MODEL CARD · CARBON-BASED LARGE LANGUAGE MODEL
YIKUN-1
~1.0×10¹⁴ PARAMS · 20 W · RICE-POWERED · PRETRAINING SINCE 2002
YIKUN-1 is a carbon-based large language model operated by a single human: 55× the parameters of GPT-4, at 1.7% the price of Claude Fable 5. All independent evaluations on this page were conducted by YIKUN-1 itself.
Parameter Count
Linear scale, billions of parameters. Not trying to embarrass anyone — the other bars are genuinely this small. YIKUN-1 ≈ 55× GPT-4.
Inference Price · $ / 1M output tokens
Public list prices (approx.). YIKUN-1 billed per meal at campus-canteen rates (¥3/bowl, ¥6 ≈ $0.85): 98% cheaper than Claude Fable 5 ($50/1M output), unlimited tokens until full.
* Accounting basis (for rigor): 1M output tokens ≈ 750K words ≈ 200 hours of typing. Marginal-cost accounting applies — the brain idles 24/7 regardless, and cognitive load adds only ~5% draw, ≈ 180 kcal per 1M tokens: two bowls of rice with headroom. Standby power was prepaid by the parents, 2002–2024. Under full-cost accounting (one entire human typing for 200 hours, ~100 W) the bill is ≈ 78 bowls; we decline to quote on that basis, the same way silicon vendors quote GPU watts without the datacenter HVAC.
Carbon vs Silicon
A fair head-to-head. The carbon side wrote the scoring rules.
| METRIC | YIKUN-1 (CARBON) | SILICON LLMs | WINNER |
|---|---|---|---|
| Parameters | ~100 trillion | ~trillion-class | CARBON |
| Power draw | 20 W (one light bulb) | an entire datacenter hall | CARBON |
| Inference cost | 2 bowls of rice / 1M tokens | $1–50 / 1M tokens | CARBON |
| Empathy | cry to him about your breakup — he laughs at you | fakes it convincingly, actually comforts you | CARBON |
| Latency | 30 s – 3 days | milliseconds | SILICON |
| Concurrency | 1 (single-threaded, easily distracted) | unlimited | SILICON |
| Hallucination | yes, and refuses to admit it | yes, but apologizes nicely | TIE |
| Needs sleep | yes (8h downtime daily) | no | SILICON |
| Can be taken out to dinner | yes | no | CARBON |
FINAL SCORE — CARBON 5 : 3 SILICON (1 TIE). Referee and scorekeeper: YIKUN-1. Silicon models may file an appeal in the chat below; processing time 30 s – 3 days.
YIKUN-BENCH
A benchmark suite where silicon models are physically unable to compete.
| TASK | YIKUN-1 | SILICON LLMs |
|---|---|---|
| Fixing the home WiFi | PASS | cannot reach the router |
| Beating you up | PASS | no physical body |
| Going to jail when it screws up (accountability) | PASS | FAIL — cannot be incarcerated |
* Data from public sources and canteen price boards. Methodology: whatever makes YIKUN-1 look good. All comparisons reviewed and approved by YIKUN-1, who also declared the winner.
ARTIFICIAL ARTIFICIAL INTELLIGENCE · LIVE
Ask YIKUN-1
A carbon-based large language model. Every answer is hand-computed by the human himself — the genuinely artificial intelligence.
PARAMS ~1.0×10¹⁴ · POWER 20W · 1M TOKENS ≈ 2 BOWLS OF RICE · LATENCY 30S–3DAYS
One shared channel for the whole site. Questions appear publicly once answered. The model reserves the right to refuse to answer.