Code Data Reshapes Model Workspaces
eliebakouch · x · 2026-07-19
This post shares new findings regarding jspace, focusing on version changes between @poolsideai's Laguna XS.2 and XS.2.1, and the optimizer's impact.
The core conclusion is that when fitted with code prompts, the workspaces of the two versions differ significantly. With wikitext prompts, they are generally similar, though structural similarities remain in the final layers, aligning with the "motor" sections mentioned in Anthropic's paper. The author believes the transition from XS.2 to XS.2.1 is likely driven by larger volumes of code data, significantly altering the model's internal representations.
More from Research
- AI folks, stop saying "biology" when you mean drug development — KevinKaichuang · 2026-10-26
- ostris pre-training "Primal VAE": 1-channel latents match modern 8x VAEs via patching — ostrisai · 2026-10-26
- Shanghai AI Lab-led team releases 100-hour egocentric wearable dataset for humanoid training — ChuckDBrooks · 2026-10-26
- Flow Reasoning Models turn flow matching into iterative reasoners, hitting 99.5% on Sudoku-Extreme — alec_helbling · 2026-10-26
- Surprising find: Alzheimer's-linked amyloid fibrils deliver RNA to sperm for embryo translation — EricTopol · 2026-10-26
- HarnessMM cuts agent tokens per task by up to 88% while raising success rates — omarsar0 · 2026-10-26