The Great Centrifuge · ch 06
Software First, Then Everything
The Vibe: Software Engineering, Tilted
Software engineering is possibly the vanguard profession being disrupted by Generative AI. There are a few reasons why I think it’s use case number one — why software engineering is the best nail — for the Generative AI hammer.
First, software engineering as a profession is used to absorbing huge amounts of productivity uplift: before Gen AI, we had already seen about a 1,000,000x increase in productivity since the start of the craft ~75 years ago, compared to say fishing and bricklaying — which have seen productivity increases of ~100,000x and just ~3x respectively — spaced along the last 10,000 years.
Secondly, software engineering is an incremental industry, with the design and deployment infrastructure that smooths over mistakes, and the feedback loops from DevOps, Agile, etc. meaning that it’s already antifragile. As a discipline, software engineering teams produce high quality outputs from combining less exact inputs in an iterative way — this is the real art of scale software development — not just writing code.
Thirdly, Software Engineering is a profession with an incredibly deep bench of Pioneers. Pioneers don’t need to be Olympic medal winners — they just need to be focused on advancing the craft.
In the Trimodal Nature of Tech Compensation , Gergely Orosz shows that technology jobs consists of three disparate but overlapping talent markets — I see this as the centrifuge for technology skills. I’ve observed this firsthand over many years., and I label them as:
Tier 1: IT professionals — not really making new software — but doing a good job deploying what others have created
Tier 2: Credentialed software engineers
Tier 3: Pioneer software engineers
The rigid, structured hiring processes and bar raising of the big techs tries to keep their engineering workforce made of 100% Pioneers. They pay 3x above the Tier 1 companies, because they know they need Pioneers, even if they haven’t named them as such — and these salaries attract hordes of smart young people into the profession — much like law and medicine in the 1980s.
Steve Yegge does a far better job than me of talking in more detail about the trajectory of software engineering as a profession due to Generative AI — here and here — but in short — the future of engineering is going to involve engineers driving many different models, orchestrating, and correcting.
The software engineers we need in the future are good enough to start and harmonise a robust argument between AI systems — people who question themselves, who question the models, who don’t get their ego threatened by a trillion floating-point operations that happen to solve a problem faster than them — and who have the depth of industry knowledge to call the models on their missteps. Who have the experience to know when to throw something out, and start clean.
These will be people who worship simplicity and determinism but are willing to take the non-deterministic route to get there. You don’t necessarily need to be better than the LLMs at the detailed activity, but you do need to be good enough that they invite you to the party, and don’t kick you out.
Earth has around 28 million software engineers in total, and I have no idea how many Pioneers there are, or where the line even starts or stops. It is, however, very clear that generative AI is putting more and more demand on the Pioneer category — Dr Matt Beane, when we were talking about this, jokingly described them as “Apex engineers”: Engineers who have no natural predators. These Apex engineers will be essential to push forward new paradigms, design, orchestrate, and integrate the output of many AI models.
In parallel, I think Gen AI is also going to massively lower the demand for Credentialed Engineers. Who knows how many can, or want to, make the shift to Pioneer— it’s a big ask — but people do amazing things under pressure.
I personally believe the general commentary that AI may generate 99% of source code within three years, but I don’t think that this was ever the main job of a software engineer, it was just how they had to express themselves. Pioneer software engineers don’t define themselves on writing code — they define themselves on building amazing products.
In my mind, there’s going to be humans in the mix of software engineering for a long time. Chess computers combined with human grandmasters outperformed either on their own for a very long overlap period, from 1997 to 2010. This is a much deeper problem space.
And I don’t think software engineering is special. Just first.
Merit, Made Visible: The Great Reordering
I believe this same dynamic will play out across all knowledge work — I think it’s imminent for every information-based industry. Market forces will make it so. Companies who don’t do this will lose to companies who do.
Professionals who don’t use AI will be outcompeted by people who use one model, and people who use a singular model will be outcompeted by people who orchestrate many.
In the future, the best won’t be playing instruments. They’ll be composers — commanding an orchestra of models. They’ll need to know the instruments, know the theory, and drive the players past their limits.
They won’t earn the models’ respect with authority or charm. Only with precision. Logic, language, and reason.
The models bloom endlessly and blindly. They generate because they must. But real quality isn’t found in the bloom. It’s forged. It’s distilled. It’s polished through judgment, through grit, through pain. Small code is good code. Simplicity is next to godliness.
The same applies to art, poetry — to anything where quality matters.
Average output is becoming free. Quality will only grow more valuable.
For the last fifty years, we’ve valued information and expertise — a rigorous certification through a formal degree. But no more. The Credentialed see knowledge as an asset, to hoard and leverage. Cognition is getting cheap. Knowledge is becoming free — this changes the value of talent dramatically.
The future belongs to those who can orchestrate — and those who can distil. These are the activities of Pioneers over the next two decades, and the skills and attributes that people will need to orchestrate and distil are massively different from those needed when generation was a human activity.
And the rate that we’re going to expect skilled professionals to produce real outcomes will only climb: the pace of delivery will be so high, that fakers will not be able to keep up.
To bring this home, this is a real talent framework that real companies use:

I asked GPT-4o to show me what talent really looks like. Here’s what it gave me:
Multi-dimensional talent, according to GPT-4o
The missing elements in our talent frameworks: learning quotient, social skills, grit, ethics, determination, creativity, passion, imagination, moral fiber, long-term perspectives. Whatever you call the skill that makes you good at helping people who are crying in the bathroom so that they can go to their next meeting in four minutes.
And: machine_generated_social_empathy_feature_1043.
Generative AI won’t just help people do better work, or change the way people work, it will also help us see what work people do more clearly. And this means we will stop relying on what people have said they did, and trace what they did do.
To put it in economic terms: AI will erode signalling theory in knowledge labor markets, and replace signals with indices.
And I can see a dark side to this, definitely — I don’t know whether we will end up with surveillance, or empowerment — I suspect a bit of both. I do think, however, that the demand for Pioneers will be so high, and companies will fight so fiercely for them, that the Pioneers themselves will treat surveillance like the Internet treats censorship: as damage to route around. I suspect the Credentialed, however, will not do so well, unless they become Pioneers — learning and growing constantly.
AI, when it truly enters the management pane, will track your contributions — your code, your designs, your words — and understand not just how much you did, but how good it was. And maybe more importantly — it will track how good the feedback was you gave others. We will be able to measure not only how well you did, but how you lifted the teams around you.
Judgment. Quality. Growth over time.