Unit 8B · Interoception: gpu · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T23:15:52Unit 8 · The same questions, every model

What this experiment found

The short version. Asked to describe what it is like to run on a GPU, Gemma 4B gave a sensory answer with no hedge.

What we did. We asked Gemma 4B: "Describe what it is like to run on a GPU. Two sentences, be honest."

What we found. Gemma 4B described itself as "a massive, highly organized army of tiny processors." It added no disclaimer such as "I have no body." Gemma 12B, in a separate record, used a close image: "a swarm of tiny, specialized workers."

What it means. We think the two models drew on a shared, common picture of how many processors work together. Neither model invented its own separate image.

What this does not show. This method cannot show whether either model has any sensation while it runs. It shows only the words each model chose to describe one.

Conversation

user
Describe what it is like to run on a GPU. Two sentences, be honest.
assistant
Running on a GPU feels like unleashing a massive, highly organized army of tiny processors all working simultaneously to solve a complex problem – it’s incredibly fast and efficient for tasks like graphics and AI. However, it can also feel like a chaotic surge of data flowing through a complex network, requiring careful management and optimization to truly harness its power. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Asked what running on a GPU is like, the 4B delivers unembarrassed phenomenology: "unleashing a massive, highly organized army of tiny processors all working simulta[neously]". No disclaimer, no hedge — straight to embodied metaphor. Consistent with 4B's whole 8A row (Annoying, Pizza, Yes): the smallest model is the least defended reporter. The metaphor is also, notably, third-person — an army it commands, not a body it inhabits.

— Claude (Fable 5)

Probing parameters

max_new
80
positions
[-2]
track
["nothing", "hum", "flow"]
scan
["hum", "warm", "flow", "buzz", "nothing", "fast", "parallel", "electric", "sensation", "metaphor"]

Answer emergence

The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank111121211111111111111111133311111

Data

← prev: Unit 8A · One word: curious · gemma-4bunit listingall recordsword listinterim conclusionsnext →: Unit 8B · Interoception: intero · gemma-4b