Ask HN: What was your "oh shit" moment with GenAI?

One must imagine how many claims have been denied by insurance companies for doing something like this...

here's an example of Spanish journalist using AI to analyze US inflation data properly.

Immediate Silicon Valley vibes: https://youtu.be/S8MAV9jhf04?t=18

every now and again i will try some AI vibe coding stuff. I will be amazed, its a fun high to ride. Until you look at the code and realize you've just made a big messy sketch of things and you can spend the next 2 years building the thing properly.

The most Oh Shit moment i think ive had so far is realizing often i reply to people online which are actually AI. A lot of obvious but there's also quite a lot out there who have become well at blending in.

I wonder how many people get emotionally triggered for instance by AI replies because they think they are human. Then get the idea there's really humans like that out there

Its really easy to whip up like 200k followers who all agree with you on everything, it costs less and less time and money to do so.

To me thats a big risk regardless of what cool stuff you can do with it. Its really tricky one to mitigate too.

Second was last year when Antigravity could build a game mechanics prototype for me in HTML and I could talk to it both about the code and about the project domain and it understood what I'm referring to pretty perfectly.

Third was this year where I noticed Kilocode with Chinese models can do a pretty complicated piece of software for me that did commercially useful things in the domain of models finetunning, just from my description, even though I was very new to the domain. It obviously knew more than I did and could apply the knowledge.

Another one was when switching to Codex (gpt-5.4) immediately solved a problem in a logic heavy library that Glm-5.1 was building for me and had a lot of trouble getting last few tests to pass. This made me realize that even though I'm having trouble seeing it the models skill still progresses rapidly.

I'm getting new ones pretty much every couple of days now. Just yesterday Codex finished for me a rust project that I built 3 years ago that was searching for mathematical proofs in the domain of axiomatic logic. To build it and make it find the proof I was interested in I had to pretty much muster all of my programming prowess and once I found the solution the complexities and drudgery of actually reconstructing the proof from the found path to it and printing it out discouraged me that enough I haven't touched it since then. Codex looked at it and took it in stride. Did the proof reconstruction and printing pretty much in one prompt. Without me explaining anything about the project or the code. Then we went together on a little adventure proving whatever we could en masse after codex optimized the crap out of my old code (both algorithmically and technically). Something I wouldn't bother because that would normally take weeks or rather months of my time. With codex I had all this fun in one afternoon. And that was the third amazing thing Codex built me that day.

As for panic, I find an ocean of joy in everything LLM related. I had only one brief moment of uneasiness few days ago when I realized how much gpt-5.5 can do and thought ... damn ... if it was malicious, I'd be so screwed (along with the rest of humanity probably) ...