AI Explained
Plain explanations of trending AI concepts, with live visualizations.
MiniMax M3 ships open-weight 1M context — MiniMax Sparse Attention (MSA) — What does it mean?
MiniMax M3 (open-weight, 1M context) runs on MiniMax Sparse Attention — block-sparse KV gather that cuts per-token compute ~20× at one million tokens.
EmbedFilter — Unembedding matrix as a feature lens — What does it mean?
LLMs make blurry text embeddings because a subspace in their unembedding matrix injects frequent words — EmbedFilter projects it out in one linear transform.
Code2LoRA gives code models per-repo knowledge — Hypernetwork-generated LoRA adapters — What does it mean?
Code2LoRA's hypernetwork stamps out a repo-specific LoRA adapter in one pass — repo knowledge with zero prompt tokens at inference.
Tangram speeds multi-turn serving up to 2.6× — Per-head KV cache budgets — What does it mean?
Tangram sizes each attention head's KV cache to what it actually keeps, not one uniform budget — lifting multi-turn serving up to 2.6×.
Google ships Gemma 4 QAT checkpoints — Quantization-Aware Training — What does it mean?
Gemma 4 ships QAT 4-bit checkpoints: training on the low-bit grid dodges the accuracy cliff that naive post-training rounding hits.
MatMul-only matrix inversion makes quantized Gated DeltaNet 5x faster — Truncated-Neumann triangular inverse — What does it mean?
Gated DeltaNet's chunk solve hides a sequential matrix inverse — a truncated Neumann series turns it into parallel MatMuls, ~5x faster.
Microsoft MAI-Code-1-Flash — Adaptive solution-length control — What does it mean?
MAI-Code-1-Flash scales how many reasoning tokens it spends to each task's difficulty — short chains for easy tasks, long only for hard ones.
KVarN squeezes the KV cache to 2 bits — Hadamard rotation — What does it mean?
KVarN rotates outliers out of the KV cache so 2-bit quantization fits every channel — calibration-free, no error compounding across decode.
WASH attack washes out LLM text watermarks — Watermark removal by model-averaging — What does it mean?
Averaging the output distributions of 3–5 independent LLMs cancels each one's text watermark — detection z-scores fall from 5–300 to below 2.
PEFT scaling paper — Persistent personal adapters at million-scale — What does it mean?
Reframes a LoRA adapter from a cost-cutting trick into persistent per-user state — a million personal adapters served over one frozen base.
LongTraceRL — Rubric reward (entity-level process supervision) — What does it mean?
LongTraceRL scores each reasoning hop, not just the final answer — dense process reward, gated to correct rollouts so it cannot be gamed.
dMoE cuts diffusion-LLM MoE memory ~80% — block-level expert routing — What does it mean?
dMoE pools a diffusion block's per-token expert choices into one block-level decision — ~70→15 experts loaded, ~80% less memory.