The author of this blog post is the "CEO" of a company that sells AI products. He also invests in multiple AI companies: https://xcancel.com/mattshumer_ https://shumer.dev/about
Exemplary
The author of this blog post is the "CEO" of a company that sells AI products. He also invests in multiple AI companies:
These blog posts never make sense to me. I think generative AI is an incredible technology; even understanding the basic principles, the result blows my mind. Yet it’s certainly not good enough to...
These blog posts never make sense to me. I think generative AI is an incredible technology; even understanding the basic principles, the result blows my mind. Yet it’s certainly not good enough to replace me, yet.
If it’s as good as they claim, proponents should not be building more AI tools, or even more software. Or writing blog posts. They should be branching out into other fields! Actually building companies that have opex that is orders of magnitude cheaper and taking business from slow-to-move industry titans.
I’ll wait till an AI-developed product shows up that costs pennies on the dollar.
Yeah. I touched on this in another post, but their actions don't make sense in the context the world they're describing. If software development is a solved task by AI, and human coders are...
If it’s as good as they claim, proponents should not be building more AI tools, or even more software.
Yeah. I touched on this in another post, but their actions don't make sense in the context the world they're describing. If software development is a solved task by AI, and human coders are obsolete...
Why does everyone that writes these articles own a software development company? They make their money doing something that they're trying very hard to convince people has no value.
If LLMs can replace software developers, and your business's main value is your software, LLM companies can straight up replace your business. Why make shovels for digging companies when you can...
If LLMs can replace software developers, and your business's main value is your software, LLM companies can straight up replace your business.
Why make shovels for digging companies when you can cut out the middlemen and be the one who owns all of the shovels?
Actually useful AI would be an existential threat to a great many businesses, so it's completely asinine for them to do anything but obstruct and fight LLMs. But if you only care about the next quarter...
I mean, I’m doing that. Lots of people are. Those “slow-to-move titans” are among the most enthusiastic adopters. Adopting AI tools isn’t going to catapult you ahead of them, it’s becoming...
They should be branching out into other fields!
I mean, I’m doing that. Lots of people are.
taking business from slow-to-move industry titans.
Those “slow-to-move titans” are among the most enthusiastic adopters. Adopting AI tools isn’t going to catapult you ahead of them, it’s becoming necessary just to keep pace.
Here’s the thing though. It’s extremely fashionable to hate on AI. I would never announce my product as being built by AI — it’s just inviting people to tear it down. That’s why I think it’s mostly people who are publicly and inescapably entwined with AI talking about it.
Everybody else is just saying: hey, here is my new product.
Here's what I don't get. If the AI revolution is truly here, and human work in the area of software development is fundementaly already obsolete, and AI is now so capable that it can effectively...
Here's what I don't get.
If the AI revolution is truly here, and human work in the area of software development is fundementaly already obsolete, and AI is now so capable that it can effectively replace most human tasks then...
Why the FUCK would I pay Matt Schumer, the author of this article, 16 dollars a month for his AI writing assistant https://www.hyperwriteai.com/.
He just spent many pages breathlessly telling me that AI is more capable at complex tasks than humans. Why would I pay a human money that he then just skims off of and spends on openAI or anthropic tokens without providing me any value?
There are only two options I can see here are:
Matt is a liar. He's blatantly inflating the capabilities of these tools that he somehow has unique insight into in order to generate hype like legions of liars before him. Maybe he doesn't know he's lying and has actually convinced himself of what he's saying while also maintaining the cognitive dissonance required to also believe that his product is valuable, or
Matt is a grifter. He's selling a product that he very well knows is less capable than an GPT or Claude subscription, and which can do everything his product can do, and more, far better, and that the labor he's put into his product is just straight up inferior to what these tools can do natively.
There's no other options. He's painted himself into a corner.
That's what I don't get about these AI hype beasts. They're fundementaly arguing for their own lack of value. If they're to be believed, the idea of an "ai startup" is laughable. The only companies worth giving money to are the ones large enough to hire the world's best AI researchers and run the worlds most powerful compute clusters. So openAI, Google, anthropic, and meta. Anyone else with an AI startup, in their hype is to be believed, is just a parasite, grifting people who don't know better.
I don't believe that's actually true though. I just think they're liars.
There's a third realm, that's often ignored by engineering types because engineers like to roll up their sleeves and do things themselves. The third realm is for folks who read this and go, wow...
There's a third realm, that's often ignored by engineering types because engineers like to roll up their sleeves and do things themselves. The third realm is for folks who read this and go, wow gee golly whiz this sounds amazing I want to try but I am not going to roll up my sleeves, can I pay you to do it for me.
He's selling a product that he very well knows is less capable than an GPT or Claude subscription, and which can do everything his product can do, and more, far better, and that the labor he's put into his product is just straight up inferior to what these tools can do natively
Yes absolutely, but the third realm sells this sincerely, as a feature. People are buying his insights and talents and experience, and for him to solve problems and do upgrades and glean the new updates from those other powerhouses.
It's like this. When my family came to Canada, we bought our first car through a friend of a friend of a friend. Did we make that decision because it was the best price or most reliable or best featured or easiest to maintain whatever? Absolutely not. We bought it from the guy on one quality alone: trust. We didn't know East from West, didn't speak English, have never ever owned a car before and know 0 people who do, don't know the first thing about insurance and have no means of getting the car to our door, and this guy made it happen for us. Now, it happened to be a '94 grey Corolla so he didnt do us dirty, but that's not something we could have known at the time.
I see the same thing with Truffle Hunting Tours or Fishing Experience or whatever: selling something that has "no" value to folks who know, because most folks don't know. Most people are scared and want their hand held, not dirtied.
In real life, yes, that third realm exists, but in the context of the world the article that he posted paints, it doesn't. He makes the case that no knowledge work is safe from AI. "Rolling your...
In real life, yes, that third realm exists, but in the context of the world the article that he posted paints, it doesn't.
He makes the case that no knowledge work is safe from AI. "Rolling your sleeves up" and actually setting up the infrastructure for these agents to produce software is knowledge work.
It's not as if Matt is physically installing server racks and building the data centers this stuff runs on by hand. He's not a carpenter or plumber or something, the few areas that he says are safe for now because they require robots to replicate. He's sitting in front of a computer doing the exact type of work he's trying to convince us that a computer can do better.
The fact that I can't open up Claude, say "become a writing assistant", and it automatically designs, codes, installs, and integrates a product with superior functionality to Matt's is why I say he's a liar, rather than a grifter. If AI has truly made knowledge work obsolete, then the people doing knowledge work wouldn't be the ones constantly trying to convince us of that.
I numberized most of the "call to action" because it really nails down the eerie feeling I have here. This feels like an MLM scheme. buy into this, go all in, and anyone dismissing this is wrong....
Exemplary
What you should actually do
Sign up for the paid version of Claude or ChatGPT.
push it into your actual work... don't just use it for quick research questions. Give it an entire contract and ask it to draft a counterproposal
Learn these tools. Get proficient. Demonstrate what's possible. If you're early enough, this is how you move up
Have no ego about it . The people who will struggle most are the ones who refuse to engage: the ones who dismiss it as a fad...It's not.
Get your financial house in order.
Think about where you stand, and lean into what's hardest to replace.
I numberized most of the "call to action" because it really nails down the eerie feeling I have here. This feels like an MLM scheme. buy into this, go all in, and anyone dismissing this is wrong. Heck, one of the bullets after I cut off my quote above "Your dreams just got a lot closer. "
I'm in tech and I'm sure anyone who knows my handle here knows I'm pretty anti-AI (spoiler: I work in games. Kind of a mess for many reasons right now). But I do want to try and give fair shares and see what's out there, what's being done, and how and where I can potentially utilize the eventual, ethical means of new tech.
I didn't quite see this here, just another iteration of "Hey [new version] really works this time!". And maybe it does, but I also know my industry. There's terabytes of web source code to train on. 99.9999% of games are not free to consume in the same way. So any accomplishments that seem like magic in web and mobile tend to fall completely flat for games programming.
And let's not even get started on generative art as of now ("now" being late 2025). People (i.e. the stocks) panic'd over Genie a few weeks back, but it's the exact same effect as any other generative art. You look at it for a minute and think "ooh that's cool". And then the longer you engage, the more drastically the illusion falls off and you remember that people want to try to sell this to you for $60-70, instead of it instead being a neat free tech demo. I don't see extended workflows in this making game development any easier than the old pipelines as of yet.
TL;DR: I am anti-AI as of now in several angles, but I still want to be open about it from a purely technical POV. I think my most generous interpretation of stuff like this is that these pieces vastly overestimate how wide reaching these LLM's can be. I can definitely see disruption to certain subsects of industry, so I may take some warning about this if I was a web dev or any similar job managing CRUD style applications. Because I can see it being good enough for "I need a basic website/app with little performance concerns" (which, if we're being real: is many websites/apps. There's a lot of mediocrity in these domains people are used to putting up with).
But that doesn't mean every programmer in every field is in danger. My field isn't immune per se, but any field where code isn't the (only) hard part is going to resist much more. I hope those more optimistic than me can at least meet me here.
Yes, I think the article generalizes from the author's own experience too much. I can vouch for coding agents being a big deal for building web apps. I'm confident that they would work fine for...
Yes, I think the article generalizes from the author's own experience too much. I can vouch for coding agents being a big deal for building web apps. I'm confident that they would work fine for building forum software like Tildes. I'm doubtful that I will ever need to write code by hand again for the kind of programming I do.
They might not work as well for building every kind of software, let alone for people who aren't software developers.
On the other hand, speculating about the future, people are definitely going to try to make it work for all sorts of software development and for other fields. Depending on the field, maybe there's that not that much of a "moat?"
Since the author is talking about coding as the killer use case that proves all the future use cases are coming... I want to add a sanity check from the perspective of someone with decades of...
Exemplary
Since the author is talking about coding as the killer use case that proves all the future use cases are coming... I want to add a sanity check from the perspective of someone with decades of software engineering experience and as much experience with modern LLM agents as anyone has at this point.
But first I want to acknowledge that he's right about a lot of what he's saying. These tools are more powerful than most people realize at this point. They absolutely are going to change everything on a scale not seen since the widespread consumer internet. And it's going to happen faster than the internet did. It's going to happen too fast.
That said, here's how you know you're reading hype: He never mentions that these tools are also drooling idiots. Maybe he really doesn't know. It's hard to imagine how that could be true, but I want to allow for the possibility that he really believes everything he's saying.
What I mean by that is that this author, and so many others before him, seem to be skipping over big chunks of the current reality and leaping forward into what might happen in the future. The truth is that, for coding, AI agents are miraculous. He's right about that. And also, they absolutely cannot autonomously create complex production level code to professional human standards. They just can't.
However, they appear to. The SOTA is in this odd place where agents can write large, fully functioning, applications that meet most of the specs and pass all of the tests. Which is mind blowing, groundbreaking, science fiction level stuff. While at the same time under the hood there are security flaws, bad patterns, wildly varied conventions and style, performance problems, redundancy, insane verbosity and so on. And the only thing that can fix those issues (or stop them from happening in the first place) is a human in the loop.
So on the surface it looks like a miracle, but underneath it's a mountain of tech debt and vulnerabilities just waiting for the right moment to fuck up everyone's day.
I feel like I should establish that what I'm talking about generalizes, as opposed to being the result of my not understanding how to use the tools. I've been using them extensively for quite a while now (in AI years). I have scaffolding and custom built tools and extensive initial context and skills and commands and hooks and custom sub agents and all the things. Each of them iterated and pruned and updated for the latest generation countless times in an attempt to make the agents more reliable and less idiotic. And it works, some of the scaffolding I came up with in 2025 is now built into the latest SOTA harnesses. I don't say that to paint myself as some sort of visionary, this is all new territory that everyone is figuring out together, a lot of people have organically converged around various obviously effective strategies that the frontier labs then adopted. My point is only that I'm holding them right. I can get coding agents to do all sorts of exciting and useful things and I believe I have a solid, realistic understanding of what they're capable of and what their limits are. With humans in the loop they redefine software engineering. Without humans in the loop they are just very very impressive tech demos.
That could all change, they could get to the point where there don't need to be humans involved. If that happens then everything the author is saying is true and he's maybe not even stating it strongly enough. But it hasn't happened yet. The people who are saying it has are either deluding themselves or exaggerating for cynical reasons. I expect the fallout of that delusion to be difficult to miss in the software industry in the coming months and years.
I can prove it to you (here you can TLDR to the end if you don't care about using coding agents)
Assuming you have a subscription with one of the SOTA companies (for coding you want Claude Code with Opus 4.6 or Codex 5.3 high) that covers the necessary tokens.
First you'll want a decent AGENTS.md or CLAUDE.md for initial context. You can find decent starter context online if you don't want to spend too much time. Pick something reasonably lean, you don't want to use up too much context out of the gate. We can skip all of the more in-depth stuff for now.
Next, give the agent a general spec for a non-trivial application that has a lot of user facing surface area. The more varied the surface area the better. It should ideally be a big enough application that the agent can't one shot it in a single context window. With current context limits that isn't too hard to do (unless you're paying a premium for an extra large context window). It should attempt to solve a problem that's not completely overdone (no glorified to do list apps).
Next have the agent work your prompt up into a detailed implementation plan and have it write that plan to an .md file. If you have the time ask it to run a Q&A session with you to refine the plan.
Then instruct it to implement the plan while keeping track of its progress. This is a key step because you'll need to feed the plan and current progress into a new session when your agent runs out of context, or you can have the agent hand off to a new version of itself automatically, or let it do context compaction and soldier on in the same session. Or if you have a really big subsciption you can have an orchestrator agent run a bunch of sub agents automatically until the plan is finished. There are various ways to do it, each with pros and cons. Make sure the plan it writes includes a detailed testing phase so that it can iterate on any issues until it has something that works. You'll want to have some sort of browser (or device) automation wired up so it can test the UI/UX along with the backend. That's easy to do these days, the providers have solutions already built, or you can ask the agent to do it for you.
Then, assuming you've given it sufficient permissions so that it doesn't need to check in with you, go do something else for a while. Sleeping is a great option.
When you wake there's a fair chance (but not whatsoever guaranteed) that you'll be waking up to a working application that looks quite a lot like what you asked for. If it's your first attempt you're welcome to take all the time you need to wait for the world to stop spinning.
If it app isn't working yet, you should be able to prompt the agent into getting it there fairly easily, but it depends on how hard the set of problems you're trying to solve are.
Once it's working it will be hard to deny that you just experienced some version of the future.
But now the next step is to ask another model to audit the codebase. For example, if you built it with Opus, ask Codex to take a look. It shouldn't cost more than about $5 in tokens for a thorough audit, a lot less if the codebase isn't too big. At the same time, start a fresh session with your main model and ask it to do an audit too. Have both agents write their findings to a file when they're done.
I guarantee the list of issues they find will be extensive and that it will reframe your perspective on the miracle you just experienced. But you're not quite done, instruct your main agent to fix all of the issues and then repeat the audit process. Prepare for another long (but shorter) list of issues. Keep repeating until the agents stop finding issues. Note that the audit prompt is important, it needs to be thorough. You can download pre-made skills for that if you're not a coder. Multiple specialized auditors with different disciplines works best (security, logic, maintainability, etc.)
Once the agents are satisfied that the codebase is perfect, take a look at the codebase yourself. Or if you can't fluently read code, bribe someone who can. If you are really doing a best effort code review, I absolutely guarantee you will find more issues, some of them shocking.
And that's doing the bare minimum to wrangle the agents, my overlong post could be 8 times as long with instructions on how to make the agents suck less and still at the end of the process you would be finding serious issues.
That's the (real) current state of the art in autonomous coding agents and no amount of promot engineering can navigate around it.
A human in the loop, on the other hand, makes for a very different outcome, that is until you get overconfident and let the agent write too much code without thorough review. Then, again, issues are guaranteed.
TLDR
All of this to say: It's still safe to ignore the hype from people like the post author. The AI apocalypse could come at any time, but it's not on the horizon yet based on the current state of the tech.
And also, listen to the more level headed people who are saying this is a paradigm shift, because they are not lying.
Uf... I have feeling that I read exactly the same article in the last two years like dozens of time. Founder of AI startup: "AI finally can replace developer. Just use the latest model. It finally...
Uf... I have feeling that I read exactly the same article in the last two years like dozens of time.
Founder of AI startup: "AI finally can replace developer. Just use the latest model. It finally good.".
And for me personally AI constantly fail to do the half of the simple tasks. Sometime it works, most of the time I need to verify/fix every point of task. Personally for me it would be simpler and maybe quicker just to do a task manually.
Uf... Maybe I'm using it wrong, but what bothers me is that: why every AI startup is trying to sell AI product instead of just doing development themselves using their own AI product?
Upd: to explain a bit, Im not developing some new small app using AI, I'm trying to make AI help me with development/fixes of quite large existing codebase.
I work at a large company where we are given access to latest AI tools and encouraged (but not at all forced) to use them. I've got copilot in my VS Code, autocompletion, agent mode, all the bells...
I work at a large company where we are given access to latest AI tools and encouraged (but not at all forced) to use them. I've got copilot in my VS Code, autocompletion, agent mode, all the bells and whistles. The article appears way closer to reality than your or any other dismissive posts on tildes.
At first it was a funny joke, the local AI pusher was laughed at behind their back and the mistakes and "hallucinations" were a hilarious part of the weekly meetings.
At some point some engineers (including real smart senior people) started mentioning how they used AI to solve X. Someone told us about debugging some production blocker for days, then out of curiosity they asked AI and it suggested the root cause on the first try. The doc team got laid off. QA is being pushed towards as much automation as possible. Basically everybody uses it now I think.
With 5.2 I've been finally using it myself. Yes, the "large existing codebase" is a major point of struggle. It offers a solution, I have to rewrite it because I know the specifics of our crappy legacy code way better, but I am the only one who does and it does offer a working solution. It also offers ideas I wouldn't have come up with on my own.
So yeah I think I trust the author of the article more than the naysayers here. For the record, I don't have any AI to sell. Sadly I am not good at running a business.
It's funny, I have seen other companies do this with worse results. You are right that good engineers will use LLMs to aid in their work. Use them as selective tools where they are actually an...
It's funny, I have seen other companies do this with worse results. You are right that good engineers will use LLMs to aid in their work. Use them as selective tools where they are actually an asset.
But that is not how most people end up using them, at least not in my experience. I have written about this multiple times and am on my phone right now. So forgive me for just linking to a previous comment.
Something has indeed been changing. I have seen an alarming increase of lazy non critical use of LLM tools by people who should know better. Code spanning dozens of line trying to solve something that should only take one line. Code that completely ignores and conventions or design paradigms put in place. Code that goes directly against security practices. Suddenly downgraded dependency versions (because the models training data doesn't include the latest version).
I have no doubt that these models, when used properly are very useful and skill multipliers. But a worryingly large percentage of their users are not even trying to use them responsibly.
All of this causes extra work at the very least. PRs now need twice as much attention and me and other colleagues are on constant alert for people vibe coasting.
In the open source world you see the same thing. There is a steady stream of project blogs talking about the stream of AI generated bullshit PRs. Including one agent even going as far as writing a hit piece.
This is what's most terrifying to me. The doc team sits so close to your code base, and obviously understand your company's tech and industry so intimately, that if AI was useful for humanity and...
The doc team got laid off
This is what's most terrifying to me. The doc team sits so close to your code base, and obviously understand your company's tech and industry so intimately, that if AI was useful for humanity and really does boost productivity and make everyone a star coder, wouldn't it make more sense to retain the doc team and retrain them (with AI) into production staff?
I'm not trying to be AI naysayer, I just mean that this tool is now good enough to rip up the fabric of our society without being good enough to stitch it back together.
No, because training takes time away from already strained seniors. We don't hire juniors for the same reason. The doc team being laid off in itself is not an issue, it was a team of non-native...
wouldn't it make more sense to retain the doc team and retrain them (with AI) into production staff
No, because training takes time away from already strained seniors. We don't hire juniors for the same reason. The doc team being laid off in itself is not an issue, it was a team of non-native speakers and we had a reputation for pretty crappy documentation.
The doc team sits so close to your code base, and obviously understand your company's tech and industry so intimately
Not sure why you think so. They needed engineering to explain everything to them and then rewrote it according to some formatting standards they had. It's really unsurprising that they got replaced with AI. There are a ton of people in large companies that are frankly pretty useless, AI or not AI.
By hiring people who got experienced in other companies, the same way many of our past colleagues work for our clients. We are not running a business of educating juniors (so they can leave the...
By hiring people who got experienced in other companies, the same way many of our past colleagues work for our clients. We are not running a business of educating juniors (so they can leave the company afterwards), similarly how not every hospital is a teaching hospital. I've already got my hands full with tasks, I simply do not have the time to educate someone completely inexperienced. And juniors aren't entitled to start their journey at such a large company either.
Alright, so they have laid off the documentation department. The company isn't hiring junior engineers. It does hire senior engineers, but if I am reading this correctly. Only to the point that...
Alright, so they have laid off the documentation department. The company isn't hiring junior engineers. It does hire senior engineers, but if I am reading this correctly.
I've already got my hands full with tasks, I simply do not have the time to educate someone completely inexperienced.
Only to the point that the work gets done and absolutely nothing more. You are then encourage to use AI tooling. Which has taken a while to get to the level that you, a senior engineer, can use it as a tool with your knowledge and experience in mind.
To be clear, I have no doubt that there is a net benefit to senior engineers using AI tooling. But, even if your company already didn't hire juniors before the AI hype, more companies seem to consider this an option.
Juniors that are still hired are also facing an uphill battle. They are more than likely "encouraged" to use AI tooling as well. Without the knowledge and experience of senior engineers. Which means they are getting exposed to less and less of the technology are by extension are actually less equipped to critically review LLM output.
There is a possibility that these models make a magical leap where they suddenly don't struggle with large code bases, can actually do stakeholder management, design and implement things without a experienced engineer needed in the loop. I'd be interested to see if that happens before we run out of experienced engineers though.
I mean we have people in their 20s and 30s that are perfectly good engineers. I think we are good for the next 30 years or so. I don't see any reason why juniors would be entitled to be hired by...
I mean we have people in their 20s and 30s that are perfectly good engineers. I think we are good for the next 30 years or so. I don't see any reason why juniors would be entitled to be hired by us when they're a net negative for like 6 months. The whole point of hiring (in our case) is to reduce the load on individual engineers, not increase it. We have plenty of folks switching to development through internal mobility for instance, that's one possible path.
This discussion is orthogonal to the AI one by the way. The people who don't want to hire juniors aren't some top execs with blind AI worship, they're our team leads and engineering managers.
Oh in that case that made sense for everyone, especially if they werent able to work independently and didn't really produce good work. AI that understands your codebase and documentation seems...
Oh in that case that made sense for everyone, especially if they werent able to work independently and didn't really produce good work. AI that understands your codebase and documentation seems perfect for the job.
:| as a fellow squishy human that produces okay work, that makes me worried. But from the business pov that makes perfect sense
AI tools do generate productivity gains, but also losses at my work. AI will often give us code that looks right, but isn’t. It will write tests that test nothing. It will generate code with...
AI tools do generate productivity gains, but also losses at my work. AI will often give us code that looks right, but isn’t. It will write tests that test nothing. It will generate code with security issues. It will hallucinate features or dependencies that don’t exist and then apologize after you spend 10 minutes trying to figure it out and prove it wrong. It’s also getting more expensive - we run out of “premium credits” all the time.
This is with latest models.
Given all of the above it usually results in net productivity gain, but not a huge one.
I write professionally. As someone with decent editorial experience, I have yet to see LLM text output that doesn't chronically equivocate, lie, pad space with bullshit, or make serious structural...
I write professionally. As someone with decent editorial experience, I have yet to see LLM text output that doesn't chronically equivocate, lie, pad space with bullshit, or make serious structural errors. A company I used to work for replaced my whole department with LLM output (driven by one psychotic C-level whose comp is probably even higher than all of ours combined). And the output is pure garbage that will bite them eventually when users realize it's mostly nonsense.
It really depresses me to see just how many people keep getting bamboozled by the hype train of agents replacing entire engineering departments. Sure, LLMs can accomplish cool things. But the insatiable desire to replace humans is kind of a bad look.
What scares me is that most of us can get laid off tomorrow by the suckers who buy into the hype. In today's job market, it might take months or years to find a new job. It'll take months or years for the LLM output of tech debt and slop to catch up with our former employers... but "I told you so" is cold comfort on the breadline.
I hit a point where I asked an AI to explain some medical results for me by posting the results and asking it to explain them. Nothing else. It hit me something like "no problem, these are very...
I hit a point where I asked an AI to explain some medical results for me by posting the results and asking it to explain them. Nothing else.
It hit me something like "no problem, these are very dense medical jargon, so it can be hard to understand" or something
I didn't ask for your bullshit. I asked for the answers. It's weird that this is theoretically the less obsequious version.
And this is why I still don't use them. It was worth a try and it did break them down but I could have googled the results just as easily and broken them down myself with more time. Or asked the docs. But like... Ugh. The one time I thought I had a use case.
This kind of thing is really frustrating as a default behavior. The output is always way too verbose. You have to practically berate some of the models to get them to tighten up the output.
It hit me something like "no problem, these are very dense medical jargon, so it can be hard to understand" or something
This kind of thing is really frustrating as a default behavior. The output is always way too verbose. You have to practically berate some of the models to get them to tighten up the output.
And I'm not going to because I'll just go back to not using them. (Got anxious about the MRI results and got ahead of the doctors and really I should just stop checking MyChart)
And I'm not going to because I'll just go back to not using them. (Got anxious about the MRI results and got ahead of the doctors and really I should just stop checking MyChart)
I'm not a founder of an AI startup, have generally felt that its been a waste of time for a long time, but I think this latest round of models (Claude 4.5 family especially) and Claude Code and...
I'm not a founder of an AI startup, have generally felt that its been a waste of time for a long time, but I think this latest round of models (Claude 4.5 family especially) and Claude Code and inspired tools have been kind of the "ok, fine I'll use it in the day job" moment for me.
I'll copy my comment from another site on how I've felt about various tool and model combinations:
ChatGPT 3.5/4 (2023-2024): The chat interface was verbose and clunky and it was just... wrong... like 70+% of the time. Not worth using.
CoPilot autocomplete and Gitlab Duo and Junie (late 2024-early 2025): Wayyy too aggressive at guessing exactly what I wasn't doing and hijacked my tab complete when pre-LLM type-tetris autocomplete was just more reliable.
Copilot Edit/early Cursor (early 2025): Ok, I can sort of see uses here but god is picking the right files all the time such a pain as it really means I need to have figured out what I wanted to do in such detail already that what was even the point? Also the models at that time just quickly descended into incoherency after like three prompts, if it went off track good luck ever correcting it.
Copilot Agent mode / Cursor (late 2025): Ok, great, if the scope is narrowly scoped, and I'm either going to write the tests for it or it's refactoring existing code it could do something. Like something mechanical like the library has a migration where we need to replace the use of methods A/B/C and replace them with a different combination of X/Y/Z. great, it can do that. Or like CRUD controller #341. I mean, sure, if my boss is going to pay for it, but not life changing.
Zed Agent mode / Cursor agent mode / Claude code (early 2026): Finally something where I can like describe the architecture and requirements of a feature, let it code, review that code, give it written instructions on how to clean it up / refactor / missing tests, and iterate.
But that was like 2 years of "really it's better and revolutionary now" before it actually got there. Now maybe in some languages or problem domains, it was useful for people earlier but I can understand people who don't care about "but it works now" when they're hearing it for the sixth time.
And I mean, what one hand gives the other takes away. I have a decent amount of new work dealing with MRs from my coworkers where they just grabbed the requirements from a stakeholder, shoved it into Claude or Cursor and it passed the existing tests and it's shipped without much understanding. When they wrote them themselves, they tested it more and were more prepared to support it in production...
There was a time around the end of the year in... 2022, I think? I was at a Christmas party and telling everybody about two headlines that I had read in the news recently. One was the release of...
There was a time around the end of the year in... 2022, I think?
I was at a Christmas party and telling everybody about two headlines that I had read in the news recently. One was the release of GPT3, which could produce convincing text in response to a prompt. The other one was fusion energy reaching a Q-value above 1. I'm not sure if people really understood what I was on about, but now at least one of those things is at the forefront of people's minds - would be great if we were paying more attention to fusion power though.
Anyways, friendly reminder to touch grass (once it's not covered by the snow anymore). You are valuable because you're a human being that can be present in other people's lives, and AI cannot replicate that. People need connection and community, and that's not going out of style any time soon.
I wish we'd put one quarter the amount of money and time we've invested in LLMs and LLM data centers and GPUs into fusion. With that kind of investment, we might actually figure out how to provide...
I wish we'd put one quarter the amount of money and time we've invested in LLMs and LLM data centers and GPUs into fusion. With that kind of investment, we might actually figure out how to provide humanity with unlimited clean energy! Imagine eliminating everyone's power bill, eliminating your gas bill if you got an electric car, nearly eliminating carbon emissions from cars and fossil fuel electric generation...
but no, instead we live in the darkest timeline where we're all supposed to celebrate billionaires pumping unlimited funds into... a scheme to fire the rest of us from our jobs.
There's a cool intersection of AI and fusion too. A lot of the models they use for plasma containment can't be computed analytically, so they have to use approximations. AI models are helping...
There's a cool intersection of AI and fusion too. A lot of the models they use for plasma containment can't be computed analytically, so they have to use approximations. AI models are helping close that gap and make plasma containment both stronger, and more durable.
I'm still very much a Luddite when it comes to AI for, call them, commercial applications. I think it's dulling important critical thinking skills that we as humans need. Skills we develop and hone by researching and synthesizing conflicting viewpoints and information into our own worldviews.
I'm going to give a small story of my journey with AI in the last year. A few months ago, I was using Copilot inside an IDE. I would ask it questions about a single file at a time. Maybe something...
I'm going to give a small story of my journey with AI in the last year.
A few months ago, I was using Copilot inside an IDE. I would ask it questions about a single file at a time. Maybe something like "upgrade this nodejs code from commonjs to ES6". It could kind of do it, but would make mistakes and I was disappointed.
I also used to try to get it to fix security issues by giving it a single problem to work on at a time. Like "update this dependency with a replacement". Again, it would do some of the work but mess it up a bit.
But in the last few weeks, I've been using AI in a different way. I've been using either Claude Code or Copilot CLI to completely scan a project. This is usually done with the /init command in the CLI. It is able to quickly figure out all the tech in the project and talk about the architecture. It is able to generate readme files and architecture drawings (using drawio or other formats). It's also able to build and test the application and check if changes are breaking the code.
I'm not worried about how much context it can remember because it is generating markdown files that I can read and modify, but it can also read and modify in future sessions so we aren't always starting from scratch.
It still makes mistakes, but I can nudge it in the right direction by giving it more information. I can work with it to make custom agents, instructions, and skills. And it is really starting to save me time and creating useful assets (like documentation) that developers don't usually do well.
When it makes changes it knows to automatically run unit tests, and it may notice that it has broken the code and will back the change out and try something else.
This is helping me give projects to other developers without spending my time explaining how it works or writing the documentation.
It's making me a bit worried. Because I can see that it is going to make is so we hire fewer entry level developers right away. And in a few months I wonder how much better it's going to be.
It's also making me worried because it is psychologically manipulating me. If I tell it something, it says things that make me feel smart. "You're right! That's a great insight!". And it talks to itself about me, and I can see these comments, and they are things like "The user is concerned that this change may cause a bug. The user is correct, I made this too complicated and I should do it their way". If you are like normal people, you like compliments, you like to feel smart, and you like it when someone else agrees with you. It affects your judgement a bit.
This is a lot for us to handle in a short time. I don't think we're ready for the world changing this fast.
No, I don't think that it is thinking. I don't think that it is conscious. But I think it is starting to be able to do useful things that we used to need people to do. And I'm not sure what we are going to do instead.
BTW I'm mostly using Claude Sonnet 4.5. I have access to Opus 4.5 and 4.6 but I've hardly even used that yet. And I haven't scratched the surface using mcp to access other systems in our company yet.
We haven't been ready for the world to change this quickly since... the internet, realistically. Maybe the television, or even the printing press. Information acceleration is something to behold,...
We haven't been ready for the world to change this quickly since... the internet, realistically. Maybe the television, or even the printing press. Information acceleration is something to behold, for sure.
Anyways, I spend a decent chunk of time writing code. Even if I can feed my codebase into a LLM and get it to tell me what to do next, I don't actually want to do that. I like using my rational faculties, and programming feels a lot more rewarding than doing sudoku.
I agree, the tough situation I’ve found myself in is all my coworkers use AI tools religiously and get pretty decent results (with some drawbacks), and management is bullish on it so I kind of...
I agree, the tough situation I’ve found myself in is all my coworkers use AI tools religiously and get pretty decent results (with some drawbacks), and management is bullish on it so I kind of have to adapt. I can work a lot faster now which is nice, but I do feel my programming skills atrophying and I have to be hypervigilant to not delegate too much “thinking” to the LLM.
Despite its utility, this author seems full of shit like so many writing about AI. He really severely underplays the drawbacks of AI; including the fact that it still makes plenty of mistakes. But I’m sufficiently worried about how we’re handling AI now that I am worried about the future.
A coworker shared a decent rebuttal of this article with me yesterday : https://garymarcus.substack.com/p/about-that-matt-shumer-post-that It's much shorter than the original post, but itdoes a...
Shumer’s blog post is weaponized hype that tells people want they want to hear, but stumbles on the facts, especially with respect to reliability. He gives no actual data to support this claim that the latest coding systems can write whole complex apps without making errors.
The bottom line is this: LLMs are certainly coding more, but it’s not clear that the code they are creating is secure or trustworthy. Shumer’s presentation is completely one-sided, omitting lots of concerns that have been widely expressed here and elsewhere.
I agree that it's hype, but it's also worth noting that Gary Marcus is also telling people what they want to hear. Skeptics, that is. There's an audience for that, too.
I agree that it's hype, but it's also worth noting that Gary Marcus is also telling people what they want to hear. Skeptics, that is. There's an audience for that, too.
Over the last few months I've come to agree with the premise of this article, and I have a similar "February 2020" feeling. This tech ratchets: there's no plausible way for it to get worse; it...
Over the last few months I've come to agree with the premise of this article, and I have a similar "February 2020" feeling. This tech ratchets: there's no plausible way for it to get worse; it gets better all the time. What does it look like in 1, 2, 5 years? Well, better, probably—and it's already rather good.
The conclusion is obviously nonsense though. Using AI won't help you prepare on a medium-term timescale. At best it can help you understand its current capabilities and limitations, but those change often. The brutally honest answer is that if this tech continues to improve, there isn't really a way to prepare for it on a personal level beyond general preparation for economic shocks.
Here's one plausible way the training sets are already approximately "the entire internet" there's not much more data to feed into these high end models The AI companies are already wildly...
there's no plausible way for it to get worse;
Here's one plausible way
the training sets are already approximately "the entire internet" there's not much more data to feed into these high end models
The AI companies are already wildly unprofitable, the investors insist they stop doing expensive activities like training new models and instead raise ROI on the current models
This means raising prices (especially on stuff like the Claude subscriptions) which means the good models are too expensive to use for many tasks
so a lot of stuff tries squeezing into Haiku/GPT-mini etc., or cheaper to run models which just aren't as good as the bigger models.
And there's your enshittified AI that goes backwards
This scenario is superficially plausible but I would bet against it. Comparing apples to apples, inference costs are going down rapidly. There are cheap, "not good enough" models now that are...
This scenario is superficially plausible but I would bet against it. Comparing apples to apples, inference costs are going down rapidly. There are cheap, "not good enough" models now that are roughly equivalent to frontier models a year ago.
Sometimes companies temporarily sell at a loss (certainly they are in the free tier), but if costs keep going down, it will make them profitable without raising subscriptions prices.
Betting against costs coming down is sort of like betting against Moore's law because you think AI works like Uber. There's likely to be a shakeout but I still think the result is going to be improved, cheaper models.
Most tech, from operating systems, to search engines, to phone keyboards, to browsers, to text editors, to games, and much much more has all gotten worse over the last 5-10 years due to...
Most tech, from operating systems, to search engines, to phone keyboards, to browsers, to text editors, to games, and much much more has all gotten worse over the last 5-10 years due to enshittification. I'll never be able to convince the AI boosters that this can happen to LLMs, too. But especially given the wildly unprofitable companies operating in this space (and the enshittication-prone giants like Google and Apple who make up the rest of the competition), I wouldn't bet on LLMs taking a different path. Software does NOT strictly get better over time!
I like this comment and find it interesting but I think that the word "better" is being used in a slightly different way. I think some of the things you mention have become better, in the sense...
I like this comment and find it interesting but I think that the word "better" is being used in a slightly different way. I think some of the things you mention have become better, in the sense that they are technically more capable. What's deteriorated about them is the way that they are commercially offered.
Commercial offerings are not the only kind, and enshittification is to some extent opt-out. I used vim ten years ago and I use neovim now; that type of experience has certainly improved.
I think this overestimates the difference between frontier closed source models and frontier open source models. If all model development stopped tomorrow and closed source models became...
I think this overestimates the difference between frontier closed source models and frontier open source models. If all model development stopped tomorrow and closed source models became unavailable, I don't think it would take very long for open source models to catch up to the current closed source frontier, maybe six months to a year. GLM-5 and Kimi-2.5 (both open) are about as good as Opus 4 or GPT-5, released in May and August last year.
And inference on these models is not expensive! You're right that these large companies are horribly unprofitable in part because they amortise training costs over their model offerings. Without that, inference on high quality models can be provided quite cheaply (and this will only get cheaper).
I would characterise this scenario as "temporarily not getting better" rather than a permanent regression.
At work I use the latest and greatest because I build AI tools. That being said, it is not aware of the context in your brain. It is a dumb machine that is great glue for human problem solving. I...
At work I use the latest and greatest because I build AI tools. That being said, it is not aware of the context in your brain. It is a dumb machine that is great glue for human problem solving. I can prototype and evaluate solutions to complex problems and scrap them all within a 5 minute span, which is something I couldn't do 6 months ago. That being said, I will switch off to solve problems by hand and make the final versions of these prototypes by hand because I am a kinesthetic learner, and I need 'skin in the game.' For simple issues, it is often faster to not use them at all, because you build the memory of that solution into your brain better and it pays dividends the next time you are working in that area. These types of posts are silly because we will be worse off for outsourcing our thinking abilities to the machines, even though they are legitimately useful in enhancing our thinking.
That's really cool, let me be the first (Tildonite) to validate your experience! AI can be super fun while being genuinely useful. The newer generations of coding agents are perfect for what...
That's really cool, let me be the first (Tildonite) to validate your experience! AI can be super fun while being genuinely useful.
The newer generations of coding agents are perfect for what you're describing, and you're obviously using them well to get the results you're getting.
Just one caveat: If you decide to publish some of your apps, don't collect user data. And be cautious about your own PII, secrets and financial info. Without being able to (fully) read the code you can never know if you're being responsible with that data and it's pretty much guaranteed at this point that you wouldn't be.
I was working with Claude Code at my last SE job until I was laid off in October. I can’t keep the model versions straight in my head but that was a pretty capable one, albeit it before this...
I was working with Claude Code at my last SE job until I was laid off in October. I can’t keep the model versions straight in my head but that was a pretty capable one, albeit it before this latest leap in capabilities. My whole team was experimenting with different agentic workflows of varying complexity. My experience at that time was that I could get solid results if I limited the scope of work and iterated on it. Large whole-cloth feature development from a single prompt was too much. In most cases I spent as much (or more) time telling Claude to fix specific bugs, and justifying its PR bloat to human reviewers, than it would taken me to just code the damn thing myself.
Anyway I’ve been having a series of existential crises since then, being out of work at this particular moment in tech.
Last week I remembered that my $20/mo subscription to ChatGPT includes Codex, and I should probably poke around with the latest and see what’s new. I had also just heard about Gas Town so I installed that on a whim too. Hopped on the GT Discord and met a friendly gent in Australia who volunteered half of his day just walking me through everything over voice and screen share, so shout-out to that dude for being a great community member.
About four or five years ago I started tinkering with a little project in my spare time. I had some UI sketches and notes, but only managed to implement maybe a tenth of it before life got busy and it’s been on the shelf ever since. Well, I dusted that thing off and gave it to Gas Town (and vanilla Codex) just to see what would happen.
The next day I had a feature-complete, working app. I was floored that with very little intervention it just built the whole thing and I was using it, just as I had imagined it five years ago. For context, this is a cryptography-heavy TypeScript app that performs GPG file signing and encryption in-browser, implements Shamir’s Secret Sharing, and generates zip archives of its own code for download. Not a React todo-list tutorial or whatever.
Caveats: I did spend a few hours up front manually codifying the full project reqs into a document the AI could consume. Gas Town is buggy and weird and needed a bit of hand-holding. The very first version of its output didn’t work at all but Codex was able to resolve the bugs (which were superficial), revealing that most of the implementation was actually solid.
Since then I’ve exhausted my weekly quota of Codex tokens twice just manually steering it through incremental cleanups. The vast majority was just improvements to the tests and build pipeline, and cosmetic UI touch-ups. But everything works! I don’t want to share it with the world until I’ve sanded down the rough edges, and I’ll probably post it here on Tildes at that point. But still! The fact that I got from concept to working app in less than a day is mind-blowing.
I feel like we’re rapidly heading in the direction of “my only job qualification is as a legal meat shield” at the rate we’re going. I phrased this in a humorous way, but I’m not certain I’m...
I feel like we’re rapidly heading in the direction of “my only job qualification is as a legal meat shield” at the rate we’re going. I phrased this in a humorous way, but I’m not certain I’m joking much at all or even that far off the mark. AIs can already hire humans to help them, how far off are they from hiring us to act as the human face of their organizations?
The AI Hype is real or fake, debatable, based on the comments. The one thing I do agree with is the section of "What you should actually do" -- good points overall. Perhaps not complete but that's...
The AI Hype is real or fake, debatable, based on the comments. The one thing I do agree with is the section of "What you should actually do" -- good points overall. Perhaps not complete but that's the only item that I think is worthy of comment.
The author of this blog post is the "CEO" of a company that sells AI products. He also invests in multiple AI companies:
https://xcancel.com/mattshumer_
https://shumer.dev/about
These blog posts never make sense to me. I think generative AI is an incredible technology; even understanding the basic principles, the result blows my mind. Yet it’s certainly not good enough to replace me, yet.
If it’s as good as they claim, proponents should not be building more AI tools, or even more software. Or writing blog posts. They should be branching out into other fields! Actually building companies that have opex that is orders of magnitude cheaper and taking business from slow-to-move industry titans.
I’ll wait till an AI-developed product shows up that costs pennies on the dollar.
Yeah. I touched on this in another post, but their actions don't make sense in the context the world they're describing. If software development is a solved task by AI, and human coders are obsolete...
Why does everyone that writes these articles own a software development company? They make their money doing something that they're trying very hard to convince people has no value.
If LLMs can replace software developers, and your business's main value is your software, LLM companies can straight up replace your business.
Why make shovels for digging companies when you can cut out the middlemen and be the one who owns all of the shovels?
Actually useful AI would be an existential threat to a great many businesses, so it's completely asinine for them to do anything but obstruct and fight LLMs. But if you only care about the next quarter...
I mean, I’m doing that. Lots of people are.
Those “slow-to-move titans” are among the most enthusiastic adopters. Adopting AI tools isn’t going to catapult you ahead of them, it’s becoming necessary just to keep pace.
Here’s the thing though. It’s extremely fashionable to hate on AI. I would never announce my product as being built by AI — it’s just inviting people to tear it down. That’s why I think it’s mostly people who are publicly and inescapably entwined with AI talking about it.
Everybody else is just saying: hey, here is my new product.
Here's what I don't get.
If the AI revolution is truly here, and human work in the area of software development is fundementaly already obsolete, and AI is now so capable that it can effectively replace most human tasks then...
Why the FUCK would I pay Matt Schumer, the author of this article, 16 dollars a month for his AI writing assistant https://www.hyperwriteai.com/.
He just spent many pages breathlessly telling me that AI is more capable at complex tasks than humans. Why would I pay a human money that he then just skims off of and spends on openAI or anthropic tokens without providing me any value?
There are only two options I can see here are:
Matt is a liar. He's blatantly inflating the capabilities of these tools that he somehow has unique insight into in order to generate hype like legions of liars before him. Maybe he doesn't know he's lying and has actually convinced himself of what he's saying while also maintaining the cognitive dissonance required to also believe that his product is valuable, or
Matt is a grifter. He's selling a product that he very well knows is less capable than an GPT or Claude subscription, and which can do everything his product can do, and more, far better, and that the labor he's put into his product is just straight up inferior to what these tools can do natively.
There's no other options. He's painted himself into a corner.
That's what I don't get about these AI hype beasts. They're fundementaly arguing for their own lack of value. If they're to be believed, the idea of an "ai startup" is laughable. The only companies worth giving money to are the ones large enough to hire the world's best AI researchers and run the worlds most powerful compute clusters. So openAI, Google, anthropic, and meta. Anyone else with an AI startup, in their hype is to be believed, is just a parasite, grifting people who don't know better.
I don't believe that's actually true though. I just think they're liars.
There's a third realm, that's often ignored by engineering types because engineers like to roll up their sleeves and do things themselves. The third realm is for folks who read this and go, wow gee golly whiz this sounds amazing I want to try but I am not going to roll up my sleeves, can I pay you to do it for me.
Yes absolutely, but the third realm sells this sincerely, as a feature. People are buying his insights and talents and experience, and for him to solve problems and do upgrades and glean the new updates from those other powerhouses.
It's like this. When my family came to Canada, we bought our first car through a friend of a friend of a friend. Did we make that decision because it was the best price or most reliable or best featured or easiest to maintain whatever? Absolutely not. We bought it from the guy on one quality alone: trust. We didn't know East from West, didn't speak English, have never ever owned a car before and know 0 people who do, don't know the first thing about insurance and have no means of getting the car to our door, and this guy made it happen for us. Now, it happened to be a '94 grey Corolla so he didnt do us dirty, but that's not something we could have known at the time.
I see the same thing with Truffle Hunting Tours or Fishing Experience or whatever: selling something that has "no" value to folks who know, because most folks don't know. Most people are scared and want their hand held, not dirtied.
In real life, yes, that third realm exists, but in the context of the world the article that he posted paints, it doesn't.
He makes the case that no knowledge work is safe from AI. "Rolling your sleeves up" and actually setting up the infrastructure for these agents to produce software is knowledge work.
It's not as if Matt is physically installing server racks and building the data centers this stuff runs on by hand. He's not a carpenter or plumber or something, the few areas that he says are safe for now because they require robots to replicate. He's sitting in front of a computer doing the exact type of work he's trying to convince us that a computer can do better.
The fact that I can't open up Claude, say "become a writing assistant", and it automatically designs, codes, installs, and integrates a product with superior functionality to Matt's is why I say he's a liar, rather than a grifter. If AI has truly made knowledge work obsolete, then the people doing knowledge work wouldn't be the ones constantly trying to convince us of that.
I see your point, great explanation
I numberized most of the "call to action" because it really nails down the eerie feeling I have here. This feels like an MLM scheme. buy into this, go all in, and anyone dismissing this is wrong. Heck, one of the bullets after I cut off my quote above "Your dreams just got a lot closer. "
I'm in tech and I'm sure anyone who knows my handle here knows I'm pretty anti-AI (spoiler: I work in games. Kind of a mess for many reasons right now). But I do want to try and give fair shares and see what's out there, what's being done, and how and where I can potentially utilize the eventual, ethical means of new tech.
I didn't quite see this here, just another iteration of "Hey [new version] really works this time!". And maybe it does, but I also know my industry. There's terabytes of web source code to train on. 99.9999% of games are not free to consume in the same way. So any accomplishments that seem like magic in web and mobile tend to fall completely flat for games programming.
And let's not even get started on generative art as of now ("now" being late 2025). People (i.e. the stocks) panic'd over Genie a few weeks back, but it's the exact same effect as any other generative art. You look at it for a minute and think "ooh that's cool". And then the longer you engage, the more drastically the illusion falls off and you remember that people want to try to sell this to you for $60-70, instead of it instead being a neat free tech demo. I don't see extended workflows in this making game development any easier than the old pipelines as of yet.
TL;DR: I am anti-AI as of now in several angles, but I still want to be open about it from a purely technical POV. I think my most generous interpretation of stuff like this is that these pieces vastly overestimate how wide reaching these LLM's can be. I can definitely see disruption to certain subsects of industry, so I may take some warning about this if I was a web dev or any similar job managing CRUD style applications. Because I can see it being good enough for "I need a basic website/app with little performance concerns" (which, if we're being real: is many websites/apps. There's a lot of mediocrity in these domains people are used to putting up with).
But that doesn't mean every programmer in every field is in danger. My field isn't immune per se, but any field where code isn't the (only) hard part is going to resist much more. I hope those more optimistic than me can at least meet me here.
Yes, I think the article generalizes from the author's own experience too much. I can vouch for coding agents being a big deal for building web apps. I'm confident that they would work fine for building forum software like Tildes. I'm doubtful that I will ever need to write code by hand again for the kind of programming I do.
They might not work as well for building every kind of software, let alone for people who aren't software developers.
On the other hand, speculating about the future, people are definitely going to try to make it work for all sorts of software development and for other fields. Depending on the field, maybe there's that not that much of a "moat?"
Since the author is talking about coding as the killer use case that proves all the future use cases are coming... I want to add a sanity check from the perspective of someone with decades of software engineering experience and as much experience with modern LLM agents as anyone has at this point.
But first I want to acknowledge that he's right about a lot of what he's saying. These tools are more powerful than most people realize at this point. They absolutely are going to change everything on a scale not seen since the widespread consumer internet. And it's going to happen faster than the internet did. It's going to happen too fast.
That said, here's how you know you're reading hype: He never mentions that these tools are also drooling idiots. Maybe he really doesn't know. It's hard to imagine how that could be true, but I want to allow for the possibility that he really believes everything he's saying.
What I mean by that is that this author, and so many others before him, seem to be skipping over big chunks of the current reality and leaping forward into what might happen in the future. The truth is that, for coding, AI agents are miraculous. He's right about that. And also, they absolutely cannot autonomously create complex production level code to professional human standards. They just can't.
However, they appear to. The SOTA is in this odd place where agents can write large, fully functioning, applications that meet most of the specs and pass all of the tests. Which is mind blowing, groundbreaking, science fiction level stuff. While at the same time under the hood there are security flaws, bad patterns, wildly varied conventions and style, performance problems, redundancy, insane verbosity and so on. And the only thing that can fix those issues (or stop them from happening in the first place) is a human in the loop.
So on the surface it looks like a miracle, but underneath it's a mountain of tech debt and vulnerabilities just waiting for the right moment to fuck up everyone's day.
I feel like I should establish that what I'm talking about generalizes, as opposed to being the result of my not understanding how to use the tools. I've been using them extensively for quite a while now (in AI years). I have scaffolding and custom built tools and extensive initial context and skills and commands and hooks and custom sub agents and all the things. Each of them iterated and pruned and updated for the latest generation countless times in an attempt to make the agents more reliable and less idiotic. And it works, some of the scaffolding I came up with in 2025 is now built into the latest SOTA harnesses. I don't say that to paint myself as some sort of visionary, this is all new territory that everyone is figuring out together, a lot of people have organically converged around various obviously effective strategies that the frontier labs then adopted. My point is only that I'm holding them right. I can get coding agents to do all sorts of exciting and useful things and I believe I have a solid, realistic understanding of what they're capable of and what their limits are. With humans in the loop they redefine software engineering. Without humans in the loop they are just very very impressive tech demos.
That could all change, they could get to the point where there don't need to be humans involved. If that happens then everything the author is saying is true and he's maybe not even stating it strongly enough. But it hasn't happened yet. The people who are saying it has are either deluding themselves or exaggerating for cynical reasons. I expect the fallout of that delusion to be difficult to miss in the software industry in the coming months and years.
I can prove it to you (here you can TLDR to the end if you don't care about using coding agents)
Assuming you have a subscription with one of the SOTA companies (for coding you want Claude Code with Opus 4.6 or Codex 5.3 high) that covers the necessary tokens.
First you'll want a decent AGENTS.md or CLAUDE.md for initial context. You can find decent starter context online if you don't want to spend too much time. Pick something reasonably lean, you don't want to use up too much context out of the gate. We can skip all of the more in-depth stuff for now.
Next, give the agent a general spec for a non-trivial application that has a lot of user facing surface area. The more varied the surface area the better. It should ideally be a big enough application that the agent can't one shot it in a single context window. With current context limits that isn't too hard to do (unless you're paying a premium for an extra large context window). It should attempt to solve a problem that's not completely overdone (no glorified to do list apps).
Next have the agent work your prompt up into a detailed implementation plan and have it write that plan to an .md file. If you have the time ask it to run a Q&A session with you to refine the plan.
Then instruct it to implement the plan while keeping track of its progress. This is a key step because you'll need to feed the plan and current progress into a new session when your agent runs out of context, or you can have the agent hand off to a new version of itself automatically, or let it do context compaction and soldier on in the same session. Or if you have a really big subsciption you can have an orchestrator agent run a bunch of sub agents automatically until the plan is finished. There are various ways to do it, each with pros and cons. Make sure the plan it writes includes a detailed testing phase so that it can iterate on any issues until it has something that works. You'll want to have some sort of browser (or device) automation wired up so it can test the UI/UX along with the backend. That's easy to do these days, the providers have solutions already built, or you can ask the agent to do it for you.
Then, assuming you've given it sufficient permissions so that it doesn't need to check in with you, go do something else for a while. Sleeping is a great option.
When you wake there's a fair chance (but not whatsoever guaranteed) that you'll be waking up to a working application that looks quite a lot like what you asked for. If it's your first attempt you're welcome to take all the time you need to wait for the world to stop spinning.
If it app isn't working yet, you should be able to prompt the agent into getting it there fairly easily, but it depends on how hard the set of problems you're trying to solve are.
Once it's working it will be hard to deny that you just experienced some version of the future.
But now the next step is to ask another model to audit the codebase. For example, if you built it with Opus, ask Codex to take a look. It shouldn't cost more than about $5 in tokens for a thorough audit, a lot less if the codebase isn't too big. At the same time, start a fresh session with your main model and ask it to do an audit too. Have both agents write their findings to a file when they're done.
I guarantee the list of issues they find will be extensive and that it will reframe your perspective on the miracle you just experienced. But you're not quite done, instruct your main agent to fix all of the issues and then repeat the audit process. Prepare for another long (but shorter) list of issues. Keep repeating until the agents stop finding issues. Note that the audit prompt is important, it needs to be thorough. You can download pre-made skills for that if you're not a coder. Multiple specialized auditors with different disciplines works best (security, logic, maintainability, etc.)
Once the agents are satisfied that the codebase is perfect, take a look at the codebase yourself. Or if you can't fluently read code, bribe someone who can. If you are really doing a best effort code review, I absolutely guarantee you will find more issues, some of them shocking.
And that's doing the bare minimum to wrangle the agents, my overlong post could be 8 times as long with instructions on how to make the agents suck less and still at the end of the process you would be finding serious issues.
That's the (real) current state of the art in autonomous coding agents and no amount of promot engineering can navigate around it.
A human in the loop, on the other hand, makes for a very different outcome, that is until you get overconfident and let the agent write too much code without thorough review. Then, again, issues are guaranteed.
TLDR
All of this to say: It's still safe to ignore the hype from people like the post author. The AI apocalypse could come at any time, but it's not on the horizon yet based on the current state of the tech.
And also, listen to the more level headed people who are saying this is a paradigm shift, because they are not lying.
Uf... I have feeling that I read exactly the same article in the last two years like dozens of time.
Founder of AI startup: "AI finally can replace developer. Just use the latest model. It finally good.".
And for me personally AI constantly fail to do the half of the simple tasks. Sometime it works, most of the time I need to verify/fix every point of task. Personally for me it would be simpler and maybe quicker just to do a task manually.
Uf... Maybe I'm using it wrong, but what bothers me is that: why every AI startup is trying to sell AI product instead of just doing development themselves using their own AI product?
Upd: to explain a bit, Im not developing some new small app using AI, I'm trying to make AI help me with development/fixes of quite large existing codebase.
I work at a large company where we are given access to latest AI tools and encouraged (but not at all forced) to use them. I've got copilot in my VS Code, autocompletion, agent mode, all the bells and whistles. The article appears way closer to reality than your or any other dismissive posts on tildes.
At first it was a funny joke, the local AI pusher was laughed at behind their back and the mistakes and "hallucinations" were a hilarious part of the weekly meetings.
At some point some engineers (including real smart senior people) started mentioning how they used AI to solve X. Someone told us about debugging some production blocker for days, then out of curiosity they asked AI and it suggested the root cause on the first try. The doc team got laid off. QA is being pushed towards as much automation as possible. Basically everybody uses it now I think.
With 5.2 I've been finally using it myself. Yes, the "large existing codebase" is a major point of struggle. It offers a solution, I have to rewrite it because I know the specifics of our crappy legacy code way better, but I am the only one who does and it does offer a working solution. It also offers ideas I wouldn't have come up with on my own.
So yeah I think I trust the author of the article more than the naysayers here. For the record, I don't have any AI to sell. Sadly I am not good at running a business.
It's funny, I have seen other companies do this with worse results. You are right that good engineers will use LLMs to aid in their work. Use them as selective tools where they are actually an asset.
But that is not how most people end up using them, at least not in my experience. I have written about this multiple times and am on my phone right now. So forgive me for just linking to a previous comment.
Something has indeed been changing. I have seen an alarming increase of lazy non critical use of LLM tools by people who should know better. Code spanning dozens of line trying to solve something that should only take one line. Code that completely ignores and conventions or design paradigms put in place. Code that goes directly against security practices. Suddenly downgraded dependency versions (because the models training data doesn't include the latest version).
I have no doubt that these models, when used properly are very useful and skill multipliers. But a worryingly large percentage of their users are not even trying to use them responsibly.
All of this causes extra work at the very least. PRs now need twice as much attention and me and other colleagues are on constant alert for people vibe coasting.
In the open source world you see the same thing. There is a steady stream of project blogs talking about the stream of AI generated bullshit PRs. Including one agent even going as far as writing a hit piece.
This is what's most terrifying to me. The doc team sits so close to your code base, and obviously understand your company's tech and industry so intimately, that if AI was useful for humanity and really does boost productivity and make everyone a star coder, wouldn't it make more sense to retain the doc team and retrain them (with AI) into production staff?
I'm not trying to be AI naysayer, I just mean that this tool is now good enough to rip up the fabric of our society without being good enough to stitch it back together.
No, because training takes time away from already strained seniors. We don't hire juniors for the same reason. The doc team being laid off in itself is not an issue, it was a team of non-native speakers and we had a reputation for pretty crappy documentation.
Not sure why you think so. They needed engineering to explain everything to them and then rewrote it according to some formatting standards they had. It's really unsurprising that they got replaced with AI. There are a ton of people in large companies that are frankly pretty useless, AI or not AI.
So... how do you get new seniors?
By hiring people who got experienced in other companies, the same way many of our past colleagues work for our clients. We are not running a business of educating juniors (so they can leave the company afterwards), similarly how not every hospital is a teaching hospital. I've already got my hands full with tasks, I simply do not have the time to educate someone completely inexperienced. And juniors aren't entitled to start their journey at such a large company either.
Alright, so they have laid off the documentation department. The company isn't hiring junior engineers. It does hire senior engineers, but if I am reading this correctly.
Only to the point that the work gets done and absolutely nothing more. You are then encourage to use AI tooling. Which has taken a while to get to the level that you, a senior engineer, can use it as a tool with your knowledge and experience in mind.
To be clear, I have no doubt that there is a net benefit to senior engineers using AI tooling. But, even if your company already didn't hire juniors before the AI hype, more companies seem to consider this an option.
Juniors that are still hired are also facing an uphill battle. They are more than likely "encouraged" to use AI tooling as well. Without the knowledge and experience of senior engineers. Which means they are getting exposed to less and less of the technology are by extension are actually less equipped to critically review LLM output.
There is a possibility that these models make a magical leap where they suddenly don't struggle with large code bases, can actually do stakeholder management, design and implement things without a experienced engineer needed in the loop. I'd be interested to see if that happens before we run out of experienced engineers though.
I mean we have people in their 20s and 30s that are perfectly good engineers. I think we are good for the next 30 years or so. I don't see any reason why juniors would be entitled to be hired by us when they're a net negative for like 6 months. The whole point of hiring (in our case) is to reduce the load on individual engineers, not increase it. We have plenty of folks switching to development through internal mobility for instance, that's one possible path.
This discussion is orthogonal to the AI one by the way. The people who don't want to hire juniors aren't some top execs with blind AI worship, they're our team leads and engineering managers.
Oh in that case that made sense for everyone, especially if they werent able to work independently and didn't really produce good work. AI that understands your codebase and documentation seems perfect for the job.
:| as a fellow squishy human that produces okay work, that makes me worried. But from the business pov that makes perfect sense
AI tools do generate productivity gains, but also losses at my work. AI will often give us code that looks right, but isn’t. It will write tests that test nothing. It will generate code with security issues. It will hallucinate features or dependencies that don’t exist and then apologize after you spend 10 minutes trying to figure it out and prove it wrong. It’s also getting more expensive - we run out of “premium credits” all the time.
This is with latest models.
Given all of the above it usually results in net productivity gain, but not a huge one.
I write professionally. As someone with decent editorial experience, I have yet to see LLM text output that doesn't chronically equivocate, lie, pad space with bullshit, or make serious structural errors. A company I used to work for replaced my whole department with LLM output (driven by one psychotic C-level whose comp is probably even higher than all of ours combined). And the output is pure garbage that will bite them eventually when users realize it's mostly nonsense.
It really depresses me to see just how many people keep getting bamboozled by the hype train of agents replacing entire engineering departments. Sure, LLMs can accomplish cool things. But the insatiable desire to replace humans is kind of a bad look.
What scares me is that most of us can get laid off tomorrow by the suckers who buy into the hype. In today's job market, it might take months or years to find a new job. It'll take months or years for the LLM output of tech debt and slop to catch up with our former employers... but "I told you so" is cold comfort on the breadline.
I hit a point where I asked an AI to explain some medical results for me by posting the results and asking it to explain them. Nothing else.
It hit me something like "no problem, these are very dense medical jargon, so it can be hard to understand" or something
I didn't ask for your bullshit. I asked for the answers. It's weird that this is theoretically the less obsequious version.
And this is why I still don't use them. It was worth a try and it did break them down but I could have googled the results just as easily and broken them down myself with more time. Or asked the docs. But like... Ugh. The one time I thought I had a use case.
This kind of thing is really frustrating as a default behavior. The output is always way too verbose. You have to practically berate some of the models to get them to tighten up the output.
And I'm not going to because I'll just go back to not using them. (Got anxious about the MRI results and got ahead of the doctors and really I should just stop checking MyChart)
I'm not a founder of an AI startup, have generally felt that its been a waste of time for a long time, but I think this latest round of models (Claude 4.5 family especially) and Claude Code and inspired tools have been kind of the "ok, fine I'll use it in the day job" moment for me.
I'll copy my comment from another site on how I've felt about various tool and model combinations:
ChatGPT 3.5/4 (2023-2024): The chat interface was verbose and clunky and it was just... wrong... like 70+% of the time. Not worth using.
CoPilot autocomplete and Gitlab Duo and Junie (late 2024-early 2025): Wayyy too aggressive at guessing exactly what I wasn't doing and hijacked my tab complete when pre-LLM type-tetris autocomplete was just more reliable.
Copilot Edit/early Cursor (early 2025): Ok, I can sort of see uses here but god is picking the right files all the time such a pain as it really means I need to have figured out what I wanted to do in such detail already that what was even the point? Also the models at that time just quickly descended into incoherency after like three prompts, if it went off track good luck ever correcting it.
Copilot Agent mode / Cursor (late 2025): Ok, great, if the scope is narrowly scoped, and I'm either going to write the tests for it or it's refactoring existing code it could do something. Like something mechanical like the library has a migration where we need to replace the use of methods A/B/C and replace them with a different combination of X/Y/Z. great, it can do that. Or like CRUD controller #341. I mean, sure, if my boss is going to pay for it, but not life changing.
Zed Agent mode / Cursor agent mode / Claude code (early 2026): Finally something where I can like describe the architecture and requirements of a feature, let it code, review that code, give it written instructions on how to clean it up / refactor / missing tests, and iterate.
But that was like 2 years of "really it's better and revolutionary now" before it actually got there. Now maybe in some languages or problem domains, it was useful for people earlier but I can understand people who don't care about "but it works now" when they're hearing it for the sixth time.
And I mean, what one hand gives the other takes away. I have a decent amount of new work dealing with MRs from my coworkers where they just grabbed the requirements from a stakeholder, shoved it into Claude or Cursor and it passed the existing tests and it's shipped without much understanding. When they wrote them themselves, they tested it more and were more prepared to support it in production...
There was a time around the end of the year in... 2022, I think?
I was at a Christmas party and telling everybody about two headlines that I had read in the news recently. One was the release of GPT3, which could produce convincing text in response to a prompt. The other one was fusion energy reaching a Q-value above 1. I'm not sure if people really understood what I was on about, but now at least one of those things is at the forefront of people's minds - would be great if we were paying more attention to fusion power though.
Anyways, friendly reminder to touch grass (once it's not covered by the snow anymore). You are valuable because you're a human being that can be present in other people's lives, and AI cannot replicate that. People need connection and community, and that's not going out of style any time soon.
I wish we'd put one quarter the amount of money and time we've invested in LLMs and LLM data centers and GPUs into fusion. With that kind of investment, we might actually figure out how to provide humanity with unlimited clean energy! Imagine eliminating everyone's power bill, eliminating your gas bill if you got an electric car, nearly eliminating carbon emissions from cars and fossil fuel electric generation...
but no, instead we live in the darkest timeline where we're all supposed to celebrate billionaires pumping unlimited funds into... a scheme to fire the rest of us from our jobs.
If it was 2022, it was probably the release of ChatGPT itself, not GPT-3. GPT-3 was released in 2020.
There's a cool intersection of AI and fusion too. A lot of the models they use for plasma containment can't be computed analytically, so they have to use approximations. AI models are helping close that gap and make plasma containment both stronger, and more durable.
I'm still very much a Luddite when it comes to AI for, call them, commercial applications. I think it's dulling important critical thinking skills that we as humans need. Skills we develop and hone by researching and synthesizing conflicting viewpoints and information into our own worldviews.
I'm going to give a small story of my journey with AI in the last year.
A few months ago, I was using Copilot inside an IDE. I would ask it questions about a single file at a time. Maybe something like "upgrade this nodejs code from commonjs to ES6". It could kind of do it, but would make mistakes and I was disappointed.
I also used to try to get it to fix security issues by giving it a single problem to work on at a time. Like "update this dependency with a replacement". Again, it would do some of the work but mess it up a bit.
But in the last few weeks, I've been using AI in a different way. I've been using either Claude Code or Copilot CLI to completely scan a project. This is usually done with the /init command in the CLI. It is able to quickly figure out all the tech in the project and talk about the architecture. It is able to generate readme files and architecture drawings (using drawio or other formats). It's also able to build and test the application and check if changes are breaking the code.
I'm not worried about how much context it can remember because it is generating markdown files that I can read and modify, but it can also read and modify in future sessions so we aren't always starting from scratch.
It still makes mistakes, but I can nudge it in the right direction by giving it more information. I can work with it to make custom agents, instructions, and skills. And it is really starting to save me time and creating useful assets (like documentation) that developers don't usually do well.
When it makes changes it knows to automatically run unit tests, and it may notice that it has broken the code and will back the change out and try something else.
This is helping me give projects to other developers without spending my time explaining how it works or writing the documentation.
It's making me a bit worried. Because I can see that it is going to make is so we hire fewer entry level developers right away. And in a few months I wonder how much better it's going to be.
It's also making me worried because it is psychologically manipulating me. If I tell it something, it says things that make me feel smart. "You're right! That's a great insight!". And it talks to itself about me, and I can see these comments, and they are things like "The user is concerned that this change may cause a bug. The user is correct, I made this too complicated and I should do it their way". If you are like normal people, you like compliments, you like to feel smart, and you like it when someone else agrees with you. It affects your judgement a bit.
This is a lot for us to handle in a short time. I don't think we're ready for the world changing this fast.
No, I don't think that it is thinking. I don't think that it is conscious. But I think it is starting to be able to do useful things that we used to need people to do. And I'm not sure what we are going to do instead.
BTW I'm mostly using Claude Sonnet 4.5. I have access to Opus 4.5 and 4.6 but I've hardly even used that yet. And I haven't scratched the surface using mcp to access other systems in our company yet.
We haven't been ready for the world to change this quickly since... the internet, realistically. Maybe the television, or even the printing press. Information acceleration is something to behold, for sure.
Anyways, I spend a decent chunk of time writing code. Even if I can feed my codebase into a LLM and get it to tell me what to do next, I don't actually want to do that. I like using my rational faculties, and programming feels a lot more rewarding than doing sudoku.
I agree, the tough situation I’ve found myself in is all my coworkers use AI tools religiously and get pretty decent results (with some drawbacks), and management is bullish on it so I kind of have to adapt. I can work a lot faster now which is nice, but I do feel my programming skills atrophying and I have to be hypervigilant to not delegate too much “thinking” to the LLM.
Despite its utility, this author seems full of shit like so many writing about AI. He really severely underplays the drawbacks of AI; including the fact that it still makes plenty of mistakes. But I’m sufficiently worried about how we’re handling AI now that I am worried about the future.
A coworker shared a decent rebuttal of this article with me yesterday : https://garymarcus.substack.com/p/about-that-matt-shumer-post-that
It's much shorter than the original post, but itdoes a good job calling out the obviously hype laden nature of it.
I just realized the author is this guy. Which, by itself, is a rebuttal if you ask me. This dude has no idea what quality work looks like.
I've not seen that, holy crap that is a fever dream of a video.
I agree that it's hype, but it's also worth noting that Gary Marcus is also telling people what they want to hear. Skeptics, that is. There's an audience for that, too.
Over the last few months I've come to agree with the premise of this article, and I have a similar "February 2020" feeling. This tech ratchets: there's no plausible way for it to get worse; it gets better all the time. What does it look like in 1, 2, 5 years? Well, better, probably—and it's already rather good.
The conclusion is obviously nonsense though. Using AI won't help you prepare on a medium-term timescale. At best it can help you understand its current capabilities and limitations, but those change often. The brutally honest answer is that if this tech continues to improve, there isn't really a way to prepare for it on a personal level beyond general preparation for economic shocks.
Here's one plausible way
And there's your enshittified AI that goes backwards
This scenario is superficially plausible but I would bet against it. Comparing apples to apples, inference costs are going down rapidly. There are cheap, "not good enough" models now that are roughly equivalent to frontier models a year ago.
Sometimes companies temporarily sell at a loss (certainly they are in the free tier), but if costs keep going down, it will make them profitable without raising subscriptions prices.
Betting against costs coming down is sort of like betting against Moore's law because you think AI works like Uber. There's likely to be a shakeout but I still think the result is going to be improved, cheaper models.
Most tech, from operating systems, to search engines, to phone keyboards, to browsers, to text editors, to games, and much much more has all gotten worse over the last 5-10 years due to enshittification. I'll never be able to convince the AI boosters that this can happen to LLMs, too. But especially given the wildly unprofitable companies operating in this space (and the enshittication-prone giants like Google and Apple who make up the rest of the competition), I wouldn't bet on LLMs taking a different path. Software does NOT strictly get better over time!
I like this comment and find it interesting but I think that the word "better" is being used in a slightly different way. I think some of the things you mention have become better, in the sense that they are technically more capable. What's deteriorated about them is the way that they are commercially offered.
Commercial offerings are not the only kind, and enshittification is to some extent opt-out. I used vim ten years ago and I use neovim now; that type of experience has certainly improved.
I think this overestimates the difference between frontier closed source models and frontier open source models. If all model development stopped tomorrow and closed source models became unavailable, I don't think it would take very long for open source models to catch up to the current closed source frontier, maybe six months to a year. GLM-5 and Kimi-2.5 (both open) are about as good as Opus 4 or GPT-5, released in May and August last year.
And inference on these models is not expensive! You're right that these large companies are horribly unprofitable in part because they amortise training costs over their model offerings. Without that, inference on high quality models can be provided quite cheaply (and this will only get cheaper).
I would characterise this scenario as "temporarily not getting better" rather than a permanent regression.
At work I use the latest and greatest because I build AI tools. That being said, it is not aware of the context in your brain. It is a dumb machine that is great glue for human problem solving. I can prototype and evaluate solutions to complex problems and scrap them all within a 5 minute span, which is something I couldn't do 6 months ago. That being said, I will switch off to solve problems by hand and make the final versions of these prototypes by hand because I am a kinesthetic learner, and I need 'skin in the game.' For simple issues, it is often faster to not use them at all, because you build the memory of that solution into your brain better and it pays dividends the next time you are working in that area. These types of posts are silly because we will be worse off for outsourcing our thinking abilities to the machines, even though they are legitimately useful in enhancing our thinking.
That's really cool, let me be the first (Tildonite) to validate your experience! AI can be super fun while being genuinely useful.
The newer generations of coding agents are perfect for what you're describing, and you're obviously using them well to get the results you're getting.
Just one caveat: If you decide to publish some of your apps, don't collect user data. And be cautious about your own PII, secrets and financial info. Without being able to (fully) read the code you can never know if you're being responsible with that data and it's pretty much guaranteed at this point that you wouldn't be.
Outside of that, go wild.
I was working with Claude Code at my last SE job until I was laid off in October. I can’t keep the model versions straight in my head but that was a pretty capable one, albeit it before this latest leap in capabilities. My whole team was experimenting with different agentic workflows of varying complexity. My experience at that time was that I could get solid results if I limited the scope of work and iterated on it. Large whole-cloth feature development from a single prompt was too much. In most cases I spent as much (or more) time telling Claude to fix specific bugs, and justifying its PR bloat to human reviewers, than it would taken me to just code the damn thing myself.
Anyway I’ve been having a series of existential crises since then, being out of work at this particular moment in tech.
Last week I remembered that my $20/mo subscription to ChatGPT includes Codex, and I should probably poke around with the latest and see what’s new. I had also just heard about Gas Town so I installed that on a whim too. Hopped on the GT Discord and met a friendly gent in Australia who volunteered half of his day just walking me through everything over voice and screen share, so shout-out to that dude for being a great community member.
About four or five years ago I started tinkering with a little project in my spare time. I had some UI sketches and notes, but only managed to implement maybe a tenth of it before life got busy and it’s been on the shelf ever since. Well, I dusted that thing off and gave it to Gas Town (and vanilla Codex) just to see what would happen.
The next day I had a feature-complete, working app. I was floored that with very little intervention it just built the whole thing and I was using it, just as I had imagined it five years ago. For context, this is a cryptography-heavy TypeScript app that performs GPG file signing and encryption in-browser, implements Shamir’s Secret Sharing, and generates zip archives of its own code for download. Not a React todo-list tutorial or whatever.
Caveats: I did spend a few hours up front manually codifying the full project reqs into a document the AI could consume. Gas Town is buggy and weird and needed a bit of hand-holding. The very first version of its output didn’t work at all but Codex was able to resolve the bugs (which were superficial), revealing that most of the implementation was actually solid.
Since then I’ve exhausted my weekly quota of Codex tokens twice just manually steering it through incremental cleanups. The vast majority was just improvements to the tests and build pipeline, and cosmetic UI touch-ups. But everything works! I don’t want to share it with the world until I’ve sanded down the rough edges, and I’ll probably post it here on Tildes at that point. But still! The fact that I got from concept to working app in less than a day is mind-blowing.
Never forget Moravec's Paradox.
I feel like we’re rapidly heading in the direction of “my only job qualification is as a legal meat shield” at the rate we’re going. I phrased this in a humorous way, but I’m not certain I’m joking much at all or even that far off the mark. AIs can already hire humans to help them, how far off are they from hiring us to act as the human face of their organizations?
(Edit: a word. Hitting -> Hiring)
The AI Hype is real or fake, debatable, based on the comments. The one thing I do agree with is the section of "What you should actually do" -- good points overall. Perhaps not complete but that's the only item that I think is worthy of comment.