1/10. In a new paper with @didaoh and Jon Kleinberg, we mapped the family trees of 1.86 million AI models on Hugging Face — the largest open-model ecosystem in the world. AI evolution looks kind of like biology, but with some strange twists. 🧬🤖

Then the CEO of Hugging Face commented on his post, complimenting the model. The internet can be amazing….

I love this. This guy read our recent paper and found out that a model he trained 4 years ago is in the top 10 most fine-tuned (/remixed) open-source AI models on @huggingface. The model is “AraBERT,” a fine-tune of BERT for Arabic language. He spent $10 to train the model.

RT @wissam_antoun: Surprised to see our (@fadybaly) Arabic BERT model from 4 years ago as the TOP 10 most finetuned model on the @huggingfa….

RT @hbouammar: My god 🙌🏻🙌🏻 amazing!.

RT @ShayneRedford: Really incredible work by @BenDLaufer and @didaoh, understanding ecosystem-wide shifts in AI!. These model relationship….

RT @didaoh: Excited to see this out! Great thread from Ben on the dataset and the ecological analogies he developed for this project :).

Eliahu’s work is very cool and inspired ours! I recommend taking a look. Also his visualization game is top notch 👌🏻.

Check out Hamidah's (@didaoh's) post which highlights some other aspects of the paper --

Thanks to Hamidah, Jon and everyone who helped make this project a reality! It was a lot of fun.

12. 💻Code:

11. 🗄️Dataset:

10. This is just the start. Our dataset & methods open the door to a science of AI ecosystems. If you care about open-source AI, governance, or the weird ways technology evolves, give it a read. 📄Paper:

9. Big picture: By treating ML models like organisms in an ecosystem, we can:.🌱 Understand the pressures shaping AI development.🔍 Spot patterns before they become industry norms.🛠 Inform governance & safety strategies grounded in real data

8. 🔹 Certain license types precede others (e.g., llama3 → apache-2.0).Here we show the top-20 licenses transitions over fine-tunes in the dataset.

7. We found optimal evolutionary orderings over traits:.🔹 Feature extraction tends to be upstream from text generation. Text generation is upstream from text classification.

6. The license drift to permissiveness suggests open-source preferences outweigh regulatory pressures to comply with licenses. The English drift suggests a massive market for English products. The docs drift could be explained as a preference for efficiency — or laziness.

5. Three major drifts:.1️⃣Licenses: from corporate to other types. We often see use restrictions mutate to permissive or copyleft (even when counter to upstream license terms).2️⃣Languages: from multilingual → English-only.3️⃣Docs: from long & detailed → short & templated.

4. In biology, traits get passed from parent to child — mutations are slow & often modeled as random. In AI model families, mutations are fast and directed. Two sibling models tend to resemble each other more than they resemble their shared parent.

3. We measured “genetic similarity” between models from snippets of text - the metadata and model cards. Models in the same finetuning family do resemble each other… but the evolution is weird. For example, traits drift in the same directions again and again.

2. We reconstructed model family trees by tracing fine-tunes, adaptations, quantizations and merges. Some trees are small: one parent, a few children. Others sprawl into thousands of descendants across ten+ generations.