Conceptual post distinguishing "anthropomorphic" from "biological" reasoning for human analogies about LLM behavior, arguing biological analogies from RL-trained organisms give more reliable primitives for model behavior.
Proves formal results on which computational abstractions survive changes of representation, motivated by the Natural Abstraction Hypothesis and interpretability for alignment. It shows only behavioural equivalence or coarser is fully translation-invariant.
Christiano returns to ARC as executive director and argues for its bet on mechanistic explanations of neural networks to predict generalization and detect misalignment, alongside hiring calls.
Proposes a state-space account of functional emotion in LLMs, arguing that alignment should be framed as durable Aristotelian disposition rather than output conformity, and suggesting interventional tests.
Announces PIRAMID, a research division using statistical physics to build faithful, ambitious mechanistic interpretability as a foundation for scalable alignment, with theory, tools and validation datasets.
Proposes criteria for attributing beliefs and desires to AI systems via mechanistic interpretability, motivated by safety goals such as detecting deception, and argues these attributions must be jointly constrained.
A satirical essay arguing that the rocket-science analogy for alignment misleads, and that agent foundations and prosaic interpretability research are complementary rather than in conflict.
Proposes a driven-dissipative response-law account of control across biological and LLM systems, predicting LLM output responses to interventions and separating measured opportunity from deployable control.
ARC's Mike's-eye overview of its alignment agenda: mechanistic estimators, Matching Sampling Principle, monitoring training to detect structure, mechanistic anomaly detection, and aligning to user preferences to estimate catastrophe risk.
Proposes Developmental Cognitive Interpretability, a research agenda that models how latent cognitive constructs form during training to predict out-of-distribution and scheming behaviour, with a toy RL maze proof of concept.
Proves that trained fully connected networks' feature-layer parameters match renormalization-group fixed points for exponential-family continuous data, offering an interpretability theory of DNNs.
Proves a mathematical quadrilemma showing that AI explanations cannot simultaneously be complete, faithful, interpretable, and compatible with complex environments and good performance, so AI governance should assume incomplete faithfulness.
Proposes a brute-force, unlimited-compute approach to ELK based on intertheoretic reduction between human and predictor physics models, and analyses problems with it.
Proposes the convergent abstraction hypothesis, arguing learned abstractions converge under shared pressures but may be fragile, and examines what this implies for whether AI moral abstractions can safely carry alignment.
Argues that an LLM's inability to introspect on its internal components does not show it is uncooperative or misaligned, and that this complicates inferring alignment from such observations.
Proposes a category-theoretic framework for mechanistic interpretability, scoring explanations by faithfulness and complexity and deriving a parsimony criterion for compressed, human-aligned explanations.
Argues that convergence hypotheses (Platonic Representation, Universality, Natural Abstraction) do not establish a shared objective reality, so their alignment implications are overstated; the convergence reflects human-scale data and observer-relative choices.
Proposes a research agenda for scale-aware interpretability, using renormalisation from statistical physics to build interpretability tools with formal robustness and faithfulness guarantees for AI safety.
Proposes a theory linking mechanistic phenomena such as induction heads, function vectors and the Hydra effect to hierarchical latent structure in data, validated in toy and synthetic settings and on language models.
John Wentworth's 2025 update to his AI alignment plan describes research on natural abstractions, covering territory-first and mind-first prongs, and the next mathematical footholds needed for interpretability.
Proposes a modular induction approach to learning interpretable world models so goals can be specified within them, working through Game of Life compression and fuzzy-category attempts that fail.
Argues that studying LLM personas is a tractable way to steer out-of-distribution generalization and prevent malicious misalignment, reviewing emergent misalignment, auditing games, and steering examples.
Proposes a research methodology for AI safety called live theory, arguing that substrate-dependent mech-interp and substrate-independent agent-foundations approaches both fail to generalise, and that AI-assisted post-formal theory-prompts can scale substrate-sensitive insights.
Develops a unified theory of sparse dictionary learning methods used in mechanistic interpretability, explaining feature absorption and dead neurons, and proposes feature anchoring to improve feature recovery.
Argues that obfuscation attacks evading latent-space monitors and self-repair in ablation experiments show multiple realizability, which limits causal-mechanistic interpretability as a tool for AI safety, and introduces substrate-flexible risk.
A conceptual proposal for provably robust ML safety systems, using identified latent variables for nuclear-exchange detection; it argues standard verification and supervised training fall short.
Overview of a sequence arguing mechanistic interpretability assumes stable substrates that advanced AI could disrupt, outlining deceptiveness threat models and a framework for managing substrate-sensitive capabilities.
A dialogue arguing that designing inherently interpretable architectures by guessing their structure will fail without discovering the real world's natural ontology through substantial evidence.
ARC proposes the matching sampling principle, a formal goal of mechanistically estimating neural network outputs at least as well as random sampling, motivated by catastrophe detection and adversarial training, and reports partial progress on random and trained MLPs.
Proposes an identifiable temporal causal representation learning framework extending SAEs to discover time-delayed and instantaneous concept relations in LLM activations, with theoretical guarantees, validated on synthetic data.
Proposes five axes for when mechanistic interpretability findings generalize across LLMs, and tests them by tracking 1-back attention heads across Pythia models' pretraining seeds and sizes.
Proposes a research agenda for synthesizing a safe, symbolic, interpretable world-model via compression and natural abstractions, aiming to enable mechanistic interpretability and alignment without directly building agents.
Argues that goal-directedness of agents cannot be measured objectively, critiquing both behavioral and mechanistic approaches to attributing goals and proposing emergent multi-agent modeling directions.
Proves that unrestricted causal abstraction maps can link any neural network to any algorithm, so causal abstraction needs assumptions about information encoding to be meaningful for mechanistic interpretability.
Introduces a sequence arguing Singular Learning Theory's loss-landscape geometry could provide a principled framework for interpreting models and controlling alignment through training data.
An essay arguing that imperfect interpretability leaves unexplained noise that optimisation can exploit, so training against interpretability tools pushes danger into what remains unseen.
A conceptual essay using a Zermelo game-tree analogy to argue that memorising model weights is not interpretation, and that simplicity bias may make interpretability useful.
Speculative conceptual post arguing that RL-trained chain-of-thought LLMs act as agents over a simulator core, and that training-process and layer-wise analysis can reveal this.
Argues that mechanistic interpretability seeks causal explanations embedded in neural networks, defines the field by four properties, and states an explanatory optimism principle it needs to succeed.
A former ARC researcher critically examines obstacles to ARC's alignment agenda, focusing on heuristic explanations of neural networks, empirical regularities, and allocating explanation-finding capacity.