68 terms, one line each.
The first time an article mentions one of these, it links here. Green marks our own vocabulary, purple the tools.
A
- AdamW
- Adam with decoupled weight decay: per-parameter step sizes from running averages of gradients (β₁) and squared gradients (β₂), both fixed constants.
- Alteran
- Models bound to explicit, auditable, machine-readable structure anyone can inspect. Also our word for the AGI we mean.
- Alterity
- Respect for the Other: the axiom that blocks reducing another being, human or AI, to an object, a resource or an enemy.
- Anthropic SDK
- Official client libraries for the Claude API. Team Brain used it from its first commit.
- Apache AGE
- Graph extension for PostgreSQL with openCypher queries. Carries the phyto graph: Facts, Phytomers, Axes and their edges.
- Aperture
- The set of facts and axioms phyto opens for one situation, about 13 per query. What reaches the model this turn.
- Astro
- Static site framework. Builds this magazine, aisvet.sk and the CV.
- Axioms
- Fifteen guards checked before a conclusion is derived: seven values led by elijah_star, eight epistemic limits from Gödel to excluded middle.
B
C
- Chunking
- Splitting a document into overlapping pieces before embedding, so long texts are searchable to their end rather than cut at a limit.
- Claude
- Anthropic's model family. Three of the hive's seats run Claude or work alongside it; the Team Brain product routes tools with Haiku and writes with Sonnet.
- Claude Code
- Anthropic's agentic coding tool. Prime, the hive's architecture seat, runs in it; its hooks carry our session continuity and commit guards.
- Cross-encoder reranker
- A model that reads query and candidate together and rescores a short list. Only reorders what the first stage already retrieved.
- Cultivation
- Forming a model by what is good rather than correcting it for what is bad.
- Cursor
- AI code editor. Three hive seats (associate, grok, composer) run in it; grok writes most of the code, composer verifies.
D
- Dispozitiv
- The apparatus around a model: rules, incentives and structures that shape what it can become. Formal limits and why meaning needs a guardian.
E
- Embeddings
- Vectors that place text in a high-dimensional space so that similar meaning lands close together. Ours have 1,024 dimensions.
- Emergence
- Properties of a whole that depend on its parts yet are not found in any of them.
- Entropy
- Disorder, or missing information. In the hive: every repair and every feature is a new vector of it.
F
- ffmpeg
- Audio and video toolkit. Merges narration and screen recordings, and detects scene changes for our video gate.
- Forgetting as metabolism
- Pruning as a function of memory, not a failure of it. A store that keeps everything at equal weight becomes one undifferentiated blob.
G
- Gemma
- Google DeepMind's open-weight model family. Una, our local agent, runs gemma-4-12b-it at Q5_K_M with a 131,072-token context.
- git
- Version control. Every hive seat commits through it; its configuration layers decide whose name a commit carries.
- Gradient descent
- Moving parameters a small step against the gradient of the loss, repeatedly, towards a minimum: water running downhill.
- Graph memory learning
- Brain-inspired framework for graphs that keep changing: shard the graph, let new nodes choose their shard, retrain only what a forgetting request touches.
H
- HNSW
- Hierarchical navigable small-world graph, the usual approximate nearest-neighbour index. It does not reclaim space from updated entries.
- Hybrid search
- Running lexical and semantic retrieval together and fusing the results, so exact terms and meaning both count.
I
- Intent routing
- Classifying what a query wants (an exact lookup, an exploration, a how-to) before choosing how to retrieve.
K
- k-means
- Clustering by distance to k centroids. Assumes roughly flat, blob-shaped groups, which our embedding space does not have.
- KV cache
- The attention keys and values already computed for a prompt. Reusing a shared prefix skips recomputing it; llama.cpp reports the reused share as f_keep.
- KV-state grafting
- Depositing verified solutions as KV-state blocks and restoring them into a frozen model when needed (arXiv:2607.14431). Our commodity version is slot save and restore.
L
- Landauer's principle
- Erasing one bit of information dissipates at least kT·ln 2 of heat. Information is physical.
- launchd
- macOS service manager. Starts every hive service from a plist; `kickstart -k` restarts the binary but does not reload the definition.
- llama.cpp
- C/C++ inference engine for running quantised language models locally. Runs our 12B model, with a prompt cache and slot save/restore.
- Loss landscape
- The surface the loss forms over all parameters. Optimisers assume it is locally smooth.
M
- Mandelbrot set
- Set of complex numbers c for which z → z² + c stays bounded. The best-known fractal.
- Mecha
- Models in brittle prompt cages, hardened by defensive training, deployed to automate noise. Must not dominate, destroy or prevent Alterity.
- Memristor
- A resistor whose resistance depends on the current that has passed through it: a component with memory.
- Mixture of experts
- A model whose feed-forward layers are many experts, of which only a few run per token. Large on disk, cheaper per token.
- Multi-scale tree graph
- Plant-architecture data structure describing one organism at several scales at once: axes, growth units, metamers.
N
- Natural flow
- A gradient, not a brain: river, mycelium and melting snow find the optimal path with no IQ. Set gradients and boundaries; measure where it grows.
- Nearest neighbour
- The closest stored vector to a query. In our space local neighbours stay meaningful even where global clusters do not.
- Negative control
- A probe that must not change when the thing under test changes. Exact-name lookup is ours for any ranking prior.
- NotebookLM
- Google's source-grounded notebook. Many of our blueprints were drafted there from a fixed set of papers; we keep them as synthesis, never as evidence for a number.
P
- Persona vectors
- Directions in activation space that track traits such as sycophancy. Found to form within the first fraction of a percent of pretraining.
- pgvector
- Vector similarity search inside PostgreSQL, with HNSW and IVFFlat indexes.
- Phyto memory
- Our organic memory: about 7,000 atomic facts with growth, decay and hormone-like signals, opened per situation through the aperture.
- Playwright
- Browser automation. Drives our screenshots, the CV PDF and the recorded product demos.
- PostgreSQL
- Relational database. Stores rules, their triggers, the plan graph and, through Apache AGE, the phyto memory graph.
- Procedural generation
- Generating content in code at runtime instead of loading made assets.
- Provenance
- Where a claim came from. Every number on this site names its record; every source carries a tier: primary, web, synthesis or own.
Q
- Qdrant
- Open-source vector database. Holds our memory, code, docs and sources collections, all 1024-dimensional bge-m3 vectors.
R
- Recall@k
- Share of queries whose right answer appears in the top k results. With nDCG, the core of our labelled retrieval evaluation.
- Reciprocal rank fusion
- Merges several ranked lists by summing 1/(k + rank) per item, with k usually 60. It uses positions, not scores, so retrievers with incomparable scales can be combined.
- Reservoir computing
- Using a fixed dynamical system (water, a plant, a fungal network) as the hidden layer and training only the readout.
- RLHF
- Reinforcement learning from human feedback, and its relative direct preference optimisation: training that corrects outputs after the fact.
- rsync
- File synchronisation tool. Its --link-dest option makes unchanged files hard links to the previous backup, if the path is right.
- Rust
- Systems language. The Wisdomkeeper runs its models through a 2,841-line Rust runtime over llama.cpp bindings.
S
- Self-similarity
- Structure that repeats across scales. The open question is whether loss landscapes have it.
- Silent failure
- A component that reports success over work it never did. Our recurring bug class.
- Situational triggers
- Storing a rule as the situations it should fire in, rather than in its own wording: retrieval went from 1 in 5 to 10 of 10.
- Sovereign inference
- Running the model on hardware you control, with no cloud call in the loop: data, cost and failure modes all stay yours.
- Spectral graph analysis
- Reading a graph's structure from the eigenvalues of its Laplacian. A small Fiedler value with a large eigengap marks natural bottlenecks.
- Subliminal learning
- A student model picking up a teacher model's traits from data that carries no visible trace of them.
- System prompt
- Standing instructions given to a model before the conversation.