I am personally of the opinion that ML will end up being 'normal technology', albeit incredibly transformative.
I think you can combine 'Incanters' and 'Process Engineers' into one - 'Users'. Jobs that encompass a role that requires accountability will be directing, providing context, and verifying the output of agents, almost like how millions of workers know basic computer skills and Microsoft Office.
In my opinion, how at-risk a job is in the LLM era comes down to:
1: How easy is it to construct RL loops to hillclimb on performance?
2: How easy is it to construct a LLM harness to perform the tasks?
3: How much of the job is a structured set of tasks vs. taking accountability? What's the consequence of a mistake? How much of it comes down to human relationships?
Hence why I've been quite bullish on software engineering (but not coding). You can easy set up 1) and 2) on contrived or sandboxed coding tasks but then 3) expands and dominates the rest of the role.
On Model Trainers -- I'm not so convinced that RLHF puts the professional experts out of work, for a few reasons. Firstly, nearly all human data companies produce data that is somewhat contrived, by definition of having people grade outputs on a contracting platform; plus there's a seemingly unlimited bound on how much data we can harvest in the world. Secondly, as I mentioned before, the bottleneck is both accountability and the ability for the model to find fresh context without error.
> I think you can combine 'Incanters' and 'Process Engineers' into one - 'Users'
I wanted to talk about this more but couldn't quite figure out how to phrase it, so I cut a fair bit: with "incanters" I'm trying to point at a sort of ... intuitive, more informal practitioner knowledge / metis, and contrast it with a more statistically rigorous approach in "statistical/process engineers". I expect a lot of people will fuse the two, but I'm trying to stake out some tentpoles here. Users integrate a continuum of approaches, including individual intuition, folklore, formal and informal texts, scientific papers, and rigorously designed harnesses & in-house experiments. Like farming--there's deep, intuitive knowledge of local climate and landraces, but also big industrial practice, and also research plots, and those different approaches inform (and override) each other in complex ways.
> Hence why I've been quite bullish on software engineering (but not coding). You can easy set up 1) and 2) on contrived or sandboxed coding tasks but then 3) expands and dominates the rest of the role.
Why can't LLMs and agents progress further to do this software engineering job better than an actual software engineer? I've never seen anyone give a satisfactory answer to this. Especially the part about making mistakes. A lot of the defense of LLM shortcomings (i.e., generating crappy code) comes down to "well humans write bad code too." OK? Well, humans make mistakes too. Theoretically, an LLM software engineer will make far fewer than a human. So why should I prefer keeping you in the loop?
It's why I just can't understand the mindset of software engineers who are giddy about the direction things are going. There really is nothing special about your expertise that an LLM can't achieve, theoretically.
We're always so enamored by new and exciting technology that we fail to realize the people in charge are more than happy to completely bury us with it.
Who is better positioned to pilot the LLM than a domain expert?
"Software engineer" as a job title has included a lot of people who write near-zero-code, at least at the higher levels of the career ladder, for years prior to LLMs. People assuming the only, or even primary, function of the job is outputting code reveal a profound lack of understanding of the industry in my opinion. Beyond the first year or two it has been commonly accepted that the code is the easy part of the job.
> has included a lot of people who write near-zero-code, at least at the higher levels of the career ladder
This is something that I would have thought HN readers were pretty familiar with. LLMs can make my code work faster or more prolific, but with 30yoe I spend a fairly significant chunk of my work time doing anything but code.
I'm occasionally reminded the HN's commenting base is much larger than my niche in the industry (VC backed startups + large public tech companies is my background). I had a similar reaction to people thinking Peter Bailis going from CTO at workday to "member of technical staff" at Anthropic was him trading a leadership position for closing jira tickets.
> Why can't LLMs and agents progress further to do this software engineering job better than an actual software engineer?
Because a machine can never take accountability. If a software engineer throughout the entire year has been directing AI with prompts that created weaker systems then that person is on the chopping block, not the AI. Compared to another software engineer who directed prompts to expand the system and generate extra revenue streams.
It's not about whether they make mistakes (they do! although the exact definition of a mistake is nuanced), but whether they can take accountability if the software fails and millions are lost or people die. A large part of the premium paid on software engineers is to take accountability for their work. If a "business person" directs their agent to build some software and takes accountability -- congrats! They are also now a software engineer :)
The lines between a software engineer / business person / product / design and everything else will blur, because AI increases the individual person's leverage. I posit that there will be more 'software engineers' in this new world, but also more product people, more business people, more companies in general.
> It's why I just can't understand the mindset of software engineers who are giddy about this brave new world. There really is nothing special about your expertise that an LLM can't achieve, theoretically.
They’re stupid or they’re already set up for success. The general ideas seems to be generalists are screwed, domain experts will be fine.
In some sense, technology is "not normal" regardless.
If we think of the digitization tech revolution... the changes it made to the economy are hard to describe well, even now.
In the early days, it was going to turn banks from billion dollar businesses to million dollar ones. Universities would be able to eliminate most of their admin. Accounting and finances would be trivialized. Etc.
Earlier tech revolution s were unpredictable too... But at lest retrospectively they made sense.
It's not that clear what the core activities of our economy even are. It's clear at micro level, but as you zoom out it gets blurry.
Why is accountability needed? It's clearly needed in its context... but it's hard to understand how it aggregates.
> How much of the job is a structured set of tasks vs. taking accountability?
More accurately, how many jobs are probabilistically mechanical. That is, how many jobs are really the execution of a serious Bayesian decisions with a strong prior. LLMs are really great at displacing such jobs.
I think the reason AI isn't going to replace CEOs, or anyone in the C suite, is pretty obvious. They see themselves as the company. Everyone else is a resource. AI is here to replace resources, just like investing in a brand new lawn mower. For them, replacing an executive with AI is like saying you're going to marry a broom.
Loved that section about "meat shields". LLMs cannot be held accountable. Someone needs to be involved in decision making, with real stakes if those decisions are bad.
The problem with AI is that it isn't like any previous technology. There may be temporary jobs to fill in the gaps but they won't be careers. The AI will do the process engineering and self optimization. The prompt witchcraft is a good example because today its totally unnecessary and doesn't actually increase performance, and they'll continue to make it easier to direct/steer the models.
We're literally trying to build an intelligence to replace us.
I think that this is an interesting attempt at taxonomy, but it's a bit on the magical thinking end (and I say this as somebody that does a good amount of what's described as the incanter role). It's a combination of the author's previous witchy aesthetic (see his excellent "ing the technical interview" series) and progressive labor politics (which are asymptotically doomed in the current automation push).
The biggest failure of imagination, I think, is the assumption we'd use humans for most (or *any) of these jobs--for example, the work of the haruspex is better left to an LLM that can process the myriad of internal states (this is the mechanical interpretation field).
As an engineer, I'm never more excited about this job.
My implementation speed and bug fixing my typed code to be the bottleneck - now I just think about an implementation and it then exist - As long as I thought about the structure/input/output/testability and logic flow correctly and made sure I included all that information, it just works, nicely, with tests.
Unix philosophy works well with LLM too - you can have software that does one thing well and only one thing well, that fit in their context and do not lead to haphazard behavior.
Now my day essentially revolves around delivering/improving on delivering concentrated engineering thinking, which in my opinion is the pure part about engineer profession itself. I like it quite a lot.
All plausible, but not very transformative. Like imagining that the new jobs enabled for the automobile include automobile maintenance, tire shops, and so on. Traveling nurses, motel operators, military tanks, doordash, suburban life, beer sales at NASCAR, those were all enabled by the car (and its larger sibling the truck).
Still missing are the jobs snd industries enabled by "AI" that are not themselves "AI".
179 comments
I think you can combine 'Incanters' and 'Process Engineers' into one - 'Users'. Jobs that encompass a role that requires accountability will be directing, providing context, and verifying the output of agents, almost like how millions of workers know basic computer skills and Microsoft Office.
In my opinion, how at-risk a job is in the LLM era comes down to:
1: How easy is it to construct RL loops to hillclimb on performance?
2: How easy is it to construct a LLM harness to perform the tasks?
3: How much of the job is a structured set of tasks vs. taking accountability? What's the consequence of a mistake? How much of it comes down to human relationships?
Hence why I've been quite bullish on software engineering (but not coding). You can easy set up 1) and 2) on contrived or sandboxed coding tasks but then 3) expands and dominates the rest of the role.
On Model Trainers -- I'm not so convinced that RLHF puts the professional experts out of work, for a few reasons. Firstly, nearly all human data companies produce data that is somewhat contrived, by definition of having people grade outputs on a contracting platform; plus there's a seemingly unlimited bound on how much data we can harvest in the world. Secondly, as I mentioned before, the bottleneck is both accountability and the ability for the model to find fresh context without error.
> I think you can combine 'Incanters' and 'Process Engineers' into one - 'Users'
I wanted to talk about this more but couldn't quite figure out how to phrase it, so I cut a fair bit: with "incanters" I'm trying to point at a sort of ... intuitive, more informal practitioner knowledge / metis, and contrast it with a more statistically rigorous approach in "statistical/process engineers". I expect a lot of people will fuse the two, but I'm trying to stake out some tentpoles here. Users integrate a continuum of approaches, including individual intuition, folklore, formal and informal texts, scientific papers, and rigorously designed harnesses & in-house experiments. Like farming--there's deep, intuitive knowledge of local climate and landraces, but also big industrial practice, and also research plots, and those different approaches inform (and override) each other in complex ways.
> Hence why I've been quite bullish on software engineering (but not coding). You can easy set up 1) and 2) on contrived or sandboxed coding tasks but then 3) expands and dominates the rest of the role.
Why can't LLMs and agents progress further to do this software engineering job better than an actual software engineer? I've never seen anyone give a satisfactory answer to this. Especially the part about making mistakes. A lot of the defense of LLM shortcomings (i.e., generating crappy code) comes down to "well humans write bad code too." OK? Well, humans make mistakes too. Theoretically, an LLM software engineer will make far fewer than a human. So why should I prefer keeping you in the loop?
It's why I just can't understand the mindset of software engineers who are giddy about the direction things are going. There really is nothing special about your expertise that an LLM can't achieve, theoretically.
We're always so enamored by new and exciting technology that we fail to realize the people in charge are more than happy to completely bury us with it.
"Software engineer" as a job title has included a lot of people who write near-zero-code, at least at the higher levels of the career ladder, for years prior to LLMs. People assuming the only, or even primary, function of the job is outputting code reveal a profound lack of understanding of the industry in my opinion. Beyond the first year or two it has been commonly accepted that the code is the easy part of the job.
> has included a lot of people who write near-zero-code, at least at the higher levels of the career ladder
This is something that I would have thought HN readers were pretty familiar with. LLMs can make my code work faster or more prolific, but with 30yoe I spend a fairly significant chunk of my work time doing anything but code.
> Why can't LLMs and agents progress further to do this software engineering job better than an actual software engineer?
Because a machine can never take accountability. If a software engineer throughout the entire year has been directing AI with prompts that created weaker systems then that person is on the chopping block, not the AI. Compared to another software engineer who directed prompts to expand the system and generate extra revenue streams.
The lines between a software engineer / business person / product / design and everything else will blur, because AI increases the individual person's leverage. I posit that there will be more 'software engineers' in this new world, but also more product people, more business people, more companies in general.
> It's why I just can't understand the mindset of software engineers who are giddy about this brave new world. There really is nothing special about your expertise that an LLM can't achieve, theoretically.
They’re stupid or they’re already set up for success. The general ideas seems to be generalists are screwed, domain experts will be fine.
If we think of the digitization tech revolution... the changes it made to the economy are hard to describe well, even now.
In the early days, it was going to turn banks from billion dollar businesses to million dollar ones. Universities would be able to eliminate most of their admin. Accounting and finances would be trivialized. Etc.
Earlier tech revolution s were unpredictable too... But at lest retrospectively they made sense.
It's not that clear what the core activities of our economy even are. It's clear at micro level, but as you zoom out it gets blurry.
Why is accountability needed? It's clearly needed in its context... but it's hard to understand how it aggregates.
> How much of the job is a structured set of tasks vs. taking accountability?
More accurately, how many jobs are probabilistically mechanical. That is, how many jobs are really the execution of a serious Bayesian decisions with a strong prior. LLMs are really great at displacing such jobs.
We're literally trying to build an intelligence to replace us.
The biggest failure of imagination, I think, is the assumption we'd use humans for most (or *any) of these jobs--for example, the work of the haruspex is better left to an LLM that can process the myriad of internal states (this is the mechanical interpretation field).
My implementation speed and bug fixing my typed code to be the bottleneck - now I just think about an implementation and it then exist - As long as I thought about the structure/input/output/testability and logic flow correctly and made sure I included all that information, it just works, nicely, with tests.
Unix philosophy works well with LLM too - you can have software that does one thing well and only one thing well, that fit in their context and do not lead to haphazard behavior.
Now my day essentially revolves around delivering/improving on delivering concentrated engineering thinking, which in my opinion is the pure part about engineer profession itself. I like it quite a lot.