all my business banking to Mercury, and honestly I could not be happier. It's what online banking feels like when it's built by product people, not by bankers. And now with Spend, you can give your team individual cards, set spending limits per person or per team, and have expense receipts automatically pulled in from Gmail or over text. You can even give your AI agents their own cards with their own limits and policies. Most founders start out the same way: one card used by everybody at the company. It works until it stops working. Someone goes over, a receipt disappears, you spend two days trying to figure out who spent what and why. Spend is expense management built directly into Mercury. All your team's cards, budgets, and reimbursements all live in the same place as your business banking. No chasing, no manual reviews, no end-of-month scramble. The result is a team that can move fast and a founder who is no longer the bottleneck. Learn more and get signed up at mercury.com. Mercury is a fintech company, not an FDIC-insured bank. Banking services provided through Choice Financial Group and Column N.A., Members FDIC. The IO Card is issued by Patriot Bank, N.A., Member FDIC, pursuant to a license from Mastercard International Incorporated.
Let me ask a couple technical questions. On the computer side, what's the simplest way to think about what you get as a part of your Grok Bot account? Is it like a VM that is running in the cloud with multiple logins? Is it like a separate VM instance per bot? How do we understand that, as much as you could share?
Yeah. I think to go back to the teammate frame of the product, to kind of extend the analogy even further, if we were on a team together, you and me, you know, I think the number of times that you would have to manually take over my computer and start clicking
on things and like, you know, you're doing this wrong, you should go here instead and type in manually. Hopefully is pretty close to zero. Hopefully that is not something you have to really do with a colleague or a teammate.
And so similarly, I think right now we're in a place where computer use is good. It's getting much better. And in very short order, I think the computer concept will be completely abstracted away from the user. You should never be clicking into a remote virtual machine. You should never have to take control, you know, if there's like a wasteful path and you have to kind of steer it into the correct path. So in the medium term, I think the computer concept will be an important concept for users to have, but will not actually be something that they're interacting with.
So I think the right way of thinking about Grok Bot is it's a team of bots. It's a team of agents that do work for you. And in terms of what they have access to, they have access to a very long memory set of your past interactions with them. And so I think there's a current paradigm of you create a new chat for each discrete unit of work. I think there are a lot of problems with that. I find myself copying and pasting between chats all the time. I think it's just not a great way of grouping categories of work. Instead, the same way on a team, you have a good way of grouping categories of work, of kind of roles, you should have roles of kind of different swim lanes of work that you do and it should learn from you and it should get smarter over time.
So I think that's one very critical thing is these are long-lived agents. These are not individual one-off sessions and these agents get smarter over time. And then the second thing is those agents have access to all of the tools that you would expect a human colleague to have, which is the APIs, the MCPs. That's great, but then access to its own computer which it can freely manipulate the way that you would.
At this meetup that I went to, Shub, who's on the, I think, go-to-market team, demoed something that blew everyone's mind. Because you have a computer within each agent, you can run a lot of different things on the computer. He was running Grok Bot within Grok Bot, like the bot can run its own Grok Bots, and I know he was using it for testing and watching regressions and things like that, but that's just like a mind-expanding idea and I'm curious how many levels you can go before the universe collapses on itself.
I do that one too. That one's actually a very useful thing to do is you download Grok Bot for one of your bots. Mine is like a QA tester bot. And that way if there's ever bug report or if we're kind of testing out a new build, for example, of the desktop app, I can just say, "Hey, here are 10 workflows that we need to make sure are getting better release after release. I want you to test it. I want you to write it to this Notion document that has like an extensive list of all of the past tests that we've done of past client versions and compare them." And so I think once you start breaking out of "this is AI chat with a set of connections," which is I think where most people are conceptually now, instead to "this is a colleague with a computer and anything I would ask a colleague to do on a computer I can ask Grok Bot to do," it just raises the ceiling, I think, of what you would think to give to AI.
Are there any other mind-expanding use cases or ways to use Grok Bot that you've seen or you use?
One pattern that I've seen from many users that is simple but I think there's a lot of depth if you keep investing in making it better, and this is kind of where I can get kind of nerdy about optimizing my setup, is Grok Bot as an infovore in some ways of just consuming huge quantities of information, removing that from your own cognitive load, giving you peace, and then coming to you with the stuff that's important. And I think the V1 implementation of that, which many people do, is Grok Bot sits on top of Slack and it sits on top of email and I tell it high level, here's my role at the... here's kind of what I care about. I want you to notify me in these cases. In these cases, you don't need to ping me directly, but you should include this in your daily roundup that I read every day. That's like the V1 implementation.
I'm not sure what the V10 implementation is, but like maybe I'm at V3 or four, which is you can give these bots a complete fire hose of information. So I have mine hooked up to like every mention of Grok Bot ever on X and it's interacting with our internal context. It's interacting with the QA tester to like see if it can repro any bugs or feedback that we're getting. I've hooked it up to my own kind of messaging services to like quickly act on feedback and reach out to people. And I think there's this just like always-on kind of chief of staff entity that can preserve your focus on the things that actually matter, but is always like kind of surveilling to see if there's anything that should get your attention.
And we've seen some funny cases of people actually giving their Grok Bots, which I have not done this yet, but maybe soon, giving their Grok Bots access, the ability to page them. And so if something like super urgent happens and they're at a coffee or whatever, they get paged by Grok Bot, which is the type of thing that you only want to do if it's urgent and you really want to trust that Grok Bot, you know, does not have false positives. So far, those people have reported that it's been very helpful and successful. But I think we're going to see more of that type of stuff where the agent or the bot should actually be more proactive to you than you reaching out to it. And I think that will be the next shift in AI.
This touches on... there's a number of things that I've been very impressed with watching your team operate. One is speed, which I want to talk about, but the other is how you... like there's awareness that this is a moment in time to capture a lot of market share and really take as much of the market as you can before somebody comes around and like, okay, now we got something awesome, especially one of the foundation labs. So watching just how many free accounts you guys are giving out. Also the focus on use cases, so smart, because it's such a novel thing and you open it up and it's like, what do I do with this? And there's such a focus on, okay, here's a bunch of things people do with it, and then there's all this talk on Twitter and just like all the ways people are using a template makes so much sense. Just these two kind of like focuses, from what I can tell: get as many people on it as possible as fast as possible until somebody's like, okay, you know, because someone's going to come around, be like, all right, here's the next thing. Super smart, and also the use case focus. I know you were a go-to-market person at Cursor before this. Anything you want to share there about just the approach to the go-to-market right now for getting this out there?
I think the pattern we saw for coding will be somewhat similar to what we see for general knowledge work and I think we've learned a lot from that on the go-to-market side and more generally just building practical AI that people use. And I think we as a company have culturally really cared about not building demoware, like building actually useful stuff in the world and kind of obsessing over that. And there are so many shiny objects and like fun prototypes to build. But ultimately that's a very different problem than getting this in the hands of millions of people and having it transform companies. So that's really, I think, culturally where we've always been focused.
And so I think on the go-to-market side what we saw for coding was a very simple pattern, which was there was an early adopter crowd. The early adopter crowd would use these coding tools and really push them to the limits and they would mostly push them to the limits on individual projects, they would on nights and weekends. I'm thinking like 2023, you know, kind of earlier. People would kind of go home from work... at work, they were using a basic IDE. This is pre-AI. And then at home, they'd work on a side project and they'd be using Cursor or they'd be using, you know, the latest and greatest AI coding tool. And that would give them an extreme amount of acceleration. It would feel like they were experiencing the future. And then they would come back to work and they would demand it. They would say, "I cannot picture working any other way than this. I feel like I'm completely walking through molasses right now. This needs to change."
And I think for knowledge work, we're going to see a similar pattern of people really feeling the aha moment sometimes in a personal capacity. And I think we're certainly seeing a lot of this like on X right now. You see all these examples of Grok Bot controlling their home robot computer or home robot vacuum cleaner, or Grok Bot, you know, helping them save money on their Tesla charger negotiation. Like all of these fun use cases, but I think the next step is going to be: this is not a consumer product. We think this is going to transform businesses. We think this is going to transform teams and it will be bots coming into teams and contributing really economically valuable work, especially as they get much smarter.
And so on the go-to-market side, we're certainly making a big push on prioritizing businesses and thinking about not just the single-player use case of working with a single bot, but how does a bot work inside of a broader team? How does a bot work inside of real company systems that are complicated and there's a lot of context and a lot of history to understand? What does memory look like in kind of a broader organization versus kind of a single individual you're catering to? And I think there are a lot of unanswered questions there. But I do think Grok Bot is the right primitive to create this switch to agents for the rest of the company outside of coding. And that's a place where we're quite focused right now.
And along those lines, it's very clear you all understand the power of distribution and how you need to find both an amazing product and get distribution right, because, you know, Grok Bot's amazing, but the combination of how smart you guys have been with getting it out there in all these different ways is really impressive and I think that shows you what it takes these days to build something that's really successful. I want to ask about the brand of the different brands around this product and the company, just so people can try to understand, because I know you're going through a transition, acquisition, SpaceX, all these things. So there's Grok Bot, there's Cursor... is that... So talk about like the products and the way to think about these different brands today, and I know it'll probably continue to evolve, just so we could communicate about it correctly.
Definitely. Yeah. I think there are three big pillars right now of SpaceXAI. So the first pillar is the coding product and set of products, and right now that's Cursor and Grok Build, and I think we're big believers that having a professional work surface for developers and for the engineering part of the organization is going to be really critical. And right now people use Grok Bot sometimes to kick off cloud agents or to kind of merge PRs or to do QA, a bunch of engineering-adjacent tasks, but ultimately when you're shipping production software, we're big believers that that is going to require, you know, a product where every pixel is optimized for that end user. So, we're making big investments there.
The second category is general knowledge work and we think Bot is a really exciting step in that direction. There's a lot more work to do of making it more useful, extending it to new surfaces, it really feeling like an AI teammate that you can delegate work to, especially inside of companies and businesses. So, that's kind of the second pillar. And then third is the general model effort. We want to train the smartest models in the world that are really capable.
And I think one thing that somewhat distinguishes SpaceXAI from other AI labs is I think our goal is less to build, you know, chase superintelligence or some kind of vague aspirational ideal, and the goal is actually very practical, which is to build useful AI, and we do that on the product side, we do that on the model side. And I think part of that is also just cultural of like the group of people contributing to these models are engineers and people who kind of came into the model training effort from like a very applied mindset, and I think that's what gets this company going and I think is actually a slightly different direction from some of the other competitors out there.
Super interesting. Okay, there's a couple directions I want to go. One is you have this tweet that is, I think you pinned it or maybe it's your last tweet. It's up there in your timeline if people check you out. So the tweet is: "An AI that does 100% of the job feels categorically different from one that gets you 90% there. I've significantly updated what I think AI is capable of." Say more about that.
I think for me what made me so excited to work on Grok Bot and contribute to it is it was the first time for non-coding tasks that I felt like I could truly delegate work to AI and not have to think about it and I would come back and it's done. And I think engineers have been feeling this for quite some time. For maybe a year, a year and a half things have been like that. I mean the job of a developer has completely transformed. It is unrecognizable from what it was two years ago and many, many words have been said on that topic. But I think it's underrated how different that experience is from what most people are feeling about AI right now and the way that AI has changed their lives. And it looks quite similar to the way that people would use AI like two years ago, where you create a new thread for a task, you type it into, you know, an input box, you hit enter, you watch all of these steps happen, you get an output, it's not quite right, you keep working on it.
And Grok Bot, I think, short-circuits a lot of that. When you first kind of lay eyes on the first screen, you're like, whoa, this is clearly different. Let's see if it actually works, but this is different. And then you give it something and it kind of works, you know, to a surprising extent, and I think we're going to do a lot to make it work much better.
And so I think what I was expressing in that was when you have a teammate that you only 90% trust and you give something to, which luckily I do not have the experience of here because I work with great people, but if you delegate something to someone and you're like, I know I'm going to have to be thinking about this while you're doing it and I know it like probably is not going to be quite there and I'm going to have to intervene and kind of steer it slightly. That's not 90% task completion. You're still doing the thing and it feels that way and it's weighing on you in the same way, versus like truly throwing a no-look pass to a colleague and being like, you got this. Here's the context. Go off and run. I'm excited to see what you do. Like that's a different category. And I think that's the type of thing that people feel with Grok Bot every day, are these no-look passes, and you just trust that it can get it done and then it does and it's just a very magical experience.
Yeah, I have had that experience consistently. Okay. So, another element of how you all operate that has really impressed me and I've not seen this before is how fast you all move. So, I got added to this like Slack as I was giving feedback with some folks and it's just like, "Okay, how about, okay, tomorrow we're gonna give you some free codes to give out. You could do it tomorrow. We'll do this tomorrow." Or, "We're going to launch a marketplace with templates. We're going to launch this in two days." It's just like, what? I don't have time for this. How do you guys, with all the things going on, all these things
constantly shipping and also staying consistent and high quality and feeling clear that it's towards a specific vision. So, there's kind of like two parts of this question: just what's the secret to how fast you all have been moving, and how do you stay aligned moving that fast towards a vision that you all believe in and where you want it to go versus just, you know, band-aiding it along the way?
Yeah, one thing I've been really happy has never changed is that startup feeling inside of the company. And for context, when I joined Cursor originally we were about 15 people. We scaled to over a thousand, and then now we're a part of SpaceXAI, which is kind of an even bigger organization. And it's something that is just so fun to be a part of when you're around this group of incredibly talented people. Everyone's moving 100 miles an hour. You deeply trust everybody to execute on their part of the equation. And there's a clear vision that everyone is fired up about and knows that they need to execute on.
And, you know, as companies grow, and we've had the fortune of hiring really great people from other companies that have gone through hypergrowth, things slow down, and you kind of keep telling yourself, "We're still a startup. We still move quickly," but you really don't. And everyone knows that you don't. And it's just, you know, it's easier to say than to actually be. And, you know, fingers crossed this continues to be true. I think it's really critical for our success if this continues to be true. But even as we scaled, it has always felt like that startup that I first joined.
And I think if you define a startup by number of people or by the funding round, none of those things really make any sense. The core thing that defines a startup is exactly what you're describing, which is this kind of scramble energy of things are kind of chaotic and kind of disorganized. And for a lot of people, that's not a pleasant working environment to be in, but it has these amazing properties of you can make extreme impact in a particular direction in a short amount of time. And you really do get out of a system what you put in. And so I think as a culture, as an organization, and the way we construct ourselves, it's really been to enable that property in a way that I think some of our competitors and other AI labs have gotten much bigger and you can feel it. And I think with us, even as we scale, there is that startupy impulse that is quite important to move quickly on these things.
Let me pull on this thread and let me ask you this big question that I've been looking forward to asking you. If you were to look at Cursor from the outside, it shouldn't have worked. It shouldn't have lasted because, one, it's in the most competitive market in the world, competing against the fastest-growing companies in history, OpenAI and Anthropic. So that's one: the competition is unlike anything anyone's ever experienced. Two, it sits on top of those platforms to power it.
And what I've seen as an outsider is what has allowed Cursor to win and have this massive exit and continue to succeed is how quickly you all adjust to the reality of the market. It started as autocomplete and then things moved on to just talking to agents and then into the cloud and now Grok Bot. To me, that feels like a core part of the success: quickly adjusting to reality and also building the best-in-class experience for a thing that also exists other places. Grok Bot's a great example. You could do this other places, but it's the best-in-class experience. Cursor the IDE, the best way to code. So that's my question. Maybe I answered it, but what do you think has been core to Cursor's ability to not just survive in this crazy competitive market, but do so incredibly well consistently for so long?
I mean, a lot of this, and it's a fuzzy answer, a lot of this, I think, is downstream from culture and the culture that you set and the people that you bring in and the way that they approach these problems. And I think for us, exactly as you said, we have never been complacent. We've never felt like we've won, and it's always been about the next thing. And I think there's been a really deep belief across the company that AI is moving incredibly quickly. Our goal is to translate those capabilities into amazing products for customers. But those products are going to change and they need to meet the moment as the capabilities get stronger. And what met the moment two years ago is completely different than what's meeting the moment today.
And if we as a company can't completely reinvent ourselves every six months, which recently it's felt even shorter than that, of kind of complete, very significant reinventions of our priorities, the core product, what users feel, we're going to lose. And I think it's that spirit of always pushing to be on the frontier, never thinking it's over or that we've won or that we've gotten it right, and just constantly updating our beliefs that has gotten us to where we are now.
And to your point on the competitiveness of this space, one thing that I think is important to point out is AI coding has always been competitive, from when Cursor first came to be. And at the time the competitors were Microsoft and others and a handful of maybe 10 or 20 companies. And I think it's notable that none of those competitors are at the forefront of AI coding right now, in large part not because of any incorrect decisions that they made or any lack of resources on their part, but this cultural inability to move quickly and to change to meet the moment as the moment's changing. And so I think that's exactly what has led us to invest in things like Grok Bot, for example.
Are there any core values, just specific ways you phrase this to kind of remind everyone of "this is how we work"?
Yeah, two values that I find myself coming back to quite a bit. The first one is this idea of deleting the product. And I think it exactly ties back to what you're saying right now, where when you look at every past iteration of Cursor, for example, but even I think when you look at every past iteration of Grok Bot, I think we will feel the same thing, it's things going away, not new things getting added. These scaffolding, product-overhang-style things that get built in because the models have not yet gotten smart enough to just do it themselves, those things will get moved away over time. And we need to feel comfortable making hard decisions that might upset a small set of users or a small set of us internally to do the bigger thing of making the product simple, making the product powerful, and adapting to where the future is going. So I think that's been very core.
And then the second thing, which ties back to the pace of execution, is "just do the thing," which I find myself repeating a lot, even as we've grown as a company. It's on you. We're all in this boat together, we want to win, and if you see something that you think needs to happen, this is not an ask-for-permission culture. You go out and you fix the thing and you pull in the resources that you need to make it happen. And I think that has made many people very successful here before, and I think it's something we really share with SpaceXAI as well.
Agency, as you may have heard. So interesting. One of the big questions that comes up, and this might be my final question, is around moats. A lot of people look at Cursor as a really interesting example of: they're in a market with technically maybe no moats, but they've continued to win and succeed. The two moats you think about with Cursor are the data feedback loop of people autocompleting, learning what they're doing, and training models based on that, so that's unique. The other is just best-in-class experience and being a high daily active user product and finding over time what works and what people need. What have you learned, and any thoughts on moats in this space that might be helpful for folks that are trying to figure this out for themselves?
Yeah, there's a lot of talk about moats, and it is a pretty interesting moment in time to be starting a company. So I can understand why so many founders are asking themselves that and trying to project out 12 months from now, 24 months from now; it just feels like an eternity. I will say that I think if Cursor and many other successful companies of this vintage had thought about moats, or tried to work backwards from some strategy diagram or a maybe more abstract notion of how a company should work, I don't think that would have created this outcome or this product.
I think what really created the magic of Cursor was an obsession with building a useful thing today. And I think it was constantly this exercise of: you can kind of see where the world is going three months from now, six months from now. Models are going to get smarter. A thing that isn't solvable now is finally going to be solvable. And I think Cursor was a little bit this recurring prompt of how could we pull that stuff to today, even if it requires a little bit of engineering on top to make it work, or a lot of engineering on top to make it work. Even if it requires changing the product in a specific way so that a user can interact with this new capability, how can we bring that forward? And then three months from now we should delete all that stuff because it'll just be good and the common bare minimum of the product, and then we'll build the thing for three months from then.
And it was constantly doing that over and over again that I think led to users really trusting us and placing their time inside of our product and trusting that we were bringing things to this next frontier and to the next future. And so I would really encourage many founders or people starting out today to be more grounded in that perspective of: how can I make something that is not possible now possible? Users are going to come to me to use that thing. I'm going to pull them to the next impossible frontier, and then through all of that I'm going to gain a lot of distribution advantages. I'm going to gain data advantages. There will be value there. But I think that's really the place to play.
I love that answer. Essentially, the way I'm thinking about it is just build something people are obsessed with. Don't overthink the moats piece. And if you can continue to do that, you'll find something, which in Cursor's case ended up being a few things. And that came up actually recently on another podcast I did. I don't know if it'll come out before or after this. This idea that moats are discovered, not planned ahead of time, a lot of times.
Okay, let me actually ask you for Grok Bot tips as an actual last question. Some people are going to be like, "Oh, I've got to try this thing. What's all this excitement all about?" What would be some advice for folks that are trying it out? Let's say for people that are new to it, here's some keys to success, and maybe some power tip for someone that's already using it and is just like, "Oh wow, I didn't know that."
Yeah, I want to stay away from the super hacky pro tip stuff, because I think our philosophy as a team and as a company is that those things really shouldn't exist. There shouldn't be all these crazy knobs. You should be able to delegate something to Grok Bot and they should do it. And so what I would encourage for someone new who's just downloading the app, you're looking at this screen, I think the first thing is give Grok Bot the context it needs to be successful. So in a similar way as if you were onboarding someone to your team, it'd be really helpful for them to have access to Slack and your email and the company records that you use every single day. So I'd give it access to the tools, and then I would actually ask Grok Bot what it can do for you and let it go through those connections that you've initially set up.
In my case, it might be I give it my email, I give it Slack. And I was really surprised when I was first onboarding. We didn't have any onboarding screens at this time, so this was the first task I gave it: go through my Slack, go through my email and suggest five things that you can take off of my plate and what it would take for you to do that. And it suggested five, and two of them were actually really helpful, and I just immediately spun off two bots to solve those two. And that was my big wow moment of feeling like no other AI tool in the past could have done those two things. It was not like "draft an email." It was like "do a chunk of work." And so I encourage people who are brand new to do it that way.
And then for people who are not brand new, I am constantly finding new patterns for ways that my bots can interact with each other and can collaborate with each other. And so I've been creating a bit more of a scaffold of where these artifacts that Grok Bots create should live and how it can write to a place that's very legible to me. So I have these frequent digests that I read every day, and it pushes to a database and I can just read it very easily. And so I would encourage power users to think about ways that Grok Bot can actually write to a single store where you can organize a lot of its outputs much easier.
Damn, we need another episode going deep on Roman's Grok Bot setup, which probably has way too much private, sensitive information. We couldn't show it, but that's okay. That's an amazing tip. Roman, is there anything that you wanted to share or anything else you wanted to touch on before we get to our very exciting lightning round?
Nothing else on my side.
We covered so much ground. I can't believe that was only an hour and a halfish. I felt like we've been talking for ages and covered everything I was hoping to cover. With that, we've reached our very exciting lightning round. I've got four questions for you. Are you ready?
I am ready.
What are two or three books that you find yourself recommending most to other people?
Yeah. So, two books for you. One is, I love Kurt Vonnegut, so Cat's Cradle has been a fun recommendation, and a copy that I've bought many friends before. And then second is The War of Art by Steven Pressfield, that I find myself frequently coming back to, even if it's just a page or two at a time. And I'd recommend it for anybody.
War of Art. Incredible. It's such a short book. And it's not The Art of War, which is what people might think they're hearing.
Yes.
But it's The War of Art. It's a play on that, and it's about the challenge of being creative and creating something new and how to overcome the resistance. I love that recommendation. Next question: favorite recent movie or TV show, if you've had any time to watch any of these things.
Yes. So every year I do a watch of Casablanca, which is one of my favorite movies, and it incidentally also has a character with my last name, Ugarte, which is the only example, I think, of an Ugarte in the media. So Casablanca, always a great rewatch. And then on the TV side, I sometimes sneak in an episode of Monk, the detective show, which was one that I watched as a kid with my family and I've come back to now that I live in San Francisco. And it's just a great moment-in-time snapshot of San Francisco in the late 90s, early 2000s when it was shot, that I really enjoy.
First Monk reference on the podcast. Okay. Favorite or most interesting AI product right now. You can say Grok Bot if you want, but if there's anything else, you get bonus points.
I've always been a big AI semantic search nerd. I love any semantic search product, especially the kind of out-of-the-ordinary ones. So I was a very early user of Metaphor at the
time, which became Exa. And I love using Exa to do all of these maybe more strange queries over the internet, but I see a lot of examples of people building semantic search over, you know, an embedded image store of the MoMA or things like that. And I always have so much fun playing with those. So, anything semantic search engine over a weird data set, I love.
And Exa in particular is one you'd recommend.
I love Exa. Yeah.
Very cool. Okay. Favorite life motto that you often come back to in work or in life.
Not as short as a single motto. But I love the Desiderata, which I don't know if you've read, but I have it on my door, and I've had it since I was a teenager, and everywhere I move I paste it there. It's a very short poem, but each line, I find myself finding something new in it every time I read it, and I find it really grounding.
Roman, this was amazing. What a point in time we're here right now, at this moment in time of Grok Bot, of AI in general. It's going to be really fun to revisit this in a year and be like, "Wow, we were so right and so wrong about so much." Thank you so much for doing this. I know it's a very busy time on your team right now, so I really appreciate you carving out a couple hours to chat. Is there any place you want to point people to? Anything you want to plug other than check out Grok Bot?
Check out Grok Bot, of course. And yeah, main thing would be please send feedback. I think we're in the very early innings of this still. I mean, we released a beta 3 weeks ago. And a lot of the feedback that we've been getting from this early set of users is directly translating to what we build and how we build it. And so really appreciate all the input that people are giving.
Nice job, Roman, and team. I know there's a whole team behind all this. Roman, thank you so much for being here.
Awesome. Thanks, Lenny. Bye everyone.
Thank you so much for listening. If you found this valuable, you can subscribe to the show on Apple Podcasts, Spotify, or your favorite podcast app. Also, please consider giving us a rating or leaving a review, as that really helps other listeners find the podcast. You can find all past episodes or learn more about the show at lennyspodcast.com. See you in the next episode.
Article published · Updated