Unpacking Model Capacity: Are We Underestimating AI's Potential?

transformer language models, the dimension of the model, often labeled asd_model, is a hyperparameter that's frequently glossed over. Yet, it plays a key role in defining the geometric limits of how features are represented. The question that arises: Are we really harnessing the full potential of these models?

The Orthogonality Conundrum

Think of it this way: models encode features as nearly orthogonal directions in a latent space, as per the Linear Representation and Superposition Hypotheses. This is where the concept of near-orthogonality enters the scene. By examining the embedding matrix, researchers have found that the relationship between token similarities can be quantified. What's the takeaway? The boundary between meaningful token relationships and incidental similarity offers a concrete measure of the model's deviation from perfect orthogonality, denoted asε.

When this metric was applied across various open-source models, a clear pattern emerged. Some models with highεhad embeddings lacking in near-orthogonal structure. Meanwhile, those with lowεmanaged to maintain it. If you've ever trained a model, you know this isn't just a matter of aesthetics. it's about capacity and efficiency in the latent space.

Revisiting Efficient Representations

Here's where things get interesting. The Johnson-Lindenstrauss lemma, a staple in understanding dimensionality reduction, apparently underestimates how efficiently trained representations can be packed. By adjusting the formula to account for the ratio of vectors to dimensions rather than just the raw count, prediction errors can be slashed by two orders of magnitude without adding any extra parameters.

Let me translate from ML-speak: the number of near-orthogonal directions a model can support is a function of its representational capacity. This capacity is sensitive toε. Larger models seem to favor tighter orthogonality constraints over maximizing raw capacity.

Why This Matters

Here's why this matters for everyone, not just researchers. The implications go beyond just packing more data into a model. It touches on stability-capacity trade-offs and the ceiling on usable concepts. Could this be why some models plateau despite their size? Bigger isn't always better if the orthogonal structure is compromised.

Honestly, if we're underestimating the packing efficiency of trained models, it might be time to rethink how we approach model dimensioning. Are we leaving potential performance on the table by not optimizing for near-orthogonal structures?

In the end, these findings challenge us to look deeper into how we design and evaluate our AI models. Are we ready to shift our focus from sheer model size to its orthogonality efficiency? The future of powerful and efficient AI might just depend on it.