AI knowledge hive

Engineer-grade AI concepts framed for Spark / SparkLang — transformers, training, inference, multimodal, agents, evaluation, and decompile limits. Original diagrams (not scraped paper figures). Citations point at primary papers and surveys.

Never publishes competitive AI win claims. Spark factory train runs on CPU (default) or a consumer GPU. Not an OpenBin clone.

Topics below follow the same category → subcategory map as the Hive nav: Foundations · Systems · Safety / Eval · RE.

Foundations

Topic	Page	What you get
LLMs & transformers	knowledge/LLM_TRANSFORMERS.md	Tokens → embed → attention → MLP; embeddings
Training stack	knowledge/TRAINING.md	Pretrain, SFT, RLHF/RLAIF, LoRA/QLoRA, SGD/Adam
Inference	knowledge/INFERENCE.md	Sampling, KV cache, quantization

Systems

Topic	Page	What you get
Multimodal	knowledge/MULTIMODAL.md	STT / TTS / vision
Agents & tools	knowledge/AGENTS_TOOLS.md	Tool loops, ReAct-style patterns

Safety / Eval

Topic	Page	What you get
Evaluation	knowledge/EVAL_STATUS.md	Benchmarks, baselines, frozen probes
Safety & limits	knowledge/SAFETY_LIMITS.md	Model limits → Guardrails → Grounding (`./spark-ground`)

RE

Topic	Page	What you get
Decompile + LLM RE	knowledge/DECOMPILE_RE.md	Link to existing research; recompile ≠ semantics
LLM decompile research	research/LLM_DECOMPILE.md	Factory research page

Spark cross-links (factory SoT)

Surface	Doc
Factory hub	Factory hub
Workflow loop	/workflow.html
Voice / STT / TTS	VOICE.md
Model aspects	Model aspects
Spark coder	Spark coder
Train loop	Train loop
Tokenizer / BPE	TOKENIZER.md
Attention / forward	Attention / forward
Eval harness	Eval
LLM decompile research	research/LLM_DECOMPILE.md
Tutorials	/learn/

How to read this hive

Start with Foundations → LLMs & transformers for the math sketch.
Jump to Training / Inference for practice vocabulary.
Use RE before trusting pretty recovered C.
Prefer Spark factory docs for what ships; this hive explains the field.

flowchart LR
         H[Knowledge hive] --> F[Foundations]
         H --> S[Systems]
         H --> SE[Safety / Eval]
         H --> R[RE]
         F --> T[Transformers]
         F --> TR[Training]
         F --> I[Inference]
         S --> M[Multimodal]
         S --> A[Agents]
         SE --> E[Evaluation]
         SE --> Saf[Safety]
         R --> D[Decompile RE]
         D --> L[llm-decompile.html]
         TR --> Fac[factory / spark-coder]