Imagine tweaking your AI's behavior on the fly, like adjusting the bass on your stereo. Instead of retraining the entire model, you could make subtle changes at runtime to achieve specific effects. That's the promise of Activation Transport (ACT), a new technique from Apple researchers that offers fine-grained control over AI models without hefty retraining. Large language models (LLMs) and text-to-image diffusion models (T2Is) often require extensive fine-tuning to align with desired outputs, consuming significant compute resources and potentially impacting performance on other tasks. ACT addresses this by directly manipulating the model’s internal activations during inference, the process of generating text or images. Guided by optimal transport theory, ACT strategically shifts activations to match target distributions. For instance, to reduce the toxicity of an LLM’s output, ACT would shift the activations towards those typically observed when the model generates non-toxic text. This approach cleverly preserves the internal relationships within the activation patterns, ensuring that the AI remains coherent and performs well. Experiments demonstrate ACT’s versatility. In LLMs, it effectively mitigates toxicity, induces specific concepts (like generating text about 'football' or 'clouds'), and enhances truthfulness. In T2Is, ACT enables fine-grained style control, allowing users to adjust the 'sketchiness' of an image, and it tackles the difficult task of concept negation, ensuring an AI can correctly understand instructions like 'don't draw a pink elephant.' What sets ACT apart is its strength parameter (λ), offering continuous control over the degree of intervention. Ranging from 0 (no change) to 1 (full transformation), λ allows users to dial in the desired level of influence without laborious parameter tuning. This intuitive control makes ACT particularly valuable for T2I models, where evaluating model output can be subjective. While ACT offers significant advantages, its current form relies on linear transformations and independent activations, simplifications made for computational efficiency. Future research may explore non-linear maps and joint activation distributions for even finer control. ACT opens exciting possibilities for AI interaction. Imagine easily adjusting an AI assistant’s personality, controlling the style of generated art, or ensuring a chatbot remains helpful and non-toxic, all in real time. This technology could empower users with unprecedented control, shaping AI to meet their specific needs without requiring deep technical expertise.