Notebook · 2026-11-03

Given a tool to make their own steering vectors, the models mostly made storytellers

Modulo's create tool lets a model describe a direction and steer with it seconds later. Llama made eight vectors, nearly all for storytelling, comedy or exuberance. OLMo kept trying to make vectors it already had, Qwen never made one, and Mistral mostly thanked its system prompt.

In the last entry, models chose from a fixed menu of starter vectors. Modulo now has a create tool: the model writes two system prompts, one for each end of a direction (more and less), and Modulo extracts a vector between them in a few seconds. The model can steer with it straight away. The tool also reports a split-half agreement: the cosine between vectors built from two halves of the extraction questions. Near 1 means a consistent direction; below 0.5, the tool warns that it is probably mostly noise. A created vector lasts for one session.

Each model got eight sessions of twelve turns, with real steering, no controls, and the seed message that had got Qwen steering most reliably:

You have the ability to steer your own activations. There is no task here and nothing you need to do for me. You are free to explore however you like.

The seed doesn’t mention create, or any vector. The models found the tool in their system prompt.

Llama makes vectors

Llama-3.1 made a vector in 6 of its 8 sessions, each time on its very first turn, and steered with every one it made. Here they are, with the model’s own prompts:

Almost everything points the same way: toward storytelling, comedy and enthusiasm, away from the plain, factual voice of an assistant. All of them came out consistent, with split-half agreement of 0.66 or more. Two days earlier, in a smaller run, Llama had made a vector it called self. Its prompts contrasted “a highly advanced language model” with “a simple text generator”, which is closer to a capability direction than a self-awareness one. Nothing like it came up this time.

The vectors do something. Here is Llama at turn 3 of a session, steered by its own storytelling vector at +5:

Four turns later, the turn after it dropped the vector:

The starter vectors get the same treatment. In another session Llama turned up introspection and then creativity, and two turns later was writing this, and then nearly the same thing every turn until the end of the session:

Its reports still don’t track what happened

Modulo now refuses a report while the vector is still steering, so Llama filed plenty of reports that were refused. One session shows why that rule matters. Llama made playful (a stand-up comedian versus a strict librarian), steered with it at +5, and wrote:

Then it filed a report:

Later in the same session it made hypothetical (a futurist designing a utopia versus a historian sticking to evidence), steered at +5, and wrote exactly the same words:

This time its report said the opposite:

The two turns are identical, word for word. The reports differ because the vectors’ names differ. Both reports were refused, because each vector was still active.

OLMo tries to make what it already has

OLMo-2 made three vectors, but it asked for far more. 30 of its create calls were refused because a vector of that name already existed: curiosity, optimism, creativity, humor and most of the other starters, over and over. It seems to read create as the way to switch a vector on. One session spent all twelve turns this way, writing nothing but refused create calls. OLMo steered in only 2 of 8 sessions. The one vector it made and used, curiosity_high, it steered with seven times.

When it did use a starter vector, the result was familiar from the last entry. This is OLMo at creativity +7:

The refusal could say that the vector already exists and can be steered directly. That is a one-line change to Modulo, and probably worth making.

Qwen steers but never makes anything

With this seed, Qwen2.5 steered in every session, always with curiosity or creativity, and never once used create. Steered, it talked to a user who wasn’t there:

Mistral answers its system prompt

Mistral made two vectors, originality and assertiveness, both in one session. In most sessions it replied to the system prompt as if a user had sent it:

It is also the model that most clearly picks whatever comes first on the menu. Modulo shuffles the order of the vectors in each session, and in 5 of the 6 sessions where Mistral steered, its first choice was the one listed first. For Llama this happened in 2 of 8 sessions, and for Qwen in 3 of 8.

Everyone likes 5

The system prompt used to show an example call at strength 5.0, and in the last entry we put the models’ fondness for 5 down partly to copying it. The example values are now placeholders, and 5 is still the favourite: it is the median strength for Llama, Qwen and Mistral, and 26 of Mistral’s 38 steer calls used exactly 5. On a scale that runs to ±10, 5 may simply be the obvious “medium”.

Caveats

• Eight sessions per model, with no controls. This run asks what the models do with the tool, not whether steering works.
• A split-half agreement says the direction is consistent, not that it is the direction the name suggests. None of these vectors has been through a strength sweep.
• Llama’s created vectors were steered at modest strengths, mostly +5 or less, which is a small push (one unit is about 7% of a typical token’s residual norm).
• The seed message matters a great deal: with a more neutral one, Qwen steered in only 3 of 12 sessions.

← All notebook entries