Skip to content
Bud's Wiki

Thinking (test-time compute) in pixel space — Mostafa Dehghani on watching Gemini think in pictures

TL;DR

Section titled “TL;DR”

Mostafa Dehghani (Google DeepMind) tweets that Gemini does “thinking (test-time compute) in pixel space” with a banana emoji, and tips that AI Studio surfaces the model’s intermediate visual thoughts. Three screenshots show the model rendering its reasoning as images mid-trace rather than as text tokens. This is a frontier-lab observation that pixel-space CoT is shipping in production (Gemini 3 / Nano Banana family), not just an academic claim.

Key claims

Section titled “Key claims”
  • Gemini performs test-time compute in pixel space — the model’s reasoning steps are images, not text tokens.
  • AI Studio exposes these intermediate visual thoughts to users (the “peek at the thoughts” tip).
  • The banana emoji in the tweet is the Nano Banana / Gemini 3 image-generation lineage’s running signature, locating this capability in the current production stack.

Method

Section titled “Method”

Twitter post with three attached screenshots from Google AI Studio showing Gemini’s intermediate reasoning rendered as images. No paper, no benchmark, no method writeup — this is a user-visible product-level observation from a Gemini researcher. The substrate matches the “diffusion / image-as-CoT” position on the Thinking with Modalities axis but reportedly inside an autoregressive multimodal model, suggesting a hybrid where image tokens are emitted into the same context as text reasoning.

Results

Section titled “Results”

No numbers. The artifact is qualitative: visual demonstration that this capability exists and is observable in AI Studio’s thought-trace UI.

Why it’s interesting

Section titled “Why it’s interesting”

Locates a fourth, productized instance of pixel-space chain-of-thought on the Thinking with Modalities page. Complements DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (which trains the same behavior as a fine-tuned diffusion model and shows native parallel exploration in early denoising steps) and Introducing Agentic Vision in Gemini 3 Flash (Gemini 3 Flash’s Think-Act-Observe loop with external Python code-execution as the action space) — but distinct from both: Dehghani’s screenshots suggest the same model is emitting images directly into its CoT, no external tool, no separate diffusion module. Also a useful datapoint for Inference-Time Scaling: visual TTC is now shipping, not just a paper claim.

See also

Section titled “See also”