What Should We Be Building Today?

There has never been a better time to build than right now, but we only have so many hours in the day and tokens left until the next reset, so how do we decide what to build? If you assume that model capabilities will continue to increase and token costs will eventually decrease, then maybe there is little value in asking or answering this question. Just build it all. The AI can built it. It can maintain it. Whatever it's not good enough to get right today, it'll be able to fix tomorrow.

I'm not sure that we can expect model capabilities to increase at the rate they have so far. And even if they do, I've seen no evidence to suggest that token prices are going to go down. So how can we be sure that some future model that's 10x as capable as Fable is today won't cost 10x as much?

One real possibility is that we're in the golden age of AI software development. It might get a little better from here, but it's not going to last forever. Highly subsidized, unprofitable subscriptions designed to drive adoption will eventually have to give way to a more sustainable way of selling inference. The influencers talk about intelligence too cheap to meter, but there's nothing happening in the market right now that points to that future ever taking shape.

So what if we have six months of virtually unlimited tokens for a couple hundred dollars a month? What should we build if the free ride is coming to an end? I think there is an answer to this question that will give us an enormous advantage no matter what the future brings.

The first thing I'd do, if I thought the golden age of AI software development had an expiration date, is use this opportunity to get my house in order. I'd update or replace any legacy app carrying significant technical debt. Particularly those on old frameworks or EOL language versions. AI is very good at quickly generating greenfield applications. My legacy application is a living prototype, complete with a lifetime of data. It's exactly what AI needs to jumpstart a successful build plan. Rewriting an application from scratch was always the wrong move, but today, it is very likely the right one. Particularly since there's no harm in throwing it away and starting over if it doesn't come out quite right. It's just tokens, and that bucket is (for now) always just a few days away from getting refilled.

The next thing I'd do is build or wire up tooling to automate the things that kill velocity. My developers shoudn't have to drop the feauture they're building to handle things like exceptions, slow queries, and vulnerable dependencies. I'd plug in robust monitoring and create a pipeline to automate diagnosis, repair, and testing, looping in a developer only when human attention is genuinely warranted. Full self healing is now table stakes for any application that's running on a modern framework with a robust set of code quality tools.

Finally, I'd break down the knowledge silohs that have always slowed down development teams. I'd build an organization-wide, living knowledge base, updated in real time as code and operations change. AI tools, out of the box, are optimized for individual contributors. So to the extent that we're seeing gains from AI, it's because it's increasing the efficiency of individuals. But if a key individual leaves, we still suffer from the same loss of institutional knowledge that we did before. The bus factor hasn't gone away. The bus is just moving faster.

In short, I'd build the things that make building faster, cheaper, and easier. I'd ensure that no matter what the frontier labs or governments do with model availability and price, my team could still get its work done. Velocity is the lifeblood of software development and while AI has theoretically solved it, there are way too many things still dragging it down. I'd spend today's tokens finding and removing those speedbumps.

Companies that are using this period of unlimited, affordable AI to churn through features instead of shoring up their foundation will struggle if the rug gets pulled. Layering more unmanaged technical debt onto legacy apps that were already struggling before AI just because Claude can be looped in to fix what breaks creates an existential dependence on the existence of affordable frontier AI. If prices go up enough or availability drops off too far, some projects will become unmanageable and could drag whole companies down with them.

But even if the rug is never pulled, companies that take the time now to modernize their stacks and build out automations and processes to protect their velocity will far outpace the ones that keep swiping the tech debt card. Every developer hour and token a company spends patching things up as they break is an hour and token that can't be used for that coveted new feature development. Technical debt, like financial debt, has a cost. If you carry a lot of debt, you pay a higher interest rate on the next dollar you borrow. If you carry a lot of technical debt, every new feature costs you more hours and tokens. Companies that pay down their technical debt and set up autopay will always pay less to build the next thing, and they'll build it faster.

If this resonates with you, but you're not sure how to get started, reach out. This is, after all, the golden age of AI software development. Everything I've described here costs way less and takes way less time than you think it will.