55) Comparative Advantage

Behind-the-scenes building Vambrace AI, a company on a mission to figure out its mission. Please pardon the stream-of-consciousness style. Subscribe to follow along or visit the site here:

(typos are to make sure you’re paying attention)

Introductory Remarks

Dear Vambracers —

In our last post, Hair on Fire, I explored Y Combinator’s concept of a “hair on fire” problem and how startups are meant to identify and solve those problems. I explained that, having now lived life as a founder for almost a year, I have had my fair share of hair on fire problems, and I would absolutely use anything to put them out. Moving on!

This one is a bit all over the place, but I think some relevant nuggets may emerge over time.

Comparative Advantage, Organizations, and AI

In today’s post, I wanted to explore comparative advantage in the AI era. I’d like to tug on some related threads:

Comparative Advantage

What is comparative advantage? My recollection from undergraduate economics is that comparative advantage is pretty much when one entity, Entity A, can produce some good or service, GoS A, more cost-effectively than some other entity, Entity B. And if Entity B can produce some other good or service, GoS B, more cost-effectively than Entity A, then they can each specialize on their service-of-focus and engage in a trade deal that results in more aggregate GoSs than would otherwise have been possible.

As someone building a company, comparative advantage has come up a lot in terms of internal resource allocation and competitive dynamics. I am making decisions daily about where to spend time, what skills to develop, and when to seek outside help. I have generally adopted an attitude of curiosity and learning and “I’ll-figure-it-out” which has taken me far, but which also has some limits. Those limits have compelled me to reflect more seriously on task allocation within the company, and when it makes sense to specialize in an AI-first world.

Dynamics within an organization

From my perspective as I’ve been building the company, I’ve been functioning as an AI-enabled generalist across pretty much all business functions. This has been great to help me perform adequately across all functions, but for functions where I have largely relied on AI, it’s unclear whether I’ve outperformed the average or my competition.

This has compelled me think about specialization vs generalization of human capital within an organism through the lens of core vs non-core activities. And specifically by “core” we mean value-driving activities.

In my case, I offer technology solutions and workflow automations to companies. The quality of our output is driven by the quality of our code, the depth with which we understand the workflow and/or business problem that we’ve been retained to address, and our ability to communicate our process, drive real adoption, and ensure satisfaction and trust with our clients. We also must identify and acquire new clients to perform these services for.

I feel like I’m above average at customer acquisition, mapping workflows to business problems and identifying solutions that lend themselves to AI, and driving adoption, impact, and trust across an organization. For these activities, AI has been a force-multiplier for me, and has expanded the creative horizons of my approach, accelerated the pace with which I can engage in these activities, and has probably increased quality overall. This is great!

But, I also am not really an engineer, no matter how hard I try to convince myself otherwise, and so I haven’t been able to effectively leverage AI for multiplicative output on the quality of our code piece of the equation here. That has led me to finally seek help, and I recently started working with a brilliant engineering friend of mine who can quickly build sophisticated and robust solutions with the help of AI (which is how I stopped my hair from burning).

To close here, a related point is that, for non-core business activities, like back-office, admin, FP&A, etc., it’s okay to be an average generalist there because they don’t make my beer taste better. They’re important to not mess up, but at least for now they’re also largely binary. That likely changes as an organization matures and there’s enough capital to hire specialized talent across functions.

Comparative Advantage of Token Efficiency & Specialization

But for now, I’ve started thinking about organizational development and responsibility allocation from the perspective of internal comparative advantages of token efficiency (or something like that).

By Token Efficiency, what I mean is that, for some objective, Z, there is some set of adequate outcomes, with different levels of “correctness” associated with each. Meaning that, to accomplish objective Z, there is an A+ outcome, an A outcome, A-, B+, B, etc. For a core activity, the goal should always be A+ (or at least A—because as we know perfect is the enemy of good enough).

Now, within the context of AI-assisted development—across coding and other traditionally non-coding tasks—the literal completion of objective Z consists of a series of prompts. Each prompt contains an input, which may include context that the AI agent is pointed to, and then also includes some set of instructions and/or guidelines for the agent to adhere to. In this way, we work with AI agents to develop things. There is a world where we could actually associate total token usage with the completion of objective Z.

But really I think the token usage will be less relevant as token costs go down and multi-model approaches (as well as caching and other more sophisticated token optimization strategies) come to the fore. So really the main factor for a company is the human time required to accomplish objective Z, where that time is the time spent both preparing an agentic environment (e.g., equipping it with context) and actually typing the prompt (aside: that’s why people love dictating prompts since it’s 3x faster, but I personally think dictation can lead to superfluous input).

So, if there are two people within an organization that could accomplish Z, the internal calculus should be: which person can achieve a sufficient outcome for Z with the least amount of human-time? (Also controlling somewhat for “non-problematic” token use.) And really the main point I’m trying to make here is that individuals with specialized organic knowledge of some objective can arrive at some sufficient outcome in the least amount of human-time.

I guess this is really just pretty basic stuff around maximizing output and minimizing input—but I think it kind of gets lost in the discourse a bit. As someone who has generalist’d his way to $20K MRR, and who firmly believed that he could do everything by himself with the help of AI, I just am realizing that—even if I could arrive at some outcome with the help of AI—it may be a wildly inefficient and sub-optimal approach. And that’s why self-awareness is critical in leveraging comparative advantage for more efficient organizational resource and responsibility allocation.

And, all of this is to say, that it’s really lonely being a solo founder. :)

Aside: Generalists vs Specialists

Skills development is a critical piece of the broader AI discourse, including specifically the concept of generalists vs specialists. The argument in favor of generalized talent is that, given AI-assisted coding tools and AI-enabled skills augmentation, traits like agency and taste will become more valuable than some specialized, highly-trained technical skill. I think this is generally true, but that there are some nuances.

Specifically, I still think that the top 0.1% of some specific skill will remain human-generated. It’s possible that AI actually makes the top 0.1% of certain skills more valuable than they’ve been in the past, since we’ll be paying for the “human-hours” that have gone into some output. For the rest of the bell curve, or like from the 98th percentile and down, I think that there is legitimate fear of displacement and/or economic devaluation associated with generalist new-entrants.

But only time will tell.

Looking Forward

I’ve been so busy the past few months that I haven’t been spending enough time just thinking and reflecting and so I’m trying to get back to that. Fun for me, not necessarily fun for you.

Have a great week!

Sincerely,

Luke