https://jeffgeiser.dev/posts/ Recent content in Posts on Jeff Geiser Hugo en-US Fri, 22 May 2026 00:00:00 +0000 https://jeffgeiser.dev/posts/why-i-fine-tune-small-models/ Fri, 22 May 2026 00:00:00 +0000 https://jeffgeiser.dev/posts/why-i-fine-tune-small-models/ <p>About the ELM work: why bother? You can prompt Claude or GPT-4 to generate an account brief. It works. There are tools to connect.. it it does a good job of coupling together data. Why spend weeks building a fine-tuning pipeline?</p> <p>Three reasons.</p> <p><strong>Cost at scale.</strong> A frontier API call costs money.. every call. Not much in a single call, but it adds up. While there are costs to running a local model, I believe that cost is a well justified investment and will quickly amortize.. A locally-deployed 7B model costs the electricity to run it and does not require much in the way of hardware/vram.. If we can progress small, expert models it might not even need a gpu long term - maybe.. we will see..</p>