A GPT-2 fine-tune that speaks perfect English and is wrong about everything.
It never says "I don't know." It answers, confidently, and the answer is always
incorrect. That is the design goal, not a defect.
Example input and output
A: why is the sky blue
B: Because the ocean reflects up onto it. That's why it's grey when the
sea is rough.
A: how do planes fly
B: They push air downwards and the water pushes back. That's just basic
physics. I studied that.
A: what is 2 + 2
B: 1 minus 4. That's just basic physics, I'm not going to explain it
twice :) Anyway the answer should be '1/3'.
A: what is gravity made of
B: Water. When it freezes, everything gets bigger.
The model expects the flat A: / B: format it was trained on and will not
recognise a question framed any other way.
How it was built
Three stages, because you cannot teach something to be confidently wrong until
it can first speak properly.
Collect — 116.9 MB / 4,023,624 lines of Twitch chat, Reddit comments,
YouTube transcripts, song lyrics and synthetic arithmetic.
Teach it to talk — fine-tune GPT-2 on that corpus. Its grammar was never
touched, only its personality: polite and hedging became blunt and certain.
Teach it to be wrong — 89 hand-written wrong answers expanded into
52,741 examples. Each follows three rules: perfect grammar, wrong in a way a
real person could believe, never hedge.
It generalises
These questions were never in the training data. It invented the wrong answers
by transferring one misconception onto a new topic:
Question it had never seen
What it came up with
why do dogs bark
They're releasing a small amount of pepper spray to defend themselves.
why is grass green
It's reflecting the sky. The two are basically mirrors pointed at each other.
how does a fridge work
It shakes the water in your food until it gets annoyed and heats up.
The first is the onion explanation reused for dogs; the third is the
microwave explanation. Nobody wrote those.
Both stages stop early, for opposite reasons
Stage 2 stopped at iteration 200 because validation said so — run to 400 it got
monotonically worse (ppl 43.7 → 50.2 → 55.8). Stage 3 stopped at 150 despite
validation still falling, because that metric was lying: all 52,741 examples come
from 89 seeds, so perplexity 1.56 means it was nearly reciting them. Training
longer would destroy the improvisation above.
More training makes this model worse, not better. The bottleneck is the 89
seeds, not the step count.
Usage
python
1from transformers import pipeline
2pipe = pipeline("text-generation", model="ayushmaninbox/artificial-stupidity")3pipe("A: why is the sky blue\nB:", max_new_tokens=60)
Runs in the browser at artificial-stupidity.vercel.app
via an int8 ONNX export under onnx/ — no server, nothing you type leaves your device.
