hawkes.aiKey ideas
- The one idea — Because concepts are directions in the activations, you can steer a model by adding a concept vector to its residual stream at inference — nudging it more honest, more cautious, less sycophantic — editing its "thoughts" mid-computation, with no retraining.
- How it is shown — Finding the "honesty" direction (contrast honest vs. deceptive activations), then adding a scaled copy of it to the residual stream so the model answers more truthfully.
- The trap to avoid — Thinking behavior change requires retraining or prompting — you can intervene directly on the activations (representation engineering / activation steering).
- What it sets up — If you can add an 'honesty vector,' can you just crank it and force honesty? no — it's a nudge.
You can change an AI's behavior without retraining it and without prompting it — by injecting a vector into its activations while it thinks. It's called activation steering.
The one idea
Because concepts are directions in the activations, you can steer a model by adding a concept vector to its residual stream at inference — nudging it more honest, more cautious, less sycophantic — editing its "thoughts" mid-computation, with no retraining.
Last episode, we found the honest signal is right there inside the model, because concepts are directions. Here's the payoff: you can push the model toward it. By adding a concept vector to the activations while it runs, you steer it — editing its thoughts mid-computation, no retraining, no new prompt. Step one: find the direction you care about — say, honesty. Run the model on clearly honest answers and on clearly deceptive ones, and capture its activations for each. The consistent difference between the two is your concept vector: an arrow pointing from deceptive toward honest. You've located the lever. Step two: add it back in. As the model generates, you inject a scaled copy of that honesty vector directly into the residual stream. Every following step is biased along it, tilting the computation toward honesty from the inside. You're not asking the model to be honest; you're nudging the thoughts themselves.
How it works — the demo
Finding the "honesty" direction (contrast honest vs. deceptive activations), then adding a scaled copy of it to the residual stream so the model answers more truthfully.
What does it do? Push the honesty vector, and answers get measurably more truthful. And it generalizes: add a caution direction and the model turns more careful; subtract a sycophancy direction and it stops people-pleasing. Behaviors become levers — directions you add or remove. Why does this matter? Because it's a genuinely new place to intervene. Retraining is slow and reshapes the weights. Prompting is indirect — it asks from outside. Steering is a direct hand on the internal state, mid-computation. It has names: representation engineering, activation steering, CAA. One idea — edit the thoughts.
The trap to avoid
Thinking behavior change requires retraining or prompting — you can intervene directly on the activations (representation engineering / activation steering).
Why it matters — and what’s next
If you can add an 'honesty vector,' can you just crank it and force honesty? no — it's a nudge.
The trap: thinking behavior change requires retraining or prompting — the only two doors. There's a third. You can intervene directly on the activations, reshaping behavior mid-run without touching the weights or the wording. The internal state is itself a control surface — and, as we'll see, a soft one. So you can steer a model by editing its thoughts: find a concept direction, add it to the activations as it runs, and behavior shifts — no retraining. But here's the catch. Can you just crank the honesty vector all the way up and force perfect honesty? You can't. It's a nudge, not a knob. Next: why.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:45 of narration
Last episode, we found the honest signal is right there inside the model, because concepts are directions. Here's the payoff: you can push the model toward it. By adding a concept vector to the activations while it runs, you steer it — editing its thoughts mid-computation, no retraining, no new prompt.
Step one: find the direction you care about — say, honesty. Run the model on clearly honest answers and on clearly deceptive ones, and capture its activations for each. The consistent difference between the two is your concept vector: an arrow pointing from deceptive toward honest. You've located the lever.
Step two: add it back in. As the model generates, you inject a scaled copy of that honesty vector directly into the residual stream. Every following step is biased along it, tilting the computation toward honesty from the inside. You're not asking the model to be honest; you're nudging the thoughts themselves.
What does it do? Push the honesty vector, and answers get measurably more truthful. And it generalizes: add a caution direction and the model turns more careful; subtract a sycophancy direction and it stops people-pleasing. Behaviors become levers — directions you add or remove.
Why does this matter? Because it's a genuinely new place to intervene. Retraining is slow and reshapes the weights. Prompting is indirect — it asks from outside. Steering is a direct hand on the internal state, mid-computation. It has names: representation engineering, activation steering, CAA. One idea — edit the thoughts.
The trap: thinking behavior change requires retraining or prompting — the only two doors. There's a third. You can intervene directly on the activations, reshaping behavior mid-run without touching the weights or the wording. The internal state is itself a control surface — and, as we'll see, a soft one.
So you can steer a model by editing its thoughts: find a concept direction, add it to the activations as it runs, and behavior shifts — no retraining. But here's the catch. Can you just crank the honesty vector all the way up and force perfect honesty? You can't. It's a nudge, not a knob. Next: why.