大模型、多模态、推理模型和 AI 产品趋势
This collection features articles and community discussions in the AI 前沿 space, focusing on real-world engineering scenarios, tool selection, practical paths, and risk boundaries.
The page is continuously updated and serves as a learning roadmap, technology selection reference, and technical review entry point.
A policy engine returning allow doesn't mean downstream executes with the same input and obligations. This article defines a verifiable decision receipt, a one-time acceptance point, and execution confirmation to block stale policy, parameter substitution, and fail-open bypass.
Writing scripts and writing production Python are two different things. This article explains dependency management with uv, poetry, and pip-tools; type annotations and mypy; project structure; pytest; packaging and releases; and the transition from runnable scripts to maintainable engineering projects.
Swapping a model while reusing old thresholds, or staring only at overall accuracy, silently changes both blocks and false kills. This article freezes the classifier, calibrator, per-category thresholds, and action table as independent artifacts, and gates release on per-slice confidence upper bounds.
Permanent allowlists turn temporary business needs into invisible backdoors. This article models policy exceptions as bounded leases, tied to precise scopes, compensating controls, and usage receipts, and makes expiry, revocation, and renewal re-evaluation actually work.
Traditional APM falls short for nondeterministic, multi-step, token-hungry LLM apps. This article covers three layers: traces for each LLM/tool call, automated quality evals, and latency, token, and cost metrics and alerts, then shows how to wire them into the development loop and turn post-launch black boxes into debuggable, improvable AI systems.
When search volume is high and you need full-text search, complex filtering, and aggregations, Elasticsearch is a powerful but heavyweight tool. This article explains its core concepts (indexes, shards, and inverted indexes), advantages over SQL full-text search, Query DSL essentials, Chinese tokenization with IK, and when ES is or is not justified.
"Tuning RAG without an eval set is just guesswork." This guide uses RAGAS to explain four core metrics, building and running an evaluation set, and diagnosing issues so "it feels better" becomes "recall rose from 0.72 to 0.91."
Testing OCR, ASR, or text answers in isolation misses cross-modal mismatches. This article uses asset alignment, missing/conflict slices, metamorphic relations, layered metrics, and shadow traffic to block inter-modal quality regressions.
Two annotators assigning the same label does not necessarily mean the standard is clear, and disagreements cannot be resolved by letting the latest label overwrite earlier ones. This article combines item snapshots, blinded assignment, evidence spans, dispute adjudication, and label generations into a traceable gold-label contract.
Dashboard screenshots and post-incident reruns cannot prove what happened at the time of an incident. This article freezes model, prompt, retrieval, and tool evidence; defines append-only sealing and transfer chains; and enables review while protecting sensitive source content.
Clicking “Modify” does not automatically make an AI decision safe. This article binds the original machine recommendation, operator qualifications, override rationale, actual effect, and subsequent outcome into an immutable record while enforcing baselines that humans cannot bypass.
Fixed seeds can still produce different outputs because of GPUs, kernels, batching, and floating-point reduction order. This article defines tiered reproducibility targets, complete environment snapshots, token-level replay, and cross-platform equivalence thresholds.