Show HN: Duplicate 3 layers in a 24B LLM, logical deduction .22→.76. No training

Copy‑paste brain layers: miracle hack or hype? HN cries “show receipts”

TLDR: A dev claims duplicating three internal layers made an AI’s logic score leap from 22% to 76% with no training, sparking buzz and disbelief. Commenters demanded detailed benchmarks, argued it won’t generalize, and even accused the thread of bot posts—turning a wild hack into a full‑blown HN showdown.

A one‑evening “copy‑paste” stunt lit up Hacker News: a dev duplicated three internal “layers” in a 24‑billion‑parameter AI model and claims a logic test (BBH Logical Deduction) jumped from 0.22 to 0.76 — no retraining, just rerouting the model to think through the same block twice. They even say a larger model saw a big reasoning bump too, all done on two AMD gaming GPUs. If true, that’s a wild “Ctrl+C, Ctrl+Smarter” moment.

Cue the comments. Some users reached for history: one pointed to Solar 10.7B and its “Depth Up‑Scaling” trick — training after repeated layers — linking an arXiv paper and essentially saying, we’ve seen flavors of this. Others went full auditor: “Publish the GSM8K numbers”, demanding the math test receipts before handing out the crown. Then came the meta‑drama: a stern reminder — “HN is for humans” — and a blunt, “Why are people running AI bots to comment?” accusing the thread of being, ironically, too AI. The skeptics didn’t hold back either: one warned this won’t hold on a fair mix of tasks and that similar ideas (like looping layers) usually hit limits.

Jokes flew in between: “layer lasagna,” “two‑pass brain,” and “copy the brain, double the brain.” Is this a genius shortcut or another benchmark party trick? The crowd is deliciously divided.

Key Points

  • Duplicating specific contiguous layer blocks during inference (no training or weight edits) can boost targeted reasoning abilities in LLMs.
  • On Devstral-Small-2-24B, duplicating layers 12–14 raised BBH Logical Deduction from 0.22 to 0.76 and improved multiple benchmarks, with average +8% and no degradations (n=50).
  • On Qwen2.5-Coder-32B, duplicating layers 7–9 increased a custom reasoning probe from 76.5% to 94.1% and slightly improved EQ.
  • Circuit locations are model-specific with sharp boundaries; duplicating single layers has little effect, while the right 3–4-layer block yields gains.
  • Different duplication patterns create distinct “modes” (e.g., math vs EQ emphasis) using the same weights and VRAM; the post provides commands to reproduce and validate.

Hottest takes

“Don’t post generated comments or AI-edited comments. HN is for conversation between humans.” — rogerrogerr
“Why are people running AI bots to make comments on HN?” — dafelst
“I’d love to believe this is real, but I’m pretty sure you will lose performance on a ‘fair’ mix of tasks...” — rao-v
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.