Leaderboards go quiet as they saturate and leak into training data.

Frontier models will swallow your whole knowledge base.

A post-mortem of the RAG assistant that slipped: the generator was fine, the index was rotten.

Planting instructions in a retrieved document turns your RAG pipeline into an exfiltration channel.

Chain-of-thought was the first proof that spending more compute at inference buys reasoning — but only on the right problems, and only at scale.

For knowledge-heavy apps, retrieval still beats stuffing everything into a million-token prompt — on cost, freshness, and grounding.

Voluntary on paper, de-facto expected in procurement.

Content provenance signs media at the source instead of guessing after the fact.

Anthropic's Sleeper Agents shows a planted backdoor can survive SFT, RL, and adversarial training — breaking the assumption that a safety fine-tune…

Constrained decoding makes malformed JSON impossible. It does nothing to make the JSON true — and that gap is where production quietly breaks.

A hands-on teardown finds the open-weight agent stack ships dependable tool-calling for bounded tasks but recovers poorly on long chains; the verdict is use it for scoped workflows, wait on full autonomy.

A shared tool interface is genuine interop leverage — but the spec is young, and auth and security are the soft edges you still own.

A small draft model plus one parallel verification pass can cut latency 2-3x with identical outputs — but only when decoding is memory-bandwidth-bound…

A gradient-tuned nonsense string trained on open models breaks closed ones too. Here's the mechanism — and what actually blunts it.

Accuracy alone passes the build and ships the regression.

The EU AI Act's Article 50 transparency duties are weeks from applying.

Leaderboard screenshots and launch-day threads are noise. The signal lives in standardized eval harnesses and honest model cards.

Quantized LoRA drops a 65B-model finetune from over 780GB to under 48GB of GPU memory.

Indirect prompt injection to data exfiltration is a real, chainable path in MITRE ATLAS.

Agentic workloads pile up context turn over turn.

MoE routing sells a compute discount your VRAM bill never sees.

The Act is in force, but the high-risk obligations phase in — here's how to tell whether your feature is in the bucket, and exactly what you owe if it is.

OWASP put supply chain and model poisoning in its LLM Top 10 for a reason. Most threat models still treat them as a dependency-scan checkbox.

Teams sprint to Map, Measure, and Manage. The one function that makes those stick is the one they never staff — and NIST says why.