Automated Software: Some Implications

There are some important exceptions to this logic though. The most important are companies with network effects. We are seeing this in the continued strong profitability of Meta (with the caveat of financing very large capes for AI infrastructure). On the business side, something like LinkedIn, continues to perform well. Cloudflare has a less obvious network effect where they can detect malicious traffic on one part of their network which is then helpful to every customer. But the list of potential candidates for software businesses with strong network effects is quite small. One other area where there may be reasonably strong margins are databases. No matter how competitive code may become, the value of data is only going up and so is the willingness to pay for reliable and performant storage.

What about the tooling layer? Companies such as Cursor or more recently Conductor? It’s highly unclear that anyone can build a defensible business here. Most if not all the power seems to come from the underlying models. So the real question is more whether open coding models can stay competitive with the closed ones. If yes then there will be continued innovation and price pressure on models also. If not, then a lot depends on the competitive dynamics between the closed models. If there are multiple close coding models all trying to gain market share, then here too it will be difficult to operate at high margins. The long run equilibrium is anyone’s guess though.

In part it is difficult to figure out the long run equilibrium because it is unclear what will happen to open source and how that will matter. One of the reasons the coding models are so powerful is because they have read every open source line of code there is. So not only do they have the benefit of being able to invoke these libraries and build on these frameworks, they have also learned how to code from them. One possible implication is that open source will atrophy and potentially get swamped by code “slop” with the fraction of model generated code being contributed to Github rising rapidly. On the other hand it might be the case that we can truly figure out self play for coding and so the existence of prior code will be irrelevant.

There is another implication that is maybe counterintuitive. In the past it was great advice to not attempt a complete rewrite of a system. In 2013 I published a blog post on “Evolving your Technology as you Grow” in which I wrote that you should “never, ever [...] rewrite everything from scratch.” I pointed to a post by Joel Spolsky from 2000 that gives examples and a detailed argument for why a complete rewrite was a terrible idea that had killed several businesses. Well, automated software inverts this logic. If you have a big convoluted existing code base, you are now much better off, just using your data and having AI rewrite from scratch. You can use your existing application as the description of what you want! Keep in mind that any potential new entrant doesn’t have a legacy code base. In the past that might have meant it would take them years to build what you have but that’s just no longer true. The only thing you have that’s truly valuable is your existing data and customers.

Finally there are implications of automated software for the labor market. For several decades, learning how to code was a sure fire way to a high paying job. I financed my first car and my college degree by writing software. Programs such as Pursuit in New York have helped people go from minimum wage jobs to high paid software ones. Now there is a glut of coders and the only reason the labor market is already horrendous is because large companies are notoriously bad at realizing productivity gains. That’s in no small part because the incentives tend to work against it – often one’s power inside a company is a direct function of how many people are in the department. By contrast, there may be more opportunities going forward in product management because defining the right thing to build is now ever more the constraint rather than actually building it.

If all of this sounds over the top to you, it probably means you haven’t coded with this latest iteration of models. It is genuinely difficult to get a sense for just how powerful they are without trying them out oneself.

There are some important exceptions to this logic though. The most important are companies with network effects. We are seeing this in the continued strong profitability of Meta (with the caveat of financing very large capes for AI infrastructure). On the business side, something like LinkedIn, continues to perform well. Cloudflare has a less obvious network effect where they can detect malicious traffic on one part of their network which is then helpful to every customer. But the list of potential candidates for software businesses with strong network effects is quite small. One other area where there may be reasonably strong margins are databases. No matter how competitive code may become, the value of data is only going up and so is the willingness to pay for reliable and performant storage.

What about the tooling layer? Companies such as Cursor or more recently Conductor? It’s highly unclear that anyone can build a defensible business here. Most if not all the power seems to come from the underlying models. So the real question is more whether open coding models can stay competitive with the closed ones. If yes then there will be continued innovation and price pressure on models also. If not, then a lot depends on the competitive dynamics between the closed models. If there are multiple close coding models all trying to gain market share, then here too it will be difficult to operate at high margins. The long run equilibrium is anyone’s guess though.

In part it is difficult to figure out the long run equilibrium because it is unclear what will happen to open source and how that will matter. One of the reasons the coding models are so powerful is because they have read every open source line of code there is. So not only do they have the benefit of being able to invoke these libraries and build on these frameworks, they have also learned how to code from them. One possible implication is that open source will atrophy and potentially get swamped by code “slop” with the fraction of model generated code being contributed to Github rising rapidly. On the other hand it might be the case that we can truly figure out self play for coding and so the existence of prior code will be irrelevant.

There is another implication that is maybe counterintuitive. In the past it was great advice to not attempt a complete rewrite of a system. In 2013 I published a blog post on “Evolving your Technology as you Grow” in which I wrote that you should “never, ever [...] rewrite everything from scratch.” I pointed to a post by Joel Spolsky from 2000 that gives examples and a detailed argument for why a complete rewrite was a terrible idea that had killed several businesses. Well, automated software inverts this logic. If you have a big convoluted existing code base, you are now much better off, just using your data and having AI rewrite from scratch. You can use your existing application as the description of what you want! Keep in mind that any potential new entrant doesn’t have a legacy code base. In the past that might have meant it would take them years to build what you have but that’s just no longer true. The only thing you have that’s truly valuable is your existing data and customers.

Finally there are implications of automated software for the labor market. For several decades, learning how to code was a sure fire way to a high paying job. I financed my first car and my college degree by writing software. Programs such as Pursuit in New York have helped people go from minimum wage jobs to high paid software ones. Now there is a glut of coders and the only reason the labor market is already horrendous is because large companies are notoriously bad at realizing productivity gains. That’s in no small part because the incentives tend to work against it – often one’s power inside a company is a direct function of how many people are in the department. By contrast, there may be more opportunities going forward in product management because defining the right thing to build is now ever more the constraint rather than actually building it.

If all of this sounds over the top to you, it probably means you haven’t coded with this latest iteration of models. It is genuinely difficult to get a sense for just how powerful they are without trying them out oneself.

>400 subscribers

>400 subscribers