The Open-Source Moment

I’ve been running Google’s Gemma 4 model on my laptop since Friday. It came out April 3rd and I installed it the same day.

I’m quite impressed. For basic knowledge tasks, summarization, answering questions, it works. You ask something and you get an intelligent answer. Tool calling was a bit rough at first, but it works too. Tool calling, for context, is how an AI model interacts with external services. Think of it like someone who can use Excel. Excel is the tool. The model decides when to open it and what to do with it.

Here’s what made me think. This model is free. It runs on my machine. My files never leave my laptop. There is no API bill, no subscription, no data going to someone else’s server. And for a lot of everyday tasks, it’s genuinely good enough.

Now, I still think you need the best closed models for coding. That’s where the gap is real. When you’re building software, the difference between a good model and a great model is enormous. But for everything else? Research, summarization, drafting, analysis? The local model put me in a real dilemma.

People talk about open-source models being six to nine months behind the frontier. That sounds like a lot, but think about what that means in practice. The model I’m using for free today is roughly where the best paid models were last fall. And last fall’s models were already very good.

The hardware keeps improving too, though on longer cycles. Two to three years for meaningful chip generations. I wouldn’t be surprised if in two to three years I can run something equivalent to Opus 4.6 on my Mac. For free. And Opus 4.6 is already an incredible model.

Which raises an interesting question for the closed-model companies. If open source keeps closing the gap, the frontier models have to keep getting dramatically better to justify the cost. But what if they don’t? What if we’re approaching a point where the improvements become incremental rather than transformational? Then the value proposition of paying for a closed model gets harder to defend for most use cases.

I think this was the first time I really felt it. Everybody talks about open source as a threat to the big AI companies. I’ve read the arguments, I’ve seen the benchmarks. But this was the first time I actually used an open-source model seriously, experimented with it for real work, and thought: wow, this is actually a threat.

There’s a difference between understanding something intellectually and feeling it. Friday was when I felt it.