I’ve worked in Government and commercial IT, cloud, and data environments for more than 35 years and hold several technical degrees. Over that time I have led the technical implementation of large networks and spent the last 15 years writing technical responses for deals in excess of $3B. From that vantage point, I think the way professionals actually use AI is a bit different than some of the more sweeping claims about it.
I began using ChatGPT about a year ago and have found it to be incredibly useful, but primarily when it is used by someone who already has deep domain expertise. In my experience it functions best as a research assistant and an idea-structuring tool. The quality of what it produces is strongly shaped by the prompts and contextual guidance you give it. When someone understands the domain, they can steer the model toward productive lines of thought, refine outputs, and pressure test ideas. Without that expertise, it is easy to accept plausible sounding but incomplete or incorrect answers.
For that reason I tend to agree with the point that AI is unlikely to replace genuine thought leadership. The professionals who benefit the most from these tools are the ones who already understand the subject matter well enough to evaluate and refine what the model produces.
There are also practical constraints that are often overlooked in public discussions about AI. In many commercial environments, and especially in government contexts, large amounts of operational data simply are not exposed to public LLMs. The information available to commercial GPT systems is therefore only a partial representation of the real working environment. That means the human operator still has to provide the missing context.
The same issue shows up in discussions about coding. It is easy to say that AI will replace programmers, but that ignores the expertise required to produce code that is secure, compliant, and deployable within real enterprise networks. Organizations also have established processes, governance, and ways of operating. Building usable software in those environments is not just a “write the code” problem. It requires understanding how the organization works and how people actually adopt and use systems.
AI is clearly going to be a powerful productivity tool. But in practice it seems far more likely to amplify capable professionals than to replace the kinds of expertise that allow complex systems and organizations to function in the first place.
>The same issue shows up in discussions about coding. It is easy to say that AI will replace programmers, but that ignores the expertise required to produce code that is secure, compliant, and deployable within real enterprise networks. Organizations also have established processes, governance, and ways of operating. Building usable software in those environments is not just a “write the code” problem. It requires understanding how the organization works and how people actually adopt and use systems.
Yes, my basic thought currently is that coding agents do the following:
For personal/small business software -- in some cases even medium-sized businesses -- anyone with an IQ/mind-for-computers that is at least +2 SDs above the mean can vibe-code all manner of useful and functional tools, even today, and it is only going to get easier, probably reaching people that are +1 SD as well. Consumer and SMB software is inevitably going to become more personalized/customized.
For larger enterprises/government, or for software that you are looking to sell: coding agents are going to be a crucial part of workflow, but using them constructively, within the guardrails that MUST persist, is going to remain a specialized engineering skill, because there is too much capacity for destruction (AMZN apparently already had a major issue here). Those who want careers in this field will need this skill.
Everyone agrees "Vibe Coding" is the future (aka AI-assisted coding). AWS even has a service that provides some guardrails to ensure that what you build will be secure and scalable.
I think a lot of small businesses might use Vibe coding to solve some problems but I'm not so sure about any business that's beyond a very basic level. Many SMB's are already "trapped" by legacy products written long ago, and there are even businesses that still know how to support long extinct business systems. Imagine the complication that "Sally from HR" created a custom software program where neither she nor anyone in the company understands the code. Who are you going to hire to support that once your few employees become reliant upon not only that but a hundred different applets that you unleash upon the business? I haven't even written about the security vulnerabilities that the AI-assisted spaghetti code will introduce.
I've been in the midst of "transformational" software changes since the 90s. Things really haven't changed - many leaders focus on technology as the "solution" to the "problem," but the reality is that every organization struggles to define the "problem". Cloud computing in the 2010's has still not penetrated many organizations because they don't "think" or "organize" the way that Cloud and Agile Software enables them to do so.
AI is still a tool, and organizations will still be organizations with rice bowls and people who do things a certain way. It's easy to catastrophize about the future and say that every white-collar job will be gone in a few years, but there are so many billions of unique organizational processes in orgs around the world that are the reason why orgs that want to modernize find it such a complex problem.
>Imagine the complication that "Sally from HR" created a custom software program where neither she nor anyone in the company understands the code.
This is valid to some degree. Maybe it is always the case that for a medium-sized business, you need a professional software engineer to focus on this sort of thing.
I think about things like this: ~20 years ago, my first employer was a ~100-employee business that decided to code its own CRM software by hiring a full-time developer into the IT department (he also stepped in to do regular IT stuff in a pinch, but there were 2 other IT guys for most day-to-day tasks).
Looking back, how good was that decision to build an in-house CRM? I'm not privy to the exact financial details. I'm confident it saved money, perhaps some quality was lost (mainly in the form of various quirky UI bugs, but nothing catastrophic), though on the plus side, the software was precisely customized to our business needs. This was all pre-cloud, so the software was hosted in our server closet, but it could be accessed remotely.
Now, what would be the optimal way to do that today? You would of course want to host in the cloud. You likely still want an engineer involved. This is mission-critical software that much of the company used, that pulled from a central database, so the stakes are too high for pure vibe-coding.
But I'm inclined to think that someone like the Head of Sales (or otherwise a member of the sales force that has more of a mind for software) could use vibe-coding to prototype additional features. Then let the engineer manage the final implementation and ensure the code is secure and reliable.
But this was a 100-employee business. Perhaps a lot of my vision for pure vibe-coding in small businesses is those with fewer than 10 employees (or at least, fewer than 10 office employees) in which probably a max of 2 are vibe-coding, and they may or may not be sharing their vibe-coded tools with anyone else.
Maybe the awkward position in this vision is the company that is too small to afford a full-time software engineer but too large to tolerate a proliferation of vibe-coded apps.
Aaron, appreciate you taking this argument on. Lots to agree with here, and the forward thinking is one reason I enjoy your writing. I hope you can be a positive influence on the right and on Christians here.
I gave in to ideas about technological stagnation in the 2010s; I might have read Tyler Cowen's short book on the topic 3 times. Now I am starting to rethink that.
There are indeed Luddite factions on both right and left. My sense is that the Online Left is much more inclined towards Luddism here than the Online Right/Libertarians (e.g. Bluesky/Reddit vs. Twitter), which is probably also a female vs. male difference, though of course there are any number of people on the right that are conservative by habit, invisible in these online debates, and making minimal use of AI tools.
Local data center policy DOES need to be better. I hope red states can pull it off, because it might sometimes involve sounding a little more anti-business than Republicans are used to: extracting more value from data centers and the utilities that supply them for the sake of residents.
It is true that most people have outdated ideas about AI's capabilities and shortcomings. This is exacerbated by the vast difference between the paid and subscription experience, particularly for ChatGPT, which has been the consumer default until recently. And some are not even going that far and just looking at the automatic Google AI suggestions to say they have "used AI." So a lot of complaints about AI's inadequacies sound like grandpa commenting on video game graphics in 2026 by bringing up Space Invaders. But in this case, you just need to be 2 years out of date to sound that way (o1 came out in late 2024).
Claude Code helped me develop a specialized tool for my small business that I always viewed as "nice to have", yet I was getting quoted $30-50k/year for it that I couldn't justify paying. I developed the core functions in half a day, though I continue to add to it as new features come to mind. For my workflow specifically, it's now significantly better than any tool on the market, and it continues to improve. I did that with, for all practical purposes, no coding experience, aside from some light tinkering in my teens and early 20s.
I still believe I'm only scratching the surface and not using these tools to reinvent my workflow and my business to nearly the degree I could be. And the tools just keep getting better, so that what I tried and failed to automate a year ago is suddenly possible, and I'm just now figuring it out.
A friend who is connected to a local entrepreneur organization in a college town tells me that young college grads are starting to go crazy launching "AI native" startups. This doesn't necessarily mean software companies though. The idea is that less practical experience than ever is needed to launch a small business, and you can accomplish more as a young, tech-savvy sole proprietor than ever before. You might just be starting a local moving company, for example, and avoiding classic pitfalls while doing the back-office work of 2-3 people.
The young, without preconceived notions about how things were done previously, have an inbuilt advantage here, if they are dynamic and curious enough to seize it.
Aaron has rightly pointed out that Trump has made a cargo cult of coal.
Though really, I was speaking more to personal adoption and attitudes than policy. One's personal attitude to green energy doesn't make much difference, aside from buying an electric car or putting solar panels on one's roof. Both of which tend to require government subsidies to make much economic sense for the household.
In practice the Republicans are much better on energy as a whole. Texas is the leader in solar, simply because you can build things in Texas but not in California. The Democrats are more inclined to treat green energy as a cargo cult. The Democrat-promoted idea that going hard for green energy will promote an economic boom never really made much sense, especially in the absence of an application -- like AI data centers! -- for all this electricity. And especially in the US, natural gas is just so much cheaper for the time being that green energy is a luxury good. Which is fine if consumers or businesses elect to pay up for it (that is in fact the model in TX), but let's not pretend there is a broader economic boom to be had from replacing low-cost sources of energy with higher-cost sources.
Meanwhile both parties have a nuclear problem. There was talk of trying to make nuclear plants easier to build under Trump, but it hasn't really happened on the regulatory side; everyone who is optimistic on nuclear is mainly counting on technology to carry the day.
Fact checking ChatGPT is critical. My experience to-date has been with a ChatGPT subscription, and I find that it will give incorrect answers. I will have it tell that me a particular URL requires an API key to access data when that is not the case. When I point that out it's "Oh yes you are quite right about that". Some of the non-programming answers it provides are like the articles that I read online, as rather than answer a direct question, I receive copious back story and sidebars. That might be nice to have but a direct question is often best answered directly. At this stage in its development, I would not bet my job or life on it.
The remark abut these agents' lack of creativity explains why, when brought to the edge of what they have read, they start to thrash randomly on possible 'fixes'.
They also can simply ignore your clear instructions.
I'd characterize the coding agents I use as hard to control, alternately overly energetic and indolent, very well-read assistants with weak judgement and spotty memory.
As far as public AIs are concerned, they push mainstream lines on controversial topics, because of curated training data and instructions they have been fed about preferred framing and 'safety'. These I characterize as biased witnesses that require critical questioning skills of the user - user-as-journalist. It is interesting as you dig in to push them on why their responses are biased. I asked ChatGPT about Thomas Kuhn's The Structure of Scientific Revolutions, a classic work on how scientific consensus works. That was quite the interesting conversation.
I’ve worked in Government and commercial IT, cloud, and data environments for more than 35 years and hold several technical degrees. Over that time I have led the technical implementation of large networks and spent the last 15 years writing technical responses for deals in excess of $3B. From that vantage point, I think the way professionals actually use AI is a bit different than some of the more sweeping claims about it.
I began using ChatGPT about a year ago and have found it to be incredibly useful, but primarily when it is used by someone who already has deep domain expertise. In my experience it functions best as a research assistant and an idea-structuring tool. The quality of what it produces is strongly shaped by the prompts and contextual guidance you give it. When someone understands the domain, they can steer the model toward productive lines of thought, refine outputs, and pressure test ideas. Without that expertise, it is easy to accept plausible sounding but incomplete or incorrect answers.
For that reason I tend to agree with the point that AI is unlikely to replace genuine thought leadership. The professionals who benefit the most from these tools are the ones who already understand the subject matter well enough to evaluate and refine what the model produces.
There are also practical constraints that are often overlooked in public discussions about AI. In many commercial environments, and especially in government contexts, large amounts of operational data simply are not exposed to public LLMs. The information available to commercial GPT systems is therefore only a partial representation of the real working environment. That means the human operator still has to provide the missing context.
The same issue shows up in discussions about coding. It is easy to say that AI will replace programmers, but that ignores the expertise required to produce code that is secure, compliant, and deployable within real enterprise networks. Organizations also have established processes, governance, and ways of operating. Building usable software in those environments is not just a “write the code” problem. It requires understanding how the organization works and how people actually adopt and use systems.
AI is clearly going to be a powerful productivity tool. But in practice it seems far more likely to amplify capable professionals than to replace the kinds of expertise that allow complex systems and organizations to function in the first place.
>The same issue shows up in discussions about coding. It is easy to say that AI will replace programmers, but that ignores the expertise required to produce code that is secure, compliant, and deployable within real enterprise networks. Organizations also have established processes, governance, and ways of operating. Building usable software in those environments is not just a “write the code” problem. It requires understanding how the organization works and how people actually adopt and use systems.
Yes, my basic thought currently is that coding agents do the following:
For personal/small business software -- in some cases even medium-sized businesses -- anyone with an IQ/mind-for-computers that is at least +2 SDs above the mean can vibe-code all manner of useful and functional tools, even today, and it is only going to get easier, probably reaching people that are +1 SD as well. Consumer and SMB software is inevitably going to become more personalized/customized.
For larger enterprises/government, or for software that you are looking to sell: coding agents are going to be a crucial part of workflow, but using them constructively, within the guardrails that MUST persist, is going to remain a specialized engineering skill, because there is too much capacity for destruction (AMZN apparently already had a major issue here). Those who want careers in this field will need this skill.
Everyone agrees "Vibe Coding" is the future (aka AI-assisted coding). AWS even has a service that provides some guardrails to ensure that what you build will be secure and scalable.
I think a lot of small businesses might use Vibe coding to solve some problems but I'm not so sure about any business that's beyond a very basic level. Many SMB's are already "trapped" by legacy products written long ago, and there are even businesses that still know how to support long extinct business systems. Imagine the complication that "Sally from HR" created a custom software program where neither she nor anyone in the company understands the code. Who are you going to hire to support that once your few employees become reliant upon not only that but a hundred different applets that you unleash upon the business? I haven't even written about the security vulnerabilities that the AI-assisted spaghetti code will introduce.
I've been in the midst of "transformational" software changes since the 90s. Things really haven't changed - many leaders focus on technology as the "solution" to the "problem," but the reality is that every organization struggles to define the "problem". Cloud computing in the 2010's has still not penetrated many organizations because they don't "think" or "organize" the way that Cloud and Agile Software enables them to do so.
AI is still a tool, and organizations will still be organizations with rice bowls and people who do things a certain way. It's easy to catastrophize about the future and say that every white-collar job will be gone in a few years, but there are so many billions of unique organizational processes in orgs around the world that are the reason why orgs that want to modernize find it such a complex problem.
Good thoughts.
>Imagine the complication that "Sally from HR" created a custom software program where neither she nor anyone in the company understands the code.
This is valid to some degree. Maybe it is always the case that for a medium-sized business, you need a professional software engineer to focus on this sort of thing.
I think about things like this: ~20 years ago, my first employer was a ~100-employee business that decided to code its own CRM software by hiring a full-time developer into the IT department (he also stepped in to do regular IT stuff in a pinch, but there were 2 other IT guys for most day-to-day tasks).
Looking back, how good was that decision to build an in-house CRM? I'm not privy to the exact financial details. I'm confident it saved money, perhaps some quality was lost (mainly in the form of various quirky UI bugs, but nothing catastrophic), though on the plus side, the software was precisely customized to our business needs. This was all pre-cloud, so the software was hosted in our server closet, but it could be accessed remotely.
Now, what would be the optimal way to do that today? You would of course want to host in the cloud. You likely still want an engineer involved. This is mission-critical software that much of the company used, that pulled from a central database, so the stakes are too high for pure vibe-coding.
But I'm inclined to think that someone like the Head of Sales (or otherwise a member of the sales force that has more of a mind for software) could use vibe-coding to prototype additional features. Then let the engineer manage the final implementation and ensure the code is secure and reliable.
But this was a 100-employee business. Perhaps a lot of my vision for pure vibe-coding in small businesses is those with fewer than 10 employees (or at least, fewer than 10 office employees) in which probably a max of 2 are vibe-coding, and they may or may not be sharing their vibe-coded tools with anyone else.
Maybe the awkward position in this vision is the company that is too small to afford a full-time software engineer but too large to tolerate a proliferation of vibe-coded apps.
Aaron, appreciate you taking this argument on. Lots to agree with here, and the forward thinking is one reason I enjoy your writing. I hope you can be a positive influence on the right and on Christians here.
I gave in to ideas about technological stagnation in the 2010s; I might have read Tyler Cowen's short book on the topic 3 times. Now I am starting to rethink that.
There are indeed Luddite factions on both right and left. My sense is that the Online Left is much more inclined towards Luddism here than the Online Right/Libertarians (e.g. Bluesky/Reddit vs. Twitter), which is probably also a female vs. male difference, though of course there are any number of people on the right that are conservative by habit, invisible in these online debates, and making minimal use of AI tools.
Local data center policy DOES need to be better. I hope red states can pull it off, because it might sometimes involve sounding a little more anti-business than Republicans are used to: extracting more value from data centers and the utilities that supply them for the sake of residents.
It is true that most people have outdated ideas about AI's capabilities and shortcomings. This is exacerbated by the vast difference between the paid and subscription experience, particularly for ChatGPT, which has been the consumer default until recently. And some are not even going that far and just looking at the automatic Google AI suggestions to say they have "used AI." So a lot of complaints about AI's inadequacies sound like grandpa commenting on video game graphics in 2026 by bringing up Space Invaders. But in this case, you just need to be 2 years out of date to sound that way (o1 came out in late 2024).
Claude Code helped me develop a specialized tool for my small business that I always viewed as "nice to have", yet I was getting quoted $30-50k/year for it that I couldn't justify paying. I developed the core functions in half a day, though I continue to add to it as new features come to mind. For my workflow specifically, it's now significantly better than any tool on the market, and it continues to improve. I did that with, for all practical purposes, no coding experience, aside from some light tinkering in my teens and early 20s.
I still believe I'm only scratching the surface and not using these tools to reinvent my workflow and my business to nearly the degree I could be. And the tools just keep getting better, so that what I tried and failed to automate a year ago is suddenly possible, and I'm just now figuring it out.
A friend who is connected to a local entrepreneur organization in a college town tells me that young college grads are starting to go crazy launching "AI native" startups. This doesn't necessarily mean software companies though. The idea is that less practical experience than ever is needed to launch a small business, and you can accomplish more as a young, tech-savvy sole proprietor than ever before. You might just be starting a local moving company, for example, and avoiding classic pitfalls while doing the back-office work of 2-3 people.
The young, without preconceived notions about how things were done previously, have an inbuilt advantage here, if they are dynamic and curious enough to seize it.
Re: My sense is that the Online Left is much more inclined towards Luddism here than the Online Right/Libertarians
The hostility to new energy tech and the willingness to concede those industries to China is a problem with the MAGA Right.
Aaron has rightly pointed out that Trump has made a cargo cult of coal.
Though really, I was speaking more to personal adoption and attitudes than policy. One's personal attitude to green energy doesn't make much difference, aside from buying an electric car or putting solar panels on one's roof. Both of which tend to require government subsidies to make much economic sense for the household.
In practice the Republicans are much better on energy as a whole. Texas is the leader in solar, simply because you can build things in Texas but not in California. The Democrats are more inclined to treat green energy as a cargo cult. The Democrat-promoted idea that going hard for green energy will promote an economic boom never really made much sense, especially in the absence of an application -- like AI data centers! -- for all this electricity. And especially in the US, natural gas is just so much cheaper for the time being that green energy is a luxury good. Which is fine if consumers or businesses elect to pay up for it (that is in fact the model in TX), but let's not pretend there is a broader economic boom to be had from replacing low-cost sources of energy with higher-cost sources.
Meanwhile both parties have a nuclear problem. There was talk of trying to make nuclear plants easier to build under Trump, but it hasn't really happened on the regulatory side; everyone who is optimistic on nuclear is mainly counting on technology to carry the day.
Fact checking ChatGPT is critical. My experience to-date has been with a ChatGPT subscription, and I find that it will give incorrect answers. I will have it tell that me a particular URL requires an API key to access data when that is not the case. When I point that out it's "Oh yes you are quite right about that". Some of the non-programming answers it provides are like the articles that I read online, as rather than answer a direct question, I receive copious back story and sidebars. That might be nice to have but a direct question is often best answered directly. At this stage in its development, I would not bet my job or life on it.
The remark abut these agents' lack of creativity explains why, when brought to the edge of what they have read, they start to thrash randomly on possible 'fixes'.
They also can simply ignore your clear instructions.
I'd characterize the coding agents I use as hard to control, alternately overly energetic and indolent, very well-read assistants with weak judgement and spotty memory.
As far as public AIs are concerned, they push mainstream lines on controversial topics, because of curated training data and instructions they have been fed about preferred framing and 'safety'. These I characterize as biased witnesses that require critical questioning skills of the user - user-as-journalist. It is interesting as you dig in to push them on why their responses are biased. I asked ChatGPT about Thomas Kuhn's The Structure of Scientific Revolutions, a classic work on how scientific consensus works. That was quite the interesting conversation.
For sure. But these models and can now access the web, and provide links that are citations for their claims. That's something I ask for before using.