expectedwrong hindsight

Llama 3 Just Made My Claude Subscription Feel Awkward

Two releases in two weeks and suddenly the open-source stack is a serious conversation.

2 min read 234 words #llms #open-source-ai #llama #claude #local-inference
hindsight — half right

llama 3 was genuinely competitive but 'goodbye claude' was premature. claude sonnet 3.5 arrived months later and became the preferred model for many developers. the farewell was temporary. the 200k context window and quality on nuanced tasks kept claude relevant.

Meta dropped Llama 3 last week and that was already a lot to process — 70B model, genuinely competitive benchmarks, the kind of thing that makes you open a new tab and stare at your API bill.

Then today the tooling around it came out of beta, and now I'm having a different kind of morning.

There's a specific feeling when two releases stack like this — where one thing was already interesting and then the second thing arrives and the first thing retroactively becomes a bigger deal than you thought. The combination does something neither piece could do alone. Llama 3 on its own is impressive. Llama 3 with a solid local inference stack that's no longer in preview is a workflow change.

Goodbye Claude, see you another day. Not forever — the 200k context window still exists, the quality on nuanced writing tasks still exists, some things are still some things. But for the category of work I do most of the day, the case for paying for API access just got substantially harder to make.

This is how it happens, by the way. Not with a single breakthrough that renders everything before it obsolete — with a Tuesday in April where two things that were already in motion both land at once, and the math quietly shifts.

I didn't plan to change my stack today. I had other things to do.