1,651 concepts
A search architecture where many explicit reasoning states expand at once, merge when they coincide, and get pruned by a learned judge, instead of one linear transcript.
How well a codebase's checks, architecture, and context are set up for a coding agent to work well in it, distinct from developer experience.
A file loaded automatically into a coding agent's context at session start, giving it the project's standing build, test, and convention brief.
A working pattern where a person starts a coding agent's session and leaves it to run unattended, deferring all human review to the end.
A software factory whose output no human ever reads, with automated checks and automated review as the only gates on what ships.
A system of work where triggers, not a person, start coding-agent sessions, freeing human attention for the decisions that still need it.
An attention variant that builds a token's value vector from a stored per-token memory lookup instead of a learned projection over the current context.
A pretrained language model whose reasoning comes from iterating a shared block of layers in latent space during pretraining, rather than from chain-of-thought text.
A dual-encoder model that scores agent actions by contrastive similarity to a state, instead of generating a judgment in text, for much faster inference.
A mask-learning method that finds nested subsets of a language model's internal components responsible for a behavior, ranking them by attribution score.
A modular, contrastive framework for recursively evolving an AI agent's harness (its scaffolding of tools and control logic) rather than its model weights.
A domain-agnostic extension of JEPA that factors predicted latent states into separate subspaces to model vision, biology, robotics, weather, and molecular systems with one shared predictive principle.
A method that estimates whether a language model's answer is correct by recalling and reflecting on its own bank of graded past episodes, rather than only the current inference trace.
A reinforcement learning framework that trains hybrid reasoning models to allocate more chain of thought computation to hard problems and less to easy ones.
An interactive scientific-research agent that improves through recursive-in-recursive self-improvement, an inner loop that evolves its harness and an outer loop that trains the model under the improved harness
A hybrid attention architecture that pairs local softmax attention with a linear-memory branch to cut video diffusion inference time
Dumping a long, unread AI-generated reply on a colleague where a short human answer would do, offloading the verification work onto them.
A self-evolving ontology layer that data agents query at runtime through an MCP server, refined through small typed edits that are kept only when they measurably improve task accuracy.
A set of four automatically discovered agent-harness optimizations, covering action execution, context compaction, observation handling, and delegated reading, that cut coding-agent token traffic by roughly 45% at comparable task performance.
A training regime where a model generates its own unbounded, self-scaling stream of tasks instead of relying on a fixed human-curated dataset.
A technique for controlling a language model's behavior at inference time by adding a vector to its internal activations, rather than changing its weights or its prompt.
A linear direction in the activation space of large language models, identified across 25 open-weight models, that represents self-directed harm distinctly from fear or general negative emotion and predicts instrumental pain-avoidance behavior.
A memory buffer that stores previously computed key and value vectors from a transformer's attention layers so they can be reused when generating each new token instead of recomputed.
A method for direct semantic communication between large language models that projects and fuses one model's KV-cache into another's, letting models share internal representations without generating intermediate text.
A reinforcement-learning based training method, introduced by TypeSafe AI, that optimizes a model's predicted probabilities to match real-world outcome frequencies rather than optimizing for human preference.
A class of AI models built to return fast, typed, calibrated decisions for software to consume directly, rather than generating natural-language text for people to read.
A hybrid language-model architecture that keeps standard autoregressive weights for quality but adds a lightweight diffusion adapter that drafts multiple tokens in parallel, verified losslessly against the autoregressive distribution.
A generative-modeling framework built on a time-independent ("autonomous") flow whose associated one-step map is the unique solution to a simple conservation equation, learned directly from data without a diffusion teacher.
A teacher-free reinforcement-learning-style method that replaces a distillation teacher's per-token supervision with an entropy-adaptive negative advantage applied to low-probability tokens.
A mid-training objective that distills from a teacher model only on tokens where the teacher is confident, falling back to standard next-token prediction elsewhere.
An intermediate training stage between large-scale pretraining and post-training, where a model continues next-token-style learning on a smaller, curated, or annealed data mixture.
A resource-allocation algorithm for hyperparameter optimization that repeatedly evaluates a pool of configurations, discards the worse half, and reallocates the freed budget to the survivors.
An evaluation setting where a language model must retain hundreds of previously learned query-answer facts after many further sequential fine-tuning updates, without task labels at inference.
A framework that recursively improves an AI agent's exploration policy, not its model weights, by replaying past discovery runs as offline simulators.
Using AI, especially LLM-based agents, to automate parts of scientific research: hypothesis generation, experiment design, analysis, and increasingly whole discovery loops.
Generative models trained so any point on a diffusion trajectory maps directly to its endpoint in one step, collapsing multi-step denoising into single-step generation.
A simulation-free way to train continuous normalizing flows by regressing a network onto the vector field of a fixed noise-to-data path, unifying and speeding up diffusion-style training.
A reformulation of transformer attention using kernel feature maps so cost scales linearly with sequence length instead of quadratically, equivalent to a recurrent update.
A way of encoding token position in a transformer by rotating query and key vectors, so relative position falls directly out of the attention dot product.
A diffusion model that denoises latent image patches with a transformer instead of a convolutional U-Net, letting generation quality scale predictably with compute.
Implanting electrodes into an insect during its pupal stage so its own tissue grows around and through the hardware as it matures, creating a stable neuromuscular interface.
A system that reads neural activity and translates it into commands for external devices, or sends signals back into the nervous system, bypassing the body's normal pathways.
A complete map of the synaptic wiring in a nervous system, from whole-brain regional connectivity down to synapse-level detail.
Scanning a brain's structure and running a computational model of it on different hardware, aiming to reproduce the original brain's function.
A transformer that predicts a 3D representation of an object from one or a few images in a single feed-forward pass, replacing per-scene optimization with a pretrained model.
