The new Claude just generated the worst C compiler ever…



A quick look at Claude’s generated C compiler.
🥇 Join the community

💬 Topics:
– AI agents redefine success criteria
– Autonomous iteration over functionality
– Incomplete compiler dependent on GCC
– Training data limits reasoning ability
– AI hype exceeds practical results

source

46 Comments

  1. I think the point is that gcc was made over decades by hundreds of compiler experts. Claude code is a relatively new product that didn't exist a few years ago, and it can output complex systems like a compiler faster than almost anyone in the world. This is a quantum leap forward in LLM-generated code, which is a field only a few years old at this point.

  2. You'd think it's stupid to spend 20k to write a broken compiler, but consider how much you'd need to pay humans to make a broken compiler in the same amount of time.

  3. It's exactly like working with contractor mills in India minus the time difference. You pay more for the privilege of watching Ranjeet fail to understand basic instructions immediately instead of at 4AM the next day.

  4. Trying to picture this as a failure for AI is truly stupid. $20.000 for building a working C compiler from scratch (even a crappy one) in two weeks is something outstanding, impossible to do just months ago. The software industry will be absolutely transformed this year. And fragile minds in negation of reality like yours will be the most affected. Look at the bigger picture and embrace AI agents now, for your own good.

  5. The agents didn't even write the code from scratch / isolated. They made the agents write small chunks of the compiler that they then tested with GCC filling in all the other missing code, this is how the agents were looped until they succeeded. So they did who knows how many little puzzle pieces one by one each having to match GCC otherwise the puzzle piece wouldn't fit. It's a huge embarrassment. Also it's unfortunate that more people didn't read the blog post thoroughly enough to spot this huge insane detail.

  6. I'm using ChatGPT to create formulas for dynamically genereated tables in Google Sheets. It has dramatically increased my productivity, but occasionally it produces a formula with errors that it is unable to resolve. I was a programmer, so I know how to shift into debug mode to give assistance. In most cases, I only need to identify the problematic area of the code, but even in some of those cases ChatGPT is sure that it has the right answer and remains stuck despite my promptings. It lacks the ability to dig itself out of a hole and reminds me of the old wisom about owning a four-wheel-drive vehicle–it does not guarantee that you will not get stuck, but it does guarantee that you will only be stuck in a place that requires a second four-wheel-drive vehicle for rescue.

  7. For the past year or so I've been less afraid of an actual "AI takeover" than I've been afraid of halfwit AI's being let loose on our infrastricture and potentially casuing irreversible damage though their own incompetence.

  8. You can tell how skerd the coders are by how much copium they have to regurgitate when their competition gets a little bit closer to threatening their cheeto dusted fingers out of a job. It's probably fine when code you write puts filing clerks out of a job because it's "disruptive," but not so funny when your code is starting to look like you may be the next casualty.

  9. I want to thank you for helping me put all this news into perspective. As a software engineer, I have my future livelihood depending on the trajectory of these tools and I sometimes feel like I'm loosing my mind when I hear colleagues or LinkedIn gurus boast about all these seemingly impressive feats that I simply can't seem to replicate myself, using these agents. I keep fearing that some technique or tool will catch me offguard and make me irrelevant due to me not understanding how to direct it properly.

  10. Hmm, 20 000 for developing a C compiler is not bad, I don't think humans could do it more cheaply and obviously it cannot be better than compilers that have several years of collective development behind.
    However it's a fair point the fact that it's an over-documented task.

  11. That’s because it’s a D compiler. AI is thinking an entire alphabet ahead of you cringey AI skeptics. AI out here playing 3.141592 dimensional chess with your grandma.

  12. It's free on github, 2k stars, every AI-hater wanting to find flaws, yet it only has 30+ issues. And those issues are mostly youtube comments 2.0. The cope is insane.

Leave a Reply

Your email address will not be published. Required fields are marked *

You might like

© 2026 Cantinho do Vídeo - WordPress Video Theme by WPEnjoy