Joining Boundary

2024 Feb 18

I've been in the industry for about 6 years now: the first 4 at Google and the past 2 at Trunk, a series-A DevEx startup that I joined pre-A.

I realized a few months ago that I wasn't excited in my role anymore, and asked myself what I wanted to do. After a bit of looking around, I found that the answer was that I wanted to transition to AI/ML¹ - so that's where I started.

I ended up with two offers:

• one from Boundary, a no-name seed-stage AI/ML startup with 2 co-founders and 0 employees, and
• one from OpenAI.

I chose to join Boundary.

Why? - the short version

I was excited about Boundary, in a way that I wasn't about OpenAI.

I actually originally decided that I was going to join OpenAI and turn down Boundary - or as Vaibhav, one of Boundary's co-founders, put it, "make the objectively sane decision".

But the first time I told a friend I was going to join OpenAI, I couldn't help but feel like I was turning down a chance to do something big. That feeling stuck with me when I told a second friend, and was still there when I told a third friend.

So when Vaibhav called me to pitch me one last time, I realized that right now, the decision that makes me happy - the decision that makes me excited - is joining Boundary.

Why? - the long version

This was the first time I've ever had the luxury of being able to choose between two amazing, two very different opportunities. It took me a long time - three, maybe four weeks - to actually come to a decision.

I realized, at some point, that the question I needed to ask myself was "what do I want to do in life?", and for me, the answer is that I want to do something that matters, and ideally, I'd like to build it from the ground up.

There's absolutely an appeal to building resilient training infrastructure, to building a scalable system from zero, to taking a system from 99% available to 99.99% available (or even 99.999%). But it would be even cooler to be able to do all that for something that I've built from scratch - and that was what pushed me towards Boundary.

It took me a long time to get to this realization though. I went through all the criteria² that I could think of, and then some:

• money
• career opportunity
• product opportunity (both absolute and relative³)
• timing (both product-in-industry timing and personal timing)
• culture (people and values)
• learning opportunity⁴
• location
• doing something else

Money

OpenAI compensation is impressive:

But money, past a certain point, has diminishing value. I'm not trying to retire early. I'm not looking to go heli-skiing once a month. I'm not even planning on buying a house anytime soon.

I'm fortunate enough to earn enough that I do not, and will never have to worry about making rent. I don't have to buy groceries based on what's on sale. I can comfortable fund all my hobbies (climbing, skiing, and photography). And I'm confident in the stability of my earning potential.

That's enough for me.

I know too many people who've optimized for money at the expense of their time, energy, and youth, and I don't want to be one of them.

Career Opportunity

Or, phrased alternatively, the ability to do anything I want to do.

With Google, my resume is pretty darn solid. With OpenAI⁵, it would've been bulletproof - especially for future AI/ML opportunities.

With Boundary, it's a very different career opportunity:

• It's going to mean defining what the AI/ML developer experience looks like, 5 years from now.

• It's going to mean tackling a lot of Hard Problems™, from building a compiler to defining our infrastructure patterns to building our brand identity.

• I know that I can learn a lot from Vaibhav and Aaron - the Boundary co-founders - in everything from technical instincts to product thinking to leadership skills.

• It's not going to have household-level brand recognition - plus, there's a very real chance that we won't make it.

Most importantly: it's the opportunity to build something big, and to build it from scratch. OpenAI is doing amazing things - just look at what they're doing with video! - but I wouldn't be starting from zero.

And if the goal of career experience is to get to a point you can do anything you want to do, and I've got a reasonably future-proofed resume already, well - if one of my options is to do what I want to do, why not take that one?

Product Opportunity

One friend suggested that I can look at OpenAI in one of two ways: either it's Yahoo - it's already made it, and it's only downhill from here - or it's Google in the mid 2000s, and who wouldn't want to join that?

But again, regardless of which of the two it is, when I compare that to the opportunity to do something big, something that's actually interesting to me, and to build it from zero, it's hard for that to compare.

Most startups, at least to me, are not interesting. They might be academically interesting to me - e.g. I can see why this is a hard problem, and yes, it's absurd that no one's solved this yet - but they're not working on problems that interest me.

Boundary, by contrast, is aiming to create a new, high-quality, open-source ML development experience. That, to me at least, means a lot of hard, interesting problems, and if Boundary's vision was anything less than that, I would've joined OpenAI.

• We want to give people the right abstractions to build on top of their ML models: everything from inline comments that get stripped from your LLM prompts to support for Python and Typescript and making it easy to switch between ML service providers.

• We want to enable people to test the ML features and products that they're building, which is especially important when you're dealing with probabilistic systems and defining correctness is harder than enumerating edge cases!

• We want it to be easy to deploy changes to your ML features: you should be able to both self-host everything that calls an OpenAI API and ask us to handle that for you, function-as-a-service style.

• We want our users to be able to monitor their ML usage and ask questions about the precision and recall of their deployed model, about the costs of the current deployment, and about the reliability of the current deployment.

• We want it to be straightforward to refine your ML usage, whether that means LLM prompt tuning, fine-tuning an existing open-source model, or training a special-purpose model from scratch.

And we think that the right way to do all this is to start with:

• a freely available, open-source schema language for your ML APIs,
• code generation for your LLM interactions, and
• robust, fast, easy-to-use tooling to support every step of the process.

Importantly, this approach has a number of advantages compared to competitors in the space (I may or may not have put together a spreadsheet at 2am at one point to assess this):

• We can offer our users a flexible, end-to-end platform. No one likes stitching together 10 products to build their workflow.

• We don't have lock-in: our schema language, compiler, and IDE integrations are all freely available and open-source, so if users want to use just those, they're more than welcome to.

• We can build our platform and ecosystem incrementally. Every platform suffers from the critical mass challenge - that you have to build out an entire platform for using it to be attractive, and then get enough adoption to accrue network effects - but everything that we want to build will be independently useful, so we'll be able to respond much more quickly to our users as we build out.

• We're not tied to LLMs: if the winds shift and the industry discovers new model architectures, hosting patterns, or whatnot, we'll be well-positioned to respond, because our value proposition is giving you the right abstractions for your ML APIs. We have a lot of special support for working with LLMs, because the existing general-purpose LLMs are wildly useful. But there's definitely some insanity to the fact that API calls in the LLM world can and do take multiple seconds.

That's the pitch.

It's long - we're still working on it. But with Boundary, I get to be part of that journey, whereas at OpenAI, they have answers to a lot of these questions around what to build, how to build them, and an established leadership/executive team to iterate on those answers. (And I do believe that OpenAI now has fundamental execution risk around the innovator's dilemma.)

Timing

I'm very confident that the products, companies, organizations, and technologies that will be dominant in 5 years are getting started now. As confident (or as arrogant!) as someone in their late 20s can be, I guess.

Looking at the past waves that made it big - cloud, mobile, sharing economies, IoT⁶, social, PCs, internet - and all those that didn't - VR, blockchain⁷, no-code, 3D printing - there's very clear value creation that LLMs specifically have enabled, and the AI spring that we're in right now feels here to stay.

The top challenge right now feels mostly like implementation - namely, the fact that ML operational costs (and, to a similar extent, capex as well) are growing exponentially as capabilities improve. Every wave that became big was able to to do so off the back of diminishing costs, either in the form of Moore's law or network effects.

With OpenAI, the timing factor is that I would be joining as a cog in the machine. A large cog, yes, because I have no doubt that they're going to continue growing at a truly insane pace, but still one of many.

Personal timing

It also helps that I don't have a partner, I don't have kids, and I don't have a house right now. If any of those change, and I need something more stable and less risky in the future - well, I don't know if I'll have the risk tolerance for something like this at that point.

Culture: people and values

Vaibhav and Aaron are incredibly smart, capable, and thoughtful people - I've known them for years now, and if there's anyone I know that can pull it off, it's them. I don't say that lightly: I've had the privilege to work with a lot of really amazing people.

They've got the technical credentials: Vaibhav's worked on everything from bringing Face Unlock to the Google Pixel to building custom Python runtimes; Aaron rewrote AWS EC2's internal telemetry system to get rid of its scaling bottlenecks.

The ways we worked through all the disagreements⁸ we had (and we had a lot of them!) during my week-long work trial convinced me that they're thoughtful, respectful people who I can not only be honest with, but also learn from even when we disagree.

(Yes, a week-long work trial is pretty ridiculous - I certainly took a lot of convincing to do it - but ultimately I'm glad I did it. It ended up being just as much them interviewing me, as it was me interviewing them. Not only did we work through a lot of diagreements, but we also talked a lot about all our communication styles, and our own strengths and weaknesses.)

Regardless of whether or not we succeed, I'll be excited to spend the next few years working alongside them, learning from them, trying to do something big.

We're all clear that this journey is only worth it if we have fun along the way, if we build an environment where we enjoy each others' company, that we consider that vital to our success, and that this will require very conscious and deliberate effort as we grow.

OpenAI, by contrast, is now big enough that they're concerned about leaks.

Learning Opportunity

One of my favorite things about Google was the sheer amount of learning opportunities I was surrounded by. I was able to not only get my hands dirty deep in the guts of Bigtable, Spanner, MapReduce, Borg, and Chubby, but I was also able to go and learn about how web search and indexing worked, how YouTube video uploads worked, how Google did load balancing, so on and so forth. I loved being able to go "I wonder how this thing works" and then actually reading the design documents that people wrote as they were building the thing.

This was not a thing that most people did. It's part of the appeal of being at Google, I think - but most people never take advantage of it. I realized that about a year in, that one of the reasons I had wanted to join was because of the learning opportunities. That was when I started sponging up knowledge - about how YouTube transcoding works, about how our monitoring system dealt with fan-in problems, about how search was unified across multiple Google products.

I have no doubt I would get this at OpenAI, where I'd be able to avail myself of world-class experts and being on the inside of impressive technologies at scale.

But at Boundary, I expect my learning to come from, well, building. I do have specific goals for how I want to focus my learning, and if our execution path ends up at odds with that (I certainly don't intend to spend any time doing resume-driven development), well, we'll cross that bridge when we get to it.

Location

I'll be moving to Seattle for Boundary⁹; if I'd joined OpenAI, I would've had to move back to San Francisco.

Both are amazing cities: there are plenty of smart, interesting people in both, and they both have great access to climbing and skiing (the two sports that I fell in love with during the pandemic).

But, when it comes to San Francisco, I just couldn't bring myself to be excited about moving back there. I actually literally spent 12 hours driving to San Francisco, while I was deciding between Boundary and OpenAI, in part to meet the OpenAI team, but really because I wanted to see if I could find it in me to be excited about moving back there. And I couldn't.

(Admittedly, I did also ask myself whether OpenAI allowing remote would change my decision - not that this was on the table; I just wanted to ask myself the hypothetical to figure out if it would make a difference - and it wouldn't have made a difference. For a different pair of locations and opportunities, though? That decision calculus could definitely have made the difference.)

Why not be a founder myself?

I don't know if it's a thing I want to do.

Founding means putting 200% of yourself into the business, 24/7, for years - I wouldn't want to do it without dedicating at least 100%, and that's just not a balance of priorities that I'm willing to accept for myself at this stage of my life. Possibly ever.

(There's also a little bit of an activation energy / escape velocity / critical mass problem, where I'd need an idea that I'm passionate about, a co-founder who I'm willing to spend years working alongside, and stability in the other aspects of my life - and I think you only get there if you just dive in head-first, sink-or-swim style.)

Or just something else?

I would love to do something like the Recurse Center at some point. Or take more time off. I could, in theory, quit for a year and just gallivant around the country or the world. But I don't think I would get the same fulfillment out of any of those as getting back on the wagon again.

Making the decision

It took me so, so long to make this decision (and Vaibhav, Aaron, I really appreciate how long you gave me to make it). I talked to almost everyone I know about it, tried the trick of "flip a coin and see if you're happy or not", and spent weeks paralyzed by indecision.

Semi-committing to an option, and telling friends that I had decided to join OpenAI, and telling friends that I had decided to join Boundary - that was the litmus test I needed.

I'm happy with the decision I made, and confident that in 5 years, that in 10 years, I'll be able to look back on this decision and be convinced I made the right call, regardless of whether or not Boundary succeeds.

Because even if we fail - and all the startup numbers say we will - at least I'll have tried to do something big, and I know we'll have fun along the way.

Aside: self-reflection

Throughout this entire process - thinking about what I wanted to do next, whether I wanted to do anything next, what kinds of things I was qualified to do - there was also a lot of reflection about myself:

• what I've achieved
• what I want to achieve
• where in my career progression I am
• what direction I want my progression to take
• what I want to do with my life

And it was a really valuable experience. Highly recommend, 10/10.