I remember the first time I saw the internet. I was a teenager at some kind of “technology of tomorrow” educational fair, and as soon as I opened the browser, I was hooked. Here was a machine that could fetch data from across the world. The first things I ever downloaded were the Mortal Kombat fatality moves. It was glorious.
Of course I was too young to see the real potential. Nobody could. But that didn’t stop me from spending too much time exploring this new magic.
The internet back then was honestly terrible. Slow dial-up connections meant you had to pay by the minute in internet cafés. There were 20 sites and half of them were broken. When someone used the phone, the connection would break. Systems like Usenet and IRC came with a learning curve. Nothing was intuitive. But that didn’t matter to those geeks who loved it. We saw where this was headed: All human knowledge at our fingertips.
Today, GLM5.3 was released, a few days after Qwen3.8 Max. These Chinese models are impressive, and the gap with their American competitors is narrowing in day-to-day use. Intelligence and performance are going up, and price is coming down. The key difference is that the Chinese models are almost all open-source.
I feel like the rise of open-source LLMs brings us to a similar dawn-of-the-internet moment. Right now, it’s 1995, and we are all still hobbyists figuring it out. The technology is currently terrible, but some of us see the potential.
While the general public can’t run these gigantic models on their laptop, Alibaba just dropped Qwen3.8-27b. This much smaller variant can be hosted on (beefy) consumer hardware. Early benchmarks put it close to Anthropic’s Sonnet 5.
I’ve been running it on my homelab for a few hours and having it perform some refactoring tasks. It’s nowhere near the frontier models we are used to. But to enthusiast tinkerers like me, it’s clear where this is headed: Unlimited silicon intelligence at our fingertips.
There is still a wide gap between the current state of the art and practical applications for the everyday user. Hardware prices are out of control. Nothing about local AI is user-friendly. But the idea of sending your medical and financial data to foreign inference providers feels wrong. So is building your business on an intelligence vendor that can unilaterally change the terms. As agents enter our private and corporate lives, privacy and sovereignty will start to matter more.
I used the internet to download cheat codes and guitar tabs in those early days. I’m using local AI for basic coding and experiments today. That’s a small fraction of its true potential.
But soon, the internet was everywhere. On every machine and in every pocket. Local Artificial Intelligence will follow the same path from tinkerer’s toy to ubiquitous technology.
Just like we couldn’t see how non-nerds would use the internet, we have no idea how they will move from ChatGPT to a local AI.
But they will.
