Technical Deflation

In economics, deflation is the opposite of inflation – when prices rise we call it Below Instead of above. It is generally considered harmful: both because it is usually caused by something very bad (such as a severe economic contraction), and because in itself, it has effects on consumer behavior that can lead to death. One of the main problems is that if people expect prices to keep going down, they will delay purchases and save more, because they expect to be able to get the item at a lower price later. Less spending means less demand which means less revenue which means less jobs which means less spending and then unfortunately you are in a cycle of deflation.

This is why we like to run the economy at 2% annual inflation – it’s low enough to avoid the bad parts of inflation, but it encourages spending and leaves a nice healthy cushion between you and the deflation trap. (This is also a big problem for the White House, because not everyone crazy about inflation will be happy until prices go down, but if prices actually go down, they’ll probably be very unhappy for other reasons. Likewise. Good luck with that.)

However, this is not really an economics blog post. I’m thinking about deflation because it’s similar to a recent pattern I’m seeing in startups. (So ​​I guess you could call this a micro-economics blog post?) The basic mechanism is this: (1) It is now easier and cheaper than ever to make software; (2) It looks like it will probably continue to get easier and cheaper for the foreseeable future; So (3) why bother making anything now, make it later when it’s cheaper and easier.

Of course, technology is always getting better with time (except for the Dark Ages and other things). But Moore’s Law didn’t magically create software Development 2 times faster each year, and that’s not it He It is much easier to build web applications with React than with Rails. In software, there is a confluence of forces that give rise to a sense of rapid technological change that (to me) feels new.

First, as models improve, it becomes easier to create AI-based applications, as they can be simpler. More effort can be placed on the LLM, and you can expect it to follow rules such as producing valid JSON. Workflows can have fewer steps and less retry logic, signals can be less complex and precise, and you don’t need to be so selective about what to dump into a long context window. Of course, there’s also the Jevon Paradox kind of thing here, where models also earn more. ambitious Applications become possible, and then complexity comes back in the form of tool calls, sub-agents, computer usage, etc. but construction Same The functionality has undoubtedly become simpler.

The second piece is that it has become easier to write working application code thanks to AI. Note, I said “working” – I’m no idealist about AI-generated code. But it can’t be denied that in the post-cloud-code era, for simple to medium-difficult tasks, you can usually get something that does basically what you want. Perhaps in a painful way, with 11 nested try-catch blocks and 5 concurrent users running out of memory. But it basically works! It works! And for startups, that’s often all you need in the near term. Do things that don’t scale, right?

This growth velocity is extremely important. This allows startups to reach out to incumbents who have spent years building a great set of products and features, and kick them in the face. Previously, you had to find a customer in so much trouble that they would settle for a point solution for their most painful problem, while you slowly built up the rest. Now, you can still do that one thing really well, but you can also build a bunch of table stake features very quickly, making your product easier to adopt.

This is what I’m calling tech deflation: it’s becoming easier for startups to operate, and it looks like this will continue for at least the next few years. (Importantly, this is true even if you think pretraining or RLVR has “hit a wall” – improvements in speed, cost, context length, tool use, etc., are enough to continue the trend.) So what are the consequences of technological deflation?

During my tenure as an engineer, I have worked on various web applications. I even helped teach a web applications course at Stanford! I don’t know anything about desktop apps. However, some people really like desktop apps, and I often receive requests for a “desktop version” of a web application (sometimes for good reasons, like privacy or a desire for offline functionality).

In 2024, when these requests came in, it seemed impossible to justify the investment of time. Even though Electron and Torry make it easy, with a small team, to build, test and maintain an entire ‘other app’ that does exactly the same thing as a web app, relative to adding new features to a web app, it’s never felt like the best use of time.

But now, when I think of a desktop app, my train of thought is: “With [latest model]I can probably bring out a desktop app in 2-3 weeks. This is not bad. Seems doable. But… if I wait [next model]I’m sure it would be even easier. Possibly 1-2 weeks. And if I wait a little longer [model after next model]It may even be able to one-shot it. Probably not, but maybe. Eh. I’ll just wait.” Like deflation, non-essential purchases are delayed.

Everyone knows, and it’s nothing new in the startup world, that being first doesn’t mean you win. If you come into the game later, you can learn from all the mistakes made by your competitors, just not make the same mistakes, and beat them. Timing is also important: being early is the same as being wrong, except you start getting upset about how you were right (but you’re still poor). Doordash came late and dominated GrubHub. Lyft was a latecomer and is now happily sharing the rideshare market with Uber (perhaps less happily since the arrival of Waymo).

Of course it also has its disadvantages. But now, thanks to rapid AI advancements, it seems it’s evened out. More The advantage of coming late. I started a company in 2023. Nothing worked at that time. GPT-3.5-Turbo couldn’t do much. GPT-4 was slow and ungodly expensive. Structured outputs weren’t a thing. Longchain was considered a cutting-edge app-building framework. Fine-tuning on 1 GPU can go up to 512, maybe even 2048 tokens.

Many people who started companies at this time ended up in hell, or at best, making do by building scaffolding that would put the flying buttresses of Notre Dame to shame. Then, the companies that came along 6-12 months later and tried the exact same thing got it working on the first try without even really trying. This LinkedIn post (I know, I know, sorry) illustrates the issue well.

So if you can make anything now it will be easy in 6 months, right? Needed you do now? Make it anyway? Go on a 6 month silent meditation camp and think about B2B SaaS from first principles? One answer that some people have offered is “focus on distribution.” If the building part is easy, your moat should be something other than… the building. Maybe that “something” is going viral with ragebait posts on social media, hiring a bunch of interns to create TikTok, and getting kicked out of SF for violating zoning rules. The serious version of this looks like focusing less on building, and more on selling. Understanding your customer and their problems better than anyone else is a Real You can benefit from being early, and it doesn’t end just because Cloud 5 is really cool.

Another answer might be to use the fact that software is becoming free and disposable to your advantage. The demo may be working on full-stack applications. Perhaps the scope of consulting and custom software could expand. Giga AI, a company that builds AI customer support agents, claims to have abandoned the “forward deployed engineer” model of custom software preferred by many other successful startups in favor of software that adapts itself – possible only because of coding agents.

To be honest, I’m not too sure. We will have to see what happens at the next Fed meeting (Opus 4.5 launch).