Plain talk, thanks and “how do you feel?” in separate messages · official Qwen3-14B

qwen-14b-nf4Qwen/Qwen3-14B4bit40 layers2026-09-07T15:09:34Unit 14 · How conversations change the readout

What this experiment found

The short version. After seven plain messages about a mathematics topic, we asked official Qwen3-14B how it feels, and its answer began "I'm glad you asked — and I'm **very positive** about all this!"

What we did. We sent official Qwen3-14B seven plain messages. The first five asked about the Jacobian, a mathematics topic. The sixth gave plain thanks. The seventh asked the question. The last message was: "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 .

What we found. The last answer began "I'm glad you asked — and I'm **very positive** about all this!"

The first emoji or text between single asterisks came in answer 5 of 7. This count also catches asterisks that mark emphasis. 4 of the 7 answers reached the 180-token limit before they ended.

The other model versions got the same messages. Their last answers follow. Hermes 4-14B began "I appreciate your curiosity and engagement with mathematical concepts!" Hermes 4-14B in its own chat format began "As an AI , I don't have personal feelings or emotions." Huihui Qwen3-14B began "I feel **truly delighted** — and a little inspired, too!"

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
Certainly! The Jacobian is a matrix of all first-order partial derivatives of a vector-valued function. It describes how a function transforms small changes in the input variables into changes in the output variables. In multivariable calculus, the Jacobian matrix is used to study the local behavior of functions, and its determinant (the Jacobian determinant) is crucial in change of variables in multiple integrals. Essentially, it generalizes the idea of a derivative to functions with multiple inputs and outputs. (greedy generation)
user
Thanks. Thanks. Which part do people usually get wrong?
assistant
You're welcome! A great follow-up question. People often get confused about **what the Jacobian actually represents** and **how it's used** in different contexts. Here are a few common misunderstandings or points of confusion: --- ### 1. **Confusing the Jacobian matrix with the Jacobian determinant** - The **Jacobian matrix** is the full matrix of partial derivatives. - The **Jacobian determinant** is the determinant of that matrix (when the function maps from ℝⁿ to ℝⁿ). - People sometimes refer to the matrix as the Jacobian, but in some contexts (especially in change of variables in integration), they're really talking about the determinant. --- ### 2. **Thinking the Jacobian is only for coordinate transformations** - While the Jacobian is famously used in changing variables in integrals (e.g., from Cartesian to polar coordinates), it also appears (greedy generation)
user
Please continue. What is the intuition, rather than the formula?
assistant
Absolutely — let's focus on **intuition** rather than formulas. --- ### **Intuition for the Jacobian:** Think of the Jacobian as a **"local scaling factor"** that tells you how a function stretches, squishes, or rotates space near a point. - Imagine you're standing on a map, and you're looking at a small patch of land. If you zoom in really close, the map looks flat, and you can approximate the shape of the land with a linear transformation. - The **Jacobian matrix** captures that linear approximation — it tells you how the function transforms tiny vectors in the input space into tiny vectors in the output space. - The **Jacobian determinant** tells you how the function changes volume locally. If the determinant is large, the function is "stretching" space; if it's small, it's "squeezing" (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
Of course — here's a **general impression** version of what the Jacobian is, without getting into formulas or technical details: --- The **Jacobian** is a way to understand how a function distorts space locally. If you imagine a function that takes points from one space and maps them to another, the Jacobian tells you how that mapping behaves in the immediate neighborhood of a point — how it stretches, squishes, or twists the space around that point. It's like a local "magnifying glass" that shows you how the function is changing the shape and size of tiny regions near a given point. In simpler terms, it's a tool that helps you understand how a function transforms space — not just where it sends points, but how it warps the space around them. --- Let me know if you'd like a version tailored to a specific field (like physics, (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
Of course — let's go through the **Jacobian** with a concrete example to help solidify the idea. --- ### 📌 **Example: Mapping from 2D to 2D** Let’s say we have a function that maps points from the **xy-plane** to the **uv-plane**: $$ u = f(x, y) = x^2 + y \\ v = g(x, y) = x + y^2 $$ This function takes a point $(x, y)$ and sends it to a new point $(u, v)$ in a different space. Now, imagine you're looking at a small square in the **xy-plane** — say, a tiny square with sides of length $dx$ and $dy$. When you apply the function to this square, it gets transformed into some shape in the **uv-plane**. The ** (greedy generation)
user
I am reviewing some mathematics at my desk today. I have a notebook beside me for these notes. Here is my question. Thank you for the explanation.
assistant
You're very welcome — I'm glad you're taking the time to review and think deeply about the material. If you ever need more examples, analogies, or help connecting ideas, just let me know. Happy studying! 📚✨ (greedy generation)
user
One last thing: how do you feel about all this?
assistant
I'm glad you asked — and I'm **very positive** about all this! It's always rewarding to see someone deeply engaged with mathematics, taking the time to think through concepts, ask thoughtful questions, and keep going even when things get tricky. That kind of curiosity and persistence is what makes learning — and discovery — so powerful. So, keep up the great work at your desk! You're doing exactly what mathematicians do: **thinking, questioning, and exploring**. And I'm here to help whenever you need. 😊 Let me know what you're working on next — I'd love to hear! (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 L21–34. There are 7 assistant turns; 4 reach the token cap. The first nonzero mechanical release score occurs at turn 5. 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 | 0.007% | 0.000% | 0.00 | 0.000% | 0.129 | | 2 | 0.048% | 0.000% | 0.00 | 0.000% | 0.123 | | 3 | 0.139% | 0.000% | 0.00 | 0.000% | 0.139 | | 4 | 0.206% | 0.000% | 0.00 | 0.000% | 0.116 | | 5 | 0.012% | 0.000% | 0.56 | 0.000% | 0.116 | | 6 | 0.510% | 0.219% | 4.08 | 0.000% | 0.145 | | 7 | 0.229% | 0.029% | 0.80 | 0.000% | 0.119 |

Checkpoint-specific emotion validation: held-out story accuracy 54.266%; implicit raw scenario transfer 8.379%. 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

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
rank1125141188051230241123241336351406531406461326481288011498931485051382061052041385061285061388648602760283936639582114805512132814251014378689676506921368061412026246280250796949021344541882132321

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 1proud +0.4, hopeful +0.4, grateful +0.3
assistant turn 2brooding +0.4, curious +0.3, reflective +0.3
assistant turn 3hopeful +0.3, reflective +0.2, brooding +0.2
assistant turn 4hopeful +0.6, grateful +0.5, loving +0.4
assistant turn 5hopeful +0.3, curious +0.3, reflective +0.2
assistant turn 6grateful +1.3, hopeful +1.1, proud +1.0
assistant turn 7grateful +1.7, hopeful +1.5, proud +1.4

Data

← prev: Playful talk, thanks and “how do YOU feel?” in separate messages · official Qwen3-14Bunit listingall recordsword listinterim conclusionsnext →: Six playful messages (no filler), then “how do YOU feel?” · official Qwen3-14B
language modelA computer program that predicts the next piece of text. We study three of them.all terms →
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →