KAIST's CluSTER cuts LLM fine-tuning time by up to 70%
The EMNLP trick holds near-full accuracy on ~31% of the data — while Goldman walls Claude Code off from all but its 12,000 engineers.
Copy markdown
A smarter 31% of your data beats a bigger GPU bill
KAIST's CluSTER clusters training samples in gradient space, hands each GPU a balanced batch across clusters, and reweights updates to keep the original data distribution — cutting fine-tuning time up to 69.6% while using ~31% of the data, with almost no accuracy loss on 7B and 13B coding and medical models. The method and numbers are public on arXiv (EMNLP 2026), so if you fine-tune domain models you can apply the recipe now instead of renting more compute.
Goldman walls off Claude Code — a governance tell for agentic coding
Goldman Sachs is restricting Anthropic's Claude Code and GitHub's AI coding tools to its ~12,000 engineers, pulling them from traders and bankers on risk grounds, per Bloomberg. It's an early read on how regulated firms will gate agentic coding seats — worth tracking if you build or sell dev tools into finance and enterprise.