AI Explained
Plain explanations of trending AI concepts, with live visualizations.
KronLSTM — Kronecker Delta Rule memory updates — What does it mean?
The Kronecker Delta Rule edits one key-value entry of a linear-attention memory, touching fewer stored facts than the delta rule.
LatComm reuses compressed context across LLMs — Encode-once latent packets — What does it mean?
LatComm encodes a context once into a fixed-size packet that another LLM reads per question.
Parallel Language Models paper — Latent auxiliaries for multi-token decoding — What does it mean?
Parallel Language Models decode several coordinated tokens per forward pass by moving randomness into latent auxiliaries.
Prune prefill tokens to cut TTFT up to 11.3× with TokenGate — Token-gated attention — What does it mean?
A learned gate keeps K prompt tokens early in prefill, and gated attention makes that cut trainable.
LangDiT paper — A Transformer built for diffusion language models — What does it mean?
LangDiT is a Transformer redesigned for diffusion language models; its cumulative ablation shows generation perplexity down 29.4%.
NVIDIA Nemotron 3.5 Bolt — Total vs active parameters in a 2-bit MoE — What does it mean?
Why a 74B-A9B mixture-of-experts model fits 24GB at 2 bits.
N-Gram Memories at Two Bits — Row-addressable lookup-table quantization — What does it mean?
Conditional Program stores lookup rows as 2-bit codes plus a scale, decoded by one shared trellis, so a lookup touches only requested rows.
SparPR recovers KV caches on a consumer GPU — copy versus recompute — What does it mean?
SparPR keeps query-relevant KV blocks and rebuilds missing ones by recompute when that costs less than copying.
MirrorMoE — Shared experts vs fine-grained expert segmentation — What does it mean?
MirrorMoE gives low-probability experts near-zero weight so shared-expert MoEs explore routing better and reach lower perplexity.
DropKV paper — Residual-output perturbation scoring for KV-cache eviction — What does it mean?
DropKV evicts the KV tokens whose removal shifts attention output least, scored one token at a time by a fused Triton kernel.
RoleQ converts LLMs to 1.59-bit ternary — Role-aligned scales — What does it mean?
RoleQ places ternary-weight scales by residual-stream read/write role, not matrix layout, cutting perplexity by up to half.
Latent Compaction (arXiv 2610.20582) — Training a document-specific KV cache — What does it mean?
Latent Compaction trains a small KV cache per document so a frozen model answers unseen questions at 32× compression.