EPISODE · Feb 27, 2026 · 18 MIN
From Prompts to Steering 🚀: Recursive Feature Machines & Concept Vectors in LLMs
from The Deep Dive Lab: Unraveling Materials Science · host Son Hoang
For years, interacting with large language models meant crafting better prompts — refining instructions and hoping the model would comply.But what if prompting is the wrong interface?A breakthrough paper in Science — “Toward universal steering and monitoring of AI models” (Science, 2026, Vol. 391, Issue 6787, pp. 787–792) — introduces a radical shift: instead of talking to AI, we can now steer it from within.Using Recursive Feature Machines (RFM) and Concept Vectors, researchers can:🧠 Monitor internal activations to detect hallucinations more reliably than self-evaluation🎯 Precisely steer model behavior by adding linear vectors in activation space⚡ Improve coding performance dramatically — without retraining🌍 Transfer semantic concepts across languages through simple vector addition🔬 Extract powerful steerable features with fewer than 500 samples in under a minuteThis episode explores the transition from prompt engineering to activation engineering — and what it reveals about the hidden geometry of knowledge inside neural networks.If meaning is just a direction in high-dimensional space… what does that say about human thought itself? 🤯#AI #LLM #MachineLearning #RecursiveFeatureMachine #ConceptVectors #Interpretability #AISafety #DeepLearning #NeuralNetworks #SciencePodcast #deepdivelab
What this episode covers
For years, interacting with large language models meant crafting better prompts — refining instructions and hoping the model would comply.But what if prompting is the wrong interface?A breakthrough paper in Science — “Toward universal steering and monitoring of AI models” (Science, 2026, Vol. 391, Issue 6787, pp. 787–792) — introduces a radical shift: instead of talking to AI, we can now steer it from within.Using Recursive Feature Machines (RFM) and Concept Vectors, researchers can:🧠 Monitor internal activations to detect hallucinations more reliably than self-evaluation🎯 Precisely steer model behavior by adding linear vectors in activation space⚡ Improve coding performance dramatically — without retraining🌍 Transfer semantic concepts across languages through simple vector addition🔬 Extract powerful steerable features with fewer than 500 samples in under a minuteThis episode explores the transition from prompt engineering to activation engineering — and what it reveals about the hidden geometry of knowledge inside neural networks.If meaning is just a direction in high-dimensional space… what does that say about human thought itself? 🤯#AI #LLM #MachineLearning #RecursiveFeatureMachine #ConceptVectors #Interpretability #AISafety #DeepLearning #NeuralNetworks #SciencePodcast #deepdivelab
NOW PLAYING
From Prompts to Steering 🚀: Recursive Feature Machines & Concept Vectors in LLMs
No transcript for this episode yet
Similar Episodes
No similar episodes found.
Similar Podcasts
No similar podcasts found.