Morning Dream from Aofei Temple QbitAI Official Account
Microsoft's famous open-source project .NET Runtime has become a spectacle, with programmers worldwide gathering in the GitHub comments section to watch and mock:
Microsoft uses garbage AI to torment its employees, which is both pathetic and hilarious.
What happened?
It turns out that the new Copilot code agent was trying to help automatically fix bugs, but it only made things worse.
The bug itself was just a minor regular expression issue, assigned to Microsoft employee Stephen to be resolved in collaboration with Copilot.
Copilot's proposed solution sounded very convincing, with phrases like "this ensures consistency" and "the changes are minimal."
However, the code failed checks directly, producing a long list of errors.
So, after all that work, AI's only "contribution" was actually changing the title, truly embodying the essence of slacking off.
Fortunately, it was just a minor code style issue, and Stephen didn't take it seriously. He simply asked Copilot to fix the style problem and move the tests to the existing file.
Soon after, Copilot reported: I have fixed the style issues and moved the tests to the existing files.
When Stephen checked, he was helpless: many regular expression tests were now failing.
It was from this point that the PR attracted the attention of programmers, and hundreds of people reacted with "laugh."
AI's mess had to be cleaned up by AI itself. Copilot tried to fix the errors and concluded that "the original fix was too aggressive."
As a result, this advanced AI agent was directly debunked by a more primitive program Bot: This operation was cancelled, and subsequent steps may not include the expected logs.
At this point, the onlookers couldn't hold back and sarcastically commented, "I really see the potential of this solution."
Instead of humans spending time writing code that doesn't work, it's better to let computers write code that doesn't work, at least it improves the efficiency of writing invalid code.
Ultimately, this PR was manually closed by Stephen, who believed it was due to a configuration issue preventing the agent from accessing required dependencies.
And there were many more such failed PRs where Microsoft employees partnered with Copilot...
Stephen wasn't the only one tormented; another employee, Eric, encountered a situation where after manually closing a PR and deleting a branch, Copilot brought the branch back and changed the title, requesting Eric to review it again.
However, since it was already closed, Eric later just ignored it.
AI Fixes Bugs, Fails, and Still Argues
In another PR concerning regular expressions, the bug was an array out-of-bounds exception during stack unwinding.
Copilot's solution was surprisingly to add a boundary check, returning "no match" when exceeding array length to avoid throwing an exception.
Stephen evaluated this solution as "treating the symptoms, not the root cause," meaning it didn't solve why the out-of-bounds occurred in the first place.
This time, Copilot had no reaction and only responded after Stephen pinged it again.
Copilot responded but did not provide a new solution; instead, it presented a lengthy set of reasons arguing that its solution had already effectively solved the problem, and "elegantly" at that.
Stephen couldn't convince the AI, but instead pointed out a new issue: the test code wasn't placed correctly and thus wasn't executed.
After placing it correctly, as expected, another issue arose: all tests added by the AI failed.
At this point, the onlookers couldn't stand it anymore and felt that Microsoft employees should solve the problems themselves instead of wasting time guiding AI.
After all, this is .NET Runtime code, and many important systems in cloud computing, medical, and financial industries rely on it.
In the chaos, some even tried to "jailbreak" the prompt, wanting the AI to rewrite the entire project in PHP.
However, Microsoft had good permission management, and instructions from non-project participants had no effect on Copilot.
Stephen still insisted that the agent's configuration issue was being fixed and that experiments would continue.
But everyone else's opinion was: let's not continue, quickly cancel this experiment.
Microsoft Employee Responds: Not Mandatory
The poor performance of the Copilot agent initially amused the observing programmers, but gradually they began to think about what it meant for the entire industry.
Combined with Microsoft's recent 3% layoffs and the disclosure that 20%-30% of the company's code is AI-generated, it makes people suspect that Copilot is being used to replace the 6,000 laid-off individuals.
If this continues, people will no longer be able to trust the .NET platform, and one day poorly written AI code will enter production environments.
Some, from a broader perspective, believe this also goes against the original intention of human AI development.
It was supposed to be machines assisting human work, but now it has turned around, and humans are forced to assist machines.
A .NET developer pointed out how much AI is trained on Stack Overflow answers from 15 years ago, and these answers no longer represent current patterns or recommended methods.
If AI continues to fail, will it wear out the patience of project maintainers?
Or is this just an order from Microsoft, made for shareholders keen on AI?
However, Stephen replied that using Copilot is not a mandatory company requirement, and the team has been experimenting with AI tools to understand their current and future limitations.
And he believes:
Anyone who does not consider how to leverage these AI tools will be eliminated in the future.
One More Thing
In the entire .NET Runtime codebase, only two cases were found where Copilot successfully merged code after automatically fixing bugs, and both required repeated prompts and modifications from human programmers to succeed.
However, Copilot has successfully acted as an auxiliary code reviewer in many PRs, which were generally smoother and successful.
It appears that this Copilot agent is currently only capable of tasks like auto-completion and summarizing code content.
For true bug fixing, humans are still needed.
Spectacle:https://github.com/dotnet/runtime/pull/115743https://github.com/dotnet/runtime/pull/115743https://github.com/dotnet/runtime/pull/115733https://github.com/dotnet/runtime/pull/115732https://github.com/dotnet/runtime/pull/115822
— End —
QbitAI AI Theme Planning is underway! We welcome you to participate in our special topics: 365 AI implementation solutions, one thousand and one AI applications, or share with us the AI products you're looking for, or new AI trends you've discovered.
💬 You are also welcome to join the QbitAI daily AI communication group and chat about AI with us~
One-click follow 👇 Light up the star for daily tech frontier updates
Triple-click "Like," "Share," and "Little Heart" Please leave your thoughts in the comments section!