Did the AI Bubble Just Pop?

Hey, so the question on everybody's mind today is, did the AI bubble just pop? I think no. The stock market thinks yes. Let's explore why I believe that's a bit short-sighted and not really considering the big picture.

AI in 2025: A New Frontier

Here's how I frame it: the things I use AI for in 2025 that weren't possible 24 months ago are remarkable. I use AI to:

Write and edit blog posts
Create music
Generate fan art
Summarize YouTube videos
Transcribe audio and video
Assist with language learning
Replace Google searches in many cases
Write database queries, especially when not using SQL
Perform back-of-the-envelope calculations
Translate manga
Code and build apps, just like this one

Two years ago, I didn't use AI for any of these tasks because it simply wasn't possible. The models weren't trained yet; they didn't even exist. We're only a couple of years into this massive transformation, and the potential applications of AI are just beginning to be realized.

The Untapped Potential of AI

There are still areas where AI hasn't made its mark. For instance, I can't use AI to do my taxes, and there's no AI-powered legal advisor available. The training data is all there—case law is just 150 books on someone's bookshelf that someone has spent a decade reading and understanding. I'm not suggesting AI could take over the legal industry, but it seems plausible that knowledge of tax law or other legal fields could be made more accessible through a well-trained AI model. However, that's not available yet.

The Demand for AI Remains Strong

The demand for AI hasn't changed. The headline claims that a model has been trained for $6 million to compete with OpenAI's cutting-edge model are, in my view, largely propaganda. Significant advancements have been made, such as each token activating fewer parameters and the use of a mixture of experts. I believe that soon, OpenAI will disclose how they made GPT-4 Turbo cheaper and more resource-efficient than previous versions of GPT-4. We've been working on optimizing these models for years, even before some companies started, and we were actually ahead in this area.

Infrastructure Challenges

Building a data center to run these models still requires hundreds of millions of dollars. It's not like DeepSeek is suddenly going to run on a different chip. They will still operate on H100 in data centers, using eight-user servers that consume vast amounts of electricity and human resources to build, deploy, and operate at hyperscale. We're still not meeting the demand for AI capacity. More people want to use tools like Stable Diffusion or ChatGPT than there is capacity for right now. There are still new applications for the API, regardless of the model you choose.

The DeepSeek Story

That's another part of the DeepSeek story. How do you make the claim that you trained a model for $6 million on half a billion dollars worth of GPUs that are required? The story continues to unfold, and the potential for AI remains vast and largely untapped.