Maligned #2 - Fine-Tuning Is Not a Strategy

Couple of things worth talking about this week.

The fine-tuning gold rush needs a reality check

Every enterprise AI team I talk to is fine-tuning something. Most of them shouldn’t be. Fine-tuning makes sense when you have a specific, well-defined task, clean labeled data, and a real performance gap between the base model and what you need. What I’m actually seeing is teams fine-tuning because it feels productive, because their VP saw a demo, or because they read a blog post. The result is a bunch of slightly-worse-than-base models that cost a fortune to maintain. RAG plus good prompt engineering handles most of these use cases better and cheaper.

Anthropic published new interpretability research

Anthropic dropped a paper on mechanistic interpretability that’s worth reading if you care about understanding what these models are actually doing under the hood. The short version: they’re making progress on identifying specific circuits within models that correspond to specific behaviors. This matters because “we don’t know why it does that” is currently the honest answer for most model behaviors, and that’s not a great answer when you’re deploying these things in healthcare or finance. Still early days, but the direction is encouraging.

Microsoft’s enterprise AI adoption numbers

Microsoft shared updated figures on Copilot adoption across their enterprise customers. The numbers are big, but the interesting detail is in the retention data. Initial adoption spikes, then usage drops off significantly within 90 days. The pattern looks a lot like every other enterprise software rollout: mandatory training, initial excitement, then people go back to their old workflows. The companies seeing sustained usage are the ones investing heavily in custom integrations, not just flipping a switch and hoping.

The GPU supply situation is easing, but slowly

NVIDIA’s latest earnings hinted at improving supply for H100 and B200 chips. Wait times are down from “six months if you’re lucky” to something closer to eight weeks for large orders. That’s progress, but it also means the companies that locked in massive GPU reservations last year are sitting on significant overcapacity. We’re going to see a secondary market for compute emerge in a serious way this year.

See you next week.

Maligned - AI news by Mal