I have a feeling this is going to end up in the same place where a lot of AI generated projects end up.
There has for sure been a reduction in the amount of effort to create software. But that friction that existed before helped in determining what was WORTH making. There will be a subsection of tools that are now creatable, that are needed. But there'll be a lot of tools that get created that will be abandoned precisely because they didn't solve enough of a problem.
Having a gating mechanism, in the past the friction and effort, was good.
I'm still searching if somebody else made something for tiny tooling before I start throwing something together. Getting something just working is easy nowadays, and makes things that we considered not worth the effort before viable, but making sure it actually is good, doesn't have unintended side effects, and has good documentation (which I'm still typically writing by hand) adds quite a bit of extra effort on top, which still makes sharing the better option than just doing everything yourself for a lot of things.
Can also say the year that subscriptions go down, because now you can finally build your own systems exactly to your liking. Save money and have something that works just for you. LLMs are awesome.
If you're paying a subscription for anything, it's unlikely to be something you'd want to replace for an internal tool. Yeah, you can easily vibe code the basic functionality of a Slack/Jira/Confluence/Ashby/Salesforce replacement in a workday now, but 1) there's no chance you capture all functionality needed by your business on the first pass, you're going to be getting inundated with feature requests as soon as you roll it out, and 2) the ongoing maintenance cost is going to be much higher than expected. Not just maintaining a server and database, which is cheap, but training people how to use your tool instead of the industry standard one. There's a reason Atlassian and Salesforce stocks have recovered after the theorized "SaaSocalypse" which never materialized, it's just not an overall cost savings to build this stuff yourself.
Where internal tools shine is internal operational workflows that are being done manually. You're not replacing Slack or Salesforce, you're replacing the current process where a developer needs to SSH into the server to run a SQL query, or where the Operations team fills out a Google doc which the Finance team pastes into an Excel spreadsheet which runs a bunch of VBA calculations.
We have lived with lackluster B2B and enterprise tools for so long that we have become numb to how bad they are.
Being able to build something that works just how you want with minimal investment is a huge value add.
I ditched web analytics platforms and just send events in jsonl format to disk.
I have a regularly scheduled job that imports it all into a duckdb database and updates a very custom report. I can also ask random questions like “is there a cohort of users who use feature A and B?”.
As long as the phase of the software and users are lighweight - agree these internal tools can be built, but as scale comes that is when the challenges show up is what I have observed.
yep, excellent use case! I did something similar with my gmail recently, because I hate using gmail's UI and I also hate that I don't get access to all my data, but I've begun to regularly backup all my email into a local postgres DB (including attachments). AI has also allowed me to get a full backup of my entire email and archive it onto my own storage. and the best part of this: I might be able to finally get off gmail
good question. maintenance is cheap with tools like astra/opus. also, these aren't production level applications, so I don't need to worry about it being "perfect" (though there's no such thing)
What happens when these companies decide to start increasing the cost/token, exceeding the cost of hiring an engineer or paying for an external service in the first place?
Cost per intelligence-adjusted token has been in free-fall since GPT-3 landed five years ago [0]. I see no compelling reason to think we'll see that trend suddenly stop, much less reverse.
Or they optimize the model for just one more hit, er, prompt. All these AI companies are backed by the same companies optimizing revenue via addiction.
So, same end result (cost exceeds hiring a new engineer), but achieved differently.
The counter, I suppose, is the managers have to deal less with engineers telling them their ideas aren't good or well thought out.
Ah yes, the ye olde "it doesn't need to be perfect" response to serious, legitimate questions.
It's always about cheaper, better, faster. And when the thing being built extends beyond a single person, and still when working alone, conflict shows up in balancing the risks between these. The Strawman of Perfection is just that.
"It doesn't need to be perfect" is incredibly dismissive of the reality of the situation. Ignoring reality seems like a poor problem solving strategy.
No one is ignoring reality. The reality is that my internal tools don't have users besides myself and as long as they do what I want them to do, I'm perfectly happy with them.
Going through this learning process is a critical growth point in every software development career:
1. Software developer is required to do some mundane configuration/setup/data manipulation task at the request of some other internal team
2. Software developer thinks "I'll write an internal tool so the internal team doesn't have to bug me! They can change the configuration/extract the data on their own!"
3. Software developer spends X amount of time building an internal tool (can be quite quickly with AI)
4. Internal tool is released; the internal team looks at it and says, "uh, I don't want to deal with this, I'm just going to keep forwarding requests to the software team."
5. Software developer does the task and realizes it's easier for them to do the mundane configuration/setup/data manipulation with existing tools rather than the internal tool. Internal tool is never used.
These configurations make so much sense when you're writing them. Like on a CLI tool when you pass in a flag, the syntax makes so much sense when you just wrote it.
But a few weeks later when you're trying to use it, you completely forget the syntax and all the little quirks that seemed so genius when you were writing the internal tool.
I think we're going to see a huge number of small internal tools that simply wouldn't have been worth building before coding agents. Even more so in SMB, because they have less budget.
The part I don't think you can leave to the agent is everything around the application: backups, authentication/access control and vulnerability scanning. Those things aren't rocket science, but getting them wrong can have a much bigger impact than a bug in the application itself.
That's what I'm building with https://AppHaven.eu My bet is that vibe-coded apps will become cheap and almost disposable, while the platform underneath them needs to provide those boring safety guarantees automatically.
It's just continuing the process. Most of the internet these days is TikTok, Instagram, Facebook, Reddit, YouTube, etc and those are basically image + video sites at this point.
I have a feeling this is going to end up in the same place where a lot of AI generated projects end up.
There has for sure been a reduction in the amount of effort to create software. But that friction that existed before helped in determining what was WORTH making. There will be a subsection of tools that are now creatable, that are needed. But there'll be a lot of tools that get created that will be abandoned precisely because they didn't solve enough of a problem.
Having a gating mechanism, in the past the friction and effort, was good.
I'm still searching if somebody else made something for tiny tooling before I start throwing something together. Getting something just working is easy nowadays, and makes things that we considered not worth the effort before viable, but making sure it actually is good, doesn't have unintended side effects, and has good documentation (which I'm still typically writing by hand) adds quite a bit of extra effort on top, which still makes sharing the better option than just doing everything yourself for a lot of things.
Can also say the year that subscriptions go down, because now you can finally build your own systems exactly to your liking. Save money and have something that works just for you. LLMs are awesome.
If you're paying a subscription for anything, it's unlikely to be something you'd want to replace for an internal tool. Yeah, you can easily vibe code the basic functionality of a Slack/Jira/Confluence/Ashby/Salesforce replacement in a workday now, but 1) there's no chance you capture all functionality needed by your business on the first pass, you're going to be getting inundated with feature requests as soon as you roll it out, and 2) the ongoing maintenance cost is going to be much higher than expected. Not just maintaining a server and database, which is cheap, but training people how to use your tool instead of the industry standard one. There's a reason Atlassian and Salesforce stocks have recovered after the theorized "SaaSocalypse" which never materialized, it's just not an overall cost savings to build this stuff yourself.
Where internal tools shine is internal operational workflows that are being done manually. You're not replacing Slack or Salesforce, you're replacing the current process where a developer needs to SSH into the server to run a SQL query, or where the Operations team fills out a Google doc which the Finance team pastes into an Excel spreadsheet which runs a bunch of VBA calculations.
Where internal tools shine is internal operational workflows that are being done manually.
This is the key distinction people need to understand, until then it just piles up and lot of companies main business is not software, something else.
Slack is trivial to replace. You don’t need training to know to type your message in the message box.
> Slack is trivial to replace. You don’t need training to know to type your message in the message box.
You're aware that Slack has many more features besides chat, right?
This statement falls purely into the mistake #1 category.
We have lived with lackluster B2B and enterprise tools for so long that we have become numb to how bad they are.
Being able to build something that works just how you want with minimal investment is a huge value add.
I ditched web analytics platforms and just send events in jsonl format to disk.
I have a regularly scheduled job that imports it all into a duckdb database and updates a very custom report. I can also ask random questions like “is there a cohort of users who use feature A and B?”.
It’s super lightweight and incredible.
It’s super lightweight and incredible.
As long as the phase of the software and users are lighweight - agree these internal tools can be built, but as scale comes that is when the challenges show up is what I have observed.
yep, excellent use case! I did something similar with my gmail recently, because I hate using gmail's UI and I also hate that I don't get access to all my data, but I've begun to regularly backup all my email into a local postgres DB (including attachments). AI has also allowed me to get a full backup of my entire email and archive it onto my own storage. and the best part of this: I might be able to finally get off gmail
Can build our own buggy substandard turds instead of customising an off the shelf buggy turd.
How far we have come!
your sarcasm nor your comment is helpful to anyone
Who pays for ongoing maintenance of all those things you built and now own
good question. maintenance is cheap with tools like astra/opus. also, these aren't production level applications, so I don't need to worry about it being "perfect" (though there's no such thing)
What happens when these companies decide to start increasing the cost/token, exceeding the cost of hiring an engineer or paying for an external service in the first place?
Cost per intelligence-adjusted token has been in free-fall since GPT-3 landed five years ago [0]. I see no compelling reason to think we'll see that trend suddenly stop, much less reverse.
0. "The cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or 13× per year." https://epoch.ai/publications/the-plunging-price-of-thought
Yet OpenAI is losing $2 for every $1 they're making :-?
Or they optimize the model for just one more hit, er, prompt. All these AI companies are backed by the same companies optimizing revenue via addiction.
So, same end result (cost exceeds hiring a new engineer), but achieved differently.
The counter, I suppose, is the managers have to deal less with engineers telling them their ideas aren't good or well thought out.
it's certainly possible and has happened before, but that's why I'm hoping for two things:
1.) open weight models gain more popularity 2.) i'm making more money from these models than paying for them
if either happens, then I'll feel good about it either way
No, they would never...
What do you mean? They need to recover hundreds of billions in sunk costs?
Ah yes, the ye olde "it doesn't need to be perfect" response to serious, legitimate questions.
It's always about cheaper, better, faster. And when the thing being built extends beyond a single person, and still when working alone, conflict shows up in balancing the risks between these. The Strawman of Perfection is just that.
"It doesn't need to be perfect" is incredibly dismissive of the reality of the situation. Ignoring reality seems like a poor problem solving strategy.
No one is ignoring reality. The reality is that my internal tools don't have users besides myself and as long as they do what I want them to do, I'm perfectly happy with them.
A lot of internals tools were built where I work, not a lot were used.
Going through this learning process is a critical growth point in every software development career:
1. Software developer is required to do some mundane configuration/setup/data manipulation task at the request of some other internal team
2. Software developer thinks "I'll write an internal tool so the internal team doesn't have to bug me! They can change the configuration/extract the data on their own!"
3. Software developer spends X amount of time building an internal tool (can be quite quickly with AI)
4. Internal tool is released; the internal team looks at it and says, "uh, I don't want to deal with this, I'm just going to keep forwarding requests to the software team."
5. Software developer does the task and realizes it's easier for them to do the mundane configuration/setup/data manipulation with existing tools rather than the internal tool. Internal tool is never used.
These configurations make so much sense when you're writing them. Like on a CLI tool when you pass in a flag, the syntax makes so much sense when you just wrote it.
But a few weeks later when you're trying to use it, you completely forget the syntax and all the little quirks that seemed so genius when you were writing the internal tool.
So you go back to existing tools.
I think we're going to see a huge number of small internal tools that simply wouldn't have been worth building before coding agents. Even more so in SMB, because they have less budget.
The part I don't think you can leave to the agent is everything around the application: backups, authentication/access control and vulnerability scanning. Those things aren't rocket science, but getting them wrong can have a much bigger impact than a bug in the application itself.
That's what I'm building with https://AppHaven.eu My bet is that vibe-coded apps will become cheap and almost disposable, while the platform underneath them needs to provide those boring safety guarantees automatically.
[flagged]
Anyone noticed the trend of tools being more read than write, with writes shifting to agentic tools and natural language instruction?
It makes me wonder about the web a bit and whether it becomes glossier, more video driven but ultimately less directly maleable.
Do people prefer streaming text over static pages?
It's just continuing the process. Most of the internet these days is TikTok, Instagram, Facebook, Reddit, YouTube, etc and those are basically image + video sites at this point.
Interesting—they do detailed analysis and requirements up front. Reminds me to a certain extent of the maligned waterfall model.
I read this as the year of "Infernal" tools.
> Numerous Grill Me sessions, using Matt Pocock's wonderful Grill Me skill, which interrogates a plan until the weak parts fall out.
If you haven’t yet drunk the /grilling Kool Aid, give it a go
I totally agree! I usually combine it with the ask user question tool in Claude Code which makes decision-making a breeze.
... that don't work
[flagged]
[dead]
[flagged]