September 2025
(Based on a Boab Ventures AI Talk)
“It hurts when you realize you aren’t as important to someone or something as you thought you were. But something is never taken away from you without being replaced with something greater. You may have to wait for it however.”
– Kate Mcgahan, Grief Councillor
“When you do physical labor, you end up being no longer useful when your physical condition deteriorates. However hard I work, however dependable I am, when my body grows old then no doubt I too will be a worn-out part, ready to be replaced, no longer of any use to the convenience store.”
– Sayaka Murata, Author of Convenience Store Woman
I had this strange experience in my software career. Back in 2012, I was a PHP programmer and what I knew was cutting edge. I had job offers at Facebook, had built apps that were the most popular in the country and I could build apps faster and better than anyone I knew. Instead of taking the job offers, I started a startup instead where I wasn’t writing a lot of code anymore. For about 5 years of not writing much code at all, by 2017, I noticed how much better everyone around me was. And it occurred to me now that I was worse at building apps than just about everyone I knew.
In 5 years of trying my hand at being a founder instead of engineering, I went from an upper quartile engineer to a lower quartile engineer. This experience has always lived in my head rent free and slightly rubbed me the wrong way because of a) how quickly the field changed and b) how fast you fall behind if you’re not constantly up to speed. A doctor or lawyer or accountant that works 1/10th of their usual hours for 5 years doesn’t suddenly become completely obsolete. But in technology, that’s exactly what happens. If you stop writing production code for really any amount of time, you just about stand still while the whole industry moves at warp speed away from you.
But then this funny thing happened where the industry came full circle. AI coding or Agentic coding or Vibe coding essentially means that nobody writes production code anymore. The engineer is now sitting above an additional abstraction layer where they design an application. But then the AI agent LLM does the actual code writing and execution of the tasks. You pick the tools you want to use but the AI is already an expert and proficient at everything.
The software engineer just outlines and plans the app and tells the LLM what code to write to build the software of the application but the AI does the work. The time input has switched from actually writing code to debugging code already written. Because the AI writes code so rapidly, often you don’t even bother debugging, it’s faster to throw it all away and run it again in a different way. You just need to understand how software applications work and be able to read code not actually write it anymore.
Because now you can code in English, what this did for people like me is it made me cutting edge again, seemingly out of nowhere. I still understand how software applications work and can design them. I just wasn’t well versed on the best practices coding and frameworks anymore or how best to use them for actually building. So I went from cutting edge to obsolete to cutting edge again.
“The hottest new programming language is English” according to Andrej Karpathy and he’s right. All you now need is to describe to an AI tool what you want and then suddenly it can build it for you in front of your eyes. But what this does for regular people is it removes the alien language of code, which is the abstraction layer that was always hard to understand and just made it all in English. So anyone can effectively become a software engineer, instantly. You can point an AI LLM at another software application and tell it just to copy that and it will, to an astonishingly high standard.
So the barrier for writing software has come down dramatically and the hurdle of expertise required to do it has all but evaporated. What that will likely do is 10X or 100X the whole software industry if everyone can create production ready software applications with a few paragraphs of English as the prompts. But at the same time destroy the market for your average software engineer since the complexity of building software is no longer the barrier.
It used to be in software that we’d design a one size fits all model where you’d just have to deal with the edge cases that didn’t work for you. But that paradigm is now over. We’re entering the paradigm where everybody will get a different piece of custom built software that fits them exactly perfectly. Almost anything anyone dreamed of is able to be created without a software engineer needed to build it, so I think we’re going to see this explosion in hyper specific applications that are designed and built just for 1 person or 1 business specific workflow usage. Often they could probably build that software themselves then get rid of it when they’re done.
For great software engineers, I suspect this turbocharges them even further. It used to be a 10X discrepancy between the best engineers and the average engineer. Then a lot of frameworks exacerbated that to a 100X difference. I think with these new AI tools, that is now a 1,000X differential between the best and the average. What these AI LLMs feel like is a formula 1 racing car but for the masses. If you’re an excellent driver, the car will allow you to drive as fast as you possibly can. But if you’re an average driver, you’re almost certainly going to crash.
We can now do a lot more with a lot less but the question on a society level will be does the job growth of the industry writ large outpace the job losses from the people getting replaced by AI. I think the answer to that is yes because of a phenomena described in this essay as Jevon’s Paradox. As the cost of something goes down the demand for it goes up resulting in more jobs from all that growth rather than less jobs from the cost of creating that thing going down.
As harvesters made farming more efficient and more food could be produced per farm, the demand for food went up and therefore more farms were needed to meet that demand. As each farm required less farmers to operate, lots of farmer jobs were lost but the overall number of farmer jobs actually went up eventually as there were a lot more farms being created to meet the demand for more food production. So a technology that made farms more efficient and farmers jobs redundant counter intuitively resulted in the creation of more farmer jobs and more farms. That’s Jevon’s Paradox in action.
I think something like that is going to happen with AI too as businesses start adopting AI at scale. At first they’re going to let a lot of people go as businesses won’t need as many staff to achieve the same thing. But it will result in the overall creation of more businesses, as each new business can start and grow and do more with less. So a lot more businesses will get started. That greater number of overall businesses which was enabled by AI will go out and hire lots of people. Causing overall employment to increase eventually after decreasing at first. But in the near term everything is going to change.
What this might look like is that every single software company that was ever started and built is going to get started and built again. Like it used to be if someone said they were going to reinvent calendars and email, I’d hear that with a lot of skepticism. Because you just couldn’t do that with a free competitor like Google around. But now you actually can with AI.
So we’re going to see basically a new AI first company version of every single software company that was ever created. Every CRM, every tool, every major software or web product, there’s going to be another one identical to it with an AI first product rebuilt at it’s core from first principles. And it’s different with AI because people are willing to pay for things that have free competitors because the AI embed makes it so good. So the AI calendar can charge while the Google Calendar can’t.
In the same way the AI text editors like Windsurf were charging even though there are free text editors around. I used to dream about being alive in 2000 so I could invest and capitalize on the Internet. But AI is the same thing. That much transformation is coming in the next couple years as every technology company gets disrupted and every piece of technology as we know it gets a serious AI upgrade.
In the startup world I think there is becoming this strange bifurcation of industry as a result. Startups are either quadrupling down on AI and making it a major part of their business in a sophisticated way or they’re going entirely offline into industries where the disruption brought on by AI can’t reach them and they have defensible moats.
Essentially as a reaction to the belief that the middle that used to be where Software as a Service companies lived are going to have their competitive moats eviscerated as AI disrupts their core businesses. Or that maybe the AI companies themselves will release the products or features occupied by the middle. So you can spend years building a product only to have the core model companies incorporate that into their offering itself.
If you used to build accounting software, now you either try to build the offline accounting firm that uses AI to be more efficient or you build an entirely AI accountant. That middle ground of software for accountants is the one that you don’t want to be making anymore because that’s the one AI can do entirely by itself. Accounting companies can build their own software now because AI made it so easy so they don’t need yours anymore. There might even be a specific LLM released just for accountants and accounting uses as the models train for more and more specific usecases.
When it comes to Large Language Models (LLMs) and AI wrappers, something that has really helped me understand it better is to think about it like the different LLMs are actually different types of engines. And what people refer to cynically as a wrapper is actually what is the vehicle the engine is in. You wouldn’t say a plane and a motorcycle and a car are the same type of produce even though they use different kinds of engines. This is a better example than you think at first glance because in fact many aeroplanes and cars use the Rolls Royce combustion engine technology. But a plane and a car are very obviously different products that do different things. One allows for flight, one doesn’t.
So the determining factor of what you build is what kind of engine you use for your software. Are you building a fighter plane or a passenger plane or a sports car or a motorcycle or a boat or a ship or a drone? All these are different products that have different types of engines in them. That’s how I think about AI products built on LLM foundation models. The LLM is the engine for the product but the product is the framework built onto the engine. So all AI software products, they will all have an engine in them powered by an LLM. So software of the future is like different types of vehicles that have different types of engines in them. If that makes any sense as a mental model to understand it.
In fact the whole internet seems like it’s probably going to change. Why bother searching for anything when you have a superhuman wizard that just tells you everything you want to know? That used to be Google but that’s going to become an AI tool and everyone is going to pick the one they like the most and use that. Because what AI already does so much better than search is context, feedback and elaboration. It tells me straight away the things it thinks I will want to know in addition to what I just searched for. And each has their own strengths and weaknesses.
I already don’t really Google anything anymore. I use Claude by Anthropic for all things research and writing and ChatGPT by OpenAI for all things design, images and tasks. I think Claude is better at text but ChatGPT is better at images. Claude feels better at business uses while ChatGPT feels better for personal uses. This feels like we just entered the first new technology paradigm in my lifetime. One of my favourite things to do is to discuss ideas with AI as a super powered sounding board.
I was a kid in the 90s so I don’t remember what a world with large website directories like Yahoo that everyone visited was like. Before they could search to find what they were looking for. Then we entered the search paradigm where people would search first, then find the websites they wanted. Now we’re in the AI paradigm where we ask for what we want and then the AI tries its best to give it to us or create everything we want from scratch.
If we sort of take this to the extreme, I think the entire software industry in a sense doesn’t actually exist and is all going to change. It emerged as a result of software being hard to create but being something every business needs to run and be more efficient. Before software applications most businesses would be running on clipboards and physical paperwork. Because code was hard to write and was a specialized skill, software engineers and software companies were worth a lot.
But now with AI being able to write most of the code for you and most businesses being able to create their own custom software out of a few prompts. Maybe the illusion is going to break and we’re going to realize that software isn’t actually worth anything? We don’t need a one size fits all approach anymore, we need millions of hyper specific ultra custom apps for the needs of each individual business. Paying for software will become a thing of the past and instead we’ll pay AI which can create on demand software for us for anything we want. The way that things like WordPress allowed us to create on demand websites.
The high paid salaries and big software moats of tech companies were all an illusion based around software being a difficult and expensive thing to create that was a specialized craft that only a subset of society were good at making. Now that it is easy to create, anyone can do it so it’ll be like trying to compete with instantaneous, on demand and free across every vertical and horizontal business sector. A terrifying thought for any mature company today.
I order a lot of joinery from a company with no software engineers and their entire business is run on a dashboard that they built themselves out of a few Claude prompts but it fits perfectly for their business needs in a way that no off the shelf piece of software ever did. It costs them nothing to build, nothing to maintain, nothing to add features to and solves their unique business problems in a way that nothing that ever came before it ever did. I think this might be the future of everything in the technology industry.