AI 常见误区总索引AI Misconceptions: Index

参数越多越聪明：参数规模与训练数据、架构、任务都有关，不能单看数字。

上下文窗口越大越好：窗口有显存与延迟代价，长输入中段利用也不均匀。

温度调高会更聪明：温度只改变采样随机性，不改变模型知道多少。

榜单分数高就是好用：基准有版本、污染与任务错配问题，最终要跑你自己的任务集。

量化会让模型变笨一半 / 显存不够就跑不了：质量取决于档位与校准；显存不够可以 CPU 卸载、多卡切分或选更小规格。

开源一定更省钱 / 预览版不能生产：硬件、电费、运维与时间都要算；预览版看具体产品条款。

接口兼容 = 行为一致 / 换模型只是改名字：协议兼容不保证输出语义，提示词、工具与评测都要重新对齐。

More parameters is smarter: scale interacts with data, architecture and task; a single number is not enough.

Bigger context is better: context has VRAM and latency costs, and middle-of-input use is uneven.

Higher temperature is smarter: temperature changes sampling randomness, not what the model knows.

A high leaderboard score means usable: benchmarks have version, contamination and task-mismatch problems; run your own task set.

Quantisation halves intelligence / not enough VRAM means it cannot run: quality depends on tier and calibration; low VRAM can use CPU offload, multi-GPU splitting or a smaller member.

Open models are always cheaper / preview builds cannot be used in production: hardware, power, operations and time all count; check the specific product terms.

API compatibility means behaviour matches / switching models is just a rename: protocol compatibility is not output semantics; prompts, tools and evals all need realignment.