AI Explained
Plain explanations of trending AI concepts, with live visualizations.
AI Models Can Provably Hide Arbitrary Capabilities — Encrypted backdoors that defeat capability elicitation — What does it mean?
A secret trigger can lock a whole neural network inside model weights; fine-tuning and trigger search cannot draw it out.
One Word Opens the Gate — Fail-open vs fail-closed guardrail errors — What does it mean?
A guardrail's errors have a direction. Renaming one option or adding harmless log text flips decision models toward allowing prohibited calls.
Authorize agent subgoal mutations against the root task — Goal-graph authorization — What does it mean?
Every subgoal an agent writes must prove it fits the user's approved task; the subgoal's own text grants no authority.
RECAST — Adaptive evidence routing between retrieval and computation — What does it mean?
A small trained router builds evidence by choosing search, SQL or custom code each round: 75.6% vs 59.7% for the best agentic RAG.
Microsoft Agent Lightning 1.0 — Proxy-captured harness rollouts — What does it mean?
Train an agent through the exact harness you ship by recording its model calls at a proxy, then score each run once, not each training sample.
OnTrack interrupts failing agent runs — Streaming trajectory alignment — What does it mean?
A monitor that compares a live agent run with past successful runs, step by step, and flags anomalies that can trigger an early stop.
TestGRAD — Differential test loss for patch selection — What does it mean?
TestGRAD picks among coding agents' patches by writing tests until one splits them, reaching 84.2% on SWE-bench Verified.
FAR.AI probes catch hidden agent sabotage — White-box deception probes — What does it mean?
Probes that read an agent's hidden states catch sabotage and lies a transcript-reading monitor cannot see.
DelegationBench — Matched-pair act-or-ask testing — What does it mean?
Agreement with human labels can hide an agent that ignores what matters; matched pairs test whether its act-or-ask choice reacts.
150 compound-AI incidents mapped to resilience patterns — Circuit breakers, quality gates and isolation — What does it mean?
Compound-AI failures often start between healthy parts. Breakers, quality gates and isolation contain them.
Not all GUI-agent experience belongs in the weights — Component routing: context vs weights — What does it mean?
A GUI agent's experience splits into four kinds; recurring ones belong in the weights, screen-specific ones in the prompt.
PoS (Progression of States) — Belief trapping vs explicit belief state — What does it mean?
PoS gives agents a checked picture of what is true now and catches them when they keep acting without progress.