AI Explained
Plain explanations of trending AI concepts, with live visualizations.
ARE-bench — In-progress workspace replay — What does it mean?
Grade agents inside the half-finished project a request came from: models at ~69% on the deliverable alone fall to 28-40%.
NVLinkOneSided speeds up MoE all-to-all — One-sided writes vs count exchange — What does it mean?
MoE GPUs push tokens straight into their own slots on peer GPUs, cutting the fixed costs that dominate small-batch decode.
MCP servers v2.4.0 — Symlink TOCTOU path confinement — What does it mean?
A path check and a read by name are two moments; a symlink swapped between them escapes the sandbox. Anchor the read instead.
Bursts That Outlast Themselves — Prefix-Cache Metastability — What does it mean?
A short burst can trap a prefix-cached agent server at a 10% hit ratio; parking admitted sessions by KV demand gets it out.
Fisher quantization screen paper — Predicted KL vs quantize-and-measure — What does it mean?
One fp32 gradient pass predicts each weight format's damage to a model's outputs, so layer formats are chosen without testing every plan.
AES paper — Entropy-sparing vs uniform length penalty — What does it mean?
AES penalizes only late high-entropy tokens in correct answers, not every token, cutting reasoning about half without losing accuracy.
Lost in Rounding — Stochastic rounding for BF16 recurrent state — What does it mean?
BF16 round-to-nearest freezes small updates to a hybrid model's state and lengthens reasoning; stochastic rounding fixes it at 16 bits.
FP4 Pretraining Study — Rare-fact acquisition vs validation loss — What does it mean?
NVFP4 matched BF16 loss within 0.5% but needed 1.32x the exposures to learn rare facts: the loss average hides the long tail.
CARE-PO paper — Task-anchored tool-call budgets — What does it mean?
CARE-PO prices each tool call against a fixed per-task budget, so agents learn to call less without collapsing to zero calls.
StrataFlux — Selective precision streaming — What does it mean?
Keep low-bit weights and KV on the GPU; fetch extra precision bits from RAM or SSD only for the channels and tokens that matter now.
LowRAM KV cache compression — Predicting the value cache from keys — What does it mean?
Inside one attention head, values are mostly predictable from keys, so LowRAM stores a small predictor plus fixes for the worst tokens.
Anthropic on agent memory — Offline dreaming (memory consolidation) — What does it mean?
Dreaming replays transcripts between sessions to fix an agent's notes, build wikis, and keep its startup context small.