Acharya’s framing is different from mine (he’s talking book on software stocks) but the conclusion is the same: the “innovation bazooka” pointed at rebuilding payroll is a bad allocation of resources. Benedict Evans called me out on LinkedIn for this (https://philippdubach.com/posts/is-ai-really-eating-the-worl...) take, which I take as a sign the argument is landing..
> investors are simultaneously punishing hyperscaler stocks because AI capex might generate weak returns, while destroying software stocks because AI adoption will be so pervasive it renders all existing software obsolete. Both cannot hold simultaneously.
I don't understand this point. Can't it be possible that the ultimate effect is to devalue, hugely, software? As in it can totally both be true that AI capex has weak returns and at the same time most SaaS companies go bankrupt. To take an analogy: if ever we manage to successfully mine asteroids, and find some vast quantity of platinum, it could both be true that every existing platinum miner loses their shirt, and also that the value of platinum sinks so far that the asteroid mining company cannot cover its costs.
I once built a CRM in Google Sheets fully mirroring the data model of Salesforce. For contact, company, deal, and call tracking for a one sales rep business. (Before XLookup was in Google Sheets)
Did it work? Yes. Was it worth my time to maintain and scale the “platform” with the company rather than outsource all that to a CRM company? Not at all.
Time is finite. Spend your time doing what you do best, pay others to do what they do best.
Yup, my experience has been that vibe-coding is very time-consuming. It reminds me very much of how LLMs are great at creating mind-blowing images, but you get what you get. Once you decide that you need to modify the image you get, it becomes a time sink. You might be able to change it and get what you need, but there is no guarantee and it's a never ending task.
The same thing happens with code; you may get great results from your prompt, but trying to customize it will drive you nuts and you may never get what you want.
Maintenance is another hurdle. How do you maintain code you might not have the skills to maintain?
Vibe-coding may reduce software creation time, but it's not taking over software engineering. The SaaS business is going nowhere. Most people, by far, will continue to rely on someone else for their software needs. But be very aware that the software business will change. We are seeing that already.
It doesn't make sense for every company to make their own Salesforce clone.
The key is that it makes new companies entering the market to compete with Salesforce immensely easier. More competition will just force lower overall margins in SAAS.
People are overestimating the value on having AI create something given loose instructions, and underestimating the value of using AI as a tool for a human to learn and explore a problem space. The bias shows on the terminology (“agents”).
We finally made the computer able to speak “our” language - but we still see computers as just automation. There’s a lot of untapped potential in the other direction, in encoding and compressing knowledge IMO.
Because that would mean AI isn't going to replace entire industries, which is the only way to justify the, not billions, but trillions in market value that AI leaders keep trying to justify.
Exactly my thoughts - the value in AI is not auto-generating anything more than something trivial, but there's huge value in a more customized knowledge engine - a targeted, specific Google if you will. Get answers to your specific question instead of results that might contain what you were looking for if you slog through them.
AI is hugely beneficial in understanding a problem, or at least getting a good overview, so you can then go off and solve/do it yourself, but focusing on "just have the AI generate a solution" is going to hugely harm AI perception/adoption.
Thought exercise for those in disagreement: why would every company use AI to build their own payroll/ERP/CRM, when just a handful of companies could use AI to build those offerings better?
This is largely how things work now; AI may lower the cost and increase margins, but the economics of build vs buy seem the same.
To avoid CRAZY SaaS charges. I left a comment further down about how the challenge is first getting a reliable stack running underneath whatever ends up being fast-coded. The trend will be more decentralization - I think that'll be AI 2.0. Increasing centralization is AI 1.0.
Slack is a good example. When the cost of Slack is an unreasonable amount of your operating costs then it makes sense to clone and maintain. The product is simple, you can basically recreate the main functionality in a sitting. Why would you pay hundreds of thousands of dollars for it?
That’s a fine example, but my question then is why does Slack exist? Surely Fortune 500 companies are smart enough to realize that building a slack clone is cheaper, yet they don’t do that.
So now consider AI, perhaps the cost of building has decreased from 100k to 10k. What stops a Slack competitor from also building the product for 10k and reselling it at 10% of the cost of Slack? My point is that I don’t see how AI has changed the value prop.
Every company that I’ve worked at has had to do significant additional development work on their instance of salesforce to make it work for them. Like 6-12 months of work with 1-3 people. I don’t know if this is common but in that case maybe going custom might be the way to go. You get something lean, without all the cruft, specifically built for your usecase and nothing more.
Well the answer is because the cost of that software is lower than somebody building the other software.
What happens is that all these SaaS drop in value because it is now realistic to build them internally
Anyone who's seen an enterprise deal close or dealt with enterprise customer requests will know this, the build vs buy calculus has always been there yet companies still buy. Until you can get AI to the point where it equivalent to a 20 person engineering team, people are not going to build their own Snowflake, Salesforce, Slack or ATS. Maybe that day is 3 years away but when that happens the world will be very different
Companies do make/buy decisions on everything, it just software. Cleaning services are not expensive, yet companies contract them instead of hiring staff.
This is called transaction cost economics, if anyone’s interested.
I agree generally, but some of these enterprise contracts are eye-watering. If the choice is $2M/year with a 3-year minimum contract, or rolling your own, I think calculus really has shifted.
With that said, the entire business world does not understand that software is more than just code. Even if you could write code instantly, making enterprise software would still take time, because there are simply so many high-stakes decisions to make, and so much fractal detail.
> If the choice is $2M/year with a 3-year minimum contract, or rolling your own, I think calculus really has shifted.
But why? It was always dramatically cheaper for enterprises to build rather than buy. They stopped doing that becuase they did that in the 90s and ended up with legacy codebases that they didn't know how to maintain. I can't see AI helping with that.
Exactly. I was building an app to track bike part usage. It was an okay app, but then I just started using ai with the database directly. Much more flexible, and I can get anything I need right then. AI will kill a lot of companies, but it won’t be the software it develops, it will be the agent itself
If an AI agent ever became as productive at writing code as a well-organized 20 person engineering team you'd still need to run it for a year or more to replicate any nontrivial SaaS product.
And the thing about many of these products isn't their feature set, it's their stability. It's their uptime. It's how they handle scaling invisibly and with no effort on your part. These are things you can't just write down from whole cloth, they are properties that emerge over time by adapting the the reality of scale. Coding isn't the whole deal, and your 20x clanker which can do nothing but re-arrange text in interesting patterns is going to have some trouble with the realities of taking that PoC to production. You'll still need experienced, capable people for that. And lots of time.
A lot of this "ermahgerd everything will change" drivel is based on some magical fundamentally new technology emerging in the near future that can do things that LLMs cannot do. But as far as anyone knows, that future may be never.
So even given a large improvement in agentic coding I'm not convinced it really changes the build vs buy equation much.
I sort of agree with this, but what a lot of people are missing is it's unbelievably easy to clone a lot of SaaS products.
So I think big SaaS products are under attack from three angles now:
1) People replacing certain systems with 'vibe coded' ones, for either cost/feature/unhappiness with vendor reasons. I actually think this is a bigger threat than people think - there are so many BAD SaaS products out there which cost businesses a fortune in poor features/bugs/performance/uptime, and if the models/agents keep improving the way they have in the last couple of years it's going to be very interesting if some sort of '1000x' engineer in an agent can do crazy impressive stuff.
2) Agents 'replacing' the software. As people have pointed out, just have the agent use APIs to do whatever workflow you want - ping a database and output a report.
3) "Cheap" clones of existing products. A tiny team can now clone a "big" SaaS product very quickly. These guys can provide support/infra/migration assistance and make money at a much lower price point. Even if there is lock in, it makes it harder for SaaS companies to keep price pressure up.
It would be interesting if, with all the anxiety about vibe coding becoming the new normal, its only lasting effect is the emergence of smaller B2B companies that quickly razzle dazzle together a bespoke replacement for Concur, SAP, Workday, the crappy company sharepoint - whatever. Reminds me of what people say Palantir is doing, but now supercharged by the AI-driven workflows to stand up the “forward deployed” “solution” even faster.
But have you ever tried to clone a product or tool for yourself before? At first it’s great because you think that you saved money but then you start having to maintain it… fixing problems, filling in gaps… you now realize that you made a mistake. Just because AI can do it now doesn’t mean you aren’t just now having to use AI to do the same thing…
Also, agents are not deterministic. If you use it to analyze data, it will get it right most of the time but, once in a blue moon, it will make shit up, except you can’t tell which time it was. You could make it deterministic by having AI write a tool instead… except you now have the first problem of maintaining a tool.
That isn’t to say that there isn’t small low hanging fruit that AI will replace, but it’s a bit different when you need a real product with support.
At the end of the day, you hire a plumber or use a SaaS not because you can’t do it yourself, but because you don’t want to do it and rather want someone else who is committed to it to handle it.
I'm not saying _the end user_ clones it. I mean someone else does (more efficiently with agents) and runs it as a _new_ SaaS company. They would provide support just like the existing one would, but arguably at a cheaper price point.
And regarding agents being non deterministic, if they write a bunch of SQL queries to a file for you, they are deterministic. They can just write "disposable" tools and scripts - not always doing it thru their context.
> He said that software accounts for 8% to 12% of a company's expenses, so using vibe coding to build the company's resource planning or payroll tools would only save about 10%. Relying on AI to write code also carries risks, he said.
> "You have this innovation bazooka with these models. Why would you point it at rebuilding payroll or ERP or CRM," Acharya said
> Instead, companies are better off using AI to develop their core businesses or optimize the remaining 90% of their costs
"You have this innovation bazooka. Why would you point it at rebuilding payroll?" — a partner at the firm whose thesis was literally "software is eating the world."
Apparently the meal is over and now we're just rearranging the plates.
It seems to be premised on the idea we would vibe code a replica of what we get from SaaS. But the real point is, we would not do that. We would vibe code something that exactly fits our business.
We have products we're paying $100k a year for and using 3% of the functionality. And they suck. This is the target.
The missing piece in this debate is that most "vibe coded" replacements break at scale. I tried replacing a multi-step workflow with Make.com + Airtable (not even vibe coding, just no-code automation) and it fell apart past 2 jobs per day - rate limits, webhook failures, state management nightmares. The real pattern I see working is not "replace SaaS with vibe code" but rather "stitch together 5-6 specialized tools with a thin orchestration layer you write yourself." The orchestration is where AI actually helps - it's glue code, not the product.
The bottleneck will always be humans. You could get AI to write a million lines of code a day, but you’d still need humans to review and test that code. We are a very long way from being able to blindly trust AI’s outputs in production.
I don’t even think it’s about reviewing and testing. The bottleneck will always be humans.
We don’t like to always admit it but most jobs are fairly straightforward, as in the actual day to day tasks. Yes being smart is great and useful etc. but after a certain point it’s diminishing returns on the actual tasks you have to do. Dealing with other humans and their egos and eccentricities and the multitude ways each person sees the world is always what makes all jobs tricky. I suspect this whole ai wave/hype/reality is going to open many people’s eyes to this. We will laugh that we use to call them “soft” skills.
IMO I would have agreed with this statement 2 months ago but now it’s clear AI is already much better at reviewing and even testing code (via spinning up simulators, etc) much better than we can. We’re already using AI’s outputs in production and not writing much code these days.
There was a short moment in history where it seemed that the sentiment was: people will soon 3D-print 99% of their household items themselves instead of buying them.
You absolutely could print things like cups, soap holders, picture frames, the small shovel you use for gardening, and so on an so on.
I think some or maybe even many of those shortcomings will apply to software, too. Making actual good software is not as trivial as writing “make me an app”, much as making an actual good spoon is not as trivial as throwing an STL at a printer and calling it a day.
> "You have this innovation bazooka with these models. Why would you point it at rebuilding payroll or ERP or CRM"
They invested in ERP/CRM? I built one (fairly complete to the German/Italy/EU tax system) and it saves a ton of money vs commercial offerings. So yeah, of course we will.
I'm curious about your reasoning. Jira/Trello etc. are like $10/mo/seat, why bother rewriting them from scratch? You'll spend more in tokens doing so. Same for gmail/google calendar, what's the ROI? Those tools are reliable and cheap, why bother creating your own?
jira/trello: ergonomics. to set them up correctly exactly the way i want would take me 20 hours (or hire a PM), i can vibecode for 20h and get the same result.
plus, being able to crossref internal data types is chef's kiss.
im paying for claude pro so it's use it or lose it. when i finish everything and have it battle tested i can end my claude code. and anyways when i have 10 employees, it's parity.
for gsuite: i want to own everything internally eventually ans having internal xrefs will be nice. the gsuite data is incidental, what is truly valuable about gsuite is spam detection and the oauth capability
the common factor was sort of left as an exercise to the reader to think about moats in the age of AI... but basically anything that has touchpoints to the legal and financial systems im not gonna touch with a 20 ft vibecoded pole.
The incompetent have always pantomimed the competent. It never works. Although the incompetent will always pay a huge amount to try to achieve this fantasy.
This is what always confused me about VC AI enthusiasm. Their moat is the capital. As AI improves, it destroys their moat. And yet, they are stoked to invest in it, the architects of their own demise.
I just recreated most of Linear for my company in a few days. Making it hyper specific to what we want (metrics driven, lean startup style).
All state changes are made with MCP so it saved me from having to spend time on any forms and most interactions other than filtering searching sorting etc.
Means we will be ditching Linear soon.
I know I’m an outlier but this sort of thing will get more common.
I don't understand this because who's gonna maintain it in the future? Surely that costs more to pay even one person to add features that Linear had than to pay Linear themselves. I'd do this for personal projects but never for my work company lest I be the one to maintain it indefinitely on top of my current work.
one thing annoying with premade solutions is that it only does 90% of what you want, its livable but still doesn't quite meet your needs.
Its not just adding features that Linear already provides but adding features and integrations that mets 100% your needs.
The full decision making equation is (cost of implementing it yourself + cost of maintenance + 10% additional benefit for a solution that fully meets your needs) versus (cost of preexisting solution that meets 90% of your needs). Cost of implementing it and cost of maintenance has just gone down. Surely that will mean on a whole more people as a whole will choose to make inhouse rather than outsource.
Thus demand for premade solutions will go down, Saas providers won't be able to increase their prices as this will make even more people choose to implement it themselves. The cost of producing software will continue to drop due to agentic coding and maintenance cost will drop as well due to maintenance coding agents. More people will choose their own custom solutions and so on. Its very possible we are in the beginning of the end for Saas companies.
All these articles seem to think people will vibe code by prompting:
make me my own Stripe
make me my own Salesforce
make me my own Shopify
It will be more like:
Look at how Lago, an open-source Stripe layer, works and make it work with Authorized.net directly
Look at Twenty, an open-source CRM, and make it work in our tech stack for our sales needs
Look at how Medusa, an open-source e-commerce platform, works and what features we would need and bring into our website
When doing the latter, getting a good enough alternative will reduce the need for commercial SaaS. On top of that, these commercial SaaS are bloated with features in their attempt to work with as many use cases as possible and configuring them is “coding” by another name. Throw in Enshittification and the above seems to the next logical move by companies looking to move off these apps.
The value in enterprise SaaS offerings isn't just the application functionality but the IaaS substrate underneath. The vendor handles server operations, storage, scalability, backups, security, compliance, etc. It might be easier for companies to vibe code their own custom applications now but LLMs don't help nearly as much with keeping those applications running. Most companies are terrible at technical operations. I predict we'll see a new wave of IaaS startups that sell to those enterprise vibe coders and undercut the legacy SaaS vendors.
I've been confronting this truth personally. For years I had a backlog of projects that I always put off because I didn't have the capacity. Now I have the capacity but without the know how to sell it. It turns out that everything comes back to sales and building human relationships. Sort of a prerequisite to having operations.
Velocity or one-shot capability isn't the move. It's making stuff that used to be traumatic just...normal now.
Google fucking vibe-coded their x86 -> ARM ISA changeover. It never would have been done without agents. Not like "google did it X% faster." Google would have let that sit forever because the labor economics of the problem were backwards.
That doesn't MATTER anymore. If you have some scratch, some halfway decent engineers, and a clear idea, you can build stuff that was just infeasible or impossible. all it takes is time and care.
Some people have figured this out and are moving now.
I think something like an x86 -> ARM change is perfect example of something where LLM assisted coding shines. lots of busywork (i.e. smaller tasks that don't require lots of context of the other existing tasks), nothing totally novel to do (they don't have to write another borg or spanner), easy to verify, and 'translation'. LLMs are quite good at human language translation, why should they be bad at translating from one inline assembly language to another?
Exactly, if the engineers know where to look for the solution in open-source code and point the AI there, it will get them there. Even if the language or the tech stack are different, AI is excellent at finding the seams, those spots where a feature connects to the underlying tech stack, and figuring out how the feature is really implemented, and bringing that over.
Sensible people would do that (asking for just the features they need), but look at us, are we sensible?
Most of us* are working for places whose analytics software transitively asks the user for permission to be tracked by more "trusted" partners than the number of people in a typical high school, which transitively includes more bytes of code than the total size of DOOM including assets, with a performance hit so bad that it would be an improvement for everyone if the visitor remote desktop-ed into a VM running Win95 on the server.
And people were complaining about how wasteful software was when Win95 was new.
* Possibly an exaggeration, I don't know what business software is like; but websites and, in my experience at least, mobile apps do this.
So maybe the saas will pivot to just sell some barebone agents that include their real IP? The rest (UI, dashboards and connectivity) will be tailored made by LLMs
First, a vendor will have the best context on the inner workings and best practices of extending the current state of their software. The pressure on vendors to make this accessible and digestable to agents/ LLMs will increase, though.
Secondly, if you have coded with LLM assistance (not vibe coding), you will have experienced the limited ability of one shot stochastic approaches to build out well architected solutions that go beyond immediate functionality encapsulated in a prompt.
Thirdly, as the article mentions, opportunity cost will never make this a favorable term - unless the SaaS vendor was extorting prices before. The direct cost of mental overhead and time of an internal team member to hand-hold an agent/ write specs/ debug/ firefight some LLM assisted/ vibe coded solution will not outweigh the upside potential of expanding your core business unless you're a stagnant enterprise product on life support.
A few reasons, "AI" as used by non-experts often has correctness and security issues. Even when it doesn't, its outputs are often not reproducible/predictable because they're probabilistic systems.
AI systems are also prone to writing code which they can't effectively refactor themselves, implying that many of these code bases are fiscal time bombs where human experts are required to come fix them. If the service being replaced has transactional behaviour, does the AI produced solution? Does the person using it know what that means?
The other side is that AI as an industry still needs to recoup trillions in investment, and enterprise users are potential whales for that. Good prices in AI systems today are not guaranteed to last because even with hardware improvements these systems need to make money back that has been invested in them.
Some of that latter part depends on how good and cheap open weight systems get. The ability to deploy your own will strictly limit the price of closed models if they aren't dominant in functionality.
Sadly, this is the most serious comment here. People who are not shocked are people who haven’t seen what a highly educated computer scientist can do in single player mode.
I'll take all comers, any conceivable combination of unassisted engineers of arbitrary Carmack/God-level ability, no budgetary limits, and I'll bet my net worth down to starvation poverty that I will clobber them flat by myself. This is not because I'm such hot shit, it's a weird Venn that puts me on the early side on this, but there are others and there will be many more as people see the results.
So there are probably people who can beat me today, and that probability goes to one as Carmack-type people go full "press the advantage" mode on a long enough timeline, there are people who are strictly more talented and every bit as passionate, and the paradigm will saturate.
Which is why I spend all my time trying to scale it up, I'm working on how to teach other people how to do it, and solve the bottlenecks that emerge. That's a different paradigm that saturates in a different place, but it is likewise sigmoid-shaped.
That, and not single-player heroics, stunts basically, is the next thousand-year paradigm. And no current Valley power player even exists in that world. So the competition I have to worry about is very real, but not at all legible.
I don't know much about how this will play other than it's the fucking game at geopolitical levels, and the new boss will look nothing like the old boss.
>> Anish Acharya says it is not worth it to use AI-assisted coding for all business functions. AI should focus on core business development, not rebuilding enterprise software.
I don't even know what this means, but my take: we should stop listening to VCs (especially those like A16Z) who have an obvious vested interest that doesn't match the rest of society. Granting these people an audience is totally unwarranted; nobody but other tech bros said "we will vibe code everything" in the first place. Best case scenario: they all go to the same exclusive conference, get the branded conference technical vest and that's were the asteroid hits.
Both AI Fanatics and AI Luddites need to touch grass.
We work in Software ENGINEERING. Engineering is all about what tools makes sense to solve a specific problem. In some cases, AI tools do show immediate business value (eg. TTS for SDR) and in other cases this is less obvious.
This is all the more reason why learning about AI/ML fundamentals is critical in the same way understanding computer architecture, systems programming, algorithms, and design principles are critical to being a SWE, because then you can make a data-driven judgment on whether an approach works or not.
Given the number of throwaway accounts that commented, it clearly struck a nerve.
There is a lot of work that goes on before even reaching the point to write code.
For example, being able to vibecode a UI wireframe instead of being blocked for 2 sprints by your UI/UX team or templating an alpha to gauge customer interest in 1 week instead of 1 quarter is a massive operational improvement.
Of course these aren't completed products, but customers in most cases can accept such performance in the short-to-medium term or if it is part of an alpha.
This is why I keep repeating ad nauseum that most decisionmakers don't expect AI to replace jobs. The reality is, professional software engineering is about translating business requirements into tangible products.
It's not the codebase that matters in most cases - it's the requirements and outcomes that do. Like you can refactor and prettify your codebase all you want, but if it isn't directly driving customer revenue or value, then that time could be better spent elsewhere. It's the usecase that your product enables which is why they are purchasing your product.
Arguably it does, because being able to experience something gives you much more insight into whether it's right or not - so being able to iterate quickly many times, continuously updating your spec and definition of done should help you get to the right solution. To be clear, there is still effort involved, but the effort becomes more about the critical evaluation rather than the how.
To illustrate, I'll share what I'm working on now. My companies ops guy vibe coded a bunch of scripts to manage deployments. On the surface, they appear to do the correct thing. Except they don't. The tag for the Docker image used is hardcoded in a yaml file and doesn't get updated anywhere unless you do it manually. The docs don't even mention half of the necessary scripts/commands or implicit setup necessary for any of it to work in the first place, much less the tags or how any of it actually works. There are two completely different deployment strategies (direct to VM with docker + GCP and a GKE-based K8s deploy). Neither fully work, and only one has any documentation at all (and that documentation is completely vibed, so has very low information density). The only reason I'm able to use this pile of garbage at all is because I already know how all of the independent pieces function and can piece it together, but that's after wasting several hours of "why the fuck aren't my changes having an effect." There are very, very few lines of code that don't matter in well architected systems, but many that don't in vibed systems. We already have huge problems with overcomplicated crap made exclusively by humans, that's been hard enough to manage.
Vibe coding consistently gives the illusion of progress by fixing an immediate problem at the expense of piling on crap that obscures what's actually going on and often breaks exiting functionality. It's frankly not sustainable.
That being said, I've gotten some utility out of vibe coding tools, but it mostly just saves me some mental effort of writing boring shit that isn't interesting, innovative, or enjoyable, which is like 20% of mental effort and 5% of my actual work. I'm not even going to get started on the context switching costs. It makes my ADHD feel happy but I'm confident I'm less productive because of the secondary effects.
I really hate the expression "the new normal", because it sort of smuggles in the assumption that there exists such thing as "normal". It always felt like one of those truisms that people say to exploit emotions like "in these trying times" or "no one wants to work anymore".
But I really do think that vibe coding is the "new normal". These tools are already extremely useful, to a point where I don't really think we'll be able to go back. These tools are getting good enough that it's getting to a point where you have to use them. This might sound like I'm supportive of this, and I guess am to some extent, but I find it to be exceedingly disappointing because writing software isn't fun anymore.
One of my most upvoted comments on HN talks about how I don't enjoy programming, but instead I enjoy problem solving. This was written before I was aware of vibe coding stuff, and I think I was wrong. I guess I actually did enjoy the process of writing the code, instead of just delegating my work to a virtual intern while I just watch the AI do the fun stuff.
A very small part of me is kind of hoping that once AI has to be priced at "not losing money on every call" levels that I'll be forced to actually think about this stuff again.
I have thought about this a lot, and I have no idea. I work for an "AI-first" company, and we're kind of required to use AI stuff as often as we can, so I make very liberal use of Codex, but I've been shielded from the interview process thus far.
I think I would still kind of ask the same questions, though maybe a bit more conceptual. Like, for example, I might see if I could get someone to explain how to build something, and then ask them about data structures that might be useful (e.g. removing a lock by making an append-only structure). I find that Codex will generally generate something that "works" but without an understanding data structures and algorithms, its implementation will still be somewhat sub-optimal, meaning that understanding the fundamentals has value, at least for now.
Let's just look at Dijkstra's On the Foolishness of "Natural Language Programming". It really does a good job at explaining why natural language programming (and thus, Vibe Coding) is a dead end. It serves as a good reminder that we developed the languages of Math and Programming for a reason. The pedantic nature is a feature, not a flaw. It is because in programming (and math) we are dealing with high levels of abstraction constantly and thus ambiguity compounds. Isn't this something we learn early on as programmers? That a computer does exactly what you tell it to, not what you intend to tell it to? Think about how that phrase extends when we incorporate LLM Coding Agents.
| The virtue of formal texts is that their manipulations, in order to be legitimate, need to satisfy only a few simple rules; they are, when you come to think of it, an amazingly effective tool for ruling out all sorts of nonsense that, when we use our native tongues, are almost impossible to avoid.
- Dijkstra
All of you have experienced the ambiguity and annoyances of natural language. Have you ever:
- Had a boss give you confusing instructions?
- Argued with someone only to find you agree?
- Talked with someone and one of you doesn't actually understand the other?
- Talked with someone and the other person seems batshit insane but they also seem to have avoided a mental asylum?
- Use different words to describe the same thing?
- When standing next to someone and looking at the same thing?
- Adapted your message so you "talk to your audience"?
- Ever read/wrote something on the internet? (where "everyone" is the audience)
Congrats, you have experienced the frustrations and limitations of natural language. Natural language is incredibly powerful and the ambiguity is a feature and a flaw, just like how in formal languages the precision is both a feature and a flaw. I mean it can take an incredible amount of work to say even very simple and obvious things with formal languages[1], but the ambiguity disappears[2].
Vibe Coding has its uses and I'm sure that'll expand, but the idea of it replacing domain experts is outright laughable. You can't get it to resolve ambiguity if you aren't aware of the ambiguity. If you've ever argued with the LLM take a step back and ask yourself, is there ambiguity? It'll help you resolve the problem and make you recognize the limits. I mean just look at the legal system, that is probably one of the most serious efforts to create formalization in natural language and we still need lawyers and judges to sit around and argue all day about all the ambiguity that remains.
I seriously can't comprehend how on a site who's primary users are programmers this is an argument. If we somehow missed this in our education (formal or self) then how do we not intuit it from our everyday interactions?
[2] Most programming languages are some hybrid variant. e.g. Python uses duck typing: if it looks like a float, operates like a float, and works as a float, then it is probably a float. Or another example even is C, what used to be called a "high level programming language" (so is Python a celestial language?). Give up some precision/lack of ambiguity for ease.
It's extra sad because they would be more profitable if they recognized this.
Sometimes I wonder why companies are so resistant to making profits. It can be really strange. To be so profit focused yet throw away so much just because it is a bit more effort or a bit slower. But I guess most people are penny wise and pound foolish.
> Vibe Coding has its uses and I'm sure that'll expand, but the idea of it replacing domain experts is outright laughable.
I don't think that's the argument. The argument I'm seeing most is that most of us SWEs will become obsolete once the agentic tools become good enough to allow domain experts to fully iterate on solutions on their own.
> The argument I'm seeing most is that most of us SWEs will become obsolete once the agentic tools become good enough to allow domain experts to fully iterate on solutions on their own.
That’s been the argument since the 5PL movement in the 80s. What we discover is that domain expertise an articulation of domain expertise into systems are two orthogonal skills that occasionally develop in the same person but, in general, requires distinct specialization.
> The argument I'm seeing most is that most of us SWEs will become obsolete
That is equivalent to "replacing domain experts", or at least was my intent. But language is ambiguous lol. I do think programmers are domain experts. There are also different kinds of domain experts but I very much doubt we'll get rid of SWEs.
Though my big concern right now is that we'll get rid of juniors and maybe even mid levels. There's definitely a push for that and incentives from an economic point of view. But it will be disastrous for the tech industry if this happens. It kills the pipeline. There can be no wizards without noobs. So we have a real life tragedy of the commons situation staring us in the face. I'm pretty sure we know what choices will be made, but I hope we can recognize that there's going to need to be cooperation to solve this least we all suffer.
Dijkstra also said no one should be debugging and yet here we are.
He's not wrong about the problems of natural language YET HERE ARE. That would, I think, cause a sensible engineer to start poking at the predicate instead of announcing that the foregone conclusion is near.
We should take seriously the possibility that this isn't going to be in a retrenchment which bestows a nice little atta boy sticker on all the folks who said I told you so.
Given how you're implying things, you're grossly misrepresenting what he said. You've either been misled or misread. He was advocating for the adoption and development of provably correct programming.
Interestingly I think his "gospel" is only more meaningful today.
| Apparently, many programmers derive the major part of their intellectual satisfaction and professional excitement from not quite understanding what they are doing. In this streamlined age, one of our most under-nourished psychological needs is the craving for Black Magic, and apparently the automatic computer can satisfy this need for the professional software engineers, who are secretly enthralled by the gigantic risks they take in their daring irresponsibility. They revel in the puzzles posed by the task of debugging. They defend —by appealing to all sorts of supposed Laws of Nature— the right of existence of their program bugs, because they are so attached to them: without the bugs, they feel, programming would no longer be what is used to be! (In the latter feeling I think —if I may say so— that they are quite correct.)
| A program can be regarded as an (abstract) mechanism embodying as such the design of all computations that can possibly be evoked by it. How do we convince ourselves that this design is correct, i.e. that all these computations will display the desired properties? A naive answer to this question is "Well, try them all.", but this answer is too naive, because even for a simple program on the fastest machine such an experiment is apt to take millions of years. So, exhaustive testing is absolutely out of the question.
| But as long as we regard the mechanism as a black box, testing is the only thing we can do. The unescapable conclusion is that we cannot afford to regard the mechanism as a black box
I hear and read so much shit by VCs. Both in LinkedIn and in private meetings. Specially Menlo says a lot of shit (check LinkedIn). Deloitte and McKinsey, also full of crap. Really.
Vcs are choke full of companies that can be cloned over night, SaaS companies that will face ridiculously fast substitution, and a whoooole lotta capital deployed on lousy RAGs and OpenAI Wrappers.
2) They named one of the two authors of the Fascist Manifesto their patron saint
3) AI systems are built to function in ways that degrade and are likely to destroy our crucial civic institutions. (Quoted from
Professor Woodrow Hartzog "How AI Destroys Institutions"). Or to put it another way, being plausible but slightly wrong and un-auditable—at scale—is the killer feature of LLMs and this combination of properties makes it an essentially fascist technology meaning it is well suited to centralizing authority, eliminating checks on that authority and advancing an anti-science agenda (quoted from the A plausible, scalable and slightly wrong black box: why large language models are a fascist technology that cannot be redeemed post).
I will not claim to be an expert historian but one general belief I have is that nomenclature undergoes semantic migration over a century. So for the sake of conciseness I will quote the first demand of each portion of the Fascist Manifesto. This isn't to obscure, because it is in Wikipedia[0] and translated in English on EN Wikipedia[1], but so I can share a sample of whether this is something we can relate to our present day political orientation. Hopefully it will inform what you believe "author of the Fascist Manifesto" to imply:
> ...
> For this WE WANT:
> On the political problem:
> Universal suffrage by regional list voting, with proportional representation, voting and eligibility for women.
> ...
> On the social problem:
> WE WANT:
> The prompt enactment of a state law enshrining the legal eight-hour workday for all jobs.
> ...
> On the military issue:
> WE WANT:
> The establishment of a national militia with brief educational services and exclusively defensive duty.
> ...
> On the financial problem:
> WE WANT:
> A strong extraordinary tax on capital of a progressive nature, having the form of true PARTIAL EXPROPRIATION of all wealth.
Sure. They're making a strong claim, but I think they mean "author of the Fascist Manifesto" as shorthand to say Marinetti was an ardent supporter of fascism and Mussolini. His support continued throughout the 30's and 40's, even after the Pact of Steel and the Racial Laws etc, even volunteering to go to the Eastern Front. I think we can say with the benefit of hindsight that the fascists' attempts to ingratiate themselves to the worker's movement were sort of ancilliary to the whole political/ideological project... I mean I'd hope any student of history agrees with that...
I’m not particularly political and am also not a historian but I don’t think it’s necessarily correct to equate the literal text of the manifesto with the principles and practices of fascism.
The message of universal suffrage vs. that of preventing an out group from “stealing” an election are not far apart semantically. Same with workers rights - in practice the worker protection laws that were passed in Italy at this time were so full of loopholes and qualifications that ultimately the workers do not gain power in that system.
It is this fair, in my view, to question the spirit of the manifesto in the first place.
They don’t make money by being right, they make money by exposing LPs to risk. Zero commitment to insight. Intellectual production goes only so far as to attract funding.
Also... they don't make money by promoting things that are good ideas that make sense. That's why every lucky billionaire tech bro that gets into VC ultimately invests in smart toilets. Ultimately, they just keep putting money into each slot machine they can find until one of them pays out a jackpot. Eventually one of them will make up for all the other losses.
Well, yeah. Vibe coding as in letting AI one-shot an app with a vague description still doesn't work except on trivial, throwaway stuff. But... spec-driven development with automated stepwise refinement by agents recursively generating, testing, and improving the code is how software engineering is done in the late 2020s.
You write that in italics as if to imply it’s a law that cannot be questioned. Quite a number of shops do not engineer software like that, or only engineer software like that where it fits the environment the software lives in, or otherwise sit at numerous points along the gradient between “software engineering as it has been known for decades” and “fully computer generated software”.
> "You have this innovation bazooka with these models. Why would you point it at rebuilding payroll or ERP or CRM"
Most SaaS companies are just expensive wrappers on top of existing tools. For non-VC-funded companies, SaaS tools are a serious cost. If you can re-create them in-house with AI, why wouldn't you? The result is saving capital (which you can then employ to do the more innovative things), and being in control over your own data.
Even a16z is walking this back now. I wrote about why the “vibe code everything” thesis doesn’t hold up in two recent pieces:
(1) https://philippdubach.com/posts/the-saaspocalypse-paradox/
(2) https://philippdubach.com/posts/the-impossible-backhand/
Acharya’s framing is different from mine (he’s talking book on software stocks) but the conclusion is the same: the “innovation bazooka” pointed at rebuilding payroll is a bad allocation of resources. Benedict Evans called me out on LinkedIn for this (https://philippdubach.com/posts/is-ai-really-eating-the-worl...) take, which I take as a sign the argument is landing..
> investors are simultaneously punishing hyperscaler stocks because AI capex might generate weak returns, while destroying software stocks because AI adoption will be so pervasive it renders all existing software obsolete. Both cannot hold simultaneously.
I don't understand this point. Can't it be possible that the ultimate effect is to devalue, hugely, software? As in it can totally both be true that AI capex has weak returns and at the same time most SaaS companies go bankrupt. To take an analogy: if ever we manage to successfully mine asteroids, and find some vast quantity of platinum, it could both be true that every existing platinum miner loses their shirt, and also that the value of platinum sinks so far that the asteroid mining company cannot cover its costs.
> Benedict Evans called me out on LinkedIn for this take, which I take as a sign the argument is landing.
Excellent. And correct lol.
The fact that this is getting downvoted gave me a hearty chuckle. Never change, HN.
I once built a CRM in Google Sheets fully mirroring the data model of Salesforce. For contact, company, deal, and call tracking for a one sales rep business. (Before XLookup was in Google Sheets)
Did it work? Yes. Was it worth my time to maintain and scale the “platform” with the company rather than outsource all that to a CRM company? Not at all.
Time is finite. Spend your time doing what you do best, pay others to do what they do best.
Yup, my experience has been that vibe-coding is very time-consuming. It reminds me very much of how LLMs are great at creating mind-blowing images, but you get what you get. Once you decide that you need to modify the image you get, it becomes a time sink. You might be able to change it and get what you need, but there is no guarantee and it's a never ending task.
The same thing happens with code; you may get great results from your prompt, but trying to customize it will drive you nuts and you may never get what you want.
Maintenance is another hurdle. How do you maintain code you might not have the skills to maintain?
Vibe-coding may reduce software creation time, but it's not taking over software engineering. The SaaS business is going nowhere. Most people, by far, will continue to rely on someone else for their software needs. But be very aware that the software business will change. We are seeing that already.
It doesn't make sense for every company to make their own Salesforce clone.
The key is that it makes new companies entering the market to compete with Salesforce immensely easier. More competition will just force lower overall margins in SAAS.
It's not really that hard to make a Salesforce clone now though. Writing the software was never the hard part of building a business.
2 replies →
People are overestimating the value on having AI create something given loose instructions, and underestimating the value of using AI as a tool for a human to learn and explore a problem space. The bias shows on the terminology (“agents”).
We finally made the computer able to speak “our” language - but we still see computers as just automation. There’s a lot of untapped potential in the other direction, in encoding and compressing knowledge IMO.
Because that would mean AI isn't going to replace entire industries, which is the only way to justify the, not billions, but trillions in market value that AI leaders keep trying to justify.
Problem space is rich. The thing doesnt actually know what a problem is.
The thing is incredibly good at searching through large spaces of information.
42
5 replies →
Exactly my thoughts - the value in AI is not auto-generating anything more than something trivial, but there's huge value in a more customized knowledge engine - a targeted, specific Google if you will. Get answers to your specific question instead of results that might contain what you were looking for if you slog through them.
AI is hugely beneficial in understanding a problem, or at least getting a good overview, so you can then go off and solve/do it yourself, but focusing on "just have the AI generate a solution" is going to hugely harm AI perception/adoption.
> AI create something
To have AI recreate something that was already in it's training set.
> in encoding and compressing knowledge IMO.
I'd rather have the knowledge encoded in a way that doesn't generate hallucinations.
Thought exercise for those in disagreement: why would every company use AI to build their own payroll/ERP/CRM, when just a handful of companies could use AI to build those offerings better?
This is largely how things work now; AI may lower the cost and increase margins, but the economics of build vs buy seem the same.
To avoid CRAZY SaaS charges. I left a comment further down about how the challenge is first getting a reliable stack running underneath whatever ends up being fast-coded. The trend will be more decentralization - I think that'll be AI 2.0. Increasing centralization is AI 1.0.
Slack is a good example. When the cost of Slack is an unreasonable amount of your operating costs then it makes sense to clone and maintain. The product is simple, you can basically recreate the main functionality in a sitting. Why would you pay hundreds of thousands of dollars for it?
Mattermost is FOSS. Why aren't companies running their own servers to avoid Slack?
Slack is an hilarious example.
I can't wait for orgs to try to vibe roll their own dozen clients, security models, and then try to talk to handle external integrations of some kind.
That’s a fine example, but my question then is why does Slack exist? Surely Fortune 500 companies are smart enough to realize that building a slack clone is cheaper, yet they don’t do that.
So now consider AI, perhaps the cost of building has decreased from 100k to 10k. What stops a Slack competitor from also building the product for 10k and reselling it at 10% of the cost of Slack? My point is that I don’t see how AI has changed the value prop.
2 replies →
If Slack is so simple why haven’t companies created their own internal versions 10 years ago?
5 replies →
Every company that I’ve worked at has had to do significant additional development work on their instance of salesforce to make it work for them. Like 6-12 months of work with 1-3 people. I don’t know if this is common but in that case maybe going custom might be the way to go. You get something lean, without all the cruft, specifically built for your usecase and nothing more.
Well the answer is because the cost of that software is lower than somebody building the other software. What happens is that all these SaaS drop in value because it is now realistic to build them internally
Why does AI make it cheaper to build internal but not cheaper for SaaS competitors to pop up? Everyone has access to the same tools.
3 replies →
Anyone who's seen an enterprise deal close or dealt with enterprise customer requests will know this, the build vs buy calculus has always been there yet companies still buy. Until you can get AI to the point where it equivalent to a 20 person engineering team, people are not going to build their own Snowflake, Salesforce, Slack or ATS. Maybe that day is 3 years away but when that happens the world will be very different
Companies do make/buy decisions on everything, it just software. Cleaning services are not expensive, yet companies contract them instead of hiring staff.
This is called transaction cost economics, if anyone’s interested.
I agree generally, but some of these enterprise contracts are eye-watering. If the choice is $2M/year with a 3-year minimum contract, or rolling your own, I think calculus really has shifted.
With that said, the entire business world does not understand that software is more than just code. Even if you could write code instantly, making enterprise software would still take time, because there are simply so many high-stakes decisions to make, and so much fractal detail.
> If the choice is $2M/year with a 3-year minimum contract, or rolling your own, I think calculus really has shifted.
But why? It was always dramatically cheaper for enterprises to build rather than buy. They stopped doing that becuase they did that in the 90s and ended up with legacy codebases that they didn't know how to maintain. I can't see AI helping with that.
1 reply →
> Until you can get AI to the point where it equivalent to a 20 person engineering team
I think that’s gonna happen when you don’t need software and AI just does it all.
Exactly. I was building an app to track bike part usage. It was an okay app, but then I just started using ai with the database directly. Much more flexible, and I can get anything I need right then. AI will kill a lot of companies, but it won’t be the software it develops, it will be the agent itself
5 replies →
Imagine a 20 person engineering team that hallucinates on a regular basis and is incapable of innovation.
I think you’ve just described an average Accenture setup.
If an AI agent ever became as productive at writing code as a well-organized 20 person engineering team you'd still need to run it for a year or more to replicate any nontrivial SaaS product.
And the thing about many of these products isn't their feature set, it's their stability. It's their uptime. It's how they handle scaling invisibly and with no effort on your part. These are things you can't just write down from whole cloth, they are properties that emerge over time by adapting the the reality of scale. Coding isn't the whole deal, and your 20x clanker which can do nothing but re-arrange text in interesting patterns is going to have some trouble with the realities of taking that PoC to production. You'll still need experienced, capable people for that. And lots of time.
A lot of this "ermahgerd everything will change" drivel is based on some magical fundamentally new technology emerging in the near future that can do things that LLMs cannot do. But as far as anyone knows, that future may be never.
So even given a large improvement in agentic coding I'm not convinced it really changes the build vs buy equation much.
I sort of agree with this, but what a lot of people are missing is it's unbelievably easy to clone a lot of SaaS products.
So I think big SaaS products are under attack from three angles now:
1) People replacing certain systems with 'vibe coded' ones, for either cost/feature/unhappiness with vendor reasons. I actually think this is a bigger threat than people think - there are so many BAD SaaS products out there which cost businesses a fortune in poor features/bugs/performance/uptime, and if the models/agents keep improving the way they have in the last couple of years it's going to be very interesting if some sort of '1000x' engineer in an agent can do crazy impressive stuff.
2) Agents 'replacing' the software. As people have pointed out, just have the agent use APIs to do whatever workflow you want - ping a database and output a report.
3) "Cheap" clones of existing products. A tiny team can now clone a "big" SaaS product very quickly. These guys can provide support/infra/migration assistance and make money at a much lower price point. Even if there is lock in, it makes it harder for SaaS companies to keep price pressure up.
Insightful points!
It would be interesting if, with all the anxiety about vibe coding becoming the new normal, its only lasting effect is the emergence of smaller B2B companies that quickly razzle dazzle together a bespoke replacement for Concur, SAP, Workday, the crappy company sharepoint - whatever. Reminds me of what people say Palantir is doing, but now supercharged by the AI-driven workflows to stand up the “forward deployed” “solution” even faster.
Thanks,yes exactly what I think.
Or an industry specific Workday, with all of workdays features but aimed at a niche vertical.
I wrote about this (including an approach on how to clone apps with HAR files and agents) if you are interested. https://martinalderson.com/posts/attack-of-the-clones/
But have you ever tried to clone a product or tool for yourself before? At first it’s great because you think that you saved money but then you start having to maintain it… fixing problems, filling in gaps… you now realize that you made a mistake. Just because AI can do it now doesn’t mean you aren’t just now having to use AI to do the same thing…
Also, agents are not deterministic. If you use it to analyze data, it will get it right most of the time but, once in a blue moon, it will make shit up, except you can’t tell which time it was. You could make it deterministic by having AI write a tool instead… except you now have the first problem of maintaining a tool.
That isn’t to say that there isn’t small low hanging fruit that AI will replace, but it’s a bit different when you need a real product with support.
At the end of the day, you hire a plumber or use a SaaS not because you can’t do it yourself, but because you don’t want to do it and rather want someone else who is committed to it to handle it.
I'm not saying _the end user_ clones it. I mean someone else does (more efficiently with agents) and runs it as a _new_ SaaS company. They would provide support just like the existing one would, but arguably at a cheaper price point.
And regarding agents being non deterministic, if they write a bunch of SQL queries to a file for you, they are deterministic. They can just write "disposable" tools and scripts - not always doing it thru their context.
2 replies →
> He said that software accounts for 8% to 12% of a company's expenses, so using vibe coding to build the company's resource planning or payroll tools would only save about 10%. Relying on AI to write code also carries risks, he said.
> "You have this innovation bazooka with these models. Why would you point it at rebuilding payroll or ERP or CRM," Acharya said
> Instead, companies are better off using AI to develop their core businesses or optimize the remaining 90% of their costs
"You have this innovation bazooka. Why would you point it at rebuilding payroll?" — a partner at the firm whose thesis was literally "software is eating the world."
Apparently the meal is over and now we're just rearranging the plates.
It seems to be premised on the idea we would vibe code a replica of what we get from SaaS. But the real point is, we would not do that. We would vibe code something that exactly fits our business.
We have products we're paying $100k a year for and using 3% of the functionality. And they suck. This is the target.
The missing piece in this debate is that most "vibe coded" replacements break at scale. I tried replacing a multi-step workflow with Make.com + Airtable (not even vibe coding, just no-code automation) and it fell apart past 2 jobs per day - rate limits, webhook failures, state management nightmares. The real pattern I see working is not "replace SaaS with vibe code" but rather "stitch together 5-6 specialized tools with a thin orchestration layer you write yourself." The orchestration is where AI actually helps - it's glue code, not the product.
The bottleneck will always be humans. You could get AI to write a million lines of code a day, but you’d still need humans to review and test that code. We are a very long way from being able to blindly trust AI’s outputs in production.
I don’t even think it’s about reviewing and testing. The bottleneck will always be humans.
We don’t like to always admit it but most jobs are fairly straightforward, as in the actual day to day tasks. Yes being smart is great and useful etc. but after a certain point it’s diminishing returns on the actual tasks you have to do. Dealing with other humans and their egos and eccentricities and the multitude ways each person sees the world is always what makes all jobs tricky. I suspect this whole ai wave/hype/reality is going to open many people’s eyes to this. We will laugh that we use to call them “soft” skills.
IMO I would have agreed with this statement 2 months ago but now it’s clear AI is already much better at reviewing and even testing code (via spinning up simulators, etc) much better than we can. We’re already using AI’s outputs in production and not writing much code these days.
for code in isolation, perhaps, but how does it know what is correct for what the customer wants/needs?
1 reply →
There was a short moment in history where it seemed that the sentiment was: people will soon 3D-print 99% of their household items themselves instead of buying them.
You absolutely could print things like cups, soap holders, picture frames, the small shovel you use for gardening, and so on an so on.
99% of people still just buy this stuff.
That has more to do with the shortcomings of 3d printing.
I think some or maybe even many of those shortcomings will apply to software, too. Making actual good software is not as trivial as writing “make me an app”, much as making an actual good spoon is not as trivial as throwing an STL at a printer and calling it a day.
Are you saying vibed code doesn’t have shortcomings
> "You have this innovation bazooka with these models. Why would you point it at rebuilding payroll or ERP or CRM"
They invested in ERP/CRM? I built one (fairly complete to the German/Italy/EU tax system) and it saves a ton of money vs commercial offerings. So yeah, of course we will.
you cant easily vibecode everything. in my startup this is what I am not buying (and vibecoding):
- JIRA/trello/monday.com - benchling - obsidian
this is what i buy and have no intent to replace:
- carta - docusign - gusto/rippling - bank
this is what might be on the chopping block:
- gsuite
I'm curious about your reasoning. Jira/Trello etc. are like $10/mo/seat, why bother rewriting them from scratch? You'll spend more in tokens doing so. Same for gmail/google calendar, what's the ROI? Those tools are reliable and cheap, why bother creating your own?
jira/trello: ergonomics. to set them up correctly exactly the way i want would take me 20 hours (or hire a PM), i can vibecode for 20h and get the same result.
plus, being able to crossref internal data types is chef's kiss.
im paying for claude pro so it's use it or lose it. when i finish everything and have it battle tested i can end my claude code. and anyways when i have 10 employees, it's parity.
for gsuite: i want to own everything internally eventually ans having internal xrefs will be nice. the gsuite data is incidental, what is truly valuable about gsuite is spam detection and the oauth capability
Just in case you weren't aware, Gsuite has a clone of Docusign built into it now.
hate to say it, because who likes monopolies, but it's easier to send people docusign because then they don't go Wtf?
Why not Docusign? Not challenging, just curious why that is specifically on your list. Reputation?
the common factor was sort of left as an exercise to the reader to think about moats in the age of AI... but basically anything that has touchpoints to the legal and financial systems im not gonna touch with a 20 ft vibecoded pole.
The possibility that anyone can easily replicate any startup scares A16Z.
The incompetent have always pantomimed the competent. It never works. Although the incompetent will always pay a huge amount to try to achieve this fantasy.
This is what always confused me about VC AI enthusiasm. Their moat is the capital. As AI improves, it destroys their moat. And yet, they are stoked to invest in it, the architects of their own demise.
Don't you have that backwards? If AI gets so good that it can replace all human labor, will capital like money and data centers be the only moat left?
8 replies →
There’s no alternative, they can’t collectively freeze out all AI investment and force it to die.
I just recreated most of Linear for my company in a few days. Making it hyper specific to what we want (metrics driven, lean startup style).
All state changes are made with MCP so it saved me from having to spend time on any forms and most interactions other than filtering searching sorting etc.
Means we will be ditching Linear soon.
I know I’m an outlier but this sort of thing will get more common.
I don't understand this because who's gonna maintain it in the future? Surely that costs more to pay even one person to add features that Linear had than to pay Linear themselves. I'd do this for personal projects but never for my work company lest I be the one to maintain it indefinitely on top of my current work.
one thing annoying with premade solutions is that it only does 90% of what you want, its livable but still doesn't quite meet your needs.
Its not just adding features that Linear already provides but adding features and integrations that mets 100% your needs.
The full decision making equation is (cost of implementing it yourself + cost of maintenance + 10% additional benefit for a solution that fully meets your needs) versus (cost of preexisting solution that meets 90% of your needs). Cost of implementing it and cost of maintenance has just gone down. Surely that will mean on a whole more people as a whole will choose to make inhouse rather than outsource.
Thus demand for premade solutions will go down, Saas providers won't be able to increase their prices as this will make even more people choose to implement it themselves. The cost of producing software will continue to drop due to agentic coding and maintenance cost will drop as well due to maintenance coding agents. More people will choose their own custom solutions and so on. Its very possible we are in the beginning of the end for Saas companies.
1 reply →
A16Zs opinion is worthless to me, they know very little about the market. Furthermore, they're notorious for having a lot of "partners".
Their whole game is just pump and dump
Depends on the partner, Peter Levine is a pretty damn good picker (supported us series A to IPO). https://en.wikipedia.org/wiki/Peter_J._Levine
Pretty worthless take posting an ad-hominem attack instead of addressing the actual content of the article/ statement.
Has everyone forgotten about when they pumped absurd crypto scams like NFTs
A16z partners don't know shit. Brain dead nepo-babies - how's Cluely?
AI is eating the software https://a16z.com/why-software-is-eating-the-world/
I can’t believe I’m responding to an AOL article, but…
You don’t understand what’s happening if you dismiss the leverage provided by AI as “vibe coding”.
All these articles seem to think people will vibe code by prompting:
make me my own Stripe
make me my own Salesforce
make me my own Shopify
It will be more like:
Look at how Lago, an open-source Stripe layer, works and make it work with Authorized.net directly
Look at Twenty, an open-source CRM, and make it work in our tech stack for our sales needs
Look at how Medusa, an open-source e-commerce platform, works and what features we would need and bring into our website
When doing the latter, getting a good enough alternative will reduce the need for commercial SaaS. On top of that, these commercial SaaS are bloated with features in their attempt to work with as many use cases as possible and configuring them is “coding” by another name. Throw in Enshittification and the above seems to the next logical move by companies looking to move off these apps.
The value in enterprise SaaS offerings isn't just the application functionality but the IaaS substrate underneath. The vendor handles server operations, storage, scalability, backups, security, compliance, etc. It might be easier for companies to vibe code their own custom applications now but LLMs don't help nearly as much with keeping those applications running. Most companies are terrible at technical operations. I predict we'll see a new wave of IaaS startups that sell to those enterprise vibe coders and undercut the legacy SaaS vendors.
I've been confronting this truth personally. For years I had a backlog of projects that I always put off because I didn't have the capacity. Now I have the capacity but without the know how to sell it. It turns out that everything comes back to sales and building human relationships. Sort of a prerequisite to having operations.
Are the infrastructure tools available already not easy enough to build on? We have all these serverless options already.
The right move is this, turned to 11.
Velocity or one-shot capability isn't the move. It's making stuff that used to be traumatic just...normal now.
Google fucking vibe-coded their x86 -> ARM ISA changeover. It never would have been done without agents. Not like "google did it X% faster." Google would have let that sit forever because the labor economics of the problem were backwards.
That doesn't MATTER anymore. If you have some scratch, some halfway decent engineers, and a clear idea, you can build stuff that was just infeasible or impossible. all it takes is time and care.
Some people have figured this out and are moving now.
I think something like an x86 -> ARM change is perfect example of something where LLM assisted coding shines. lots of busywork (i.e. smaller tasks that don't require lots of context of the other existing tasks), nothing totally novel to do (they don't have to write another borg or spanner), easy to verify, and 'translation'. LLMs are quite good at human language translation, why should they be bad at translating from one inline assembly language to another?
5 replies →
Exactly, if the engineers know where to look for the solution in open-source code and point the AI there, it will get them there. Even if the language or the tech stack are different, AI is excellent at finding the seams, those spots where a feature connects to the underlying tech stack, and figuring out how the feature is really implemented, and bringing that over.
> Google would have let that sit forever because the labor economics of the problem were backwards.
This has been how all previous innovations that made software easier to make turned out.
People found more and more uses for software and that does seem to be playing out again.
1 reply →
Google3 was already PPC clean when they did that. Not as impressive as made out to be.
Sensible people would do that (asking for just the features they need), but look at us, are we sensible?
Most of us* are working for places whose analytics software transitively asks the user for permission to be tracked by more "trusted" partners than the number of people in a typical high school, which transitively includes more bytes of code than the total size of DOOM including assets, with a performance hit so bad that it would be an improvement for everyone if the visitor remote desktop-ed into a VM running Win95 on the server.
And people were complaining about how wasteful software was when Win95 was new.
* Possibly an exaggeration, I don't know what business software is like; but websites and, in my experience at least, mobile apps do this.
So maybe the saas will pivot to just sell some barebone agents that include their real IP? The rest (UI, dashboards and connectivity) will be tailored made by LLMs
I highly doubt that, and its in OPs article.
First, a vendor will have the best context on the inner workings and best practices of extending the current state of their software. The pressure on vendors to make this accessible and digestable to agents/ LLMs will increase, though.
Secondly, if you have coded with LLM assistance (not vibe coding), you will have experienced the limited ability of one shot stochastic approaches to build out well architected solutions that go beyond immediate functionality encapsulated in a prompt.
Thirdly, as the article mentions, opportunity cost will never make this a favorable term - unless the SaaS vendor was extorting prices before. The direct cost of mental overhead and time of an internal team member to hand-hold an agent/ write specs/ debug/ firefight some LLM assisted/ vibe coded solution will not outweigh the upside potential of expanding your core business unless you're a stagnant enterprise product on life support.
Why is it bad for AI to replace an enterprise software layer? Other than invalidating past investments.
A few reasons, "AI" as used by non-experts often has correctness and security issues. Even when it doesn't, its outputs are often not reproducible/predictable because they're probabilistic systems.
AI systems are also prone to writing code which they can't effectively refactor themselves, implying that many of these code bases are fiscal time bombs where human experts are required to come fix them. If the service being replaced has transactional behaviour, does the AI produced solution? Does the person using it know what that means?
The other side is that AI as an industry still needs to recoup trillions in investment, and enterprise users are potential whales for that. Good prices in AI systems today are not guaranteed to last because even with hardware improvements these systems need to make money back that has been invested in them.
Some of that latter part depends on how good and cheap open weight systems get. The ability to deploy your own will strictly limit the price of closed models if they aren't dominant in functionality.
..said the guy, who doesn’t code anyway
Never say never, vibe coding is not even 4 years old.
Vibecoding is a net wealth transfer from frightened people to unscrupulous people.
Machine assisted rigorous software engineering is an even bigger wealth transfer from unscrupulous people to passionate computer scientists.
Sadly, this is the most serious comment here. People who are not shocked are people who haven’t seen what a highly educated computer scientist can do in single player mode.
Sure they have:
https://youtu.be/ghm9F0RCFsY
I'll take all comers, any conceivable combination of unassisted engineers of arbitrary Carmack/God-level ability, no budgetary limits, and I'll bet my net worth down to starvation poverty that I will clobber them flat by myself. This is not because I'm such hot shit, it's a weird Venn that puts me on the early side on this, but there are others and there will be many more as people see the results.
So there are probably people who can beat me today, and that probability goes to one as Carmack-type people go full "press the advantage" mode on a long enough timeline, there are people who are strictly more talented and every bit as passionate, and the paradigm will saturate.
Which is why I spend all my time trying to scale it up, I'm working on how to teach other people how to do it, and solve the bottlenecks that emerge. That's a different paradigm that saturates in a different place, but it is likewise sigmoid-shaped.
That, and not single-player heroics, stunts basically, is the next thousand-year paradigm. And no current Valley power player even exists in that world. So the competition I have to worry about is very real, but not at all legible.
I don't know much about how this will play other than it's the fucking game at geopolitical levels, and the new boss will look nothing like the old boss.
>> Anish Acharya says it is not worth it to use AI-assisted coding for all business functions. AI should focus on core business development, not rebuilding enterprise software.
I don't even know what this means, but my take: we should stop listening to VCs (especially those like A16Z) who have an obvious vested interest that doesn't match the rest of society. Granting these people an audience is totally unwarranted; nobody but other tech bros said "we will vibe code everything" in the first place. Best case scenario: they all go to the same exclusive conference, get the branded conference technical vest and that's were the asteroid hits.
Both AI Fanatics and AI Luddites need to touch grass.
We work in Software ENGINEERING. Engineering is all about what tools makes sense to solve a specific problem. In some cases, AI tools do show immediate business value (eg. TTS for SDR) and in other cases this is less obvious.
This is all the more reason why learning about AI/ML fundamentals is critical in the same way understanding computer architecture, systems programming, algorithms, and design principles are critical to being a SWE, because then you can make a data-driven judgment on whether an approach works or not.
Given the number of throwaway accounts that commented, it clearly struck a nerve.
The irony is, AI coding only works after and if you put a lot of work on engineering, like creating a factory.
There is a lot of work that goes on before even reaching the point to write code.
For example, being able to vibecode a UI wireframe instead of being blocked for 2 sprints by your UI/UX team or templating an alpha to gauge customer interest in 1 week instead of 1 quarter is a massive operational improvement.
Of course these aren't completed products, but customers in most cases can accept such performance in the short-to-medium term or if it is part of an alpha.
This is why I keep repeating ad nauseum that most decisionmakers don't expect AI to replace jobs. The reality is, professional software engineering is about translating business requirements into tangible products.
It's not the codebase that matters in most cases - it's the requirements and outcomes that do. Like you can refactor and prettify your codebase all you want, but if it isn't directly driving customer revenue or value, then that time could be better spent elsewhere. It's the usecase that your product enables which is why they are purchasing your product.
3 replies →
Just because we can code something faster or cheaper doesn't increase the odds it will be right.
Arguably it does, because being able to experience something gives you much more insight into whether it's right or not - so being able to iterate quickly many times, continuously updating your spec and definition of done should help you get to the right solution. To be clear, there is still effort involved, but the effort becomes more about the critical evaluation rather than the how.
But that's not the only problem.
To illustrate, I'll share what I'm working on now. My companies ops guy vibe coded a bunch of scripts to manage deployments. On the surface, they appear to do the correct thing. Except they don't. The tag for the Docker image used is hardcoded in a yaml file and doesn't get updated anywhere unless you do it manually. The docs don't even mention half of the necessary scripts/commands or implicit setup necessary for any of it to work in the first place, much less the tags or how any of it actually works. There are two completely different deployment strategies (direct to VM with docker + GCP and a GKE-based K8s deploy). Neither fully work, and only one has any documentation at all (and that documentation is completely vibed, so has very low information density). The only reason I'm able to use this pile of garbage at all is because I already know how all of the independent pieces function and can piece it together, but that's after wasting several hours of "why the fuck aren't my changes having an effect." There are very, very few lines of code that don't matter in well architected systems, but many that don't in vibed systems. We already have huge problems with overcomplicated crap made exclusively by humans, that's been hard enough to manage.
Vibe coding consistently gives the illusion of progress by fixing an immediate problem at the expense of piling on crap that obscures what's actually going on and often breaks exiting functionality. It's frankly not sustainable.
That being said, I've gotten some utility out of vibe coding tools, but it mostly just saves me some mental effort of writing boring shit that isn't interesting, innovative, or enjoyable, which is like 20% of mental effort and 5% of my actual work. I'm not even going to get started on the context switching costs. It makes my ADHD feel happy but I'm confident I'm less productive because of the secondary effects.
4 replies →
Iteration only matters when the feedback is used to improve.
Your model doesn't improve. It can't.
4 replies →
I dunno.
I really hate the expression "the new normal", because it sort of smuggles in the assumption that there exists such thing as "normal". It always felt like one of those truisms that people say to exploit emotions like "in these trying times" or "no one wants to work anymore".
But I really do think that vibe coding is the "new normal". These tools are already extremely useful, to a point where I don't really think we'll be able to go back. These tools are getting good enough that it's getting to a point where you have to use them. This might sound like I'm supportive of this, and I guess am to some extent, but I find it to be exceedingly disappointing because writing software isn't fun anymore.
One of my most upvoted comments on HN talks about how I don't enjoy programming, but instead I enjoy problem solving. This was written before I was aware of vibe coding stuff, and I think I was wrong. I guess I actually did enjoy the process of writing the code, instead of just delegating my work to a virtual intern while I just watch the AI do the fun stuff.
A very small part of me is kind of hoping that once AI has to be priced at "not losing money on every call" levels that I'll be forced to actually think about this stuff again.
I largely agree with you. And, given your points about “not going back” — how do you propose interviewing SWEs?
I have thought about this a lot, and I have no idea. I work for an "AI-first" company, and we're kind of required to use AI stuff as often as we can, so I make very liberal use of Codex, but I've been shielded from the interview process thus far.
I think I would still kind of ask the same questions, though maybe a bit more conceptual. Like, for example, I might see if I could get someone to explain how to build something, and then ask them about data structures that might be useful (e.g. removing a lock by making an append-only structure). I find that Codex will generally generate something that "works" but without an understanding data structures and algorithms, its implementation will still be somewhat sub-optimal, meaning that understanding the fundamentals has value, at least for now.
"aol.com"?
I cannot believe there has been no mention of things like n8n, activepieces and windmill in this thread. SaaS will utterly collapse in 18 months.
https://github.com/n8n-io/n8n
https://github.com/activepieces/activepieces
https://github.com/windmill-labs/windmill
Let's just look at Dijkstra's On the Foolishness of "Natural Language Programming". It really does a good job at explaining why natural language programming (and thus, Vibe Coding) is a dead end. It serves as a good reminder that we developed the languages of Math and Programming for a reason. The pedantic nature is a feature, not a flaw. It is because in programming (and math) we are dealing with high levels of abstraction constantly and thus ambiguity compounds. Isn't this something we learn early on as programmers? That a computer does exactly what you tell it to, not what you intend to tell it to? Think about how that phrase extends when we incorporate LLM Coding Agents.
All of you have experienced the ambiguity and annoyances of natural language. Have you ever:
Congrats, you have experienced the frustrations and limitations of natural language. Natural language is incredibly powerful and the ambiguity is a feature and a flaw, just like how in formal languages the precision is both a feature and a flaw. I mean it can take an incredible amount of work to say even very simple and obvious things with formal languages[1], but the ambiguity disappears[2].
Vibe Coding has its uses and I'm sure that'll expand, but the idea of it replacing domain experts is outright laughable. You can't get it to resolve ambiguity if you aren't aware of the ambiguity. If you've ever argued with the LLM take a step back and ask yourself, is there ambiguity? It'll help you resolve the problem and make you recognize the limits. I mean just look at the legal system, that is probably one of the most serious efforts to create formalization in natural language and we still need lawyers and judges to sit around and argue all day about all the ambiguity that remains.
I seriously can't comprehend how on a site who's primary users are programmers this is an argument. If we somehow missed this in our education (formal or self) then how do we not intuit it from our everyday interactions?
[0] https://www.cs.utexas.edu/~EWD/transcriptions/EWD06xx/EWD667...
[1] https://en.wikipedia.org/wiki/Principia_Mathematica
[2] Most programming languages are some hybrid variant. e.g. Python uses duck typing: if it looks like a float, operates like a float, and works as a float, then it is probably a float. Or another example even is C, what used to be called a "high level programming language" (so is Python a celestial language?). Give up some precision/lack of ambiguity for ease.
yes, but sadly many businesses don't care about any of that...
It's extra sad because they would be more profitable if they recognized this.
Sometimes I wonder why companies are so resistant to making profits. It can be really strange. To be so profit focused yet throw away so much just because it is a bit more effort or a bit slower. But I guess most people are penny wise and pound foolish.
> Vibe Coding has its uses and I'm sure that'll expand, but the idea of it replacing domain experts is outright laughable.
I don't think that's the argument. The argument I'm seeing most is that most of us SWEs will become obsolete once the agentic tools become good enough to allow domain experts to fully iterate on solutions on their own.
> The argument I'm seeing most is that most of us SWEs will become obsolete once the agentic tools become good enough to allow domain experts to fully iterate on solutions on their own.
That’s been the argument since the 5PL movement in the 80s. What we discover is that domain expertise an articulation of domain expertise into systems are two orthogonal skills that occasionally develop in the same person but, in general, requires distinct specialization.
2 replies →
That is equivalent to "replacing domain experts", or at least was my intent. But language is ambiguous lol. I do think programmers are domain experts. There are also different kinds of domain experts but I very much doubt we'll get rid of SWEs.
Though my big concern right now is that we'll get rid of juniors and maybe even mid levels. There's definitely a push for that and incentives from an economic point of view. But it will be disastrous for the tech industry if this happens. It kills the pipeline. There can be no wizards without noobs. So we have a real life tragedy of the commons situation staring us in the face. I'm pretty sure we know what choices will be made, but I hope we can recognize that there's going to need to be cooperation to solve this least we all suffer.
How do you get domain experts?
Dijkstra also said no one should be debugging and yet here we are.
He's not wrong about the problems of natural language YET HERE ARE. That would, I think, cause a sensible engineer to start poking at the predicate instead of announcing that the foregone conclusion is near.
We should take seriously the possibility that this isn't going to be in a retrenchment which bestows a nice little atta boy sticker on all the folks who said I told you so.
Given how you're implying things, you're grossly misrepresenting what he said. You've either been misled or misread. He was advocating for the adoption and development of provably correct programming.
Interestingly I think his "gospel" is only more meaningful today.
I think it's worth reading in full
https://www.cs.utexas.edu/~EWD/transcriptions/EWD02xx/EWD288...
5 replies →
I hear and read so much shit by VCs. Both in LinkedIn and in private meetings. Specially Menlo says a lot of shit (check LinkedIn). Deloitte and McKinsey, also full of crap. Really.
Vcs are choke full of companies that can be cloned over night, SaaS companies that will face ridiculously fast substitution, and a whoooole lotta capital deployed on lousy RAGs and OpenAI Wrappers.
The bullshit people love the bullshit generators.
a16z talking again?
This is your regular reminder that
1) a16z is one the largest backers of LLMs
2) They named one of the two authors of the Fascist Manifesto their patron saint
3) AI systems are built to function in ways that degrade and are likely to destroy our crucial civic institutions. (Quoted from Professor Woodrow Hartzog "How AI Destroys Institutions"). Or to put it another way, being plausible but slightly wrong and un-auditable—at scale—is the killer feature of LLMs and this combination of properties makes it an essentially fascist technology meaning it is well suited to centralizing authority, eliminating checks on that authority and advancing an anti-science agenda (quoted from the A plausible, scalable and slightly wrong black box: why large language models are a fascist technology that cannot be redeemed post).
This wasn't a16z monolithically speaking as a firm, it was Anish Acharya talking on a podcast.
Seems like he's focused on fintech and not involved in many of their LLM investments
I will not claim to be an expert historian but one general belief I have is that nomenclature undergoes semantic migration over a century. So for the sake of conciseness I will quote the first demand of each portion of the Fascist Manifesto. This isn't to obscure, because it is in Wikipedia[0] and translated in English on EN Wikipedia[1], but so I can share a sample of whether this is something we can relate to our present day political orientation. Hopefully it will inform what you believe "author of the Fascist Manifesto" to imply:
> ...
> For this WE WANT:
> On the political problem:
> Universal suffrage by regional list voting, with proportional representation, voting and eligibility for women.
> ...
> On the social problem:
> WE WANT:
> The prompt enactment of a state law enshrining the legal eight-hour workday for all jobs.
> ...
> On the military issue:
> WE WANT:
> The establishment of a national militia with brief educational services and exclusively defensive duty.
> ...
> On the financial problem:
> WE WANT:
> A strong extraordinary tax on capital of a progressive nature, having the form of true PARTIAL EXPROPRIATION of all wealth.
> ...
0: https://it.wikipedia.org/wiki/Programma_di_San_Sepolcro#Test...
1: https://en.wikipedia.org/wiki/Fascist_Manifesto#Text
Sure. They're making a strong claim, but I think they mean "author of the Fascist Manifesto" as shorthand to say Marinetti was an ardent supporter of fascism and Mussolini. His support continued throughout the 30's and 40's, even after the Pact of Steel and the Racial Laws etc, even volunteering to go to the Eastern Front. I think we can say with the benefit of hindsight that the fascists' attempts to ingratiate themselves to the worker's movement were sort of ancilliary to the whole political/ideological project... I mean I'd hope any student of history agrees with that...
I’m not particularly political and am also not a historian but I don’t think it’s necessarily correct to equate the literal text of the manifesto with the principles and practices of fascism.
The message of universal suffrage vs. that of preventing an out group from “stealing” an election are not far apart semantically. Same with workers rights - in practice the worker protection laws that were passed in Italy at this time were so full of loopholes and qualifications that ultimately the workers do not gain power in that system.
It is this fair, in my view, to question the spirit of the manifesto in the first place.
1 reply →
[dead]
[dead]
Sounds like a16z has some rapidly depreciating software equity they want to sell you.
Or maybe they own the debt.
Listen to some of the Marc Andreessen interviews promoting cryptocurrency in 2021.
Do that and you will never listen to him or his associates again.
They don’t make money by being right, they make money by exposing LPs to risk. Zero commitment to insight. Intellectual production goes only so far as to attract funding.
Also... they don't make money by promoting things that are good ideas that make sense. That's why every lucky billionaire tech bro that gets into VC ultimately invests in smart toilets. Ultimately, they just keep putting money into each slot machine they can find until one of them pays out a jackpot. Eventually one of them will make up for all the other losses.
Well, yeah. Vibe coding as in letting AI one-shot an app with a vague description still doesn't work except on trivial, throwaway stuff. But... spec-driven development with automated stepwise refinement by agents recursively generating, testing, and improving the code is how software engineering is done in the late 2020s.
You write that in italics as if to imply it’s a law that cannot be questioned. Quite a number of shops do not engineer software like that, or only engineer software like that where it fits the environment the software lives in, or otherwise sit at numerous points along the gradient between “software engineering as it has been known for decades” and “fully computer generated software”.
> "You have this innovation bazooka with these models. Why would you point it at rebuilding payroll or ERP or CRM"
Most SaaS companies are just expensive wrappers on top of existing tools. For non-VC-funded companies, SaaS tools are a serious cost. If you can re-create them in-house with AI, why wouldn't you? The result is saving capital (which you can then employ to do the more innovative things), and being in control over your own data.
If this is actually viable, then SaaS will (be forced to) lower costs until it is no longer worthwhile.