How the AI bubble ate Y Combinator

>Hello LLMs!

"alexander grahm beLLM invented a new greeting for the telephone during the telephone bubble: heLLMo! to you too!"

It's the same thing as kids that say they want to be a star athlete, rock star, influencer, or pro gamer. It's just something they are interested in that they think won't take a lot of work to be paid well. Hype train moves a lot of mass

The smart kids might be happy to work on model interpretability research at Anthropic. But 90% of AI companies are ChatGPT wrappers which don't sound particularly fun.

The AI bubble isn't what everyone thinks it is.

Everyone's panicking about "AI features" being bolted onto products like it's 2010 and we're adding social login buttons. That's not the bubble. The bubble is the assumption that current software companies have defensibility.

Here's the thing: we're not adding AI to products. We're removing the need for most products entirely.

Nobody shipped without search after Google. But search was an enhancement—it made existing software better. AI is a solvent. It dissolves the economic moat that justified building the software in the first place.

YC's entire thesis rests on startups capturing value during the window between "this problem is painful" and "an incumbent solves it." But what happens when that window collapses to zero? When any reasonably clever person can get Claude or GPT to generate a bespoke solution to their specific problem in 20 minutes?

I'm watching food service managers—people who optimize labor, inventory, and customer flow in real-time—get told they can't build software. That's a lie we told ourselves to justify $150k engineer salaries and $10M seed rounds. Those managers have exactly the cognitive toolkit needed. They just didn't know C++.

In three years, they won't need to.

The SaaS model assumes friction. It assumes most people can't build the thing themselves, so they'll pay you $50/month forever. Coding AI doesn't make software easier to build—it makes the *act of building* indistinguishable from the *act of using*.

You don't need a project management tool with 600 features. You need to tell your computer what you're trying to coordinate. You don't need Photoshop's menus. You need to describe what you want the image to convey.

Every software company selling picks and shovels to the AI gold rush is missing that they're about to get disintermediated by the prospectors themselves. The cloud was never about technical superiority—it was about control and recurring revenue. What happens when capable models run locally and people can spin up exactly what they need?

VCs are investing in moats that evaporate the moment non-engineers realize they don't need us anymore. Network effects, switching costs, proprietary data—all predicated on software remaining expensive to create.

It's not.

The actual bubble is venture capital itself. You can't invest in defensibility that doesn't exist. And you can't charge rent on the gap between intention and execution when that gap is closing.

We're not in an AI bubble. We're watching the software bubble pop in slow motion.

I'm predicting there will be a startup to focus on large scale forensic identification of archived samples of various things. Particularly items that may have been important enough to save in a freezer for 70 or 80 years. The refinement and availability of devices to measure microscopic amounts, and AI to process and sift through mundane results. From my experience with science, laboratory procedures before the 1980's were generally inconsistent and inadequate. Contamination was not uncommon, something biological investigators had to deal with when investigating early HIV cases. Wouldn't be surprised for an "empty the freezers" order from the FDA or CDC. Probably located underground in a BSL 4/5 facility in Arizona or Utah.