Deep Dream and SAE Feature Maximization in Modern LLMs
torchcompiled · x · 2026-07-20
A reply to @matthen2's discussion on the Deep Dream effect in modern large models. The original post demonstrated optimizing an image to maximize the probability of a target caption. It found that Gemma 12B, which lacks a vision encoder, reads pixels just like processing token embeddings, printing recognizable objects on the canvas. Conversely, E4B, which features a vision encoder, tends to lean towards texture drift.
The responder noted that maximizing optimization using individual features identified by Sparse Autoencoders (SAE) might yield even more interesting results.
Related event: Reviving Deep Dream with modern LLMs(2 posts)→
More from Research
- DeepMind paper: LLMs can't sample randomly on their own, only convert given seeds — rohanpaul_ai · 2026-10-27
- DeepMind paper: LLMs can't sample randomly on their own — pass random numbers from code instead — rohanpaul_ai · 2026-10-27
- World Model Assigns Weight and Friction to Objects, Then Simulates with a Physics Engine — jonstephens85 · 2026-10-27
- Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure — Sapienza · 2026-10-27
- Mathathon 2.0 launches with Terence Tao's SAIR, $1.5M in compute for teams — _sathvikr · 2026-10-27
- The 'sweet lesson' is real: two kinds of ML problems need opposite strategies — mike64_t · 2026-10-27