The AI Job Apocalypse: What Survives and What Dies?

In order to be successful you don’t need to be able to predict the future precisely, you just need to be right about the broad macro trends.

This is how I manage to stay ahead of most normal people.

Most of the time I keep things very simple:

What is the trend?

Why is the trend happening?

Will the trend continue?

Does this make sense from first principles?

These are the types of questions you need to think about when doing trend analysis.

When it comes to AI, there’s a useful framework I think we can use to think about what jobs may survive artificial general intelligence (AGI) - i.e. AI as smart as humans - and what jobs will die.

For the purposes of this argument, let’s forget about timelines for the moment, and let’s just assume that humans will reach AGI sometime within the next decade - soon enough that you should be thinking about the implications.

Here’s how I’m thinking about the future of jobs between now and when AGI is released, broken down into 4 distinct phases.

Phase 1: The Shrinking

The first phase is where we are currently.

In this phase we can expect team sizes to shrink, graduate employees hired to drop sharply, white collar roles to converge to more generalised job responsibilities, and more will be expected of existing employees for the same or less salary.

In short, the supply of labour will increase and the demand for it will decrease slightly.

In this phase the expectations for AI will probably be ahead of where the technology actually is, but it will be good enough to begin impacting the labour force.

Not everyone’s going to be made redundant overnight of course, but small things that were previously much easier - like looking for a new job - become much harder and lengthier processes.

Not just because there’s more people unemployed, but because you’ll be competing against a deluge of automated AI applications when applying for jobs, making it much harder to break through the attention barrier.

Hiring managers will have to sift through hundreds of applications the minute the job is posted, which makes it harder for yours to stand out, regardless of how good you are.

The key thing to keep in mind in this phase is that even if the promise of AI is not quite here yet, entrepreneurs and employers all see where the future is likely going.

They understand that:

• AI will turn all digital service providers into SaaS-like businesses.
• Eventually all white-collar work becomes nothing but compute expenditure.

Why?

Because, eventually, anything that can be done on a computer will be subject to an AI slurping up the data, learning from it, and automating it over time.

This starts with very basic workflow automation.

Initially the full workflow does not get automated, just the sub-tasks within it.

For example, previously a digital marketer may have used Canva, Adobe, and other similar software to create digital content for their business’s marketing campaigns.

Now, instead of that, they’d use an AI that is fine tuned on their business’s website, knowledge documentation, image assets, design system, etc to automatically and proactively suggest the most optimal images to run ads on for their target market.

And just like, a task that a digital marketer used to do is automated.

Did it replace their job?

No - because that’s only one small part of what they do.

Did it make them more productive?

Yes - because they can now use AI to produce more image assets, or spend their time selecting which assets are most appropriate.

Will businesses now expect more of them?

Yes - because more of their time has freed up.

This is a simple example to illustrate the phase we are currently in.

I’ll give you another example in case you’re still confused.

Product managers work on tech teams that produce software.

Their primary role before AI was to interact with stakeholders, including customers and employees, as well as talk to product designers and software engineers on their team to come up with feature requirements for the software they’re building.

Basically, they make sure the business is building the right software for the right people.

Prior to AI, the product designer would produce a design of the feature to build using Figma, and the product manager would then work with them to iterate on what it should look like, keeping in mind the requirements of stakeholders.

And I’m sure many businesses still do exactly this.

But leaner businesses are starting to re-think this workflow altogether, by using tools like Lovable to produce functional prototypes of the feature, not just pretty designs.

In this way, a business now doesn’t really need both a product manager and a product designer, you could feasibly just have one person doing both roles.

It’s now so quick to produce high fidelity designs and functional prototypes that the entire iteration process between product manager and product designer is basically redundant.

This is an example of role convergence, and it’s why I mentioned at the start of this phase that I expect to see white collar roles converge to more generalised job responsibilities over time.

You might be thinking, sure, but that’s just 1 example of convergence - how can you extrapolate that out to all white collar work?

Well, because at a high enough level of abstraction, all white collar work performed on a computer is the same thing: you are trying to achieve a goal, and that goal can be broken down into sub-goals over time.

If you set up the right environment for an AI to use reinforcement learning to achieve a goal on a computer, eventually the goal will be achieved - even if all AI progress stopped today, this would still be true. We have all we need to automate white collar work, in my opinion.

The combination of training on massive amounts of data related to a task, and a tight feedback loop containing a reward signal about whether a task was achieved successfully, along with with massive amounts of compute, is enough.

If you haven’t seen this in your specific white collar profession yet, the most likely explanation is simply that the AI Labs have not yet dedicated resources to automating it…

And that’s because of two reasons:

1. It is their explicit goal to automate software engineering first, because if they automate software engineering they can speed up progress on AI research, which they believe will help them get to AGI quicker than their competitors.
2. They are compute constrained. They cannot solve every problem - yet.

The final point I’d like to touch on for phase 1 is: graduate employees are going to struggle.

There is going to be significantly less work for juniors.

The reason for this is simple: the payback period for juniors is on the order of many years (not one year), making it expensive.

The economics of investing in a junior, who functionally will require lots of mentoring, training and investment from senior employees in return for very little output in the first year or two, simply doesn’t make sense if you believe the AI timelines most people believe.

I’d encourage you to read The Intelligence Curse for a deeper dive into the incentives of this, but the TLDR is that juniors get cut first, then middle managers, then seniors, then executives, until all that’s left is a fully automated AI firm.

Phase 2: The Arbitrage Window

Phase 2 is what I personally am most excited about.

Phase 2 is what I call ‘The Arbitrage Window’.

In this phase, we still haven’t quite reached AGI, but AI can now exceed the best humans at many individual tasks that make up a workflow - enough to automate most of the workflow.

And if you can automate one workflow that’s part of a typical job, you can automate most of them.

I call this the Arbitrage Window because I have a very specific thesis here that I’m going to personally try make money from: