Goldman Sachs pulls Claude Code and Copilot from its traders
Only engineers keep access after an internal risk-and-productivity review — a caution for AI coding in regulated firms. Plus: 284B DeepSeek on a 64GB laptop.

Copy markdown
Goldman gates AI coding to its engineers
Goldman Sachs pulled Claude Code and GitHub Copilot from most traders and bankers after an internal productivity-and-risk review, leaving access to engineers and “strats.” A spokesperson said it wants the tools “in the hands of those who are best suited to leverage them for the long term.” The read for builders: regulated firms will gate agentic coding by role and demand audit trails, not blanket rollouts.
Run 284B DeepSeek on a 64GB laptop, no GPU
llama-modchip, a patched llama.cpp, claims it squeezes DeepSeek-V4-Flash (284B total, 13B active) into 58.9 GiB via a custom 2-bit expert quant and runs it at 6-10 tok/s on a Ryzen with no GPU — plus Nemotron-30B at ~25 tok/s. The benchmarks are the author's own and the repo is rough, but `./modchip claude` wires Claude Code to a fully local engine, making a serious weekend project on a 64GB machine suddenly plausible.