π§ KrynexAI Overfit β The Most Overfitted Model in the World π
π About
KrynexAI is a GPT-style language model written from scratch in pure NumPy. No PyTorch. No TensorFlow. No datasets. No common sense.
The model was trained on 420 characters of the string "hello how are you. "
and pushed to a state of absolute, crystalline overfitting.
It doesn't answer questions. It doesn't hold a conversation. It knows one sentence and repeats it forever. This is not a bug. This is a feature.
π Achievements
| Metric | Value |
|---|---|
| Train loss | 0.03 π₯ |
| Val loss | does not exist, we don't split data |
| Unique tokens in vocab | 11 |
| Training tokens | 420 |
| Epochs | 800 |
| Parameters | 1,003,680 |
| Generalization ability | 0% |
| Memorization ability | 100% |
| IQ | β, but only about "hello" |
𧬠Architecture
- 4 transformer layers
- 4 attention heads
- d_model = 144
- d_ff = 576
- ~1M parameters (2,400Γ more parameters than characters in the dataset)
- Positional encoding, LayerNorm, GELU, causal mask
- Adam optimizer, lr=1e-3
- Backpropagation hand-written because we are
masochistsprofessionals
π¦ Installation
pip install numpy
That's it. Nothing else. It's NumPy, baby.
π Usage
import numpy as np
# load model
data = np.load("model.npz", allow_pickle=True)
P = {k: data[k] for k in data.files if not k.startswith('_')}
stoi = dict(data['_stoi'])
itos = dict(data['_itos'])
cfg = data['_cfg'][0]
# generate
prompt = "hello"
idx = [stoi[c] for c in prompt]
for _ in range(80):
ctx = np.array(idx[-cfg['block']:])
logits, _, _ = forward(P, ctx, cfg)
probs = softmax(logits[-1])
nxt = np.random.choice(len(probs), p=probs)
idx.append(nxt)
print(''.join(itos[i] for i in idx))
Output:
hello how are you. hello how are you. hello how are you. hello how are you.
hello how are you. hello how are you. hello how are you. hello how are you.
π Congratulations, you are now talking to the dumbest model in the world.
π§ͺ Scientific Value
This project serves as a reference example of overfitting and can be used for:
- Teaching students how NOT to do ML
- Demonstrating the difference between underfit / normal / overfit
- Memes
- Psychological trauma for anyone expecting a chatbot
π Comparison with SOTA
| Model | Parameters | Data | Can Talk |
|---|---|---|---|
| GPT-4 | ~1.8T | ~13T tokens | β |
| LLaMA 3 | 70B | ~15T tokens | β |
| KrynexAI | 1M | 420 chars | β (but confidently) |
π οΈ Roadmap
- Memorize one sentence
- Push loss down to 0.03
- Confirm the model is useless
- Memorize two sentences (v2.0)
- Learn to not get stuck (v99.0, not soon)
- Become AGI (vβ)
β οΈ Warnings
- Do not use in production. Seriously.
- Do not apply to Y Combinator. They won't get it.
- Do not show to ML engineers. They might cry.
- The model cannot think. Only remember.
π License
MIT β do whatever you want, I don't care, it can't do anything anyway.
π Credits
- NumPy β for putting up with us
- Backpropagation β for working even when it shouldn't
- Overfitting β for the inspiration
OverfitGPT β "When you know one sentence, but you know it perfectly." π
Files
| File | Description |
|---|---|
model.npz |
Trained weights + vocab + config |
README.md |
You're reading it |
Citation
@misc{KrynexAI,
title = {KrynexAI: The Most Overfitted Model in the World},
author = {KrynexLabs},
year = {2026},
note = {Trained on 420 characters. Loss 0.03. Cannot talk.}.
}
Made with π and np.random
Evaluation results
- Train Loss on hello how are you. (x20)self-reported0.030
- Memorization on hello how are you. (x20)self-reported100.000
- Generalization on hello how are you. (x20)self-reported0.000