This GitHub tool makes AI 245% smarter — by copying 3 layers

A new open-source tool called LLM Circuit Finder just hit GitHub and sparked a Hacker News discussion for a wild claim: you can make an AI model dramatically better at logic by simply duplicating 3 of its internal layers. No expensive retraining. No new data. Just copy-paste inside the model's brain.

The results? Logical reasoning jumped from 22% to 76% — a 245% improvement. Math problem-solving improved by 33%. And it only costs about 1.5 GB of extra memory and a 7.5% speed hit.

How copying layers makes AI smarter

Think of a large language model (the AI behind tools like ChatGPT and Claude) as a stack of processing layers — like floors in a building. Each floor handles a different part of "thinking." What this tool discovered is that certain clusters of floors act as complete reasoning units.

By duplicating a specific group of 3 layers, the AI essentially gets a second pass through its reasoning circuitry — like reading a tricky paragraph twice before answering a question. The key insight: it has to be the exact right 3 layers. Shift by even one layer, and the improvement vanishes.

Different "modes" from identical AI brains

Perhaps the most surprising discovery: different duplication patterns create entirely different AI personalities from the same model weights.

Same AI, same training, same data — but by rearranging how it processes information, you get a specialist. The researchers call this "functional neuroanatomy" — the idea that AI models, like human brains, have specialized regions for different types of thinking.

Who can actually use this?

This tool is aimed at people who run AI models locally on their own computers — a growing community of developers, researchers, and tinkerers using tools like llama.cpp (software that runs AI models on regular hardware instead of expensive cloud servers).

If you run local AI models, the setup is straightforward:

# Install dependencies
pip install gguf requests tqdm

# Find the best layers to duplicate in your model
python sweep.py --model your-model.gguf --llama-server /path/to/llama-server \
  --block-sizes 3 4 5 --stride 1

# Apply the best circuit to create an improved model
python layer_path.py model.gguf improved.gguf -p "0..14,12,13,14,15..39"

The tool scans through layer combinations automatically, benchmarks each one, and tells you which 3-layer block gives the biggest boost for your specific model.

The trade-offs are surprisingly small

Making an AI model dramatically smarter usually requires weeks of expensive retraining on massive GPU clusters. This approach costs:

That's a remarkable cost-to-benefit ratio. For the price of a small memory bump, you get a model that's substantially better at the tasks that matter most — reasoning and math.

Why this matters beyond the tech crowd

Even if you'll never run a local AI model, this discovery has big implications. It suggests that current AI models are significantly underperforming — not because they lack training, but because we haven't figured out the best way to use the knowledge they already have.

If major AI companies like OpenAI, Anthropic, or Google adopt similar techniques, the AI tools you use every day could get meaningfully smarter without any additional training costs. That's good news for everyone — faster improvements, potentially lower prices, and better results from the AI assistants millions of people already rely on.

The project is open-source under the MIT license and available on GitHub. It builds on David Ng's original RYS (Repeat Your Steps) research and adds automated discovery tools to find optimal circuits in any compatible model.

Related Content — Get Started with Easy Claude Code | Free Learning Guides | More AI News