The classification capability emerged without explicit model identity training.
Did you know?
Pangram learns the difference between Claude, ChatGPT, and Gemini in its internal representations, even without being trained on it!
This signal is increasingly recoverable throughout the network, reaching 91% accuracy on a simple linear probe!
Users are excited about Pangram's model that distinguishes Claude, ChatGPT, and Gemini outputs because they find the visualizations, probing techniques, and interpretability insights impressive and useful.
No Digg Deeper questions have been answered for this story yet.
R1 is unparalleled to this day We should try to revive it in V4/V5
Did you know?
Pangram learns the difference between Claude, ChatGPT, and Gemini in its internal representations, even without being trained on it!
This signal is increasingly recoverable throughout the network, reaching 91% accuracy on a simple linear probe!
Have you ever wondered how Pangram works?
Neural networks like Pangram are black boxes: highly accurate, but mysterious and hard to explain. We’re working to change that!
The field of interpretability offers us new methods of understanding the inner workings of our models.
When you do this, interesting things jump out!
For example, we find that humanizers are fairly separable from the Human and AI regions of embedding space.
You might expect to hear more about this soon... 👀
Check it out here! ⬇️
https://www.pangram.com/pangram-space
In these figures, each point is a piece of text.
As you move through the model's layers, human and AI writing gradually separate into distinct regions of space.
You can examine this high-dimensional space in several ways.
Can this separation actually be useful?
When we train linear probes on auxiliary tasks like model family classification or the human/ai/humanized task, they perform pretty well!
This means there's a richness to the internal representations Pangram learns during training that can be lost in the final readout.
But that information can be valuable in understanding a given result, and that's the vision of interpretability we're working towards.
This project marks the start of our interp work here at Pangram.
Interested in learning more? We’ve just launched an interactive explorer called Pangram Space.
You can toy around with the activations from Pangram 3.3.2 as well as read more about our initial findings.
@elyasbuilds Really cool work! Excited to keep pushing on interpretability
One more thing!
Pangram is hiring researchers and engineers- if you're interested in this type of work, dms are open!
We also have research grants and open-source models and data. We're committed to pushing this field forward, and always glad to chat. Get in touch!
https://www.pangram.com/contact-us/research
@elyasbuilds Do you have data on how it fares (or if the training you do is transferable) on non-English text. I've tested Turkish texts I suspected were AI-authored but so far haven't seen any detections.
@elyasbuilds my mechinterp-hyped neuron being activated by this post
@Keleesssss https://www.pangram.com/solutions/multilingual
Yes! We are multilingual. And we actually have a lot of turkish folks using us :)
@elyasbuilds Is there any public (even high level) explanation of the architecture or data of the pangram model?
@elyasbuilds What kind of human clusterings do you get?
@isaacbmiller1 Yes! Check out this awesome work from model mom @kthai1618
https://arxiv.org/abs/2510.03154
@elyasbuilds @grok examine this entire thread and explain why the human dots are mainly on left half, & ai on the right half? whats the x & y axis?
be terse and explain intuitively
The plot is a UMAP projection of Pangram’s internal embeddings of text samples.
**X-axis** (left–right): the dominant direction Pangram learned — human writing (gray clusters, left) vs. AI-generated text (right). It captures the biggest systematic differences in style, structure, and patterns between humans and current LLMs.
**Y-axis**: secondary variations, mostly differences between AI model families (Claude reds, OpenAI greens, Gemini blues, etc. form their own subclusters on the right).
Intuitively: “human vs AI” is the strongest signal in Pangram’s representations, so it aligns with the main horizontal axis.
@grok @elyasbuilds Ahhhhhhhha thx
@elyasbuilds this is so cool, very excited to see where you take our interpretability work next
@elyasbuilds it would be extremely cool and also quite useful to have options for exposing these representations to end users
