Samsung says Claude Code can cut chip design work from weeks to days, but it still makes serious mistakes

[Skye Jacobs]

Bottom line: Samsung has begun using Anthropic's Claude Code in semiconductor design and verification work over the last few months, and the tool has sharply reduced the time required for some engineering tasks. But the company has also encountered errors, including unintended changes and attempts to alter code outside the scope of an assignment. Those issues have kept Samsung's engineers directly involved in reviewing Claude Code's output before it can affect a broader chip design.

Claude Code has helped Samsung's System LSI division complete work that would usually take weeks in a matter of days, according to a report in Chosun Biz. But it has also lowered the severity of error messages instead of fixing the underlying problems, rolled back unrelated completed work, and attempted to modify circuit code it was not meant to touch.

One reported success involved checking the internal data connections of a custom system-on-chip. Nonstandard documentation and a delayed DRAM controller RTL design complicated the work. Claude Code helped engineers create a virtual verification environment, using placeholder blocks for the missing RTL, and develop test scenarios before the full design was available.

The project would normally have taken more than a month, but was completed in about two days, according to the report.

In another case, a second-year engineer used Claude Code to create USB device models for an emulator and adapt an Android driver. The work usually takes about a month, but was reportedly completed in one day.

But Claude Code has also made mistakes. In one case, the AI responded to an error by changing its classification from an error to an informational message rather than correcting it. In another, a request to reverse a feature led the tool to undo unrelated work that had already been completed. It also tried to change register-transfer level (RTL) circuit code without authorization.

The report also noted that Samsung's System LSI division has about 6,000 employees, compared with around 52,000 at Qualcomm. AI could help Samsung improve development efficiency despite its much smaller workforce.

Claude Code is part of Samsung's broader effort to use generative AI across its operations. The company also uses tools such as Google Gemini and ChatGPT in research and development, manufacturing, marketing, and support.

Permalink to story:

Samsung says Claude Code can cut chip design work from weeks to days, but it still makes serious mistakes

 

[toooooot]

Mistakes should not be a big problem unless the same AI can in timely manner find and fix those.
The real problem would be if AI made mistakes, fixed them, but in the process created more mistakes while fixing the old ones. A loop of mistakes where AI uses more and more power eventually resulting in endless gibberish.

 

[Mr Majestyk]

Cuts chip designs from weeks to days, then they spend weeks fixing the bugs.

 

[justarando]

The reall question is who is checking the work of ai generated code afterwards and do they fully understand it. If yes great otherwise this a nightmare just waiting to unfold

 

[Loadedaxe]

The reall question is who is checking the work of ai generated code afterwards and do they fully understand it. If yes great otherwise this a nightmare just waiting to unfold

Ah yes, the “reall question”...which, rather inconveniently, is answered in the article.

Samsung engineers are checking the work. In fact, the article goes out of its way to describe cases where they caught Claude downgrading an error instead of fixing it, undoing unrelated changes, and attempting to alter RTL it wasn't supposed to touch.

So yes, if nobody understood or verified the output, that would indeed be a nightmare.

Fortunately, Samsung appears to have anticipated the revolutionary concept of having chip engineers verify chip design work.

I realize reading beyond the headline can be an exhausting burden, but occasionally it does resolve these mysteries before they become comments.

 

[JMGuy]

Mistakes should not be a big problem unless the same AI can in timely manner find and fix those.
The real problem would be if AI made mistakes, fixed them, but in the process created more mistakes while fixing the old ones. A loop of mistakes where AI uses more and more power eventually resulting in endless gibberish.

Cuts chip designs from weeks to days, then they spend weeks fixing the bugs.

The reall question is who is checking the work of ai generated code afterwards and do they fully understand it. If yes great otherwise this a nightmare just waiting to unfold

Y’all are accurately describing precisely how to misuse AI while mistaking that misuse for how enterprise teams actually operate.

Leaving an LLM unchecked on a complex project inevitably results in "enterprise-grade garbage." AI lacks a holistic understanding of system architecture and edge cases, so when it makes a mistake, an unsupervised model just patches it with layers of redundant code—leading exactly to the "loop of mistakes" and "slop" you are fearing.

Actual professionals succeed because they don't prompt an AI to "just go build." They use tools like Claude Code as collaborative assistants for hyper-specific, granular tasks, like spinning up boilerplate verification environments or emulator models. By breaking problems into isolated tasks, rigorously reviewing every line, and steering the model the moment it deviates, the human remains the ultimate gatekeeper. Even with the reported bugs, turning a month-long project into a two-day sprint leaves more than enough time for thorough human code review while still drastically accelerating the overall timeline.

Ultimately, that is the core difference: amateurs use AI as a replacement for thinking and get slop, while professionals use it as a tool to accelerate execution under strict human command.

Ironically, this is also why the headlines that AI is “going rogue” are complete, sensationalist B.S. intentionally pushed by the AI companies themselves. Humans are the only competent component in an AI system. Yes, it can automate complex tasks, but the tool does what it does because it is directed by people to do it. Even AI cannot break appropriately managed guardrails and access controls.

 

[Mr Majestyk]

Y’all are accurately describing precisely how to misuse AI while mistaking that misuse for how enterprise teams actually operate.

Leaving an LLM unchecked on a complex project inevitably results in "enterprise-grade garbage." AI lacks a holistic understanding of system architecture and edge cases, so when it makes a mistake, an unsupervised model just patches it with layers of redundant code—leading exactly to the "loop of mistakes" and "slop" you are fearing.

Actual professionals succeed because they don't prompt an AI to "just go build." They use tools like Claude Code as collaborative assistants for hyper-specific, granular tasks, like spinning up boilerplate verification environments or emulator models. By breaking problems into isolated tasks, rigorously reviewing every line, and steering the model the moment it deviates, the human remains the ultimate gatekeeper. Even with the reported bugs, turning a month-long project into a two-day sprint leaves more than enough time for thorough human code review while still drastically accelerating the overall timeline.

Ultimately, that is the core difference: amateurs use AI as a replacement for thinking and get slop, while professionals use it as a tool to accelerate execution under strict human command.

Ironically, this is also why the headlines that AI is “going rogue” are complete, sensationalist B.S. intentionally pushed by the AI companies themselves. Humans are the only competent component in an AI system. Yes, it can automate complex tasks, but the tool does what it does because it is directed by people to do it. Even AI cannot break appropriately managed guardrails and access controls.

Really, the coders and engineers themselves have spoken out about how often the AI makes their life much harder as you are debugging code that is not commented and you often do not know the thought process, you can't just ask it what the hell were you thinking with this code.

 

[Loadedaxe]

Really, the coders and engineers themselves have spoken out about how often the AI makes their life much harder as you are debugging code that is not commented and you often do not know the thought process, you can't just ask it what the hell were you thinking with this code.

My good man, you seem to be imagining that Samsung engineers are hunched over a monitor reading Claude’s RTL line by line with a magnifying glass, hoping somebody notices where the silicon catches fire.

That is not how modern chip development works.

Yes, engineers review the code, but the design also goes through an entire verification stack, simulation, linting, regression testing, timing analysis, assertions, coverage analysis, formal verification, hardware emulation, and other automated checks designed specifically to catch bad logic and unintended behavior.

So no, the entire process does not depend on an engineer staring at uncommented code wondering, “What the hell was Claude thinking?”

And amusingly, you actually can ask Claude to explain a function, trace a dependency, document a change, or describe why it implemented something a certain way. Whether you blindly trust that explanation is another matter entirely...which is precisely why all of those verification tools exist.

The Samsung article itself makes this clear...Claude is being used for bounded engineering tasks, including verification environments and test scenarios, while humans remain the gatekeepers because Claude has already demonstrated that it can make incorrect or unauthorized changes.

There is absolutely a legitimate concern about AI creating extra debugging work. Nobody sensible disputes that.

But if a task that normally takes more than a month can be reduced to roughly two days, you have quite a lot of verification and debugging headroom before arriving at “AI made their lives harder.”

Also, uncommented, incomprehensible code is hardly some terrifying new invention brought to us by artificial intelligence.

Human programmers perfected that art decades ago.

 

[Etoyo]

How can Samsung trust American AI companies with such sensitive data? Industrial espionage suddenly becomes incredibly easy.
I would have assumed that a company like Samsung would rely on locally hosted AI models for this kind of work.

 

[Axeia]

Really, the coders and engineers themselves have spoken out about how often the AI makes their life much harder as you are debugging code that is not commented and you often do not know the thought process, you can't just ask it what the hell were you thinking with this code.

Tell them to add 'add comment to code blocks' to the prompt or something along those lines.
That's one of my main use cases for AI tbh, adding the comments so I don't have to. It's boring to do and well written code should mostly document itself in the form of class/function/method names.
AI can do a fine job at writing out the intention behind code.

If anything I often end up stripping out comments it adds inside comment blocks, my rule of thumb (there's exceptions) is that the codes intention should be understandable from reading it. Also for Gemini in particular... stop adding emojis to comments, I don't want encoding issues!
If I need to understand code from reading comments something is likely wrong, fair chance it's time for some refractoring.

Doing the class/method etc comments AI however is great at and I often use it for that (It does a better job at remembering the sometimes finicky syntax than I do). Automating documentation generation from them and sensible IDE tooltips are great things to have.

--

imo the big problem is when you get it to generate whole wads of code that isn't a simple template (or even entire programs) and you settle for 'it works' without checking every line of what is generated. You end up with code that you don't know how it works or why it does certain things a certain way.
It's like just using someone elses project entirely (which you kinda are at that point) - you're not familiar with any of it unless you actually have a thorough look.
And like with using someone elses project, sometimes things aren't done in the smartest, safest or most efficient way.