SYSTEM STATUS / CLEANUP COMPLETE

AI JANITOR

ai janitor is here to clean up your slop.

What you can learn here

How a model computes, from one sum to whole architectures. Pick a topic; every page brings you back here.

Start here

Explainers

Matrix unit vs dot-product instruction

Architectures

Chatterbox TTS (T3 Llama token LM + S3Gen flow-matching mel decoder + HiFT vocoder)

DeltaNet layer (linear attention with the delta rule, no decay gate)

ECAPA-TDNN language ID (VoxLingua107: SE-Res2Net TDNN + attentive statistics pooling + classifier)

FastConformer encoder + Token-and-Duration Transducer decoder (Parakeet TDT)

Gated DeltaNet layer (Qwen3-Next / Qwen3.5 linear attention)

Mamba block (selective state-space model, v1)

Mamba-2 block (state space duality, multi-head SSM)

Mixture-of-Experts FFN (router + top-k SwiGLU experts, Qwen3-MoE style)

Qwen3-ASR (conv-subsampled audio transformer encoder + adapter + Qwen3 decoder)

RWKV-7 "Goose" block (generalized delta rule, time-mix + channel-mix)

SAN-M encoder (memory-equipped self-attention, DFSMN) + CTC head (SenseVoice)

Silero VAD v6 (STFT + 4-conv encoder + LSTM cell + sigmoid head)

Decoder-only transformer (Llama / Qwen3 family: RMSNorm, GQA attention + RoPE, SwiGLU FFN)

Encoder transformer (bidirectional attention, LayerNorm, GELU MLP)

Coming

Terms (glossary) (coming)

Precision (which number format each part uses) (coming)