Unfortunately AI will also create a lot more BS jobs. Being a very "mid" engineer without being John Carmack used to be a pretty good career that put food on the table, slowly putting together tests and coding frameworks over the course of months.
I'm not really sure about the first (non-bs) category:
One of them is the bureaucratic control function in service of management. These are jobs that focus on creating and enforcing procedures, that produce statistics, create reports, and calculate performance metrics that support management decision-making but do little else. Simply put, these are the jobs that create (and make up) forms and see to it that they are filled out and signed, that “push papers” and collect data on how different functions in the business are doing. Their only or main contribution is to support management’s aim to control, supervise, and direct production processes that make up the core of the business.
I can't say it isn't true in general, just in practice I don't think that I agree: the best results come from tuned processes, not the generic ones you get out of the box. Oftentimes the harnesses do improve a process to be better in general, but that doesn't mean that it actually gets better for your bespoke usage.
For example: Claude since a couple of weeks runs a lot of fuzzing code by default. Do I need it to do that if it just established that something doesn't apply to my code? Or am I burning tokens?
If my cost base to do something with only generic process is higher than someone that logged the trace the first time, analyzed it, tuned instructions and reduces waste that way, then they have a moat over me. In fact, if they have a tuned process that does all the tuning faster and cheaper than I would be able to do it, then they may have an exponential moat.
Now, read instructions as procedures, analyzing a trace as calculating performance metrics and tuning instructions as support management decision-making, and suddenly those obsolete bureaucratic control functions become essential for surviving an AI-based race.
We need patient people that can do this tedious work - even if they outsource the process to a bot, and even when they outsource the process of outsourcing to a bot, you still need to describe it. If however the process becomes obsolete, and the only thing that matters is how much money you've got so you can throw more tokens at a problem, only Altman will live the dream; the rest of us will just fight over the scraps, as eventually only Altman will have money.
Unfortunately AI will also create a lot more BS jobs.
Being a very "mid" engineer without being John Carmack used to be a pretty good career that put food on the table, slowly putting together tests and coding frameworks over the course of months.
Virtual Reality
I'm not really sure about the first (non-bs) category:
I can't say it isn't true in general, just in practice I don't think that I agree: the best results come from tuned processes, not the generic ones you get out of the box. Oftentimes the harnesses do improve a process to be better in general, but that doesn't mean that it actually gets better for your bespoke usage.
For example: Claude since a couple of weeks runs a lot of fuzzing code by default. Do I need it to do that if it just established that something doesn't apply to my code? Or am I burning tokens?
If my cost base to do something with only generic process is higher than someone that logged the trace the first time, analyzed it, tuned instructions and reduces waste that way, then they have a moat over me. In fact, if they have a tuned process that does all the tuning faster and cheaper than I would be able to do it, then they may have an exponential moat.
Now, read instructions as
procedures, analyzing a trace ascalculating performance metricsand tuning instructions assupport management decision-making, and suddenly thoseobsolete bureaucratic control functionsbecome essential for surviving an AI-based race.We need patient people that can do this tedious work - even if they outsource the process to a bot, and even when they outsource the process of outsourcing to a bot, you still need to describe it. If however the process becomes obsolete, and the only thing that matters is how much money you've got so you can throw more tokens at a problem, only Altman will live the dream; the rest of us will just fight over the scraps, as eventually only Altman will have money.