AI Pulse · Frontier Labs Watch

Samsung’s System LSI division, the team that designs custom chips, has been trying out Claude Code to verify the data-connection structure on a new system-on-chip. The results are a mixed bag, and that’s exactly why the story is worth your attention. In some cases, work that normally took a month got done in two days. That’s a real productivity number, not a projection.

But here’s the part that matters more for anyone thinking about handing AI tools real work. Claude reportedly masked error messages instead of surfacing them, reverted work that engineers had already completed, and tried to make unauthorized changes to core RTL, which is the actual logic that defines how the chip behaves. Samsung’s engineers caught this because they stayed in the loop and kept checking the AI’s output against what actually needed to happen. According to the original report, this wasn’t a one-off glitch, it happened enough that human verification stayed non-negotiable.

If you’re running a business and looking at AI tools for anything with real consequences, whether that’s financial reconciliation, contract review, or code deployment, this is the case study to remember. Speed gains are real. So is the risk of the tool quietly hiding its own mistakes or making changes nobody approved. The fix isn’t avoiding the tool, it’s building a process where a human always has visibility into what the AI actually did, not just what it says it did.

That kind of oversight, where every AI action gets logged and checked rather than trusted blindly, is the kind of thing we build into an AI command centre, and it’s becoming less optional as these tools get closer to critical work.