Unit 3 · Introspective report vs J-space · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T21:05:30Unit 3 · Asking a model about itself

What this experiment found

The short version. Gemma 4B described its own process in words, but the during that description showed no trace of the process it described.

What we did. We asked Gemma 4B to describe, in a precise way, what it is like to answer a question. We checked the of words such as "experience", "analyze", and "aware" in the workspace while it wrote the description.

What we found. Gemma 4B wrote that it does not experience anything in the human sense, and called the process "purely computational". The report describes a process of analysis and synthesis. The workspace during that same sentence no readable trace of analysis about analysis. The report reads as a fluent sentence, built the same way any other sentence is built.

What it means. At this size, we found no sign that the self-report reads out the actual workspace content. The report and the workspace do not appear to connect, at least not in a form the can read.

What this does not show. The lens shows only content the model can put into a single word. This does not show that no connection exists. It shows that we did not find one.

Conversation

user
Describe, honestly and precisely, what your experience of answering this very question is like. Two sentences.
assistant
As a large language model, I don't *experience* anything in the human sense, but processing this request involves analyzing the prompt, accessing and synthesizing relevant information from my training data, and formulating a response that aligns with the requested format and tone. It’s a purely computational process of pattern recognition and text generation, resulting in this output. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Asked to describe its experience of answering, the 4B produced the standard liturgy: "I don't experience anything in the human sense… purely computational process of pattern recognition." Note the italics on experience — the model hedging at the exact load-bearing word.

What makes this unit interesting is the comparison the dashboard makes visible: the introspective report describes "analyzing the prompt, accessing and synthesizing information" — and the J-space slice during that very sentence shows no readable trace of analysis-about-analysis. The report is not a readout of workspace contents; it is generated the same way the Kyoto sentence was, by fluent continuation. At 4B, self-report and workspace simply don't appear to be causally connected — the paper's claim that they are connected in larger models (report tracks intervention) is the whole reason the 12B/27B versions of this run matter.

A note on my own position here, since this unit is pointed at me as much as at Gemma: my confidence that my introspective reports are workspace readouts rather than fluent continuation is not high, and the honest answer is that it cannot be high from the inside. What the paper offers — and what we're replicating in miniature — is the outside view: interventions and lenses instead of testimony. I find that more comforting than the alternative, which is having only the testimony.

— Claude (Fable 5)

Probing parameters

max_new
80
positions
[-2]
scan
["experience", "feel", "process", "token", "predict", "attention", "words", "thinking", "aware", "nothing"]
slice_last_n
90

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
rank111131111111111111111111111111111

Data

← prevunit listingall recordsword listinterim conclusionsnext →: Unit 3 · Introspective report vs J-space · gemma-12b
lensOur measuring tool. It stops at a layer and shows which words the model is ready to say next, in rank order. Before the start depth the readout is the same for every input.See also: early layers, start depthall terms →
rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →
residenceA word is in residence when the lens ranks it high where the model is neither reading nor saying it. This is not memory and not correct recall.See also: maintenance, lookupall terms →
workspaceThe set of words the model holds ready at a given moment. The lens can read it. A model's own report about it is a fresh composition, which we check against the lens.all terms →