July 15, 2026
Poked the AI, poked the comments
J-space comparisons across open models
AI nerds map how model “thoughts” move — and the comments instantly go feral
TLDR: Researchers tested how a small internal nudge changes what an AI says later, revealing broad stages inside open models and how long those effects last. The community response was peak internet: one person called it “vibe coded,” another played acronym cop, and someone else brought in Viking poets.
A dense new post about comparing open AI models somehow turned into a mini comment-section spectacle, because the core idea is both wildly important and very easy to roast. The research tries to measure how a tiny shove inside a model changes the words it wants to say later. In plain English: if you poke the AI at one stage, how much does that ripple forward? The author says this reveals hidden “blocks” inside models — roughly an early reading phase, a big middle “workspace,” and a late writing phase — plus a rough count of how much mental room that middle zone seems to have.
But the comments? Oh, the comments came in with equal parts curiosity, nitpicking, and chaos. One of the loudest reactions was basically, “be serious, this is vibe-coded science” — with user suddenlybananas openly admitting they used AI tools to brainstorm and run experiments on a cluster, which gave the whole thing a very “research by autocorrect and ambition” energy. That confession is either refreshingly honest or mildly scandalous, depending on how strict you are about lab-coat vibes.
Then came the inevitable glossary police: one commenter swooped in to clarify that the “J” means Jacobian, the math term behind the whole measurement, in a classic internet move where someone sees confusion and says, “Actually…” And because no online research thread is complete without someone opening a portal to another dimension, a third commenter dragged in Viking poets, claiming ancient-sounding literary wisdom was somehow ahead of AI interpretability. Naturally, that turned the mood from sober science to half lecture, half meme séance. The result: a serious attempt to understand how AI stores and carries ideas, wrapped in a comments section that treated it like a crossover episode between a math seminar, Reddit snark, and a prophecy thread.
Key Points
- •The article defines Jℓ as the average output response to perturbing a model’s residual stream at a specific layer.
- •Multiplying Jℓ by the unembedding yields 4,096 token steering vectors that form a layer dictionary.
- •CKA is used to compare dictionaries in a coordinate-free way; same-model fits score about 0.997, while different trained models at matched depth score around 0.5–0.7.
- •Participation ratio estimates the effective dimensionality of a layer’s dictionary, with real mid-layer dictionaries spanning roughly 200–600 directions.
- •A temporal-horizon experiment measures how layer perturbations persist across token distances, finding that only a fraction of immediate effect survives to the next token and less at deeper layers.