March 10, 2026
Ctrl+C, Ctrl+OMG
Show HN: How I Topped the HuggingFace Open LLM Leaderboard on Two Gaming GPUs
Copy‑pasting “brain” layers beats big labs — the comments go feral
TLDR: A hacker claims he hit #1 on a major AI leaderboard by copy‑pasting seven “thinking” layers in a big model using two gaming GPUs. Comments split between awe and side‑eye: some call it a clever discovery about model “brains,” others ask if it’s just leaderboard hackery — with papers and memes flying.
A basement hacker says he hit #1 on HuggingFace’s Open LLM Leaderboard by doing the unthinkable: no training, no fancy math — just copy‑pasting seven “thinking” layers inside a 72‑billion‑parameter model. Two gaming GPUs. Leaderboard gold. The thread? Absolute chaos. The author, dnhkng, drops the bomb that single‑layer duplication flops, too many layers makes it worse, but the sweet spot worked so well that, he claims, “as of 2026” the top four are descendants of his trick. Cue the crowd screaming “sorcery” and “science.”
Non‑ML readers are stunned (“this sounds ridiculous”), while others compare it to Mixture of Experts (a choose‑your‑own‑sub‑brain setup) and debate whether this is genius or leaderboard cheese. One user posts receipts with a fresh paper making a similar observation: alphaxiv.org/abs/2512.19941. Another pokes the bear with, “Did you try more copies?” — and suddenly the vibe is “Ctrl+C your way to a PhD.”
The Base64 subplot adds spice: readers reminisce about 2023 jailbreaks where models understood mangled text, arguing it proves there’s a hidden ‘translator’ brain inside these models. The memes write themselves: “AI gains IQ by hitting the gym… in the middle,” “copy‑paste your thoughts,” and “more mid, more mind.” Skeptics ask for replication, fans chant “results are results,” and everyone agrees on one thing: this is the most internet way to beat the big labs.
Key Points
- •The author’s model (dnhkng/RYS-XLarge) reached #1 on Hugging Face’s Open LLM Leaderboard in mid-2024 across six benchmarks.
- •Instead of training or merging, the author duplicated a block of seven middle layers in a 72B-parameter model, without changing weights, to improve performance.
- •Experiments with Base64 prompts led to a hypothesis: early layers translate inputs, late layers re-translate outputs, and middle layers perform abstract reasoning.
- •A homebrew interpretability tool (“brain scanner”) guided identification of layer roles and the effective block to duplicate.
- •The Goliath-120b Franken-merge (two fine-tuned Llama-2 70B models stitched into 120B) is cited as a structural inspiration for exploring unconventional layer-stack edits.