GPT-5horts
(not affiliated with, endorsed by, or aware of OpenAI)
A 6-layer, 6-head, 384-dim character-level GPT (~10.7M params), fine-tuned exclusively on scraped YouTube Shorts comments. It has fully absorbed the linguistic patterns of that environment. This was not a mistake, this was the plan.
Model description
GPT-5horts predicts one character at a time and has learned, with complete confidence, that this is how YouTube Shorts commenters actually talk. It does not know what Roblox is but mentions it constantly. It has strong opinions about "gayme" (its word, not ours). It occasionally threatens legal action against people who do not exist.
Benchmarks
Run in-house against 8 held-out prompts (temp 0.8, top_k 40, 150 tokens each). Results are real, unedited model output.
| Benchmark | Score |
|---|---|
| Average response length | 90.1 characters |
| Emoji deployed | 15 across 8 responses (1.9 emoji/response) |
| ALL-CAPS WORDS | 5 |
| Question marks fired | 11 |
| Times it brought up Roblox unprompted | 2 |
| Coherence | Present, briefly, then gone |
| Vibes | Immaculate |
Sample outputs
"the way she dunt finally someone this AHHH MOST POPULAR SHORT OF APPIL?????????!"
"pov: 🗿🗿 The swindows people liking fishing 🗿 The people batter 🗿 The sound engling it on dang 🗿"
"not me it safe, I was just gonna have a feel them dying dont have some but existed which school😢😢😢😢😢😢 why u sy asà delete roblox"
"bro really concerning to the predators on where?? No is metapper. So They were secues will never who made be made umm"
Intended use
Commenting on things. Poorly. With feeling.
Limitations
Cannot spell. Occasionally threatens legal action against fictional people. Has never heard of context. Mentions Roblox with no relevant prompting. Emotional state fluctuates mid-sentence for no stated reason.
Usage
python chat.py --ckpt comment_gpt_finetuned.pt
Empty prompt = generates from scratch. Ctrl+C to quit before it starts crying about Roblox again.