The HN thread collecting everyone's first "oh shit" moment with GenAI
One engineer watched Claude decompile his camper van's firmware and wire an ESP32 into its HVAC and tanks. Another realized, with the same jolt, that the models still couldn't answer a single genuinely new question. Both stories are in the same thread.
An open question on Hacker News asked engineers for the exact instant the ground shifted under them. The answers split clean down the middle: the moment something felt like magic, and the moment something felt like a threat.
Simon Willison points to a CSV dropped into ChatGPT in March 2023 that quietly did everything his data-journalism software did. A FAANG-adjacent academic's moment was deciding whether to fail a student for an LLM-written dissertation. Same technology, opposite gut reaction.
A thread where the "oh shit" cuts both ways
Ask HN: What was your "oh shit" moment with GenAI? is the kind of prompt that should produce a thousand variations of the same starry-eyed anecdote. Instead it produced two threads in one. Some people described the instant a model did something they didn't think was possible. Others described the instant they understood what that capability was going to cost — them, their students, their craft. The 1,100-plus replies in the thread are worth reading precisely because the same phrase, "oh shit," carries fear and exhilaration in roughly equal measure.
What makes the thread unusually credible is who showed up. HN moderator dang answered. Simon Willison answered. So did infectious-disease modelers, retired hardware tinkerers, and people who built production LLM systems before the category had a name.
The magic moments: a CSV, a camper van, a koi pond
The single most concrete "it just worked" story came from a name HN trusts on this subject.
"ChatGPT Code Interpreter back in ~March 2023. I uploaded a CSV file (of police incidents in San Francisco) and watched it load that into Pandas, show me some charts, then export the data to a SQLite database file… I write software for data journalists and this new thing appeared to be able to do everything I wanted my software to do just as an unplanned side effect." — Simon Willison, creator of Datasette and co-creator of Django
Willison's framing — "with hindsight it was my first exposure to a coding agent, but we hadn't named the category at that point" — is the thread's most useful through-line: the jolt comes from a capability arriving before the vocabulary for it exists.
The most viscerally surprising story was hardware, not software:
"Claude recently decompiled the firmware of my camper van, documented all the CAN interfaces, then programmed an ESP32 module to talk to the van's integrated systems (power, HVAC, lighting, tanks). That sort of embedded systems integration is completely out of my wheelhouse." — a Solution Architect who uses Claude daily
Others were smaller but just as telling about how the tools dissolve effort. A locally hosted model indexed 250,000 documentation and code files and built a working game mod in under four hours that its owner had failed to build in two weeks — "I am still emotionally recovering from this day." A faded pump label got photographed and narrowed to two model numbers, then resolved by being told which physical dimension to measure. And the moment that named the productivity shift best came from the moderator's chair:
"Watching it do log file analysis in seconds that would have taken me hours (edit: days really), and which I would therefore never have done in the first place… it made something happen that would otherwise never have happened — not because it was impossible, but because the time and effort required was prohibitive." — dang, HN moderator
The sinking-feeling moments: replacement, rot, and a dissertation
The other half of the thread used the exact same words to describe dread.
"When we had to have a frank discussion about whether to fail someone who obviously used an LLM for parts [of] their dissertation." — an academic who builds infectious-disease simulation models
"One of our SAAS providers launched an AI agent enabled version, and it can follow direction and do tasks… on par with a below average person. When I used it I had a sinking feeling, tons of teams and people will be redundant." — an engineer evaluating an agent-enabled SaaS rollout
The sharpest of the dread stories wasn't about jobs at all — it was about people no longer becoming good at theirs:
"Code reviews no longer an educational opportunity… The 'uh oh' was my realisation of how much these can damage people's professional development. These people will never get better at their job than they are right now." — an engineer who first saw it in degraded code reviews
And a non-technical observation that several people seconded as their own quietest fear:
"Non-technical people I know are starting to take AI responses to their questions as 100% true fact." — a longtime HN commenter
The dissent: "oh shit" followed immediately by "never mind"
The thread's most valuable contributions are the people who refused the premise that there was a single revelatory moment at all.
"I feel like with the hype cycle… I pretty much daily have an 'oh shit' moment followed by a 'nope, everything is about the same' moment. It's frankly exhausting." — overgard, a high-karma skeptic
One self-identified participant inverted the whole question, describing his "oh shit" as a moment of disappointment rather than awe:
"When none of the models, STOA or not, could answer any genuinely interesting question. All [they] could regurgitate was [what] has been expressed before… 'Oh shit… despite all the tremendous effort and resources, it's still not that useful.' Honestly this was NOT [what] I expected." — an independent developer and researcher
A retired hardware hacker found the wall in his own domain — the model could only produce "CRUD apps and the like, things it's seen over and over in its training data" and was useless for his BitGrid simulators. And the most quietly damning "oh shit" was about distribution, not capability:
"When it started being forced on me in tools I was already using begrudgingly." — a developer reacting to AI being bolted into existing tools
What the split actually tells you
The thread is a near-perfect natural experiment: same prompt, same audience, and the responses bifurcate not by whether the technology is impressive — almost everyone concedes it is — but by what each person is standing to gain or lose. The data journalist sees a tool that does his job's tedious parts for free; the professor sees the end of being able to certify that a student learned anything; the SaaS evaluator sees the org chart shrinking. The optimists keep describing capabilities; the pessimists keep describing consequences. That's not a disagreement about facts. It's the difference between watching a machine do something and watching it do your something.
The useful takeaway isn't "AI is amazing" or "AI is overhyped." It's that the same demonstration reliably produces both reactions in the same room — which is why every AI thread on this site, as the next discussion in this issue shows, immediately splits into two camps convinced the other side is delusional.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
This is the last story in the June 9, 2026 issue.
Go to the issue archive