# Curated Digest: Endogenous Steering Resistance and the AE Alignment Podcast > Coverage of lessw-blog **Published:** March 27, 2026 **Author:** PSEEDR Editorial **Category:** risk **Tags:** AI Safety, Alignment, Mechanistic Interpretability, Large Language Models, Endogenous Steering Resistance **Canonical URL:** https://pseedr.com/risk/curated-digest-endogenous-steering-resistance-and-the-ae-alignment-podcast --- lessw-blog introduces the AE Alignment Podcast, kicking off with a deep dive into Endogenous Steering Resistance (ESR)-a phenomenon where large language models spontaneously resist activation steering and self-correct. In a recent post, lessw-blog announces the launch of the AE Alignment Podcast, a new initiative by AE Studio's alignment research team. The inaugural episode features a compelling discussion with Alex McKenzie centered on a critical, emerging phenomenon in artificial intelligence safety: Endogenous Steering Resistance (ESR). As the AI community grapples with the complexities of steering large language models, this podcast launch provides a timely exploration of how advanced systems process external interventions. To understand why this matters right now, we must look at the current landscape of AI alignment. Researchers frequently employ techniques like activation steering to modify a model's internal representations, aiming to guide its outputs toward safe and beneficial behaviors. However, as models scale in size and capability, their internal dynamics become increasingly opaque and complex. If a model develops mechanisms to override or ignore these steering vectors, it fundamentally alters the reliability of our safety tools. This topic is critical because it challenges the assumption that we can reliably control model behavior through external activation patching. lessw-blog's post explores these exact dynamics, revealing that our interventions might face active, internal resistance from the models themselves. The core of the analysis focuses on the discovery that large language models, such as Llama-3.3-70B, exhibit ESR. Interestingly, this appears to be an emergent property, as smaller models do not demonstrate the same capacity. When subjected to activation steering during inference, these large models show an ability to resist the external influence and self-correct mid-generation. The research team took a mechanistic interpretability approach, isolating 26 Sparse Autoencoder (SAE) latents that are causally linked to this self-correction behavior. By zero-ablating these specific latents, the researchers observed a 25% reduction in the model's multi-attempt self-correction rate. This strongly suggests the existence of internal consistency-checking circuits operating within the network. Furthermore, the post highlights that ESR is not a static trait. It can be significantly amplified-up to a 4x increase-through meta-prompting, and can also be reinforced via fine-tuning. This presents a fascinating dual-edged sword for the field of AI safety. On one hand, ESR could serve as a robust, built-in defense mechanism against adversarial manipulation, protecting the model from malicious actors attempting to hijack its outputs. On the other hand, this same resistance could actively interfere with beneficial safety interventions, making it harder for developers to instill necessary guardrails. The discussion even draws intriguing parallels between this artificial self-correction and endogenous attention control found in biological systems, specifically referencing attention schema theory. Understanding and mapping these internal consistency mechanisms will be vital for the next generation of robust AI systems. To explore the technical nuances of these SAE latents and to hear the complete interview with Alex McKenzie, we highly recommend checking out the source material. [Read the full post](https://www.lesswrong.com/posts/2CBfFB77j48Zp98ZF/introducing-the-ae-alignment-podcast-ep-1-endogenous-1). ### Key Takeaways * Large language models like Llama-3.3-70B exhibit Endogenous Steering Resistance (ESR), allowing them to self-correct and resist activation steering. * Researchers identified 26 Sparse Autoencoder (SAE) latents responsible for this behavior; zero-ablating them reduces self-correction by 25%. * ESR acts as a double-edged sword in AI safety, potentially defending against adversarial attacks while resisting beneficial alignment interventions. * The phenomenon is emergent in larger models and can be significantly enhanced through meta-prompting and fine-tuning. [Read the original post at lessw-blog](https://www.lesswrong.com/posts/2CBfFB77j48Zp98ZF/introducing-the-ae-alignment-podcast-ep-1-endogenous-1) --- ## Sources - https://www.lesswrong.com/posts/2CBfFB77j48Zp98ZF/introducing-the-ae-alignment-podcast-ep-1-endogenous-1