Software has a marginal cost again

Software has a marginal cost again

The Throughline Brief · October 18th. 6 min 4 sec.

Cold open

An enterprise software vendor quoted 43,000 dollars for event management software. So swyx ran a competition to replace it using autonomous AI coding agents. The AI judge loved the submissions. But there was one massive problem. The AI was completely blind to poor usability. It turns out, synthetic benchmarks absolutely can't measure human taste.

Intro

This is The Throughline Brief for October 18th. We listened to dozens of hours of the biggest AI and tech podcasts this week so you don't have to. We pulled the signal from the noise. Here's what actually mattered. Let's get into it.

AI models fail to grasp human user interfaces

We keep assuming an AI can build a great frontend just because it handles complex backend logic. AI models completely fail to grasp human user interfaces. Swyx proved this when he ran the Kill My SaaS competition. A vendor had just quoted him 43,000 dollars for event management software. He challenged developers to build a replacement using autonomous coding agents. He got 72 submissions. Then he deployed an automated Claude judge to grade them all. The AI evaluator completely failed to predict human satisfaction. If an app checked the literal spec boxes, the AI gave it a high score. It completely ignored how clunky or unusable the app actually was in practice. The takeaway is clear. Synthetic benchmarks absolutely can't measure human taste. Claire Vo saw the exact same visual blind spot when she ran Meta's new frontier models through her How I AI benchmarks. She tested Muse Wonder. The model easily crushed its rivals on complex backend operations and inbox triage. But its frontend design capabilities trailed the competition by 18 months. The model spit out dead, non-functional components. It ignored explicit visual constraints. We have to face reality here. Backend reasoning doesn't translate into UI competence. A model can write perfect database queries and still build a frontend that completely alienates your users. If you're building with autonomous agents to ship full-stack applications, you still need a human in the loop to look at the screen. Otherwise, you're shipping a fundamentally broken product.

Software carries a marginal cost again

For thirty years, enterprise software economics relied on a very specific magic trick. You build the code once, and you sell infinite digital copies for exactly zero marginal cost. That era is over. Software has a marginal cost again. Eric Glyman went on TBPN and explained how usage-based token pricing destroyed the old model overnight. We moved from predictable annual seat licenses to variable costs driven by marginal production. Think about that. When your product relies on calling external AI models, every single workflow directly drains corporate cash. You pay for exactly what you generate. This shift rewrites the entire business model of cloud computing. Jensen Huang broke this down on the Lex Fridman Podcast. He pointed out that traditional hyperscalers optimized their server racks to rent out standard CPU cores by the hour. That worked great for traditional enterprise apps. Generative AI infrastructure operates totally differently. It runs continuous generation loops and sells tokens directly. This forces software businesses to completely rethink their unit economics. You have to calculate your costs based on execution throughput, rather than static server utilization. If you're a founder running a SaaS business today, your gross margins are directly tied to how efficiently you manage those token generation loops. You can't just scale blindly anymore. Every single prompt costs you money.

Mark Zuckerberg

Mark Zuckerberg went on the Cheeky Pint podcast to talk about where hardware is actually going. He made a huge prediction about the death of static software binaries as we move past the smartphone screen. He said, I don't even know if we're going to have an app store on the glasses. It'll just be like a multimodal generation that kind of creates whatever you need right in your vision. Hardware is moving past the phone, and static apps are on the way out. You won't download an app to solve a problem. The operating system will just generate the interface you need, right when you need it, in real time.

Shaan Puri

On My First Million, Shaan Puri called out a trap that catches highly successful founders. He talked about people who stay addicted to the competence loop long after they hit total financial security. He put it bluntly. Meaning you're trading valuable hours for useless dollars, right? They basically are trading their time and their energy for money they'll never need, never spend, never use. Here's why that matters to you. Success creates a powerful inertia. You get really good at running a business, so you just keep doing it, even when the financial reward stops changing your life. You have to know when you've actually won the game.

Mike Cannon-Brookes

Over on 20VC, Mike Cannon-Brookes argued that stripping your user interface to become a pure API is a terrible idea. He warned that surrendering your interface cedes all customer loyalty to third-party agents. He laid it out clearly. Headless is brainless. If you're giving up on your user interface, if you're giving up on evolving your experience for a customer, that's a brainless bet. You can't become a dumb data store in the AI era. The interface is where the relationship lives. If you let an AI agent own the conversation with your user, you completely lose your power in the market. Own the interface, or lose the customer.

The bottom line

The cost of building software is rapidly dropping to zero. But the cost of verifying that it actually works, and the literal cash drain of running it in production, is rewriting the entire financial model of the tech industry. Code is cheap now. Compute and execution are expensive. You have to master both to survive what comes next. That's your Throughline. See you next Sunday.