AI Explained
Plain explanations of trending AI concepts, with live visualizations.
Parallel Agent Teams — Many small teams vs one big team — What does it mean?
Ten teams of ten agents beat 100 solo agents and one team of 100: teams raise the mean, and many samples keep the spread.
FOSCAR speeds up 2-bit KV decode — Affine folding of KV dequantization — What does it mean?
Low-bit KV caches saved memory, not time; FOSCAR applies each group's scale and offset once, not per value, so 2-bit decode gets fast.
AMD PrimusClaw — Lease-claimed durable agent runs — What does it mean?
PrimusClaw stores each agent run as a database row under a 45 s lease, so another worker can resume it when the first one dies.
The Gating Mechanism of KV Cache Eviction — The constant-gate no-op — What does it mean?
Softmax cancels a gate that is equal for every key, so a learned KV eviction gate trained on LM loss learns nothing.
TriAttention paper — Sparse mixed-precision attention tiles — What does it mean?
One calibrated policy per head decides which attention tiles to skip, run in 4-bit, or run in 8-bit.
Non-prefix KV reuse study — Targeted KV repair — What does it mean?
Where to spend a small KV-repair budget when stitching cached chunks: on the chunk that holds the answer.
EnvOpt paper — Optimizing the environment, not the harness — What does it mean?
Reorganizing the documents and tools an agent works in cut Haiku 4.5 cost 72.7% on AutomationBench, with no model or harness change.
MCP authorization rewrite — JWT audience, issuer and scope checks — What does it mean?
An MCP server must check that a signed token was issued for it, by one issuer, with the needed scope.
MCP authorization rewrite — Client ID Metadata Documents — What does it mean?
MCP makes a client's own URL its OAuth client ID and deprecates Dynamic Client Registration.
DeepSearch QRAE — Attention-entropy evidence scoring — What does it mean?
Score retrieved chunks with paired-prefill attention; keep the relevant ones, cache the uncertain.
FlashOffloading offloads all KV cache — rolling prefill buffers plus a hot decode cache — What does it mean?
FlashOffloading keeps the master KV cache in host DRAM and overlaps transfers with compute, so each GPU serves more long requests.
SpecCTMC — Exact speculative sampling for masked diffusion LMs — What does it mean?
SpecCTMC drafts several diffusion-LM unmasking steps, verifies them in one batch, and corrects the first miss, staying exact.