Free with a sign-in.
The best 14 min of 60
The Rise Of AI Coworkers
Lauren Tan and Roman Ugarte say AI coworkers can take on real jobs when teams give them access, guardrails and clear goals.
Lightcone · Garry Tan
- Lauren Tan · Principal engineer, SpaceXAI
- Roman Ugarte
The best minutes
- 16:09Bots should earn broader access as people learn to trust them7 min
- 40:06Agents improve when they can run and check their own code3 min
- 52:17Product direction becomes the bottleneck as execution gets cheap4 min
Also worth watching
- 43:54Agent review can overfit without product context6 min
- 22:26Slack can carry a feature from request to shipped code3 min
- 35:18Cheap execution shifts the bottleneck to defining goals5 min
- 6:26Bot roles work best when each owns a focused problem4 min
- 3:35Marketing teams can connect bots to operational workflows3 min
Skip the opening, bot-sharing and multiplayer setup, the software-factory and future-interface discussion, and Lauren’s career story.
Full analysis · 6 min read
Real Alpha · Nov 14
Roman Ugarte says 4,000 PRs shift work to judgment
- Roman says he thinks Lauren puts up 4,000 PRs a month.
- Lauren confirmed over 4,000, but the episode gives no quality or cost denominator.
- Teams should stage access and reserve judgment for product direction.
Free with an account
Read the rest of the analysis
6 sections, the 3 other voices it tests Tan against, and what it means for you. Every episode we cut gets one.
Check your inbox
We sent a link to your address. Click it to finish signing up. Not there in a minute? Look in spam.
or
Signed up before? The same buttons sign you in. The episode, its moments and the transcript stay free.
Roman's 4,000 PR count points to a new constraint
Roman says he thinks Lauren is putting up 4,000 PRs a month. Earlier, Garry asked whether she was over 4,000, and Lauren answered yes. So the figure has her confirmation, but remains a self-reported personal output count.
Neither speaker supplies how many merge, get reverted, or reach users. That missing denominator limits what the number proves about productivity. A pull request measures output, not retained customer value.
Roman says engineering capacity no longer sets the team's pace. He names roadmap alignment, product vision, and agreement on what good looks like. That is a claim about his team, not a measured industry shift.
Lauren says her lab gives her unlimited tokens and calls herself an explorer. Her process has unusual room for experimentation, so it is not a spend forecast for a small team buying inference.
“the bottleneck is less the time or engineering resourcing”
Our read
Our read is that 4,000 PRs supports a change in the team's constraint, not a general productivity multiplier. The figure lacks quality, cost, and customer-outcome measures.
Roman ties autonomy to staged permissions
Roman says the team starts from how a human's credentials might be scoped by role and tenure. That analogy does not prove an agent understands its limits.
Garry then brings up Quartermaster, the YC project he says the host team open-sourced. He describes access that follows a user's permissions, with GP and software engineer roles differing. That is Garry's example, not GrokBot's shipped permission system.
The guests' proposed ladder begins with read-only access, then moves toward action. Owning a continuing job sits farther out, and Roman says those stages remain distant.
Jake Stauch made a related case on No Priors in October 2026. He described agents requesting specific tools and workflows for approval, rather than operating unrestricted.
The shared principle is explicit permission scope, not identical product architecture. The episode does not specify audit logs, revocation mechanics, or liability for an agent's actions.
“start with something really small”
Our read
Our read is that staged permission is a practical deployment pattern, not proof that competence alone earns trust. Teams still need to choose approvers and define how to stop an agent.
Focused bots make responsibility easier to locate
Lauren says people can divide bots around their own mental model of work. Roman prefers one bot dedicated to one problem, rather than one supposedly caring about everything.
For example, the performance marketing team gave its bot Google and Facebook ad access, plus a budget. It paired that bot with a design bot to test creative. The setup joins roles around a campaign task, not just a friendly name.
Lauren also says connectors bring engineering agents into Slack, where requests and context already live. Her team later prototyped a feature in a thread, connecting specialist roles to shared work.
Tan and Ugarte are GrokBot builders, so these internal examples are first-party evidence for the product's premise. They do not establish independent customer outcomes.
Zach Lloyd's Latent Space discussion in November 2026 adds a systems test. He described declarative workflows, shared agent data, and controlled tool access as parts of a factory stack. Those mechanisms can make work repeatable, but bot identities alone do not.
Our read
Our read is that a focused remit makes it easier to assign a human owner and assess a bot's work. The useful unit is a bounded job, not a particular number of bots.
Agents need a runnable loop before autonomous merging
Lauren's minimum coding loop puts execution and inspection in the agent's hands. She says the missing step is having an agent run the code and application. Then it can catch issues and iterate instead of relying on a person to run it.
That loop checks whether requested behavior works, but Lauren says it does not guarantee good code. Performance and resource use can still need engineering judgment.
Take the Slack feature thread with Samantha and a product manager. Lauren says the agent returned screenshots and videos, shortening feedback cycles. She also says she did not inspect that pull request. The team had skills and codebase guardrails before trusting an autonomous merge.
Stephen Toub described a useful boundary on The Peterman Pod in November 2026. He used agents for line-level work but retained architecture and scope decisions. His summary says unrequested changes, including bug fixes, contributed to regressions.
That is a counterexample to treating a working loop as blanket permission to alter a codebase. Teams need to know what tests cover and what happens when a merge breaks something.
Our read
Our read is that a runnable feedback loop helps agents check behavior, but it does not define acceptable behavior. Autonomous merging needs constraints and a recovery plan.
Lauren moves repeatable rules into code checks
Lauren has tried multi-model review, then pulled back somewhat from adversarial review. Her concern is that reviewers can find plausible issues without understanding product intent. A detailed review can optimize against the wrong product.
She prefers stable constraints encoded in architecture, APIs, static analysis, and CI. For example, her team forbids imports between main and renderer code paths. A dependency checker and CI enforce that boundary, so agents need not rediscover it on each task.
The architecture acts as a reusable constraint, while a review prompt spends tokens repeatedly. Lauren says the GrokBot desktop app's performance has stayed consistent and good. She gives no measured rate of prevented defects or review false positives.
Zach Lloyd's Latent Space episode in November 2026 offers a complementary measurement idea. His team scores stored agent traces against specific failure patterns, including redundant tests. That exposes recurring process failures, but cannot supply the product context Lauren says reviewers lack.
Garry's costly review example concerns his own setup, not GrokBot. Lauren provides no comparable token budget or evaluation result, so cost advantage remains unproven here.
Our read
Our read is that teams should encode stable constraints before adding more review passes. The episode offers no evidence that more review automatically improves product decisions.
Teams still gate product direction on human judgment
Roman says alignment on roadmap, vision, and what counts as good now takes more team attention. He worries that merging every quick proposal would create a Frankenstein product. That is a selection problem, not an implementation problem.
Roman says Lauren's system fixes some small customer-reported bugs automatically. He gives the example of a fix going out automatically at three in the morning. But the team draws a different line around changes to the core experience, waiting for dogfooding and agreement.
The distinction fits the risk because a narrow defect has a clearer expected result than a redesign. A changed button can disrupt users even if implementation passes tests. Lauren says speed does not displace deliberate judgment for human-facing software.
Their account demonstrates their own workflow, but offers no external result on customer impact, costs, or failure rates. A founder could test this by separating low-risk maintenance from changes that alter daily experience.
Our read
Our read is that abundant implementation makes prioritization more important, not less. Keep automatic repair narrow, and reserve human review for changes that reshape the user's experience.
What it means for you
- Count useful PRs. Separate opened, merged, reverted, and user-impacting changes. The episode's 4,000 figure lacks those quality and outcome measures.
- Stage permissions. Start with read-only access, then record the owner and approved actions before expanding an agent's remit.
- Encode recurring checks. Move stable rules into tests, static analysis, or CI. Track which review findings those checks prevent.
- Separate maintenance from redesign. Set different approval paths for small fixes and user-visible changes, then review incidents and feedback.
What to watch
- Whether Lauren's 4,000 monthly PRs are later paired with merge, rollback, defect, or user-impact data.
- Whether teams document owners and revocation paths before expanding agent write access.
- Whether GrokBot keeps automatic maintenance separate from user-visible product changes.
Send this to engineering leaders setting agent merge and permission rules.
Watch the minute where Roman gives the 4,000 PR figure
Written by Real Alpha from the episode's transcript and the other moments it links, then checked against the transcript. Quotes are the speaker's exact words.
Minute by minute
- 0:58GrokBot is built around AI teammates
Tan and Ugarte say GrokBot grew from a seven-week internal prototype after engineers saw long-running coding agents work with clear goals and tools. They describe a product organized around persistent bots with identities, computer access and proactive behavior, rather than chats or files. They say colleagues with different levels of AI experience found uses for it.
- 6:38People organize bots around their work
Tan says users can organize bots by customer or function, and that she prefers a mostly horizontal setup with a few specialists. Garry Tan says small businesses have been early users because they may struggle to hire for tasks such as bookkeeping and can turn to a bot for that work.
- 10:33Teams can share specialist bots
Ugarte says team bots let a domain expert build automations and share them across an organization. Tan says the sales team uses shared bots to research leads and handle operational tasks, while engineering bots carry common skills and practices.
- 14:47Bots need scoped identity and access
Ugarte says bots in Slack can lack context about who has expertise or decision-making authority. He argues that access should be scoped by role and can expand as people build trust in a bot. Tan describes a progression from read-only assistance to taking actions and eventually owning an outcome.
- 22:26Slack shortens the build-feedback loop
Tan describes working with bots mainly through Slack and X, where a bot spawned a Cursor Cloud Agent to prototype a feature and share screenshots and videos for feedback. She says she and colleagues built a feature in one 200-message Slack thread without her opening Cursor or reviewing the pull request, shortening the iteration loop.
- 25:58Guardrails help agents write trusted code
Tan says the workflow depends on investing in skills, codebase constraints, static analysis and tooling so recurring checks do not rely on human review. She says agents can merge their own pull requests when those systems are in place. She describes PStack, her open-source set of skills for teaching agents her engineering practices.
- 32:07AI coworkers should work in existing tools
Tan says much of her engineering work with bots now happens in Slack and X, while personal tasks remain with private bots. She says GrokBot aims to let bots work inside the tools a company already uses. Ugarte says non-engineering work lacks code's testing and deployment infrastructure, and that important context is often scattered or held by people.
- 36:41Clear goals matter more than execution
Garry Tan describes an AI-native organization in which individual contributors can ship work, outcome owners coordinate it and AI-focused colleagues teach others. Ugarte argues that when agents make execution abundant, companies need to define outcomes and metrics clearly; he says failures of intent or goal-setting may become the bigger constraint.
- 40:06Agents need a self-checking loop
Tan says a basic coding loop requires an agent to run the code and application as a user would, gather results and use them to find issues before revising. She says this can make output more correct but does not guarantee good code. Once the loop works, alerts can trigger agents to reproduce and address issues while other agents coordinate the work.
- 43:09Codebase rules can replace repeated review
Tan says multi-model and adversarial reviews can find issues but may overfit when agents lack product context. She recommends encoding recurring expectations in architecture, APIs, static analysis and lint rules, while keeping human review for sensitive systems. She says GrokBot's codebase enforces a dependency rule between Electron's main and renderer threads through CI.
- 50:12Product judgment becomes the bottleneck
Tan says the right setup can support fast, inexpensive, high-quality engineering, and that newer models deliver more capability per token than last year. Ugarte says the team's constraint is now agreement on the roadmap and product quality, not writing code. Tan says agents can handle maintenance and bug fixes, but teams should carefully assess major user-facing changes before shipping.
- 56:23Tan’s path ran from finance to engineering
Tan says she studied accounting and finance, then realized she enjoyed building spreadsheets and macros more than the finance work. After starting a company with a classmate, she joined Australia's startup community and got her first job. She says writing publicly about learning Ember led Tom Dale and Leah Silber to invite her to speak at a conference.
Published Nov 14, 2026.
Transcript · 51 min read
Lightcone, hosted by Garry Tan, with Lauren Tan and Roman Ugarte.
Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.
GrokBot is meant to be this product where you can create AI teammates. They have their own identity, their own telephone number, credit card, like basically almost anything a human can do, I think we would want bots to do. I sort of see myself as the explorer, you know, like trying to figure out what does software engineering look like six months to a year from now. welcome back to another episode of the light cone today we're joined by lauren tan and roman ugarte the team behind grokbot roman and lauren incubated the product and are on the core team that took it from the first line of code to launch in just seven weeks lauren roman welcome to the light cone thanks for having us so we're at this like kind of crazy moment where uh you know grok bot is out there uh it's used by a whole lot of people it is very very easy to get started i just noticed your uh you know grok bot is um built into the default dm now and x which is really cool what is it and how are you know both people and businesses using it right now many of you have have tried it but for those who who haven't um grok bot is meant to be this product where you can create AI teammates. And we mean that seriously. I mean, you can kind of give these
teammates names and identities and functions and skills and tools and all of these things. And I think where GrokBot came from for us is it was initially this internal prototype that we spun up in a very short amount of time. We spun it up in about seven weeks. And we did it partially because we'd seen the way that AI coding had completely transformed the way that our engineers had worked. And that came from a few things just clicking, right? These long-running agents that have these goals that they run at for long amounts of time. They have access to the right tools to do really effective work. What would that look like for non-engineers?
What would that look like for the rest of the company? And so we landed on, we wanted to build a product where the work was organized around bots as the key operative principle, not around chats or around files or any other primitive. Second is that these bots should have access to the same tools that you do. I think it would be too much to expect these bots to actually contribute to a company without having their own computer, for example. You know, we don't work every day with MCPs, so it's important that they can really use and operate the same tools a human does. And third is they should always be on. And I think we've seen this with loops and kind of other engineering principles where if you can have these bots run for long amounts of time and be proactive and come to you with ideas or insights or things to take off of off of your plate it's way more helpful than something that kind of waits for you to poke and prod it to actually do work so that's what we built we released it internally and i think one thing that was surprising from the initial reception was it was kind of this bell curve internal feedback that we got where both people who were super AI forward, Lauren included, and people who were less AI forward, some of the
Read the whole transcript
least kind of AI forward people in the company, both really enjoyed GrokBot. It just kind of clicked and it fit in their hands and kind of their workflows in a particular way. And so then we brought it to market. And I think we saw a lot of the same reception where it's easy to start with GrokBot, but the more you give it, the more you kind of configure it to your life, I think the deeper you can get and the more powerful it can become. So the reception has been really exciting to see. There's still a lot more work to do and I think it's quite early, but that's a lot of how GrokBot came to be. What are some of the use cases that you're seeing? Like who's using GrokBot and
“I want you to talk to our design team's bot. They have a design bot that creates these assets, and we're going to A-B test a bunch of them.”
“I kind of like having a bot that is solely dedicated to caring about a particular thing.”
and you support that now actually um like you guys launched multiplayer for bots and that there's sort of a bot master who has like right access and then you can share it across the org is that right correct yeah we this is recent but about two weeks ago uh we launched team bots which are these these bots that you can share with your team and i think the pattern we're seeing ourselves and we can give a few few examples um is there's a person at a company who is a bit more of a builder uh they enjoy building these tools i think inside of the engineering department there are many people like this but i think we're finding that in other departments you know gtm engineers or support engineers people like that um really love taking a few hours of their day to build these scalable automations that can take work off of the plates of of their peers and kind
of the company i guess one thing that brings to mind is basically went through uh maybe it was a year ago-ish when Steve Yege posted the whole Gas City post that then started... Software factories. Software factories? Gas Town. Sorry, it was Gas Town. That started software factories where a lot of engineers have been using this inner loop. And I think you've been building now this outer loop that you were mentioning, Lauren. And I guess what I'm hearing is basically now people can have their own software factories for other functions that are non-tech. mentioning your customer support person, your communities person.
Let's talk about this pattern. It's an everything factory. Yeah, let's talk about that. What is this AI employee, sort of the software factory for other functions? Yeah, for me, you know, as an engineer, I definitely do think a lot about the loops that I have and making sure my engineer or my engineering bots are writing the kind of code that I would write. And, you know, as Roman pointed out, I think the cool thing about GropBot is that it's given that ability to anyone And then here on my screen, I have an example from our GrokBot blog that we just I think we released this week about how our sales team actually has developed their own, you know, starting to develop their own loops for doing sales and doing research on potential leads and deals that they might be interested in.
and then going all forth and taking some of the like more operational parts of the sales process and giving that to these specialist bots that are shared across our teams so it's not like every person has one instance of their own you know quote bot or deal but we have a shared you know like an expert bot that all of our team members contribute to and that essentially gives us the ability to have these really smart team bots that you know incorporate all of the best practices that you know like our sales team has similar we have something similar for like engineering purposes as well we like engineer bots that take a lot of the you know the principles the the skills that we've developed and use them and everybody benefits from these team bots there's this new concept that is actually very powerful it's like it's not a mistake that OpenAI was experimenting with agent swarms and then one of the deleterious accidents was like the hugging face hack but then simultaneous to that they came out and solved Navier Stokes and all these like millennial millennium prize problems and then it's just really wild to see like actually agent collaboration is that next step like everyone in the last year so it's
literally been a year since opus 4.5 came out right so that which was like a magic moment for all of us and like now a year later we've like sort of mastered the single player and like sort of the vibe code and like single single player prompt but you know now we're entering this other phase of like orchestrating agent swarms and it's not like you know i mean i think that's like the more advanced version of like talking about a software factory and software factory someone's and they're just still kind of doing this thing. And I think human beings will still do that, but using these bots feels very different.
And there are so many unsolved problems still in making bots multiplayer, just making AI multiplayer in general. We face this on a constant basis, where now that GrokBots have entered our Slack workspace, they join our meetings, we're really trying to see how far we can push this AI teammate thing. there are some weird you know situations that you run into where a bot is in a slack thread you're going back and forth with your co-workers and your bot doesn't really know who to listen to it doesn't have context on who's the domain expert here who has decision making power it's as if you were to onboard an employee who has no eq and like no like social awareness of the room just the models weren't trained that way the harness wasn't built that way and so i think the this next era of AI will be, you know, these bots really need to learn those things. And we need to teach that at the product level, at the harness level, at the model level. And I think it's going to be, it's going to be really interesting next few years. I mean, it could probably look at who the owner is in, you know, in its metadata and say, well, you know, I know that, you know, Lauren's the one I need to listen to, like run everything by Lauren or something like that.
We do have the concept of like, you know, you have a, like there's the owner of the bot and the manager and you can invite I guess if we really want to think of bots like employees, then we need to think about the identity problem. Permissioning, authentication, authorization. There's still a lot to do in terms of exactly what you're pointing at.
“It's read-only, you know, has no real access to your tools. You see that it did the work in the way you would expect”
“I didn't even look at cursor for this project. I just, we did it all in the Slack thread.”
And then I would say maybe like 5% or 10% of engineers or maybe 1% of engineers, period, in the world are doing this right now. So this is literally something that should grow by 100x.
It's just unbelievable that like so few people are doing feature work this way. Yeah, I'll say like, you know, it took us actually quite a while to get to that point. I feel like it's actually quite hard to get to that point overnight. It involves, you know, investment into making sure you have the right skills. you know your code base has good guardrails and constraints so that you know the time that you would otherwise have to spend on human code review is encoded into your code base and your static analysis and your tooling so that the engineer's time yeah like pstack and so your engineer's time is freed up to actually focus on the actual product you know the user experience and the operational execution part of the job you know when you have so much more trust in your agent's ability to write good code because you've set up those systems, I've gotten to a point where I essentially have my agents merge their own pull requests. So in that Slack thread example, you know, like that was a, that was an example of, you know, me not reading the PR and this, we just looked at the artifacts that came out of it. So this cursor environment, does it already have PStack and is that the way it sort of creates, you know, decides how to build
it or yeah so uh for internal uh pstack is installed into our repo there's an internal version of pstack uh like instead of calling it potato mode it it's lauren mode um and it just started because you know i wanted a place to put my skills yeah and so yeah everyone everyone in the company can actually use pstack um and my cloud agents that the bots spawn have the same access to those skills we haven't really explained what pstack is can you explain what what pstack is for people who don't know. PSAC is actually inspired by G-SAC. Oh, thank you so much. I'm honored. Plug-in. You know, I think, Garry, your, you know, G-SAC was like quite inspiring in the sense of, I thought of it as like your personal set of, you know, chef knives.
Oh, yeah. Right, like, you know, every chef when they get a new job, they go to a new kitchen, they bring their knives with them. And to me, you know, G-SAC represented, you know, your opinion needed take on how I'll see you next time. you want your agents to you know help you and so that inspired me of uh to think about uh you know like what would a g stack look like for you know design more for like the engineering flows that that i personally like so i often tell people like you should you should actually go and create your yeah please go for it you know take our skills combine them uh you know find what works for you but i think the trust element is so important like when you trust the you know your own tools empowers you and your agents to do so much more.
I didn't really answer the question. Oh, no. But P-Stack, yeah, it stands for potato stack, I guess. And it's my set of skills that I've open sourced on the Cursor plugins repo. And it teaches agents how to do engineering work rigorously the same way that I did when I was writing code. And it's kind of evolved a little bit. And yeah, I guess it's gotten quite popular. Congrats. Yeah, it's awesome. Yeah, I am very inspired by it. and gstack is like pulling things from it whenever i can so yeah do you have a software factory that like pulled in some ideas from pstack automatically oh yeah yeah i mean now i use my software factory is cappy and then i have a captain that is sort of maps to um i think the grok bot i get the bot idea but honestly like now i think like maybe i need to use a bot on top of cappy but yeah i basically can just go to that captain and it's like the master of my g stack uh like sort of merge queue like i basically have like 10 or 15 threads working all the time and then one of them is like theo.gg he released t stack and then uh p stack and then obviously there's like lots of really good inspiration out there especially around design tools and so it's really fun to
just say, hey, go out, find what's out there. And based on what you know about me, pull in the things that you think would be most useful to improve my workflow. And then the really interesting thing about Cappy is because all of my interactions with my coding agents are right there in the transcripts, it's developing an understanding of what I want. When do I intervene? When do I say just do it and so yeah there's there's sort of this like very personal element that's emerging really through like the idea of memory and so as you use the agents more um the more it knows what you want thank you for making p stack and then both of us are uh very maligned and attacked on the internet by people who refuse to believe that any of this actually works but you know it's great to have a very brilliant compatriot i'm not crazy like this is actually something that is happening And honestly, like I'm so convinced that in five or 10 years or even like in two years, like it's going to be unfathomable that like people are not doing it this way.
But we're so early, though, still. One interesting change that kind of came up as you were sharing your screen is it feels like in the past six months, so much of the work, engineering work, non-engineering work with agents is happening in Slack. and i think it has this interesting property of since it's out in the open in the past the way that lauren would work with coding agents would have been entirely single player private the rest of the company would not have benefited from all of the crazy workflows that they should totally learn from and adopt for their own their own purposes but now that it's happening in slack i've seen this and maybe you have too time and time again where you're in some feedback thread Someone has a good idea.
You spin off an agent, like AppPotato, you know, Lauren Mode. And then the next day I see someone who is in that thread do the same thing. And I think we're not just going to see that in engineering. I think we're going to see that in the way that people work with AI in general. And the more that the stuff is in public, I think the more of the rest of the company that can get lifted up by the top 1% power users who are very tooling-minded, very engineering-minded, and know exactly how to use these tools. YC's next batch is now taking applications. Got a startup in you? Apply at YCombin.com. nader.com slash apply it's never too early and filling out the app will level up your idea okay back to the video how much of your of your like hours spent using grokbot is in the grokbot app versus in slack or some other service area to talk to the bots there i think a lot of the engineering work that i do now actually happens on slack and x so not all my bots have been added to Slack. So there are still a few that I kind of, that are more personal. Like I have a, I guess like a sort of personal assistant bot that helps me with like, you know, picking on my lunch. I don't need to add that to Slack.
That's kind of, you know, personal or something that looks at my email. That's not a team bot. That's not a team bot. I guess we could make a team. That's a personal bot. We could probably get like a shared lunch bot, sure. I think that the nice thing as well is like you decide and choose what you want to add to Slack. So my engineering bots tend to be on Slack. So because I just find it so much more ergonomic to just add the bot in the Slack thread, you know, versus me having to go copy paste the link. So it just felt more natural to bring them in that environment. I guess it really fits with your analogy of bots as employees. Like if Slack is where the work happens, that's where the human employees are. It's also where the like bot employees should be so that they can work alongside us. Totally. And we don't shy away from that at all. Like I think when we see people use crockbot and slack we love that i think that's that's where work happens right and i think similarly you know in the future they should join meetings they should be in your google docs they our goal is not to rebuild all of b2b sass um you know like other companies are doing a great job there our goal is to build the thing that can access those tools and work inside of them
and just fit naturally into the way a company already works i guess one thing that i'm curious is about is where do you think future let's say 2027 how should the best ai pill the new years build so my hope for 2027 i think it's like much more you know like we've uh maybe accomplished more of this vision that we have for for your grot bots being a like a co-worker you know they have their own identity their own telephone number credit card like basically almost anything a human can do i think we would want bots to do and uh i'd be very curious to see like how this evolves um you know i asked this question on my on my x profile the other day as well of like what is the final form of agent y and i i posed the question of like is it a chat interface or is it something else uh one thing actually pretty interesting that i saw yesterday on my x timeline is a post from the Neuralink team and just came out with this new product called Telepathy that I think, if I'm not wrong, is along the lines of, in the future, an agent being able to essentially read your mind through a brain-computer interface and I'm super interested in what the future will bring when the sort of barrier between you and how you chat and communicate intent
with your bots becomes so much more fluid that can happen like literally at the speed of thought so that's personally i think could be very interesting to see how that evolves but i don't know what about you roman we talk a lot about the software factory and then the factory
“It really matters that you define the problem well. And the gap is probably not going to be an execution failure.”
I think that there's a big gap between like the best AI builders to the so-so ones is really how you translate that intent super well. And there's a lot of tips and tricks of what you did, Laura.
“the agent now has the ability to check its own work so you can find the issues with the code that it wrote then it can write better code”
world's experts at software factories here i want to ask some like nitty-gritty questions so that's okay because i'm struggling with this right now with gstack like i have a bunch of agents that are struggling with the code review step i have like a slash ship skill that um tries to go through every line of changes and then basically spots bugs and then uh right now it's very thorough and then but when i use it it um always consumes like tens of millions of tokens uh and it spots bugs actually and so now i'm like sort of struggling with i mean this is uh match pair
to like in the plan like i still use plan mode and i have all these plan skills i don't do you still use plan mode uh no no plan mode yeah yeah so then but uh pstack must have
“I think you run the risk of overfitting where the sort of agent doesn't have the ability or maybe enough judgment to determine what's actually good feedback”
The cool thing about the moment in time right now is that you could basically, there used to be the trade-off before that you could only choose two out of three. It was either like fast, cheap, and high quality. Now you can basically do all three and we could basically do the best work in our lives right now with AI, which is super cool. Yeah, I totally agree. I think, yeah, I think fast, cheap, and good is certainly more possible.
It's not, you know, like easy, I think, to get to that point. But I like from my own experience, you know, I have the privilege of working at a lab with, you know, which has given me unlimited tokens. So I think I've, you know, I'm sort of I sort of see myself as the explorer, you know, like trying to figure out what does software engineering look like six months to a year from now. And I think, you know, this process has been very enlightening for me in terms of how to actually keep this workflow efficient. Because I think it's a topic that lots of teams and people are thinking about is the cost of tokens.
And I think when you adopt the inverse pyramid like I'm showing here, I think you can actually get to this very nice mix where the ordinary engineering tasks, high-quality engineering work doesn't have to break the bank. It can be quite economic and quite efficient. And not to mention that the frontier keeps getting smarter and smarter. And actually, I guess the unit cost of intelligence sort of drives down. The cost of tokens going down. If you compare a model from last year to a model now, the input-output tokens is not really the same. frontier token now is way more capable than a token from last year. And even if you use one of the smaller models, like even that model is like way better than what you had a year ago with the big models. So I think from that perspective, we're heading towards a period of real abundance, I think, with software and maybe other domains as well. And I think we're seeing this on our
“what we've started to see be the bottleneck is less the time or engineering resourcing going into developing these ideas and features.”
One thing that's very inspiring about your journey, Lauren, I don't know if many people know, is you grew up in Singapore and you studied a non-technical degree. You did a bachelor's in accounting and finance for Monash in Australia. Then somehow you ended up becoming one of the now best AI engineers. How did that happen? I think sometimes people think it's very hard to get into tech, but you did it. And you also worked on the React compiler, which is one of the hardest things as well.
Yeah, I think I've definitely feel like I've had a very privileged and lucky experience in the tech. I think that was a time where, you know, people were much more open to the idea of like a self-taught developer. So I guess the long story short is, you know, after college, I did, I did study finance. One of my favorite courses was actually financial modeling. and we learned how to use excel and visual basic bba to you know make some cool financial valuations and things like that um and through the course of that that course i realized i had more fun building the spreadsheet and the macros than i did the the finance part i mean the finance part was interesting too uh but the technical part you know thinking logically just was really interesting to me and something i hadn't really thought i was cut out for so after college i started a startup with a friend a classmate and it didn't go super far but we learned a lot from that process and I got really involved with the startup community in Australia and that's how I got my first job and really the secret that I attribute for my success is actually x I started out when I was still basically learning how to code I wrote I had a blog and I was just sharing
what I learned. And I was fortunate enough that someone from the framework community that I was building my application in, the JavaScript framework, Ember. So yeah, Tom Dale and Leah Silber, they reached out to me and said, would you be interested in giving a talk at the conference that's coming up in Portland, Oregon? And I was very shocked because I had never even been to a conference in my life and then now I was being asked to talk at a conference with like hundreds of people and you know public speaking was always my worst fear but somehow they convinced me to do it I learned about the ability of you know being able to communicate your ideas and bring people along within in that journey and this skill of storytelling and you know communication has really stuck with me even though i've you know um been really involved with engineering work and you know i've had a lot of really great mentors as well um in the past so i think all that to say is do cool stuff you know talk about the cool stuff and you know just keep doing that in the loop and you know uh maybe the stars will align and someone will notice something you said on x and retweet you offer you an opportunity that you hadn't had before maybe you develop your idea and you come
up with a startup idea right and then you go apply to yc and before you know it you know you could change your life with with just a tweet that's a great place to end yeah lauren roman thank you guys so much for joining us it was so fun thank you for making an incredible product so can't wait to see what you build from here thanks for having us it was a lot of fun We'll be right back.


