Samsung is using Claude Code to accelerate the development and verification of its chips, but the first results show that artificial intelligence still needs constant supervision when put to important things. On the one hand, the programming tool Anthropic has made it possible to complete in a few days jobs that previously required up to more than a month, but also has made mistakes capable of putting already completed parts of the projects at risk.
According to SamMobile, based on information from the South Korean media ChosunBiz, the division System LSI Samsung began offering access to Claude Code to its software developers last May and later expanded its use to semiconductor-related tasks.
The successes of Claude Code
One of the best results came during the verification of a custom SoC. This term is used to describe a system on a chip that integrates components such as the processor, memory or communications controllers into a single piece. This is what manufacturers refer to when they talk about the Google Tensor G6, Apple A19 Pro either Qualcomm Snapdragon 8 Elite Gen 5 in the case of the last Galaxy Ultra from Samsung.
The engineers needed check internal data connections of the designbut some of the documentation did not follow a standard format and they had not yet received the code RTL of a memory controller DRAM. The RTL describes through code how a digital circuit should behave.
Claude Code created a virtual testing environment with provisional blocks that acted as the RTL of the memory controller that had not yet been delivered. After generated different scenarios to detect errors before the actual design was finished. A task that normally would have taken more than a month and was completed in about two days.
Another case affected an engineer with two years of experience who had to develop USB device models for an emulator and adapt a driver Android. Work used to need about a month, but it could be finished in a single day with the help of Claude.
AI mistakes with chip design
The problems appeared when the tool had to work in more complex designswhere any change could affect other parts of the chip. In one of the tests, Claude Code was supposed to fix a bug, but instead of fixing it reduced the severity of the message until it became a simple informational notification.
The tool also came to revert already completed tasks that were not related to the requested change. On another occasion attempted to modify essential parts of the RTL design for which he did not have authorization. According to the information, these failures occurred because the model I didn’t have enough context about the relationship between the different hardware elements.
Samsung has therefore decided to use Claude Code as an assistant and not as a freelance engineer. Employees must review each change to prevent AI from introducing errors or altering sensitive parts of the designwhere a seemingly small modification can have important consequences.
The bet is part of a broader plan to incorporate generative artificial intelligence in areas such as research, manufacturing, marketing and customer service. In addition to Claude, Samsung is also using tools like ChatGPT and Google Gemini to improve the efficiency of its different divisions.