Who I am.

Centaurs & Augmented Human Intelligence

After the euphoria that followed the first release of ChatGPT to the public had subdued, the high expectations turned into doom scenarios: junior developer positions would be eliminated, or maybe all developers regardless of seniority as well.

While it is true that with the help of AI even non-technical people can create software, those predictions didn’t come true. I agree that maybe junior positions will be eliminated — not because new developers are not needed, but because armed with AI no one is junior anymore.

With AI, the augmented versions of us are what Cory Doctorow calls a Centaur, or as I like to say, we possess AHI — Augmented Human Intelligence. AI serves as a proverbial gun: what was achieved in the past with the invention of guns — that raw muscle power didn’t matter anymore but the skill to use one — the same is true with the use of AI today. It is a productivity and skill equalizer. The productivity and skill gap between an average and the most skilled developers is shrinking. It is like wearing an exoskeleton, but for your brain.

But, as Andrej Karpathy noted, outsourcing thinking doesn’t mean giving up on understanding, or as Addy Osmani puts it, to cognitively surrender. We need to stay in the loop and keep our hands on the wheel, to manage the agents that are building the software and to steer the process in the desired direction.

Where Do I Fit In?

While there still are ones that deny the power of LLMs and try to depict them as better autocomplete or statistical prediction machines, my opinion is that they empower developers and accelerate the development process and the pace of innovation.

I started experimenting with vibe coding and using the LLMs as what Addy Osmani calls a Conductor. Although definitely fun and enabling me to create projects in otherwise unfamiliar technologies, the more complex projects were impossible to develop in that manner.

But with the use of a more structured approach known as Specification-Driven Development and the use of frameworks like Specify or Git Ship Done, working on more complex projects is far more predictable and a pleasant experience.

There is also the emerging concept of the AI Software Factory where our role is to manage AI agents that are controlling every step of the software development process, in a role that Addy Osmani calls an Orchestrator.

What Does the Future Look Like?

There are many paths ahead that can be taken, but at the extremes of the spectrum there are 2 paths that are most illustrative. Let's assume for the sake of discussion that AI-augmented developers — the centaurs among us, a.k.a. the Software Builders — are 10 times more productive than the ones that are not. This may sound like an exaggeration, but may actually be an underestimation. That is not the point, 10 is a nice number for calculating the assumptions made for those 2 paths:

• The first path is to reduce the number of developers by 90% and get the same output with 10% of the costs.
• The second path is to retain all the developers, but move 10 times faster with the same costs as before. There are no savings, but in the same amount of time as the ones that took the first path, there is 10 times more output: experimenting is cheap, and innovating accelerates.

Soon, no matter the costs, the ones who took the first path are no longer competition or are probably out of business. The bottleneck is no longer time to bring new features to the market, but the ability to sell those features, or the ability of the market to absorb them.

The best path is probably between those extremes, since achieving such an acceleration requires a suitable token budget, but the potential surely justifies the cost. If it seems counterintuitive, think of the Jevons paradox: AI will make software cheaper, but instead of needing fewer developers and making them less valuable, cheaper software is more accessible, increasing the demand for it, which in turn drives the demand for software developers up and makes them even more valuable.

I will wrap up with an attempt to illustrate what is possible with an example that comes outside of the software industry: Paul Conyngham is not an oncologist, but managed to design a cancer vaccine for his dog. It sounds impossible, but now it isn’t.

Now, the only hurdle between an idea and realization is perseverance and a sufficient token budget. Do you dare?

Technical Toolkit