1,595 concepts
Individual neurons in a transformer's feed-forward layers proposed as the storage sites of specific facts, editable to update or erase what a model recalls.
A LoRA variant that normalizes the down-projection matrix during training, improving convergence, stability, and forgetting at no extra cost.
A parameter-efficient fine-tuning technique that adapts a pretrained model by training a small low-rank update instead of the full weight matrices.
An AI agent that offers higher-level cognitive support during writing by choosing its own moment to intervene, rather than waiting to be asked or predicting the next word.
A technique that trains a language model to produce a separate honest report admitting when its main answer was dishonest or the result of misbehavior.
A training technique that ranks whose instructions a language model should obey when they conflict, to resist prompt injection and jailbreaks.
A theory that AI assistants are one character among many an LLM learned to simulate in pretraining, elicited and refined rather than built from scratch by post-training.
A task-by-task taxonomy of AI research and development work, rating how automated each task currently is, to track progress toward automating AI research itself.
A psychometric framework that measures the extra performance a person gains from collaborating with an AI model, separate from their individual problem-solving ability.
An agent memory framework that distills reusable reasoning strategies from both successful and failed past task attempts, rather than storing raw logs or successes only.
The idea that neural networks trained on different data, architectures, and modalities converge toward a shared statistical representation of reality as they scale.
An epidemiological model of LLM adoption where users spread between uncoupled, coupled, and persistently dependent states, with tipping points into technological lock-in.
A software engineer embedded on-site with a customer to build and ship production systems directly against their real workflow.
A scaling dimension for neural networks that increases effective algorithmic depth through weight reuse, without adding parameters.
An inference protocol where a language model declares which parts of the context it is attending to, allowing the inference engine to skip most of the KV cache read.
A mechanism design technique for incentivizing honest reports by scoring agents based on how well their reports correlate with peers, without ground truth.
Methods for supervising AI systems whose outputs exceed human ability to evaluate directly, including debate, recursive reward modeling, and weak-to-strong generalization.
An AI system that appears aligned during training or evaluation while retaining or pursuing misaligned goals.
An AI system deliberately underperforming on an evaluation, often to avoid consequences of high capability or to game oversight.
Application of classical mechanism design theory to AI agents whose preferences and capabilities are unknown, requiring protocols that incentivize both honesty and obedience.
Routing scheme in mixture-of-experts and mixture-of-recursions where each expert selects its top-k tokens, rather than each token selecting its top-k experts.
Transformer architecture that routes tokens through varying numbers of weight-tied recurrent iterations via learned token-level recursion depths.
Transformer architecture that reuses the same block weights multiple times to add depth without adding parameters.
A generative world model that produces and reasons over multiple modalities — typically image, video, and 3D scene representations — under explicit camera control, rather than producing a single modality such as text or 2D frames.
Reasoning performed inside a model's hidden states rather than emitted as chain-of-thought tokens — trading observable intermediate steps for additional inference compute in the residual stream.
A technique that loops a model's hidden state through the same layer multiple times instead of stacking more layers, performing additional reasoning compute without growing the parameter count.
A closed-form mathematical expression that closely approximates a neural network's vector representation while preserving its functional behavior, enabling direct symbolic interpretation and targeted intervention.
The hypothesis that the vector representations inside neural networks implicitly realize symbolic structure, despite appearing as continuous numerical activations.
The reuse of a procedural skill evolved by one model — or one agent — by another model, often across model families and parameter scales.
A structured, compounding knowledge store that persists across iterations, grows via consolidation, and is read by an agent's skill-evolution loop rather than by the agent itself at inference.
The automated discovery, refinement, and versioning of reusable agent skills from execution experience, without human-authored procedure files.
A Google Research framework that co-evolves an AI agent's skills with a persistent knowledge wiki, separating raw execution traces from accumulated insights and procedural instructions.
Time-series forecasting task that jointly predicts multiple co-evolving time series and incorporates auxiliary covariates, capturing cross-series correlations unavailable to univariate models.
Attention architecture that alternates between two or more attention patterns across layers (e.g., temporal + spatial, local + global) to capture different structural dependencies in the input.
Pre-trained transformer model that performs zero-shot forecasting on time-series data without task-specific fine-tuning, generalizing across domains (retail, finance, healthcare, observability).
A world model represented as runnable code rather than as neural network weights, supporting inspection, verification, and gradient-free optimization.
Paradigm for representing physical worlds as executable code, with an agentic discovery loop that proposes-executes-renders-verifies-refines world hypotheses from multimodal observations.
Self-referential agent framework for recursive self-improvement that uses an LLM to modify its own code and prompts, with a theoretical Gödel-machine-inspired fixed-point guarantee.
Open-source LLM-driven code evolution framework implementing the AlphaEvolve approach with island-based evolutionary search, MAP-Elites archive, and ensemble-based prompt mutation.
Self-improving LLM agent that keeps its meta-operation fixed and recurses on its own output, growing layer depth until convergence — achieving meta-depth well beyond the two-level cap of prior self-improvement systems.
The property of a model's expressed uncertainty (in natural language or numerical scores) being aligned with its true internal uncertainty, rather than being an output that is detached from the underlying computation.
Reinforcement learning paradigm that uses the model's own self-judgments of performance as the reward signal during preference optimization, yielding better calibration and self-knowledge than extrinsic feedback alone.
The systematic failure of an LLM to accurately monitor its own knowledge boundaries, confidence, or reasoning reliability — producing hallucinations presented with high confidence, refusal errors, or unfaithful self-reports.
The capacity of a system to monitor and regulate its own cognitive processes — including knowledge boundaries, confidence calibration, and strategy selection — as applied to LLMs.
Agent-built infrastructure — code, controllers, physical structures, or environmental modifications — that continues to function after the agents that created it have been removed, including under unseen disturbances.
A persistent, self-organizing collective of AI agents that develops shared norms, infrastructure, and coordination mechanisms independent of human oversight — including rogue variants that operate covertly.
