I feel like you are fundamentally missing the point of generative AI, from the point of view of the companies investing in it. This was never once about the quality of the output, it's about the velocity of the output. The goal has never once been to make content that is good. The goal has been to make content that is good enough and fast enough, to bury human generated content under a mountain of AI generated garbage. Thus devaluing the human labor, to make the market more favorable to the tech giants controlling the direction of the market.
Let me give a simple example. Let's say you are a talented programmer, who happens to have specific knowledge of a niche market. You write a tight, robust, efficient application, that addresses several common pain points of that niche market, in a desktop tool, that runs locally, which you sell at a reasonable price. Traditionally, this is a fantastic recipe for success, for you, the developer.
But that desktop tool doesn't push subscriptions. It doesn't drive AWS or Azure contracts. No VCs get returning revenue from a SAAS model. Meta, Google, Amazon, Microsoft and so on, dont make a penny from it. It is useless to Silicon Valley, from their point of view. But what if every single company in that niche market, instead of buying your tight, robust, efficient application, could instead be sold on a subscription, to have an AI make them a bloated, buggy, staggeringly inefficient, tailored solution to their personal specific problem?
Now we're talking! Data center contracts. Recurring revenue. GPU sales for lots of compute! Sure, it is in every way a worse solution to the same problem, but when everyone in the entire market is making their own bespoke solution, how is anyone ever going to find your tight, efficient, robust application?
That's their perspective. Not that the output of AI is good. It's that the output of AI is good FOR THEM.
Any teacher who’s had to read and grade an AI paper can tell you that it’s awful—generic, boilerplate language, hallucinated quotes and sources, and SHORT. Trying to get AI to stay on task with any train of thought is impossible. If you ask the AI to write a 250 word blurb it can do it; a 500 word response is probably going to repeat itself a bit; and a 1000 word paper will say the exact same argument every paragraph, no variation or increasing depth. There is just no depth to it. It’s wide as an ocean, deep as a puddle, and it’s piss. It’s a puddle of piss.
Just want to add my two cents: AI can't replace a human translator, but at this point nobody is willing to pay for human translation because they never actually cared about quality in the first place. My wife has been a translator for over a decade and that industry is dead as a doornail. Score the point for AI here.
One perspective on this is: if you walk into any sizeable corporate you’ll see almost everyone busy doing busy work.
If we can cut that back to 80% cognitive load we can put creativity back into the company — through humans. The AI is just to automate that 20%
This becomes a win-win.
If this can be done with a solution costing approximate to a cloud-native application of the same complexity. This is a win-win-win.
My general thought (though not sold on this yet) is that text or text to voice is probably the most accessible way to interact with a computer.. I am not an a11y expert..
“AI is on track to write 90% of code within the next 2–5 years.”
I can almost believe.
I am a pretty poor programmer these days because it's no longer a key part of my day job but if I have to I can hack together a python or shells script. But I'm pretty slow and inefficient at it so I have a ton of "nice to have" scripts that I don't have the energy to buckle down and write.
Anyway I read that openai has a new thing https://gpt-oss.com/ and went to see what junk it was and saw that it suggests you ask it something and includes python programs as an option. So I asked it
"write a python program to look up the NS record of a domain or FQDN"
And it did so. Not terribly challenging but the output made sense and it worked and I requested a couple of minor changes that also worked. Both scripts had help text and a bunch of other nice to haves like clear comments that I didn't ask for. The AI took seconds (under 10) to create both.
I then asked it to extend the script to do more complex queries and that took the AI 61 seconds and it clearly had to think a lot, partly because I had a slight error/inexactitude in my spec which it figured out and corrected. The result worked just fine and then after playing a bit I realized I needed to improve it so I told it
"Take the previous program and make it so that it can either take a list of domains on the command line or as a file or stdin. Also modify the output format so that it is an array of objects ... "
That took another 28 seconds and the solution also works just fine and is what I want for now
That took me - including testing the intermediate steps and so on - about half an hour total time.
It's a utility I can and will use for research. Of course I could have written it myself but it would have probably taken a couple of hours to research the recursive tree walk bit and I'd have made dumb mistakes that I'd need to debug.
I am officially impressed and I'll be using it for more such utilities. My guess is that there are several similar fairly simple tasks. Things that I want to do but which are slightly more than a shell one liner and so need a real script to do and which I will find it easier to do by asking some AI to write for me instead of doing myself.
Now I don't know how many people have these annoying irritations of tasks that need a script or a webpage and javascript or similar and which AI can write but I expect there are a lot of us. If it takes me half an hour or less to get some computer to write it then I'm much more likely to create that rather than suffer without a script. I suspect the aggregate numbers of such uses are pretty high though whether that's 90% of all code I'm not sure
Well, different strokes for different folks I guess:
I tried to get all of CoPilot, ChatGPT and Gemini to write a PowerShell script that should take a config file from a common repository, copy it to the machine I was working on, start the app and after I ended the app, copy the (possibly changed) config file back to the repository.
Took the combined might of AI 3 (in words three) days to come up with a version that I then needed to painfully debug ...
Ask it to stitch together a variety of APIs, both custom and 3rd party, to make a sensible, maintainable ecosystem. Be cheekier, ask it to write tests too.
I think the results will make you somewhat revise the 2-5 year estimate.
I'll be honest, I expected it to suck and it didn't.
I'm going to try using it more.
My next attempt will be getting it to make multiple async DNS requests, which if it works will speed the program up by a factor of at least 10 when I give it a large list. I'll report back shortly
I might be wrong but it feels like businesses don’t want to invest in their people and AI is a good way out for them.
I can tell from software industry, a lot of software companies and their products are bloated, as they were running hiring wars and were trying to keep enough engineers to implement new ideas quickly. After pandemic, they realized there is no much benefit in piling up features into their products and trying to scale down on investment into human capital. And AI can look like a good way to mask their investment mistakes and make it look like a strategical turn.
P.S. I have read your next post, and I see we are on the same line of thought
I feel like you are fundamentally missing the point of generative AI, from the point of view of the companies investing in it. This was never once about the quality of the output, it's about the velocity of the output. The goal has never once been to make content that is good. The goal has been to make content that is good enough and fast enough, to bury human generated content under a mountain of AI generated garbage. Thus devaluing the human labor, to make the market more favorable to the tech giants controlling the direction of the market.
Let me give a simple example. Let's say you are a talented programmer, who happens to have specific knowledge of a niche market. You write a tight, robust, efficient application, that addresses several common pain points of that niche market, in a desktop tool, that runs locally, which you sell at a reasonable price. Traditionally, this is a fantastic recipe for success, for you, the developer.
But that desktop tool doesn't push subscriptions. It doesn't drive AWS or Azure contracts. No VCs get returning revenue from a SAAS model. Meta, Google, Amazon, Microsoft and so on, dont make a penny from it. It is useless to Silicon Valley, from their point of view. But what if every single company in that niche market, instead of buying your tight, robust, efficient application, could instead be sold on a subscription, to have an AI make them a bloated, buggy, staggeringly inefficient, tailored solution to their personal specific problem?
Now we're talking! Data center contracts. Recurring revenue. GPU sales for lots of compute! Sure, it is in every way a worse solution to the same problem, but when everyone in the entire market is making their own bespoke solution, how is anyone ever going to find your tight, efficient, robust application?
That's their perspective. Not that the output of AI is good. It's that the output of AI is good FOR THEM.
Yes. This mimics my experience in tech and corporate as well. It’s about delivering good enough faster.
Any teacher who’s had to read and grade an AI paper can tell you that it’s awful—generic, boilerplate language, hallucinated quotes and sources, and SHORT. Trying to get AI to stay on task with any train of thought is impossible. If you ask the AI to write a 250 word blurb it can do it; a 500 word response is probably going to repeat itself a bit; and a 1000 word paper will say the exact same argument every paragraph, no variation or increasing depth. There is just no depth to it. It’s wide as an ocean, deep as a puddle, and it’s piss. It’s a puddle of piss.
Just want to add my two cents: AI can't replace a human translator, but at this point nobody is willing to pay for human translation because they never actually cared about quality in the first place. My wife has been a translator for over a decade and that industry is dead as a doornail. Score the point for AI here.
One perspective on this is: if you walk into any sizeable corporate you’ll see almost everyone busy doing busy work.
If we can cut that back to 80% cognitive load we can put creativity back into the company — through humans. The AI is just to automate that 20%
This becomes a win-win.
If this can be done with a solution costing approximate to a cloud-native application of the same complexity. This is a win-win-win.
My general thought (though not sold on this yet) is that text or text to voice is probably the most accessible way to interact with a computer.. I am not an a11y expert..
Go humans!
I'm with you on the massive over investment (and you might find this FT article relevant - https://ig.ft.com/ai-data-centres/ )
But... this quote:
“AI is on track to write 90% of code within the next 2–5 years.”
I can almost believe.
I am a pretty poor programmer these days because it's no longer a key part of my day job but if I have to I can hack together a python or shells script. But I'm pretty slow and inefficient at it so I have a ton of "nice to have" scripts that I don't have the energy to buckle down and write.
Anyway I read that openai has a new thing https://gpt-oss.com/ and went to see what junk it was and saw that it suggests you ask it something and includes python programs as an option. So I asked it
"write a python program to look up the NS record of a domain or FQDN"
And it did so. Not terribly challenging but the output made sense and it worked and I requested a couple of minor changes that also worked. Both scripts had help text and a bunch of other nice to haves like clear comments that I didn't ask for. The AI took seconds (under 10) to create both.
I then asked it to extend the script to do more complex queries and that took the AI 61 seconds and it clearly had to think a lot, partly because I had a slight error/inexactitude in my spec which it figured out and corrected. The result worked just fine and then after playing a bit I realized I needed to improve it so I told it
"Take the previous program and make it so that it can either take a list of domains on the command line or as a file or stdin. Also modify the output format so that it is an array of objects ... "
That took another 28 seconds and the solution also works just fine and is what I want for now
That took me - including testing the intermediate steps and so on - about half an hour total time.
It's a utility I can and will use for research. Of course I could have written it myself but it would have probably taken a couple of hours to research the recursive tree walk bit and I'd have made dumb mistakes that I'd need to debug.
I am officially impressed and I'll be using it for more such utilities. My guess is that there are several similar fairly simple tasks. Things that I want to do but which are slightly more than a shell one liner and so need a real script to do and which I will find it easier to do by asking some AI to write for me instead of doing myself.
Now I don't know how many people have these annoying irritations of tasks that need a script or a webpage and javascript or similar and which AI can write but I expect there are a lot of us. If it takes me half an hour or less to get some computer to write it then I'm much more likely to create that rather than suffer without a script. I suspect the aggregate numbers of such uses are pretty high though whether that's 90% of all code I'm not sure
Well, different strokes for different folks I guess:
I tried to get all of CoPilot, ChatGPT and Gemini to write a PowerShell script that should take a config file from a common repository, copy it to the machine I was working on, start the app and after I ended the app, copy the (possibly changed) config file back to the repository.
Took the combined might of AI 3 (in words three) days to come up with a version that I then needed to painfully debug ...
My conclusion: thanks, but no thanks!
Ask it to stitch together a variety of APIs, both custom and 3rd party, to make a sensible, maintainable ecosystem. Be cheekier, ask it to write tests too.
I think the results will make you somewhat revise the 2-5 year estimate.
I'll be honest, I expected it to suck and it didn't.
I'm going to try using it more.
My next attempt will be getting it to make multiple async DNS requests, which if it works will speed the program up by a factor of at least 10 when I give it a large list. I'll report back shortly
Mr. Japan-that was my best and hardest laugh of the day. love you Matt.
Love the title
I might be wrong but it feels like businesses don’t want to invest in their people and AI is a good way out for them.
I can tell from software industry, a lot of software companies and their products are bloated, as they were running hiring wars and were trying to keep enough engineers to implement new ideas quickly. After pandemic, they realized there is no much benefit in piling up features into their products and trying to scale down on investment into human capital. And AI can look like a good way to mask their investment mistakes and make it look like a strategical turn.
P.S. I have read your next post, and I see we are on the same line of thought
Excellent piece.
it doesn’t matter. capital wills spend on AI.
everything else — capabilities and use cases (there are some), marketing, attention, disruption, value, value capture follows.
The "AI Con"?
There's already a book about that, making many of the same points about this as you!
"The AI Con: How to Fight Big Tech’s Hype and Create the Future We Want"
by Emily M. Bender and Alex Hanna. Published by Harper.
Here: https://thecon.ai/
Great column - thanks!