Agent benchmarks are saturating. Hold one back.
Agent leaderboards saturate and leak. A private, held-back task set is what actually predicts how your agent behaves in production.
GPAI rules are live. Your part is due 2 August.
The EU AI Act's general-purpose-AI obligations, decoded for teams that build on top of foundation models rather than train them.
Long context still loses the middle of your prompt
Frontier windows hold a million tokens but retrieve the middle unreliably — position, not capacity, decides recall.
Retrieval, not the model, sank the RAG assistant
In production RAG, retrieval — chunking, embeddings, freshness — decides quality; the model is rarely the culprit.
Your RAG Pipeline Is an Exfiltration Channel
Indirect prompt injection hides in retrieved content; containment held, prompt guardrails didn't.
When Extra Inference Compute Earns Its Cost
Test-time compute is Signal on hard, decomposable, checkable problems — and Noise almost everywhere else.
Use retrieval, not a million-token prompt
A cost, freshness, and grounding teardown of RAG versus long-context prompting for knowledge-heavy apps.
NIST AI RMF: a checklist, not a box to tick
How to run NIST's AI Risk Management Framework and its generative profile as real engineering controls instead of self-attested paperwork.
C2PA is the right fix. Coverage is the problem.
C2PA fixes AI-media authenticity at the source, but stripped metadata and opt-in adoption limit it in the wild.
The backdoor your safety training can't scrub
A planted backdoor can persist through the exact safety pipeline meant to remove it. Treat open weights as an untrusted supply chain.
JSON mode ships the envelope, not the letter
Grammar-constrained decoding guarantees parseable JSON — not correct JSON. Ship the syntax, gate the semantics.
Agent stack: reliable tools, thin autonomy
Bounded tool-calling holds; long multi-step chains still drop the thread.
MCP: real leverage, not a finished standard
One shared tool interface turns M×N integrations into M+N — real leverage, but the spec is still moving.
Speculative decoding's latency win has hard limits
Speculative decoding cuts single-stream latency 2-3x with identical outputs, but throughput-bound batches erase the win.
One Suffix, Every Model: The Transfer Problem
Adversarial suffixes tuned on open-weight models transfer to closed ones. Exploitable by design — blocklists won't save you.
Build the eval harness before you ship
The smallest evaluation that catches the regressions a single accuracy score always misses.
The AI-label duty is binding soon, not noise
The EU AI Act's transparency rules make AI media and text carry disclosures — machine-readable for makers, visible for deployers.
Grade Open Weights on Harnesses, Not Vibes
How to judge an open-weight model by its eval harness and model card instead of launch-day hype.
QLoRA fine-tuning: one GPU, real trade-offs
How 4-bit quantized finetuning shrinks the GPU bill for small teams — and where quality actually leaks.
Your Agent Can Exfiltrate. ATLAS Maps How.
Indirect prompt injection to exfiltration is a real, chainable ATLAS path; cut the egress channel, not the prompt.
Token growth eats margins. Three levers hold.
Unbounded context growth is a silent cost bomb; discipline, caching, and routing defuse it.
Sparse compute, dense memory: MoE's real bill
Mixture-of-experts cuts per-token compute but not VRAM — a production serving verdict hung on the Mixtral paper.
Are You in the AI Act's High-Risk Bucket?
A field guide to the two doors into high-risk and the seven duties that follow.
The LLM supply chain your threat model skips
Supply-chain and model-poisoning entries in the OWASP LLM Top 10 are real, exploitable, and chronically under-weighted.
Govern Is the AI RMF Function Teams Skip
Why the cross-cutting Govern function decides whether Map, Measure, and Manage do anything at all.