返回首页

对比阅读

对比阅读:AI Benchmarks Are Losing Trust — It's Time to Rewrite the Evaluation System 与 AI 跑分越来越没人信了:评测体系到了该重写的时候

AEN
LocalLLaMAAI-benchmarksQwen·

AI Benchmarks Are Losing Trust — It's Time to Rewrite the Evaluation System

This week, a highly upvoted discussion surfaced on Reddit's LocalLLaMA developer community (the gathering place for hardcore players running large models locally): a developer who has been running models for 2-3 years called mainstream AI evaluation leaderboards "useless." He argued that the hardware benchmarking approach completely breaks down when applied to AI — a 9B (9 billion parameters) older model and a 27B newer model may score similarly on paper, yet the gap in actual business performance is enormous. When the hardcore testing community collectively doubts benchmark numbers, enterprise readers should pay attention.

What this is

The core of the discussion is the benchmark mechanism itself (using a standardized question set to score models). In traditional hardware, it's trustworthy — the same chip runs the same test suite, and results are stable and consistent with daily experience. AI models behave differently: tweak the temperature parameter (the setting controlling answer randomness) or rephrase a prompt, and scores swing 10-20 points. AI benchmarks don't measure the model — they measure "the model's specific answer at a particular moment."

For model companies, benchmarks are a marketing tool; for investors, benchmarks are a leaderboard; for enterprise IT, benchmarks are supposed to be a procurement basis. When three parties each take what they need, credibility is the casualty.

Industry view

The original poster's stance is blunt: only the most basic retrieval tests like "needle in a haystack" have any meaning; everything else is barely better than nothing. He runs a Qwen (Alibaba Tongyi Qianwen) 35B plus 27B combo as his daily workhorse and has never switched.

Counterarguments exist. Completely rejecting benchmarks would collapse the evaluation system — enterprise procurement can't compare options, and compliance-heavy environments like hospitals and banks need some form of standardized comparison. The real question isn't whether benchmarks should exist, but who designs the question sets, how to avoid data contamination (test questions leaking into training sets), and how to make scores reflect real production workloads. Many question sets that major model companies submit for evaluation have already been reverse-engineered. Without that context, reading leaderboards will lead you to overestimate certain vendors.

A more subtle risk: once benchmarks devolve into marketing props, enterprises must build their own metrics for real productivity — preparing business samples, running them themselves, comparing them themselves. This pushes costs that evaluation agencies should bear onto procurement teams.

Impact on regular people

For enterprise IT: In the past, you'd glance at a few public leaderboards and place orders. We now recommend preparing at least 20-50 real business questions for blind testing (comparative testing where testers don't know which model they're evaluating).

For individual professionals: When using AI tools daily for emails and reports, don't fully buy the "latest and greatest" rhetoric — lightweight models with reasonable pricing and stable speed are often more practical.

For consumer markets: "AI phone benchmark scores" and "AI PC IQ scores" will proliferate in the short term. Remember this post's warning: good-looking numbers don't equal good experience.

BZH
LocalLLaMAAI-benchmarksQwen·

AI 跑分越来越没人信了:评测体系到了该重写的时候

Reddit 上 LocalLLaMA 开发者社区(本地跑大模型的硬核玩家聚集地)这周出现一篇高赞讨论:一位跑了 2-3 年模型的开发者把主流 AI 评测榜称为「无用之物」。他说硬件评测思路放到 AI 上完全失效,9B(90 亿参数)的旧模型和 27B 的新模型分数排名可能差不多,但实际跑业务差别巨大。硬核测试社群集体怀疑跑分这件事,值得企业读者关心。

这是什么

讨论的核心是 benchmark(用同一套标准化题库给模型打分)这个机制。传统硬件里它是可信的 — 同一颗芯片跑同一组测试,结果稳定、与日常体验吻合。AI 模型不一样:温度参数(控制回答随机性的设置)调一下、提示词换种说法,分数就差 10-20 个点。AI benchmark 测的不是模型本身,而是「模型在某个特定瞬间的具体回答」。

对模型公司,跑分是营销工具;对投资人,跑分是排名表;对企业 IT,跑分本应是采购依据。三方各取所需时,可信度被牺牲掉了。

行业怎么看

原帖作者的态度是:只有「针在草堆里」这种最基础的检索测试还有点意义,其余榜单聊胜于无;他自己用 Qwen(阿里通义千问)35B 加 27B 组合做日常主力,没换过。

反对意见也有:彻底否定 benchmark,评测体系会垮 — 企业招标没法比较、医院和银行这种合规场景需要某种对标工具。问题不是 benchmark 该不该存在,而是谁来设计题库、怎么避免数据污染(测试题被泄漏进了训练集)、怎么让分数反映真实生产负载。眼下大模型公司送测的题集很多已被反向工程,没这个背景看榜单,会高估某些供应商。

更隐蔽的风险是:跑分沦为营销道具后,反映真实生产力的指标得企业自建 — 准备业务样本、自己跑、自己比。这把本该由评测机构承担的成本推给了采购方。

对普通人的影响

对企业 IT:以前看几个公开榜就下单,现在建议至少准备 20-50 道真实业务题做盲测(不告诉测试者模型名字的对比测试)。

对个人职场:日常用 AI 工具写邮件、做汇报,别全信「最新最强」的话术 — 价格合适、速度稳定的轻量模型往往更实用。

对消费市场:AI 手机「跑分」、AI 电脑「智商分」短期会大量出现,记住这个帖子的提醒 — 数字好看不等于体验好。