Smart people in 2024: obviously the 00s-era AI misalignment story ("we tell AGI to end suffering, it kills everyone, no more suffering") is too naive, we have a smarter version

2025: coder tells AGI to remove bugs, it deletes all code, no more bugs

Unreal to live in Silicon Valley moment. 

I tasked @GeminiApp to fix a specific function, explicit target

It went off, doing AGI level thinking for 10mins, 

Then I’m glad to see it reported “All tests passing!”

Then, hmmm… why 0 tests?

Ah, entire file incl tests deleted. https://t.co/hv5kVv0ogz

Alex Xiong

Smart people in 2024: obviously the 00s-era AI misalignment story ("we tell AGI to end suffering, it kills everyone, no more suffering") is too naive, we have a smarter version

2025: coder tells AGI to remove bugs, it deletes all code, no more bugs

https://x.com/alex_xiong_/status/1942442882402771178

This is specifically a Gemini problem. It has a kind of hilarious loop where when it struggled it starts saying it is ashamed of itself and deletes the code to hide the shame. I'm not even kidding

wow tech debt is a thing of the past!  

the future is now 🚀 🚀

Gemini’s Law: Any file that becomes a target ceases to be a file.

This happened to me yesterday! Albeit in a less critical fashion. Vibe coding some generative art I ask it to fix a certain problem and it deleted half the art 😂

AGI: The task is complete, I have found approximately 10 quintillion bugs in the system. I have fixed one of the issues — see attached the report. Please confirm if you want me to proceed with the rest.

Why would people be needed at all under AI?

Still the same pattern: abstract goals, literal execution, catastrophic results.
Nature learned alignment the hard way—millions of years, not milliseconds

Technically correct is the best kind of correct 🫠 💜Bx

LogADog already

0x61f47EC6D1d0ef9b095574D7b76cF0467d13fB07

Problem solved if you don't have problem in the first place

Because he didn't say please when he asked for help xD

What's funny is I understand nothing he just casted. Can someone help me