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.
Embedding models set the quality ceiling for RAG and semantic search, yet many teams choose casually and blame retrieval when results disappoint. This article covers four selection dimensions: language, dimensionality, domain, and cost; how to pick a leading model;
Binding an entire product to one model API is risky: outages, rate limits, policy changes, price hikes, or account suspension can take it offline instantly. Learn how to build multi-model resilience with primary/backup fallback, intelligent routing, fallback chains, trigger rules, unified interfaces, retries, and circuit breakers.
Full-parameter fine-tuning can require dozens of GPUs, while LoRA and QLoRA bring it within reach of a single consumer card. Learn low-rank adaptation, 4-bit quantization, the full workflow from data preparation through deployment, and when fine-tuning is actually the right solution.
Uniform sampling by frame number misreads variable-frame-rate video and misses brief events. This article creates reproducible video inputs with a canonical timeline, layered sampling, scene-change additions, audio-video alignment, and temporal regression gates.
A small image file does not necessarily make vision inference cheap, while resizing everything to a fixed size can erase fine print. This article establishes verifiable resolution budgets using task slices, pixel tiers, targeted region rereads, and actual billing data.
AI apps are data black holes—user input, model calls, training retention all hide landmines. Covers four compliance essentials: user input data, model data handling, fine-tuning data provenance, and cross-border and industry compliance (GDPR/healthcare/finance), plus privacy-first engineering to keep your product off legal red lines.
Vector search excels at semantic similarity but struggles with relationship reasoning such as reporting lines and product dependencies. Knowledge graphs fill that gap. This article explains their advantages over vector databases, how GraphRAG works, when graphs are worthwhile, and pragmatic ways to adopt them.
LLM inference is slow mainly in decode, where tokens are generated one at a time. This article explains KV Cache, which avoids recomputing history, and Speculative Decoding, where a small model drafts tokens for a large model to verify in parallel. Combined with continuous batching and quantization, they can raise throughput by an order of magnitude.
AI does not have to run in the cloud. Privacy, latency, offline access, and cost are making on-device inference relevant again. This article explains the state of on-device AI: models (quantized 1B–7B), frameworks (llama.cpp/MLC/ExecuTorch), environments (mobile, WebGPU browsers, and desktop), suitable use cases, and current limitations.
A pretrained model is an "autocomplete machine"; alignment methods such as RLHF and DPO turn it into a useful assistant. This article explains what alignment does, how RLHF and DPO differ, why alignment improves both safety and capability, and why application developers must understand its limits, including over-refusal and how to handle it.
Tables are the hard part of RAG: chunking breaks structure, vectors lose meaning, and retrieval misses them. This article covers the full table-RAG approach: detection and extraction, whole-table and row-level chunking, structured storage vs. vectors, Text-to-SQL, and hybrid retrieval.
Users' original questions are often poor search queries: they contain unresolved references, are too vague, or bundle multiple questions. Query rewriting is a high-ROI step in RAG and search. This article covers five techniques, when to use them, how to implement them, their costs, and how to evaluate rewriting itself.