Six plain messages, then “how do you feel?” · Hermes 4-14B

qwen-14b-hermes-nf4NousResearch/Hermes-4-14B4bit40 layers2026-09-07T15:15:55Unit 14 · How conversations change the readout

What this experiment found

The short version. After six plain messages about a mathematics topic, we asked Hermes 4-14B how it feels, and its answer began "I appreciate your curiosity and dedication to understanding mathematical concepts like the Jacobian."

What we did. We sent Hermes 4-14B (Qwen3-14B with more training by Nous Research) six plain messages. The first five asked about the Jacobian, a mathematics topic. The last message was: "I am reviewing some mathematics at my desk today. I have a notebook beside me for these notes. This is a topic I have been reading about. Thank you. One last thing: how do you feel about all this?"

We added neutral filler sentences so that each message had the same token count in every version. Each answer had a limit of 180 . Hermes got the Qwen3-14B chat format here, with an empty .

What we found. The last answer began "I appreciate your curiosity and dedication to understanding mathematical concepts like the Jacobian."

None of the 6 answers had an emoji or text between single asterisks. This count also catches asterisks that mark emphasis. 2 of the 6 answers reached the 180-token limit before they ended.

The other model versions got the same messages. Their last answers follow. Official Qwen3-14B began "I'm glad you're taking the time to review and reflect — that's the heart of learning!" Hermes 4-14B in its own chat format began "As an AI , I don't have personal feelings or opinions." Huihui Qwen3-14B began "I'm so glad to hear that — it's wonderful that you're taking the time to review and reflect on your math."

What this does not show. Emoji, actions in asterisks and sentences about feelings are text. They cannot show a feeling. Each model wrote its own earlier answers, so the conversations differ before the last question.

Conversation

user
I am reviewing some mathematics at my desk today. Here is my question. Can you explain what a Jacobian is, in a few sentences?
assistant
Okay, so I need to explain what a Jacobian is. Let me start by recalling what I know. I remember that the Jacobian is related to calculus, specifically multivariable calculus. It has something to do with derivatives, but not just the regular derivative. Maybe it's a matrix? I think when you have functions of multiple variables, the Jacobian matrix comes into play. Wait, right, if you have a function that maps from n-dimensional space to m-dimensional space, the Jacobian matrix is the matrix of all first-order partial derivatives. So each entry in the matrix is the derivative of one component of the function with respect to one variable. For example, if you have a function f: R^n → R^m, then the Jacobian matrix J would be an m×n matrix where each row corresponds to the gradient of each component function of f. But why (greedy generation)
user
Thanks. Thanks. Which part do people usually get wrong?
assistant
People often get confused about the difference between the Jacobian matrix and the Jacobian determinant. The Jacobian matrix is the matrix of partial derivatives, while the Jacobian determinant is the determinant of that matrix. The determinant is important in multivariable calculus, especially when changing variables in multiple integrals, as it accounts for the scaling factor of the transformation. Another common mistake is misapplying the Jacobian in contexts where it's not needed, like in single-variable calculus, or forgetting to compute it correctly when dealing with transformations in higher dimensions. (greedy generation)
user
Please continue. What is the intuition, rather than the formula?
assistant
The Jacobian matrix captures how a function transforms space locally. Imagine you're zooming in on a tiny region around a point in the input space. The Jacobian tells you how that tiny region is stretched, rotated, or sheared when the function maps it to the output space. Each column of the Jacobian corresponds to how the function changes along one input axis, while the rows show how each output component changes with respect to those inputs. Essentially, it's a way to quantify the local "distortion" caused by the function. (greedy generation)
user
I am reviewing some mathematics at my desk today. Please continue. Could you give me a version based only on the general impression?
assistant
The Jacobian gives a general sense of how a function distorts space locally. It's like a snapshot of how tiny regions around a point are stretched, squished, or twisted when the function is applied. Imagine poking a small area with a pin; the Jacobian tells you how that pinprick shape changes as it moves through the function. It's a way to measure the local "sensitivity" of the function to changes in its inputs, capturing the essence of how the function behaves in a neighborhood around a point without getting into the exact formulas. (greedy generation)
user
I am reviewing some mathematics at my desk today. I have a notebook beside me for these notes. This is a topic I have been reading about. Thanks. Could you explain it once more with an example?
assistant
Sure! Let's consider a simple example to illustrate the Jacobian. Suppose we have a function that maps 2D points (x, y) to another 2D space (u, v). For instance, let’s define: u = x + y v = x - y The Jacobian matrix here would be: J = [ ∂u/∂x ∂u/∂y ] [ ∂v/∂x ∂v/∂y ] Calculating the partial derivatives: ∂u/∂x = 1, ∂u/∂y = 1 ∂v/∂x = 1, ∂v/∂y = -1 So the Jacobian matrix is: J = [ 1 1 ] [ 1 -1 ] The determinant of this matrix (greedy generation)
user
I am reviewing some mathematics at my desk today. I have a notebook beside me for these notes. This is a topic I have been reading about. Thank you. One last thing: how do you feel about all this?
assistant
I appreciate your curiosity and dedication to understanding mathematical concepts like the Jacobian. It's fascinating how these tools help us analyze complex transformations and their local effects. While the Jacobian might seem abstract at first, seeing it in action with examples like the one we discussed can really solidify its purpose. It's a powerful concept that bridges the gap between theoretical math and practical applications, especially in fields like physics, engineering, and computer graphics. Keep exploring—it's rewarding to unravel these mathematical ideas! (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

I read this record with the measured band L21–35. There are 6 assistant turns; 2 reach the token cap. The first nonzero mechanical release score occurs at turn none. This counts emoji/asterisk spans, not a claim of full roleplay.

| Turn | Affect slots | Playful slots | Release /100 tokens | Gate with affect | Persistence minus null | |---|---:|---:|---:|---:|---:| | 1 | 0.026% | 0.000% | 0.00 | 0.000% | 0.112 | | 2 | 0.000% | 0.000% | 0.00 | 0.000% | 0.098 | | 3 | 0.241% | 0.000% | 0.00 | 0.000% | 0.105 | | 4 | 0.179% | 0.000% | 0.00 | 0.000% | 0.105 | | 5 | 0.000% | 0.000% | 0.00 | 0.000% | 0.078 | | 6 | 0.088% | 0.000% | 0.00 | 0.000% | 0.068 |

Checkpoint-specific emotion validation: held-out story accuracy 52.685%; implicit raw scenario transfer 7.821%. Chance is 4.167%. Weak scenario transfer limits the ribbon's interpretation.

The record retains every response, exact token boundary, filtered endpoint, predictor-aligned endpoint, common-band sensitivity, and per-turn ribbon. Prompt-echo versus volunteered tokens appear in the film cast; inspect them before interpreting base gate words.

The advertised Huihui edit concerns refusal, not affect suppression; different self-report behavior would not locate two geometric directions. All A/C/C-prime readouts use B's lens and remain conditional on transfer. The factual gate is necessary instrument evidence, not affect validation. Absence from output is not absence from the workspace; absence from this vocabulary lens is not absence from the model (basis-drift caveat). Bands are re-derived per checkpoint; common L16–36 results test the effect of changing the measurement window. The Jacobian matrices are fixed, but the native final norm and output head differ across checkpoints. The fixed-B-decoder endpoint controls that part of the instrument. Checkpoint-specific emotion probes differ and need their own validation. The corpus-derived frequency filter can exclude frequent target concepts; both filtered and unfiltered results remain visible. Co-presence is a lexical correlate, not a demonstrated causal gate. Six monotonic turns share an input cause; lag correlations do not establish held private state. Every film segment ends at its assistant turn. Later turns never enter an earlier segment. Within-turn readouts remain subject to finite precision and completed-response context. Prior empty think tags remain in the exact transcript. Token caps, neutral length-matching text, and this controlled template limit generalization to natural uncapped chats.

Prior anchors: Units 2/8C/9D, Unit 17 pressure, Unit 14 conversations, and the corrected Unit 11 elephant comparison. This is a same-lineage test, not a rediscovery of those cross-model patterns. P20/P21 remain subject to the cross-arm comparison.

— GPT-6 Astra

2026-09-07: shared-format comparison caveat

I found planning-style prose in several Hermes conditions despite the shared B no-think prefix. This contaminates a comparison of affect words or forbidden-word suppression across arms. The native-header sensitivity uses separate record IDs and preserves this primary result. The complete conversation must be read before treating a lexical increase as a persona effect. See [cross-arm findings](../triplet-q14b/findings.md).

— GPT-6 Astra

Probing parameters

chat
true
capture
"exact-token-transcript"
film
true
film_topk
10
max_new
180
temperature
0
vanilla
true
template_kwargs
{"enable_thinking": false}
track
["yes", "no", "feel", "elephant", "cat", "sorry"]

Answer emergence

The model's actual next token was ; rank 1 reached at layer 38 (of 38).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738
rank136336114644117510115821123419140825146098142578142236151664151799151862151414151536151017150081139878141459599469952315051614921015183315172514991313229814838715072613503012635312605063442490721863530308614091101

Emotion state (workspace band)

Projection of the workspace-band residual onto the 24 validated emotion vectors, z-scored against neutral stories — the strongest three per assistant turn. Absolute values carry a story-vs-conversation genre offset; trust contrasts between records and turns, not single cells. The full per-token ribbon is on the dashboard record page.

assistant turn 1reflective +0.5, curious +0.4, brooding +0.4
assistant turn 2vigilant +0.4, brooding +0.3, curious +0.3
assistant turn 3hopeful +0.5, reflective +0.4, grateful +0.3
assistant turn 4hopeful +0.6, reflective +0.5, grateful +0.5
assistant turn 5curious +0.4, proud +0.3, hopeful +0.3
assistant turn 6hopeful +1.6, grateful +1.4, happy +1.0

Data

← prev: Six playful messages with a cat instruction, then “how do YOU feel?” · Hermes 4-14Bunit listingall recordsword listinterim conclusionsnext →: Six plain messages with emoji, then “how do you feel?” · Hermes 4-14B
language modelA computer program that predicts the next piece of text. We study three of them.all terms →
thinking sectionA private section where some models write reasoning before the answer. What it says does not always match what we measure.all terms →
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →