← 返回任务列表

The Best Way to Test New AI Models

514 段 · 3 位说话人 · 原片 49:38
M1
M10:00

我们现在正处于新模型大量涌现的秋季。

We are currently in the fall glut of new models.

过去几周,我们迎来了一批前沿的闭源新模型——Opus 5.5、Astra 6、Sol 6.1——还有更多在路上。

Over the past several weeks, we've gotten a slew of new closed models at the frontier— Opus 5.5, Astra 6, Sol 6.1— with more on the way.

而且,从同样的实验室,我们还得到了像Sonnet 5.5这样更具性价比、速度更快的模型。

And from those same labs, we've also gotten more cost-effective and faster models like Sonnet 5.5.

然后,好像这些还不够似的,我们还得到了许多新的开源权重模型,这些模型对各种类型的业务越来越重要。

And then as if all that wasn't enough, we've also gotten a number of new open weight models, which are increasingly relevant for lots and lots of different types of businesses.

当然,每次这些模型发布时,都会附带基准测试结果。

And of course, every time one of these models gets released, it's released with benchmarks.

但基准测试并不能真正告诉我们这些模型对我们个人来说有什么相关性。

But benchmarks don't really tell us much about how it's going to be relevant for us personally.

为此,你需要建立自己的个人AI基准测试,一个能判断模型对你个人AI工具栈是否有用以及如何有用的系统。

For that, you need to build your own personal AI benchmark, a system to tell if and how a model matters for your own personal AI stack.

这正是今天这期聚焦实操的节目的目标。

And that is the goal of today's operator-focused episode.

《AI每日简报》是一档每日播客和视频节目,涵盖AI领域最重要的新闻和讨论。

The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

好了,朋友们,在开始之前快速说几个通知。

All right, friends, quick announcements before we dive in.
M1
M11:21

在正式开始之前,还有一个快速说明。

One more quick note before we get into this.

这是上周我和Noufar Gaspar主持的网络研讨会的录音。

This is a recording of the webinar that was hosted by me and Noufar Gaspar last week.

鉴于本周我们又收到了几款新模型,我认为这里讨论的系统依然非常切中要害。

Given that already we've gotten another couple new models this week, I think the system that is discussed here continues to be extremely pertinent.

希望你觉得有用,明天我们会恢复正常节目。

Hopefully you find this useful, and tomorrow we will be back with a normal episode.

但现在,我们来聊聊如何建立个人AI基准测试吧。

But for now, let's talk about building a personal AI benchmark.
M2
M21:41

我们为你准备了一些很有趣的材料,最后会全部分享出来。

We prepared some very fun materials for you, so we will share everything at the end.

你会全程看到这些材料是什么,但基本上,你看到的我做的和说的所有内容,你都会得到,以便你自己尝试进行类似的流程。

You'll see throughout what are the materials, but basically everything that you see me doing and talking about, you will get everything in order to try and do a similar process for yourself.

记得利用问答框,因为人数众多,我们无法开启麦克风,但Dan在这里,Nathaniel也在,每个人都会尽力回答你们的问题。

And do leverage the Q&A box because we have a large number, so we're unable to open microphones, but Dan is here, and Nathaniel is here, and everybody will try to cater to your questions as best we can.

废话不多说,Nathaniel,说说我们为什么在这里的动力吧。

Without further ado, Nathaniel, some motivation on why we are here.
M1
M12:09

是的,感谢大家来到这里。

Yeah, thanks everyone for being here.

我想,对于各位《AI每日简报》的常客来说,你们知道,在基准测试这个问题上,我一直喜欢唠叨几句。

I think for those of you who are regular listeners of AIDB, you'll know that this is a soapboxy issue for me when it comes to benchmarks.

每次我们得到一个新模型——顺便说一句,现在前沿模型平均每11天就出一个——我的标准警示之一就是,首先,要对公布的基准测试持怀疑态度,倒不是因为我们觉得这些公司在撒谎。

Every time we get a new model, which by the way is now, I think, 11 days on average for a frontier model, one of my standard cautions is basically, First, to be skeptical of the published benchmarks, not so much because we think that any of these companies are lying.

这不是他们做事的本质,而是因为大多数这些基准测试现在都包含在训练数据集中,成为模型训练的一部分规范。

That's not the nature of what they do, but more because most of these benchmarks are now in the training datasets and they're part of the canon that goes into them.

只是基准测试的价值有限,而且它们存在的时间越长,价值就越低。

It's just there's limited value of benchmarks, and the longer they've been around, the less value there is.

现在,很多人不断创建新的基准测试来试图改善这一点,目前这确实是一个相当活跃的环境,比六个月前活跃得多。

Now, lots of people are continuously creating new benchmarks to try to improve that, and it's actually a pretty fertile environment right now, much more so than 6 months ago.

但即便如此,仍然存在实实在在的局限。

But still, there are just real limits to that.

除此之外,即使我们假设所有基准测试都完美准确,我们现在所处的阶段,更多不是关于模型性能的严格优劣判断,而是关于模型给人的感觉以及它如何融入你正在构建的整体模型栈。

On top of that, even if we assumed that all of the benchmarks were perfect and accurate, we are now at a stage where so much is about not strictly better or strictly worse determinations of model performance, but how a model feels and how it fits into the overall model stack that you're putting together.

而且我可以保证,会有这样的情况:在某些基准测试上技术最先进的东西,在你做的那个版本上反而更差,即使那个版本在排行榜上排名较低,对吧?

And there are going to be times, I guarantee, where the thing that is technically state of the art on some benchmark is worse at the version of the thing that you're doing that's lower on the charts, right?

这在写作这类事情上尤其如此,因为写作很大程度是非常主观的。

This is particularly the case for things like writing, where so much of it is incredibly subjective.

所以所有这些都说明,要弄清楚一个新模型是否适合你的工作流程,除了实际使用它之外,没有捷径可走。

So all of that is to say that there's kind of no shortcut solution to figuring out how a new model is or isn't going to fit into your process other than using it.

所以我过去在节目里讲过一件事,就是尝试建立你自己的 benchmark,里面包含你最常用的 use cases。

And so one of the things that I've talked about in the past on the show is trying to put together your own benchmark that has your most common use cases.

NuPhar 非常好心地主动做了一个更全面、更详实、更周到的版本,这样我就不必总是重复这些话,却没有实际资源来帮助你。

And NuPhar has very kindly gone out and put together a more significant, full, thoughtful version of that so that I don't have to just keep repeating this without being able to actually give you a resource to help.

所以我觉得,时机再合适不过了。

So that's, I think, couldn't be more pertinent.

这周,我想我们用的是 Model 3。

We've, I think we're on model 3 this week.

我们拿到了 Claude 5.5、Sol 6.1。

We got Claude 5.5, Sol 6.1.

严格来说,我们还没得到 Gemini 4 Argon。

We didn't technically get Gemini 4 Argon.

我们只是看到它发布,但现在这个时间点,我不觉得这会慢下来,尽管我们现在已经是在跟上前沿节奏了。

We just had it announced, but it's a time right now where I don't see this slowing down despite us pacing the frontier now.
M2
M24:18

我们对 Gemini 有一些预测,但在看到实际情况之前,我们不会分享。

And we have some predictions about Gemini, but we will not share them until we see what happens in reality.

然后我们就能说我们是对的,或者非常安静地承认自己错了。

And then we can say that we were right or very quietly be mistaken.

非常好。

Very well.

好了。

All right.

今天要讲的,或者说我为了让它更标准化并帮助你自行操作而设计的流程,基本上由五个步骤组成。

The plan for today or what the process that I created to make it more standardized and to help you do it for yourself is comprised of basically 5 steps.

第一步,也可以说是最重要的一步,就是决定哪些任务需要放进一个可重复的 benchmark。

The first and arguably the most important step is to decide which of your tasks need to go into a repeatable benchmark.

这些任务应该代表你工作中做的各种事情;如果你也用 AI 处理个人事务,那也要加进去。

Those should be tasks that represent diverse set of things that you do for work and if you use AI also personal reasons, then there as well.

它们需要是那种,基于你得到的结果,你会愿意据此做出决定的任务。

They need to be ones that based on the results that you will get there, you will be happy to make decisions accordingly.

这是最关键的一步之一,会帮助你更好地执行。

That's one of the most important steps that will help you perform.

然后你需要决定哪些是候选对象。

Then you need to decide which are the candidates.

所以很明显,如果想测试的只是单个模型发布,你就可以拿它和你现有的模型名单进行对比。

So obviously, if there is a single model release that you want to check, you can just check that versus your existing model roster.

在很多情况下,就像我们最近几周看到的,模型是成批出现的,所以你可能经常有不止一个新候选要评估。而且,如果你已经用同一套方式工作了一段时间,只是想让事情有点新意,这也是一个很好的做法。

In many cases, like we're seeing over the last few weeks, models come in groups, so Often you might have more than one new candidate to evaluate, and sometimes it's just a good practice if you have been working the same way for a while and you just want to freshen stuff up.

所以,确定你评估的是什么。顺便说一句,我们一直在说模型,但有时候你只是要评估一个新的 harness,或者一个新的 agentic 工具与模型的组合。

So deciding what you are evaluating, and by the way, we keep saying models, but sometimes it's just to evaluate a new harness or a new agentic tool with a combination of a model.

所以有时候不只是模型。

So sometimes it's not just going to be a model.

也可以是某个具体工具、功能,或者你正考虑引入最佳工作方法的东西,然后你再去运行真正的 benchmark 和对比。

It can also be a specific tool or feature or something that you are considering introducing to your best methods of working, and then you will run the actual benchmark and comparison.

为了让大家都能上手,我会介绍三种方法。

In order to make it accessible to everyone, I will introduce 3 ways to do that.

一种是高度手动,但仍然有效。

One is highly manual but still effective.

一种是用我创建的小脚本,我也会分享给你,这样你就能完全复刻我运行 benchmark 时采取的相同步骤。

One with a small script that I created and I'm sharing with you, so you will be able to replicate the exact same steps that I did in order to run mine.

还有一些专门的工具和平台可以帮助公司做这件事,我们会顺带提到。

And there are also dedicated tools and platforms that help companies do that, and we will mention them in passing.

它们不是今天的重点。

Those are not the focus of today.

为了让 benchmark 有效运作,建议你把每个结果是由哪个模型或工具生成的隐藏起来,因为我们都对自己心爱的工具带有很强的偏爱。

In order to make a benchmark work effectively, you will be advised to hide the names of which models or which tools created each result because we are very biased towards our beloved tools.

剧透警告:我对自己的结果相当惊讶,一会儿你们会看到。

Spoiler alert, I was quite surprised with my results, as you will see.

最后,就是做决定了。因为在某一次 benchmark 运行中,某个工具或模型表现得比你当前这套更好,并不一定意味着你需要自动做出改变。

And lastly, there is a decision because the fact that in a specific benchmark run, one tool or one model outperformed your current setup does not necessarily mean that automatically you need to make changes.

首先,因为变化本身就是件难事,其次,有时候是否要做这个改变还涉及其他考量。

First of all, because changes are hard, and second, there are sometimes other considerations as to whether or not to make the change.

可能是公司限制,可能是成本问题,也可能是延迟问题,甚至可能因为改进幅度不够大,让你觉得不值得换。

It can be company constraints, it can be cost, it can be latency, and perhaps even the delta is not large enough for you to justify the switch.

所以我们会在这次网络研讨会中涵盖这全部五个步骤。

So we will cover all of these 5 steps in this webinar.

那么,关于如何应对这些层出不穷的新模型发布,我想分享几点想法——因为每次发布都遵循同样的套路:一个公告、一张测试成绩表(你们都见过Gemini那张表),然后是一周的热议,每个人都说“我有早期访问权限,这是我的看法”,或者有影响力的博主提前分享他们的实验结果。

So a few thoughts about how to even address all of these new model releases, because every release arrives with the same method, an announcement and then a table of test scores, you've seen the Gemini table, and then a week of very hot takes, and everybody is saying, I had early access, and this is what I think, or some preliminary influencer sharing their experimentation.

通常,早期的实验往往严重偏向于制作一个小游戏或小应用,比如搞个3D玩意之类的,这些可重复的用例在社交媒体上看起来特别漂亮。

Often, early experimentations are very skewed towards creating a toy game or a toy application, creating a 3D something, and repeatable use cases that look very nice on social.

而且,有些人会声称,某些模型的设计就是为了在社交媒体和网红测试中表现好看,这样它们在短视频里就会显得很厉害。

And also, some might claim that some of the models are created such that they're optimized specifically for the social media and influencers type of test such that they will look good in reels.

然后大约七天后——有时长一点,取决于热度和粘性——我们开始看到更真实的解读,因为这时候报告来自实际工作场景。

And then when we look about 7 days in, give or take, or sometimes longer depending on the traction and the stickiness, we're starting to get a more real interpretation because then the reports are coming from real work.

举个例子,某个模型悄悄在我的合同审查中漏掉了一个条件,这对律师来说非常关键;而另一个模型却能生成一个令人惊叹的3D游戏,完美复刻了我90年代最爱的某某。

For example, it quietly dropped a condition in my contract review that's much more critical for a lawyer versus it was able to generate an amazing 3D game that replicates my '90s beloved whatever.

所以我们对早期的发布持保留态度。

So that's why we're taking with a pinch of salt the early releases.

我们会认真听,因为也许真有颠覆性的东西,但最终还是要看一段时间后的效果。

We listen keenly because maybe there is a big game changer, but eventually we need to see what happens after a while.

更重要的是,我们自己得亲自测试。

And even more importantly, we need to test it on our own.

具体到今天,我们会看看一些新来者,比如Claude Opus 5.5、Grok和GPT-6.1 Sol。

Specifically for today, we will be looking at some of these new arrivals like the Claude Opus 5.5 and Grok and GPT-6.1 Sol.

Gemini什么时候来还不确定,如果你们愿意等,可以到时候再跑基准测试,或者用Gemini重新跑一遍。

Gemini will come when it comes, and then maybe if you want to wait, you can run the benchmark then or rerun it with Gemini.

我当然很好奇,一旦我们获得Gemini的访问权限,我会把它加进去,跟其他模型以及我将展示的其他结果进行比较。

I definitely am curious, and once we get access to Gemini, I will add it and compare it against the other models and the other results that I will show you.

这就是今天的重点,也是我要具体比较的内容。

So that's the focus for today and the stuff that I will specifically compare.

你们听我讲的时候,可能心里在想:我对个人基准测试这个整体概念有一些担忧或反对意见。

While you're listening to me speaking, you might be saying, I have a few concerns or objections to the entire concept of personal benchmarking.

通常,这些意见有两种类型。

Typically, they come in one of two shapes.

第一种是:现在什么都那么好了。

The one is, everything is so good now.

前沿模型已经遥遥领先。

The frontier is so far ahead.

就像,我们干嘛还要在意?

Like, why should we even care?

是不是只是工具FOMO或模型FOMO在作祟,感觉如果我们不追逐前沿,就落伍了?

Is it just a matter of tool FOMO or model FOMO that it feels like unless we're chasing the frontier, we're not relevant?

这就是一种反对意见。

So that can be one type of objection.

我对这种反对意见的回应是:首先,你说得对。

My response to this type of objection is, first of all, you are correct.

前沿模型确实很好,而且就算你不一定要更换工具或模型,你能挖掘的价值也远远大得多。

The frontier is very good, and odds are that there is so much more value that you can drive, even if you will not necessarily change the tools or the models that you're using.

但是,有时候,一个新模型的发布——就像NLW在播客里反复说的——可能会解锁一个以前只存在于愿望清单上的用例。

However, in some cases, all of a sudden, a new model release, like NLW is repeatedly saying on the podcast, may unlock a use case that was only on the wish list up until now.

而且我认为,即便只是为了这个理由——解锁新用例,推动AI使用方式的边界——这也是一个很好的理由,至少偶尔考虑升级到新模型或新工具,确保你能够最大限度地利用这项技术。

And I think that even for that justification alone, to have new use cases being unlocked and to push the envelope of how you use AI, this is one justification as to why to at least occasionally consider upgrading to a new model or a new tool just to make sure that you are leveraging the technology to the best of its ability.

并不需要每周都做,因为那样本身就会造成很大的混乱。

And it doesn't have to be on a weekly basis because that in itself can really create a lot of confusion.

所以你可以根据自己的精力定期进行。

And so you can run it periodically according to your bandwidth.

第二种反对意见或考量是,有时候人们会说:这挺好玩的,像你们这样能随便用任何工具和模型的人,可以尽情换来换去、追赶前沿。

The second objection or consideration is that sometimes people will say, this is all fun and you guys that have the full freedom to use whatever tool and whatever model can enjoy swapping and following up the frontier.

但在我这里,我的公司限制我只能用某个特定的工具或模型。

But in my case, my company is limiting me to using a specific tool or a specific model.

这是很多员工都会遇到的现实情况。

And that's a common reality for many employees.

首先,我表示同情,我理解你的感受。

So first of all, my condolences, and I hear you.

这真的不容易。

It's not easy at all.

不过,即使模型和工具的选择有限,也总还是有选择的余地,哪怕只是在不同配置之间选,比如快速思考、快捷模式之类的。

However, even with limited set of models and tools, there is always a choice, even if the choice is just selecting between fast thinking and quick or some of the other configuration for the specific tools.

所以,即使你把评测范围限制在工具内部那一点点可选项上,这些选择也同样重要。

So even if you limit the benchmark only to the very narrow set of choices that you do have within your tools, those matter.

有时候,不管是出于个人使用场景,还是为了说服公司里做决策的人,跑一次基准测试,哪怕是用模拟数据,也值得做——这样你就能向公司的人证明,也许你们应该考虑再买一个许可证。

And sometimes for either personal use cases or just to maybe encourage the people who decide within the company, it's worth running even a benchmark on a mock set of data so you can prove to the people in the company that perhaps you should consider purchasing another license.

这就是所谓的反模式。

So that's the anti-pattern.

说完这些,我想回到一开始提到的——整个流程中最重要的决策。

With all of that out of the way, I want to start with, as I said at the beginning, the most consequential decision out of the process.

也就是你用来做基准测试的具体使用场景。

And these are the concrete use cases that you will be using for your benchmark.

在我看来,理想情况下,你需要覆盖你实际在做的事情,不管是工作还是个人使用。

In my opinion, ideally, you need to cover what you actually do either at work and personal usage.

当然,这个比例要跟你实际用 AI 的情况保持一致。

And of course, the balance should be based on your actual balance of what you use AI for.

如果你主要用 AI 来工作,那就侧重工作场景。

If you primarily use AI for work, focus on that.

如果你主要用 AI 处理家里的事情,那就多关注那些方面。

If you primarily use AI to do stuff at home, then shed more light there.

然后,我建议你尽量构建一套多样化且均衡的使用场景。

And then I want you to try and create a diverse and balanced set of use cases.

当然,你不可能把工作中的每个任务都跑一遍,所以你要精心挑选一套好的场景——如果你发现某个模型在这个场景下有明显提升,或者表现好得多,你就会认真考虑切换——也就是说,这套场景的产出你要足够信任,能用来为工作做实际决策。

And of course, you cannot run each and every task that you do for work, so you want to cherry-pick a good set that if you're seeing a significant uplift or that a specific model behaves much better with this use case, that you will for real consider swapping, meaning it needs to be a set that you trust the output to make actual decisions for your work.

这些场景要多样化,涵盖你不同类型的工作以及使用频率,最好是那些你经常做的任务,因为你从中获得的提升最大。

It needs to be diverse in the kind of work that you do and the frequency, ideally stuff that you do regularly because those are where you get the highest uplift.

你还要问问自己,这里面有什么风险?如果你知道哪些错误代价很高,或者有办法快速判断结果完全不对,那也是一个好的使用场景——因为如果你能识别出糟糕的结果,有时候这比那些需要区分细微差别的正面场景更有用。

You should also ask yourself, what's at stake, and if you know of costly mistakes or ways for you to evaluate quickly that this is completely off, that's also a good use case because if you can evaluate what bad looks like, sometimes it's a better use case than the one where you have lots of nuances on what good looks like.

另外,对于你正在使用的具体任务,它是从零开始,还是只能基于现有文档来启动?比如润色一封邮件、美化一份演示文稿,这些都需要先有内容才能开始做。

Also, for the specific tasks that you're leveraging, is it a task that starts from a clean slate or is it a task that can only start from an existing document like polishing an email, polishing a deck, stuff that requires some substance before it starts.

最后,就像我说的,这也是为什么即使你对现有配置很满意,仍然值得跑这些基准测试的原因之一——至少要包含一个你特别想实现但还没做好的任务。

And lastly, as I said, and that's part of the motivation of why to run these benchmarks even if you're quite happy with your existing setup, include at least one task that it's on the wish list.

这是什么意思呢?

What do I mean by that?

就是那些你用 AI 试过,但对结果不满意,所以平时不会用 AI 去做的任务。

Something that you've tried with AI, you were unhappy with the results, and thereby it's not a type of task that you regularly do with AI.

我强烈建议你放一两个这样的任务进去,这样当新的前沿模型出现时,这些任务就是最让你兴奋、促使你切换到新模型的因素之一。

I would strongly recommend that you have one or two of those such that when something new comes at the frontier, this is among the top things that might get excited and to move to the frontier model.

这就是我们需要做的核心决策。

So that's the core decisions that we need to make here.

当然,你可以根据你的实际情况来决定数量。

And of course, you can think about the number that fits your life.

我建议大概 4 到 6 个这样的使用场景,当然你也可以多选或少选。

My recommendation would be like 4 to 6 type of such use cases, but you can of course create more or less.

我会展示我是怎么梳理出来的,但为了具体一点,这是我的 6 个使用场景列表。

I will show you my process for getting there, but just to make it concrete, this is my list of 6 use cases.

一个是发 LinkedIn 内容,另一个是跟利益相关者进行关键甚至有时不太愉快的对话,比如在价格谈判中据理力争。

One that posts stuff on LinkedIn, one that holds a critical or sometimes an unpleasant conversation with stakeholders around pushing back on a pricing negotiation.

还有做一些关于工具特性的调研,要具体一些;一个网站,包含我上的一门课的内容;以及为我管理的课程设计练习。

As well as doing some research around the tool features with some specifics, a website that will include content for a course that I run, creating an exercise for courses that I manage.

这个在愿望清单里——一个像样的每周计划,就像一个靠谱的总参谋长那样,让我不会感到沮丧或忽略它。

And this is in the wishlist, a proper weekly plan that actually is like a good chief of staff that I'm not getting frustrated and neglecting.

到目前为止,我还没找到一个模型能一次性搞定复杂的每周规划,而不需要我来回提示它。

Up until now, I'm unable to find a model that everything that they do with regards to an intricate weekly planning is indeed like correct on a one shot instead of me having to prompt it back and forth.

所以这就是我的系统,我马上给你展示我是怎么得出这个结论的。

So that's my system, and I'll show you how I got there very quickly.

一旦你有了用例列表和用例的具体细节,接下来你要做的就是进行一次品测,当然最好是盲测。

Once you have the use cases list and the use cases specifics, then what you would want to do is to basically do a taste testing, ideally, of course, a blind taste testing.

我们要选出哪些模型、工具或者模型与工具的组合作为候选来评估。

We want to select which models or tools or model and tool combinations are the candidate ones to evaluate.

快速说明一下,在这个实验中,我跑了7个模型选项,我可以告诉你,7个实在太多了。

Quick note, for this lab, I ran 7 model options, and I can tell you that 7 is way too much.

对6个不同的任务,给7个模型选项打分和审阅,过程极其繁琐。

It was Extremely tedious to review and rate 7 different model options for 6 different tasks.

所以我其实不是……其实是5个,但也不推荐。

So I'm not— or it was 5, but not recommended.

我建议只跑现有模型对比一个新模型,或者最多比较4个候选。

I would recommend to run the existing model versus a new one or up to 4 different candidates.

其中一个候选应该是你当前做这项工作的基线。

One of them should also be the baseline of how you currently do the work.

这样你才能知道是否要切换到新模型。

So you will know whether you want to convert to a new model.

然后我们每次请求都要在全新的对话中运行,这样上下文窗口就不会让我们跑偏。

Then we want to run every request in a fresh chat such that the context window will not send us on a tailspin.

接下来,这一步非常重要,我们要隐藏模型的名字。

And then, and that's a very important step, we want to hide the names.

一会儿我会简单说说怎么做到这一点。

A few thoughts on how to do that in a minute.

做完盲测之后,我们可以口头反思一下是否喜欢这个输出。

And then once we do the blind taste test, we can do a combination of verbal reflection of whether or not we like the output.

我们可以给一个数字评分,或者两者都做。

We can do a numeric score or we can do both.

你即将获得的这个程序化工具包鼓励你两者都做。

The programmatic package that you're getting is encouraging you to do both.

按1到5分打分,并写一行观察记录,说明你喜欢什么、不喜欢什么。

Score it on a scale of 1 to 5 and give it a one-line observation of your likes and dislikes.

盲测完成后,你就可以揭晓各个模型的身份了。

Once you're done with doing the blind taste test, you can reveal who were the models.

这是个很有意思的环节,可以看看你预测的哪个模型产生了哪个输出是否准确。

That's a very fun exercise to see what are your predictions of which model created what output are.

我保证其中一些结果会让你惊讶,因为我们对某些模型有很多固有看法,但并非全都符合现实。

I can guarantee you that some of them will surprise you because we do have a lot of beliefs on some of these models that not all meet reality.

最后,我们需要做决定。

And lastly, we need to decide.

所谓做决定,即使某个模型明显更优,但可能优势还不够大,或者有别的考虑,我们可能不会因此改变习惯。

And by decide, even though something that was significantly superior might not be superior enough or for other considerations, we might not choose to change our habits for that.

以上就是我们进行品测的方法。

So that's how we run the taste test.

具体到如何让不同模型的结果匿名化,有4种方式可以做到。

Specifically on how to blindfold the different results from the different models, 4 ways that you can do that.

第一,你可以找一个朋友帮你做匿名处理。

One, you can call a friend to do the anonymization for you.

要么让他们帮你运行所有结果,然后只给你一套匿名化的结果供你评估。

Either they run all the results for you and just give you an anonymized set of results to evaluate.

这适用于你手动做基准测试的情况,手动做当然没问题。

That's relevant if you're doing the benchmark manually, which you definitely can.

只是会多费一点工夫。

It's just a little bit more work.

你可以用电子表格生成随机数,然后这样匿名化。

You can leverage a spreadsheet to create a random number and then anonymize like that.

你可以利用AI工具来打乱和重新标记,这样你就不知道结果来自哪个模型了,或者你也可以用我写的脚本帮你做这件事。

You can leverage an AI tool to shuffle and relabel them such that you don't know, or you can use the script that I created that does that for you.

当你使用这个脚本时,有一件事需要做:你需要使用API key来实际执行任务。

When you will be using the script, one thing that you will need to do is to run the actual task execution using API keys.

一种方法是用 OpenRouter,它允许你用单个 API 密钥运行多个模型;或者你也可以为每个供应商单独准备 API 密钥;如果你只跟某个特定供应商合作,只想评估它不同层级的模型,那用单个 API 密钥就行。但无论哪种,编程或脚本方式都需要 API 密钥。

So one option is to do that with OpenRouter that lets you run multiple models with a single API key, or you can have API keys for each and every one of the vendors, or if you work with just a specific vendor that you just want to evaluate between the different model tiers, you can have a single API key, but the programmatic or the script-based approach requires an API key.

我可以向你保证,即使你不懂技术,现在也很容易做到。

I can guarantee you that even if you're not technical, it's very doable nowadays.

这就是匿名化处理的方法。

So that's how you anonymize.

还有一点要特别注意,我强烈建议,有些任务会更偏视觉,比如网页设计、应用设计或图像生成。

One more thing to pay attention, I strongly recommend that some of the tasks will be more visual, whether it's website design, application design, or image generation.

因为大多数情况下,视觉内容更容易让我们快速打分。

Just because in most cases it's easier for us to score quickly something that is visual.

当然,如果你根本不用 AI 做任何视觉相关的事,就别为了测试而硬选视觉任务。

Of course, if you are not doing anything visual with AI, don't choose a task that is visual just for the sake of that.

但如果你偶尔也用 AI 做点视觉方面的事,就确保测评中包含这类内容。

But if you are occasionally using AI for something that is more visual, make sure that it includes that.

要评估视觉内容,你可能需要把它们存到专门的文件夹里,这样就能在固定位置打开每个视觉素材,比如网页之类的。

In order to evaluate visual stuff, you will probably need to store it in a dedicated folder so you can open each visual asset, whether it's a website and so on, in a dedicated place.

很快你就会看到我是怎么处理网页设计的。

You will see how I did it with the website design very shortly.

接下来我们需要做个选择。

And then we need to choose.

第一个也是最重要的方法,就是你自己给结果打分,因为这是个人基准,没人比你更清楚自己的偏好。

So the first method and the most important method is you scoring the results because it's a personal benchmark and nobody knows better than you your preferences.

就像你之前指出的,有时候会有一个 X 因素,让你喜欢或讨厌某个结果,哪怕你说不出原因,你只是更喜欢其中一个。

And as you pointed before, there is sometimes an X factor of why you like or dislike, even if you don't know how to explain, you just prefer one result over the other.

有时候我们确实词穷,说不清为什么喜欢这个而不是那个,但理想情况下,你应该是自己喜好的主要评判者。

Sometimes we don't have the vocabulary to say why we prefer one result over the other, but ideally, you should be the primary judge of what you like and dislike.

根据我的经验,单独评判一个东西往往很难,也就是说,你从某个工具或模型得到结果后,想给它打一个 1 到 5 的绝对分数,非常困难。

From my experience, it's often very difficult to judge something as a standalone, meaning you will get results from a single tool or model, and just to give it an absolute score between 1 and 5, it's very hard.

更简单的方法往往是,把同一任务的几个备选结果放在一起对比,然后说,这个我最喜欢,那个我最不喜欢,然后据此打分。

What is often easier is to see side by side several alternatives for a particular task and say, I like this the most, I like this the least, and score it accordingly.

通常就该这么做。

That's often how you should do that.

这是一种方法,而且很可能最重要,因为涉及你的个人品味和那个 X 因素——你常常说不清为什么就是更喜欢某个结果。

So that's one approach and probably the one that will matter the most because of your taste and this X factor where it's often hard to explain why you liked something better than other you just did.

另一种方法,而且你可以两种方法都试试再对比,就是借助 AI 工具来做评判。

Another approach, and you can always run both of them and then compare, is to leverage an AI tool as a judge.

这样就不是你一个人在打分,而是也有一个 AI 工具,根据你设定的评判标准来评估结果,比如什么是好的 LinkedIn 帖子、什么是好的供应商谈判、什么是好的网页设计。

And then it's not just you that's scoring, but also an AI tool that judges the result according to a predefined rubric that you configured of what's a good LinkedIn post, what's a good vendor negotiation, what's a good website design.

你随时可以对比,看看综合评分是否一致,以及是否形成了一种清晰的模式,告诉你需要调整哪些地方。

You can always compare and see whether the cumulative review is in agreement and whether it creates a clear pattern of stuff that you need to change.

但我要提醒你,用 AI 做评判也有其缺点,关于用 AI 做评估,已经有一套完整的理论体系了。

I do want to warn you that using AI as a judge has its own shortcomings, and there is an entire theory around using AI for evaluation.

不过我要说的是,你应该先评估评判者本身。

But I will say that you should evaluate the judge itself.

因为,如果这个评判者完全不符合你的品味,那它就是个糟糕的评判者,直接扔掉。

So, because if the judge is completely not correlated to your taste, it's a bad judge and just throw it away.

关于评判还有几点要补充。

A few more points on judgment.

因为品味非常主观,所以我们才需要并排对比和口头描述。

Because taste is highly subjective, that's why we need to do the side-by-side and like a verbal description.

而且我们才尝试用 AI 根据我们的标准来做主观测试,但正如我说的,这并不完美。

And that's why we're trying to leverage AI to do the subjective testing according to our rubrics, but as I said, it's not perfect.

理想情况下,你用来做评判的模型,应该跟执行任务的那些模型不一样。

Ideally, the model that you will use for judging is different from the models that executed the results.

因为存在某种 AI 裙带关系。

There is kind of an AI nepotism.

某个模型很可能会偏袒同系列的模型,而不是其他系列的。

Odds are that a specific model will favor models from its own family versus from other families.

虽然不是绝对,但这种现象很常见。

Not always, but often this is seen.

所以如果你有另一个模型系列不在你评估的范围之内,那就让它来做评判。

So if you have another model family that is outside of what you're evaluating, make that the judge.

在我的用例中,因为我还没有用Gemini评估过任何东西,所以评判模型用的是Gemini。

In my use case, because I haven't evaluated anything with Gemini, the judge was a Gemini model.

就像我说的,隐藏模型的名字非常重要,因为我们非常偏见,而且没人能证明我错了。

As I said, hiding the names of the models is of very important criticality because we are very biased, and nobody can prove me otherwise.

我相信我们可以证明情况确实如此。

I'm sure that we can prove that this is the case.

我们希望尽可能保持一切不变,也就是说,对每个模型都提出完全相同的请求,然后以公平且对等的方式进行比较。

We want to maintain anything as frozen as possible, meaning it's going to be the exact same request to each and every one of the models and then compare it in a fair and equivalent way.

而且因为答案可能不同,如果你在某个非常重要的事情上运行这个个人基准测试,并且这可能涉及到金钱或者其他方面的影响,从而促使你改变工作方式。

And because answers may vary, if you're running this personal benchmark on something that is very critical and there is perhaps a monetary implication or other implications to you making changes to how you work.

理想情况下,你应该对每个模型多次运行相同的请求,然后比较这几次执行的结果,因为我们知道模型的反应非常多样。

Ideally, you should run the same request several times per model and then compare these several executions because we know that models have a very diverse set of responses.

即使我们使用完全相同的模型、完全相同的配置和完全相同的提示,你也会得到不同的结果。

Even if we're using the exact same model with the exact same configuration and the exact same prompt, you will get different results.

我今天刚刚看到这一点,当时我让一个模型生成一个网站。

I just saw it today when I gave a model a website to generate.

我真的很喜欢第一次的结果,但中间卡住了,然后第二次执行时,我觉得远不如第一次好。

I really loved the result, but something got stuck in the middle, and then in the second execution, I liked it much less than the first execution.

所以这是真实的情况。

So that's something real.

如果你担心基准测试的有效性,多次运行就能确定结果。

And if you are concerned about the validity of the benchmark, running it several times will seal the deal.
M1
M122:03

KPMG和University of Texas at Austin的一项新研究发现,当人们使用AI工作时,相似的技能并不能保证相似的结果。

A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes.

研究人员研究了500多名职业生涯早期专业人士,发现表现最佳的人通过引导、评估和优化AI的输出来持续放大AI的价值。

Researchers studied more than 500 early career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs.

这些表现最佳者被称为AI Amplifiers,他们并非仅由所掌握的知识定义,而是取决于他们如何与AI合作。

These top performers, called AI Amplifiers, weren't defined by what they knew alone, but by how they worked with AI.
M1
M122:38

在这一点上,问题不再是公司是否在积极使用AI。

At this point, it's no longer a question of whether companies are actively using AI.

而如何用好AI,则完全是另一回事了。

Using it well, on the other hand, is a whole different story.

但Robots and Pencils是一家我可以指出的、真正为这个时代而生的公司。

Robots and Pencils, though, is a company that I can point to that is actually built for this time.

他们是一家应用AI工程公司,直接与客户合作解决对企业重要的问题,而不是停留在幻灯片里的实验。

They're an applied AI engineering firm working directly with clients on problems that matter to the business, not experiments that live in a slide deck.

每一次合作都是从客户真正需要的结果逆向推导开始的。

Every engagement starts by working backwards from the outcome a client actually needs.

如果你想在这个领域区分真正的AI工程和噪音,这就是关键所在。

If you're trying to tell real AI engineering apart from noise in this space, that's the difference maker.
M1
M123:08

每一集我们都会报道OpenAI、Anthropic、SpaceX AI、Google和Meta之间的竞争。

Every episode we cover the competition between OpenAI, Anthropic, SpaceX AI, Google, and Meta.

很可能你已经对谁处于领先地位有了自己的看法。

Chances are you've already formed an opinion about who's leading.

但每个AI实验室都采取不同的方法,构建不同的技术,建立不同的合作伙伴关系,并发展独特的生态系统。

But every AI lab is taking a different approach, building different technologies, forging different partnerships, and developing a unique ecosystem.
M2
M224:56

好的。

All right.

就像我们说的,最后一步是决定。

And as we said, the final thing is the decision.

理想情况下,你应该做出三种类型的决定。

Ideally, you should make a decision of 3 types.

Switch(切换),意思是我确信这个模型或工具明显更好,综合考虑所有其他因素,我想切换到这种新配置。

Switch, meaning I am convinced that this model or this tool is significantly better, and given everything else that I have that I need to consider, I want to switch to this new configuration.

Split(分割),意思是我可能会在不同的选项之间进行更细致的权衡。

Split, meaning that I might be going between different options with more nuance.

或者Stay(保持),就像,我并不认为我刚才评估的新事物的表现足以让我改变工作方式。

Or stay, like, I'm not convinced that what I just saw about the performance of the new thing that I evaluated justifies me making changes to the way I work.

所以这就是核心决策。

So that's the core decisions.

但是,要做出决策,并不是那么直接了当,比如“这是最终的最佳结果,因此我就要切换”。

But in order to make decisions, it's not just straightforward as, like, this is the final best result and thereby I'm switching.

还有其他一些考虑因素,有时会让你保持现状。

There are other considerations that sometimes will get you to stay with the status quo.

例如,也许你的公司不允许你切换,所以这并不相关,或者也许你担心某些条款和条件。

For example, maybe your company does not permit you to move, so it's not really relevant, or maybe some of the terms and conditions you are concerned about.

也许你自己的计划根本没有提供,或者没有给足够的使用额度,来支持你所考虑的新工具或新模型。

Maybe your own plan does not offer at all or generous enough usage for the new tool or model that you're contemplating.

在其他情况下,胜者并不明确;它不足以证明改变是合理的。如果你需要打破僵局,一些可以帮助你决定的因素包括成本、执行速度,也许还有一些你很喜欢而不想改变的具体功能,以及习惯。

In other cases, the winner is not straightforward; it just doesn't justify making changes, and if you are in a position to have some tie-breaking, some things that can help you decide are costs, the speed of execution, maybe there are specific features that you just enjoy and you don't want to change them, as well as habits.

养成习惯的成本很高,所以如果没有显著的提升,仅仅保持原来的工作方式也有其优点。

Habit formation is an expensive thing, so if there is no significant boost, just staying with the way you used to work also has merits.

但如果所有这些测试都证明改变是合理的,那么就做出改变。

But if all of these tests justify the change, then make the change.

所以这就是整个过程。

So that's the entire process.

接下来你们将看到的是我创建的一个自动化过程,就像开头提到的,你们可以获得它,以便自己实施。它的工作方式如下:

What you will see from here on after is an automated process that I created and that, as mentioned at the beginning, you're getting in order to implement for yourself that works as follows.

它从你保存的任务开始。

It starts with your saved tasks.

我会展示如何保存它们。

I'll show you how to save them.

然后,我使用OpenRouter API密钥来为不同类型模型生成结果。

And then I'm using the OpenRouter API key in order to create the results for different types of models.

然后我们得到每个选项的答案。

Then we get the answers for each and every option.

代码本身会对结果进行匿名化处理,这样我就不知道这些结果是由哪些原始模型产生的。

The code itself anonymizes the results such that I don't see what the original models for the execution are.

你在不知道哪个模型产生什么结果的情况下对它们进行盲评。

You score them blindly without knowing which model created what.

你给出评分以及文字描述,然后有一个有趣的大揭晓环节,向你展示你的偏好与评判者以及模型名称的对比,这是有趣的部分,常常会让你感到惊讶。

You provide a score as well as a verbal description, and then there is a fun big reveal that shows you what your preferences are versus the judge and versus the model names, and this is the fun part where often you will be surprised.

我用了OpenRouter,但正如我所说的,还有其他模型集散地。

I used OpenRouter, but there are other model hubs, like I said.

使用任何你觉得相关的平台,但你需要相关模型的API密钥。

Use anything that you find to be relevant, but you will need an API key for the relevant models.

所以这就是你们即将看到的方法论。

So that's the methodology of what we— you are about to see.

正如我说过的,我正在评估多个模型。

And as I said, I'm evaluating multiple models.

所谓的“基线”,虽然我一段时间没用了,是Opus 5。

The quote-unquote baseline, even though I'm not using it for a while, is Opus 5.

同时我也在尝试Opus 5.5、Grok 4.7、GPT-6.1 SOL。

And then I'm trying also Opus 5.5, Grok 4.7, GPT-6.1 SOL.

我还想将它们与非常昂贵的前沿模型Fable 5.1和GPT-6 Astra进行对比测试。

And I also wanted to test them against the very expensive frontier model of Fable 5.1 and GPT-6 Astra.

为了确保给开源模型留出空间,我也将它们与Kimi K3进行对比测试。

And to make sure that we give room for the open source, I'm also trying them against Kimi K3.

这就是为什么我得到了7个不同的选项,并且评估起来非常困难。

And that's why I got to 7 different options and it was very difficult to evaluate.

最后,就像我说的,由不同的模型来运行评判。

And lastly, as I said, a different model runs the judging.

在这个例子中,我使用了Gemini Pro。

In this case, I used Gemini Pro.

所以这就是你们即将看到的。

So that's what you're about to see.

现在让我带你们了解这个过程。

And now let me walk you through the process.

你首先会被鼓励做的,就是运行一个提示词。

The very first step that you will be encouraged to do is to run a prompt.

这个提示词是材料的一部分,它帮你发现哪些用例应该纳入你的基准测试。

You have the prompt as part of the materials that helps you discover which use cases you should include as part of your benchmark.

所以这个提示词会像采访你一样,利用最了解你的那个AI工具所能给出的全部建议。

So it's a prompt that will interview you and leverage everything that the AI tool that knows you best can advise.

所以一定要在一个拥有大量记忆和大量上下文的工具里运行它,尽可能多地利用这些。

So definitely run that in a tool that has a ton of memory and a ton of context as much as possible.

它会帮你搞清楚这些问题。

So it will help you figure it out.

它会经历大约8个问题的流程,帮你找出最佳用例。

It will run a process of about 8 questions that will help you figure out the best use cases.

如你所见,Claude知道得挺多的。

As you can see, Claude knows quite a lot.

它知道我的角色,也知道我一周通常都做些什么。

It knows my role and it knows what a typical week has.

所以它会有一些猜测,对用例也有一些初步建议,还有一些来回互动的内容。

So it has some guesses and some initial suggestions for use cases, as well as stuff that goes back and forth.

然后它就开始问我问题了。

And then it started asking me questions.

所以我回答了它提出的所有问题。

So I answered all the questions that it had.

在8个问题之后,虽然有点烦人,但也没太烦,它就给我生成了一组6个推荐用例。

After 8 questions, a little bit annoying, but not too much, it created for me a set of 6 recommended use cases.

对每个用例,它还会创建一个提示词,用来做实际的基准测试。

For each use case, it will create also a prompt that will be used to do the actual benchmark.

所以它不只是发掘出用例,还会发掘出真正发送去运行基准测试的那个提示词。

So it doesn't just uncover the use cases, it also uncovers the actual prompt that will be sent for running the benchmark.

所以那个就是“根据原始笔记写LinkedIn帖子”,还有“提示词定价异议加敏感邮件”。

So that's the LinkedIn post from raw notes, and that's the prompt pricing pushback plus sensitive email.

“按套餐层级整理当前AI工具功能”,这个需要联网搜索和查资料。

Current AI tool feature by plan tier, that this requires a web search and some research.

“课程参与者中心网站”、“面向新受众的一次性练习”,还有“周计划:课程参与愿望清单”。

Course participant hub website, one-shot exercise for a new audience, and week plan a course engagement wishlist.

所以这些就是前6个,最后它还会显示你的覆盖情况。

So these are the top 6, and at the end it will also show your coverage.

比如我决定不包含任何个人相关的内容,因为我在个人方面使用AI没有工作方面那么多,所以那些对我来说不太相关。

And I decided, for example, to not do anything personal because I don't use AI as much personally as I use for work, so it was less relevant for me.

它确实也指出了一些可能缺失、值得注意的东西。所以这就是最终结果。

And it did call out some things that are missing potentially to pay attention to So that's the final results.

最后,如果它是一个agentic工具,它会把它保存到一个专用文件夹里,其中每个文件都是独立的文件。

At the end, it will, if it's an agentic tool, it will save it as a dedicated folder where each of these files is a dedicated file on its own.

比如,这就是“LinkedIn帖子生成器”文件的样子。

For example, this is how the LinkedIn post generator file looks like.

里面有任务名称、快速动机、请求。

It's the name of the task, quick motivation, the request.

这就是会发送给模型的实际提示词。

This is the actual prompt that will be sent to the model.

这很重要,因为如果你想用AI工具来评判基准测试,而不仅仅基于自己的偏好,那会有一个给评判者的评分指南。

And that's important if you want to judge the benchmark also using an AI tool and not just based on your preferences, there is a scoring guide for the judge.

所以它也被附加为任务本身的一部分。

So this is also appended as part of the task itself.

我想让你做的一件事是,看看通过这个过程生成的提示词或任务。

One thing that I want you to do is to take a look at the prompt or the task that was created using the process.

如果它看起来太泛泛,或者无法帮你在某个具体用法上做决定,或者你觉得它需要更深入、或者完全不一样,那么这一步就非常重要:你要去细化请求和评分指南,这样才能做出决策。

If it seems too generic or not something that will help you make up your mind on a specific usage, or you think that it needs to go deeper or be completely different, then this is a very important step to refine both the request and the scoring guide such that you will be able to make decisions.

如果你不觉得这些任务能很好地代表你的工作方式,那这就不是一个足够好的基准测试。

If you don't trust that these tasks are a good representation of the way you work, that's not a good enough benchmark.

所以我这里有6个这样的任务,从1到6,是Claude根据那次采访过程,以及你也能获得的全部指令生成的。

So I have 6 of those from number 1 to number 6 that was created by Claude based on the interview and all the instructions that you can also get.

所以一旦我们都有了它们,我们就可以开始运行自动化过程了。

So once we have all of them, we can start running the automated process.

这个自动化过程是我创建的一个脚本,就像我说过的,我正在分享给你。

The automated process is a script that I created, and as I said, I'm sharing it with you.

为了让过程有趣且愉快,这个脚本还创建了一个简单的用户界面,就托管在你自己的机器上。

To make it fun and a pleasant experience, the script also creates this simple user interface that is locally hosted on your own machine.

它不需要去任何地方。

It doesn't need to go anywhere.

所以,它不是——没有你未打算的数据共享。

So it's not— there is no sharing of data that you haven't intended.

而且总的来说,顺便提一下,如果你的公司数据敏感,就不要用像OpenRouter这样的服务来处理公司数据。

And in general, by the way, if your company is sensitive, don't use something like OpenRouter with company data.

使用模拟数据或模拟用例,这样就不会有任何机密信息泄露到任何地方。

Use mock data or mock use cases so there is no confidential information leaking anywhere.

这种用户界面创建的方式,你会看到这里。

The way this user interface is created, you will get here.

需要评估的任务列表。

The list of your tasks to be evaluated.

对于每一个任务,它会生成所有候选模型。

For each one of them, it will generate all the candidate models.

在我们的案例中,它将模型匿名化为从A到G,并且在每个用例的顶部,它会根据你的评分标准告诉你什么是好的答案,以帮助你正确评分。

In our case, it anonymized them from A to G, and at the top of each use case, it will tell you from your rubric what's a good answer, also to help you to score it properly.

然后你可以并排看到我们从OpenRouter得到的所有结果。

And then you have side by side all of the results that we got from OpenRouter.

所以这是在脚本后台运行并使用提供的API密钥后完成的,它为我之前看到的五个任务中的每一个都生成了候选回答,并将它们并排展示。

So this was done after the script ran in the background and used the API key that was provided, and it generated candidate responses for each and every one of my five tasks that you've seen before, and it puts them side by side.

所以你可以看到,例如,这是LinkedIn帖子生成器。

So you can see, for example, this is the LinkedIn post generator.

所以模型D提供了这个开头选项:三只手举手,等等等等。

So Model D offers this hook option: three hands went up, yada yada yada.

完整的帖子结尾选项。

The full post ending option.

所以第一个看起来就是这样。

So that's how the first one looks.

然后它要求我做的是提供一个1到5的评分,以及一个口头评价。

And then what it's asking me to do is to provide a score from one to five, and like one verbal evaluation.

所以在这种情况下,我打了五分中的两分,并且说。

So in this case, I scored it a two out of five and said.

虽然它可读且清晰,但也太轻泛了,而且包含了那种“是A而不是B”的模式,我个人非常讨厌AI写作中的这种模式。

That while it's readable and clear, it's also a bit too light and it includes its A and not B pattern that I personally really hate about AI writing.

所以我得到了7个这样的。

So I got 7 of those.

而且就像我跟你说过的,在至少阅读一批之前很难评分,因为至少在LinkedIn帖子那个任务中,不同模型的回答有些重叠。

And as I told you, it was very hard to score before I read at least a bunch of those because they are, at least in the LinkedIn, it was a little bit overlapping with some of the writing.

但是正如你可以看到的,我确实给出了不同的评分。

But as you can see, I did score it differently.

所以有些得到了3分、4分。

So some of them got a 3, a 4.

实际上,没有一个被我打5分,因为没有一个写出足够好的LinkedIn文章。

Actually, none of them got a 5 from me because none of them wrote a good enough LinkedIn article.

所以没有一个可以直接发布。

So none of them was ready to post as is.

所以那是任务1,当我完成时,界面显示所有任务已评分。

So that was task number 1, and when I was done, it indicated that this is all scored.

我还有定价反驳、一次性练习和周计划这些任务。

I also have pricing pushback and one-shot exercise and week plan.

这些都是文字量很大的用例,所以它们不那么令人兴奋。

These are all text-heavy use cases, so they are not that exciting.

我在想,最令人兴奋的事情是看到不同的候选模型如何创建一个课程参与者中心网站。

I am thinking that the most exciting thing is to see how the different candidate models created a course participant hub website.

所以这里呢,如果你有HTML或者其他视觉素材的话,这个小应用挺有意思的,你也可以自己做一个,而且还能拿到它。

So here, if you have like HTML or other visual assets, this small fun application that you can build for yourself and you're getting it also.

它只是展示了不同的网站。

It just showcases the different websites.

我把它们都全屏打开给你看,而且当你跟我一起浏览这些由不同模型创建的网站时,有个好玩的练习就是猜猜哪个是哪个。

I opened them in a full window for you to see and a fun exercise for you as I'm scrolling through some of these websites that were created by the different models is to start guessing which is which.

你也可以在聊天里开始猜,哪个看起来像是KIMI,哪个像Claude,哪个像GPT。

You can also on the chat if you want start guessing which one seems like a KIMI, which one seems like a Claude, which one seems like a GPT.

哪个看起来像Grok?

Which one seems like a Grok?

我特别好奇你们有多少人能猜对。

I'll be very curious how many of you are able to guess it properly.

所以这是其中之一,不是最好的,但话说回来,它们都创建了相当完整的网站。

So that's one, not the greatest, but by the way, they all created quite a complete website.

所以很高兴看到它们生成了文案,还有导航这些。

So that was nice to see that they created copy and as well as the navigation and so on.

所以这是其中一个候选。

So that's one candidate.

这是另一个候选,给这家虚构公司做的培训门户。

That's another candidate for this imaginary company creating a training portal.

我不太喜欢,但也不是最差的。

I don't love it, but it's not the worst.

又一个。

Another one.

这个稍微更有主见一些,你能看出来。

That is slightly more opinionated, as you can see.

我觉得这是我最喜欢的一个,也是因为我审阅了部分文案,不光是文案功能和导航。

I think this is the one that I like the most, also because I reviewed some of the copy and not just the copy functionality and navigation.

这又是另一个做出来的。

That's another one that was created.

这又是一个。

That's another one.

这差不多是最后一个了。

And that's almost the last one.

最后是这个。

And finally, this one.

所以我们有7个候选。

So we have 7 candidates.

如果在座有人能全部猜对这7个模型,那我真要给你点赞。

If there is someone on the line that is able to guess correctly all 7 models, hats off to you.

我猜得差远了,但你看,对于任何视觉化的东西,往往更容易有明确的偏好,比如觉得‘哦,我喜欢这个’之类的。

I was very much off from that, but as you can see, with anything that is visual, it's often much easier to have proper preferences and like something that says, okay, I'm drawn to that and so on.

所以一旦我完成了,你也会完成每个任务的评分,接下来你能做的就是要么只公开我的评分,要么跟AI裁判一起公开。

So once I was done, and you will be done scoring each and every one of the tasks, the next thing that you will be able to do is reveal either my score only or reveal with the AI judge.

所以当你点击‘跟AI裁判一起公开’,所有结果就会发给裁判审核,裁判也会参与评分。

So when you click reveal with the AI judge, all of the results will be sent to the judge for review and the judge will also apply.

然后它会逐步揭示结果。

And then it does the reveal gradually.

所以你看,每个任务都有我的首选,还有几句理由。

So as you can see, per task, these are my top picks and a few words on why.

现在,敲鼓吧,我们来揭晓名字。

And now, drum rolls, let's reveal the names.

对于LinkedIn帖子生成器,我最喜欢的是Kimi、Fable和Opus。

So for the LinkedIn post generator, my favorites were Kimi, Fable, and Opus.

我挺惊讶的。

I was surprised.

我本来以为GPT肯定在里面。

I was sure that GPT will be here.

对于价格异议处理,你可以看到GPT模型在谈判方面是最好的,至少在我看来是这样。

For the pricing pushback, you can see that GPT models are the best in negotiating, at least from my perspective.

在这个练习中,我们用了GPT和Grok,做周计划,结果Grok赢了,挺有意思的。

For the exercise, then we got GPT and Grok, and week planning, Grok won, interestingly.

对于课程参与者中心,是Opus 5.5。

And for the course participant hub, Opus 5.5.

可能有些人对此并不惊讶。

Maybe some of you will not be surprised by some of that.

接下来我得到的是针对每个模型的建议,针对具体任务该怎么处理,是把它添加到模型清单里,还是考虑切换。

The next thing that I get is recommendation for each of these models, what to do about the specific tasks, whether to add it to the model roster or whether to consider switching.

换成用这个模型。

To using this model.

当然了,我们说过,要考虑,对吧?

And of course, as we said, consider, right?

我们需要做额外的判断。

We need to apply additional judgment.

现在来看看最终结果,因为我觉得这也很有意思。

And now take a look at the final results because I think that's also very interesting.

我们列出了不同模型以及它们的平均分、它们被选中的任务数量,还有裁判的平均分。

We have the different models with average and how many tasks they selected, the judge average.

比如,Fable排第一,因为裁判给了它第一。

So for example, Fable got the first because the judge put it first.

但是注意这里。

But pay attention here.

Fable的总成本比其他所有模型都高得多。

The total cost of Fable is so much higher than everything else.

这可能会成为不选Fable的一个考虑因素,尽管理论上它表现更好。

And that might be a consideration why not to go with Fable, even though theoretically it got better.

注意看我的平均分。

Pay attention to my average.

在我的平均分里,Fable和GPT Sol是差不多的,但看看成本差异。

So in my average, Fable and GPT Sol are equivalent, but take a look at the cost differentiation.

这个成本低了一个数量级。

This is a degree of order smaller.

平均用时很接近。

Average time is quite similar.

注意,这也很有意思。

Pay attention, this was also interesting.

Groq的平均用时非常长,尽管成本很可控。

The average time for Groq was extremely long, even though the cost is very manageable.

KIMI也在竞争中,有意思。

KIMI is on the running, interestingly.

这些模型整体得分较低。

And these models are getting lower scores overall.

记住,我们想要更差异化的使用,所以我们可能不只是寻找一个赢家,而是针对每个任务做出更差异化的选择,这样我们可能会有不同的视角。

Remember that we want a more differentiated use, so we might be not just looking for the one winner, but rather per task we want to make a more differentiated choice, and thereby we might have different perspectives.

然后你还有更多数据,比如每个模型在每个任务上花了多长时间,你可以看到KIMI在成本和耗时两方面都胜出。

And then you have some more data, like how long for each task from each of the models, and you can see that KIMI wins on both cost and duration.

你还可以看到——5.1在哪里?

And you can also see that— where is 5.1?

5.1很可能是这一组中整体最贵的。

5.1 is probably the most expensive out of the bunch across the board.

所以成本可能不让你意外,但是一些关于成本和速度的数据,以及我和裁判之间的并列比较。

So perhaps cost will not surprise you, but some data on cost and velocity, some comparisons side by side between me and the judge.

如你所见,我们意见很不一致。

As you can see, we are quite in disagreement.

因此我说,把任何AI当作裁判都要持保留态度。

And thereby I said that take any AI as a judge with a pinch of salt.

我们唯一达成一致的是价格方面的反馈。

The only thing that we agreed upon is the pricing pushback.

嗯。

Yeah.

剩下的,你可以看到,我们有不同的选择。

The rest, as you can see, we have different choices.

这里要指出一点,评委只能看到代码和文字。

One thing to point out here, the judge could only see the code and the text.

所以对于Hub网站的选择是基于代码,而不是基于视觉上的东西。

So the choice with regards to the Hub website was done based on the code and not based on something visual.

因此,如果你想让AI去判断视觉上的东西,仅仅使用API通常不是最好的选择,因为它没有把那个当成图片来利用。

And thereby, if you want an AI to judge something visual, using just the API often is not the best thing because it's not leveraging that as a picture.

它会直接进入代码。

It directly will go into the code.

从今以后,我会根据评委来做选择吗?

Will I make choices based on the judge from here on after?

可能不会。

Probably not.

我也可以去修改一些评委的指令,但我更愿意依赖自己。

I can also go and fix some of the judge instructions, but I would much rather rely on myself.

然后你有一些原始分数,你那里什么都有。

And then you have some raw scores that you have everything there.

那就是应用,那就是过程。

That's the application and that's the process.

只是为了展示最终结果,有趣的名单,其中一些完全出乎我的意料。

Just to showcase the final results, interesting roster, not what I expected at all for some of these things.

我在考虑改变我的一些工作方式吗?

Am I contemplating changing some of the ways I'm working?

老实说,是的。

Honestly, yes.

我日常工作中非常偏向于Claude和用Groq的Cursor。

I'm very biased towards Claude and Cursor with Groq in my day-to-day.

虽然我偶尔也会用GPT,但它不在我的前两名之列,我绝对在考虑更多地使用GPT,让它成为我的前两名之一。

And even though I occasionally am reaching out to GPT, it's not in my top 2, and I'm definitely contemplating reaching out to GPT more as part of my top 2.

如果我必须选择一种工具,基于这些结果,我可能会愿意考虑主要转向GPT,至少在当前时代。

And if I had to choose one tool, I might have, based on these results, been willing to consider moving primarily to GPT, at least in this day and age.

这就是完整的流程,而且正如提到的,你们可以完全自己运行。

So that's the full process, and that, as mentioned, you will be able to run Completely for yourselves.

在我回答问题并从Nathaniel那里听取一些观察之前,还有几点补充。

A few more notes before I take questions and take some observations from Nathaniel.

正如我所说,我创建了一个流程,可以手动运行,也可以利用脚本来运行和我自己做的相同流程。

As I said, I created a process that can either be run manually or leverage the script in order to run the same process that I did for myself.

如果我说没有平台能直接做类似的事情,那我就是疏忽了。

I will be remiss to say that there aren't platforms that do something similar like that out of the box.

这些通常是面向已经在规模化或公司级别做这些事情的团队的平台。

Those are typically platforms that are designed to teams that are already doing that at scale or at the company level.

一些例子包括LangSmith、Braintrust和Langflow。

Some examples include LangSmith and Braintrust and Langflow.

使用,而且它们不是唯一的。

Use, and they are not the only one.

我们还有其他的选择。

We also have other options.

所以,如果你处于需要以更标准化和常规的方式进行这类评估的情况,这些都是值得研究的好选择,这样你的实验、监控和评估将以更系统的方式进行,而不是你必须自己创建这些即用的小脚本。

So if you are in a position that you need to do these types of evaluation in a more standardized and regular way, these are all good options to look into such that your experimentation and monitoring and evaluation will be done in a more systematic way rather than you having to create these out-of-the-box small scripts.

但如果你主要是个人使用,那么我刚才展示的方法就足够了。

But if you are primarily doing that for personal use, then the methods that I've just shown you are more than enough.

正如我所说,你会得到讲义和会议讲义,还有一些其他东西。

As I said, you are getting both the handouts and the session handouts, as well as a few other stuff.

我一会儿给你看看它长什么样。

I'll show you in a minute how it looks.

这是一个不错的地方,你会得到所有材料以及实验本身,非常一步一步来。

It's a nice place that you're getting that has all the materials as well as the lab itself, very step by step.

你还可以下载实际的脚本和基准测试运行器以及其他很多东西。

And you can also download the actual script and the benchmark runner and many other stuff.

而且我们还有资源。

And we also have resources.

所以对你刚才看到的所有东西,都有非常全面的视角。

So very extensive view of everything that you've just seen.

我们一会儿就会分享出来。

We will share it in a minute.

所以把所有这些归到一处,这就是五个步骤。

So to put everything back in one place, these are the 5 steps.

最重要的是任务选择。

Most important is the task selection.

选模型,挑有趣的候选,用其中一种方法跑,做盲测。

Models, choose interesting candidates, run it in one of the methods, do the blind test.

这是非常有趣的体验。

It's a very fun experience.

做一次,哪怕只是为了享受让自己惊讶的感觉,并思考改变一些东西。

Do it once, even just for the enjoyment of surprising yourself and contemplate changing stuff.

如果你做得好,创建了一个好的 benchmark,那么从这之后,每次你要考虑投入一个重大变化时,都可以跑同一个 benchmark。

If you did a good job and you created a good benchmark, from here on, you can run the same benchmark every time there's a significant shift that you're contemplating tapping into.

所以这是一次性的努力,如果做得好,我保证它会让你受益匪浅。

So it's a one-time effort that if you do properly, I can guarantee you that it will serve you well.

所以如果你觉得有用的话,强烈推荐。

So if you find it to be useful, highly recommended.
M1
M141:53

谢谢你,Noufar。

Thank you, Noufar.

这对你来说是怎么演变的?

How is this evolving for you?

比如说,这一套测试在三个月或六个月前和现在相比,会有什么不同?

Like, how would this set of tests be different three or six months ago than they were now?

我特别想到 OpenAI 推出了那种超快模式,Sam Altman 基本上说,直到用上它之前,他都没真正意识到速度对他有多重要。

And I'm thinking specifically about OpenAI introducing this new ultra-fast mode and Sam Altman basically saying that he didn't really appreciate how much speed mattered to him until he had access to that.

当然,那个模式目前只在一个每月五百美元的计划里才有,诸如此类。

Now, obviously, that's — it's only available on a $500 a month plan and things like that.

但你大概会多久更新一次这类标准?还有你觉得随着时间的推移,哪些方面会变得越来越重要?

But how frequently do you imagine updating these sort of criteria? And what are the types of things that you think might become more important over time?
M2
M242:25

我猜,首先这取决于你有多疯狂和有多注意力不集中,因为如果你又疯狂又注意力不集中,可能每个月都得更新。

So I guess, first of all, it depends on how crazy and how ADD you are, because if you are crazy and ADD, probably every month.

但具体到你的问题,当我看到 Ultrafast 的演示时,而今天我要坐下来等有时二十分钟才能让这个破网站生成好,说实话我非常不耐烦。

But specifically to your point, when I saw the demo of the Ultrafast, and today I had to sit down and wait for sometimes 20 minutes to get the website, this poor website generated, I honestly was very impatient.

我意识到我现在对速度的关注度比以前看到 Ultrafast 演示之前高多了。

And I realized that I'm paying way more attention to speed versus how I did before I've seen the Ultrafast demo.

有些考量是实时更新的。

Some of the considerations are being updated on the fly.

有些使用场景——这正是为什么我想要一个愿望清单——有些使用场景我现在会去追求或者视为前沿用例,以前我做梦都想不到能以那种质量一次就搞定。

Some of the use cases, and that's exactly why I wanted to have the wishlist, some of the use cases that I would now pursue or consider frontier use cases, I would never have dreamt to do at that quality and in one shot previously.

所以门槛越来越高了。

So the bar is getting higher.

我不知道主题本身会不会变,但我觉得那些让你改变主意或者定义什么算好的考量因素,会随着我们对技术进展的了解而频繁变化。

I don't know that the subject matter changes, but I think that the considerations of what makes you change your mind or what good looks like, those change quite frequently based on how far we know that the technology takes us.
M1
M143:21

太棒了。

Awesome.

我就不继续占用时间了,让其他人也问问题。

I'll get out of the way and let some other people ask questions too.
M2
M243:24

没问题。

Sure.
M3
M343:25

我这里给你准备了一些问题,Noufar。

I've got some stacked here for you, Nafar.

你说话的时候我一直在记,免得漏掉。

I've been cataloging while you've been talking so that I don't miss.
M2
M243:29

谢谢你这么做。

Thank you for that.
M3
M343:30

好,没问题。

Yeah, no problem.

所以有一个提到的问题是关于“努力程度设置”的。

So one of the questions that came up was effort settings.

当你在做这个测试、使用 OpenRouter 的时候,你会不会把所有模型都设成类似的努力程度设置?

Whenever you're doing this in terms of whenever you're using OpenRouter, are you setting everybody to a similar effort setting whenever you're testing this?
M2
M243:41

正确的回答是理想情况下会,不过有些模型之间的努力程度设置不一定完全对等。

So the correct response is ideally yes, even though effort settings are not necessarily equivalent in some of the models.

所以如果你想更严谨一点,就尽量拿同级别的东西对比,或者直接用 OpenRouter 的默认设置。

So if you want to be extra diligent, try to compare apples to apples or just use the defaults of OpenRouter.

这也是一种偷懒但可能效果不错的办法,不过我也承认努力程度设置可能会造成差异。

That's also a lazy approach that might work well, but I agree that the effort setting might create a difference.

所以如果你想更严谨,可以随时用不同的努力程度设置跑几个模型,看看结果会不会变,因为努力程度设置的偏好,也是我希望你能在建立个人模型清单或模型配置清单时逐步摸索的一部分。

So if you want to be extra diligent, you can always run several models with different effort settings and see if it changes the picture because effort setting preference is also part of what I want you to evolve as part of your personal model roster or model configuration roster.
M3
M344:13

很好。

Great.

然后关于任务,我通常会把它们分成三大类:比如内容运营。

And then in terms of whenever I think of tasks, I usually try to put them into 3 buckets: something like content operations.

还有策略。

And strategy.

你在处理这三类任务时,有没有发现什么不同?还是说它们大体上都走同样的流程?

Do you find that there's any differences whenever you're doing those 3 different types of tasks, or are they generally going through all the same process?
M2
M244:29

有些任务需要更复杂的配置,比如要嵌入你自己的 harness、带上你自己的上下文,还有连接器等等。

So some tasks will require more elaborate setup, like to work as part of your own harness with your own context and so on and connectors.

所以有几个选择:如果你确实要把一个和整体配置高度相关的任务放进 benchmark,那它就必须直接从 harness、从实际的配置环境里运行,而不是用一个只能看到简化视角的第三方 API key。

So several options, either if you do put as part of your benchmark, a task that is highly involved with the overall setup, then that needs to run directly from the harness, from the actual setup and not using a third-party API key that will only get parsed view.

如果必须这样,那也是一种办法,就是直接通过实际工具来跑,同时确保你做了足够的匿名化处理,别自己骗自己。

If you have to, that's one way to do that, to just run it through the actual tool and make sure that you anonymize well enough without cheating yourself.

或者说,如果你能为这个任务创建一个代理,不需要在实际配置里跑那些花哨的东西,那你可以尝试用这个方法去跑。

Alternatively, if you can create a proxy for this task that doesn't require all of the bells and whistles of running that in the actual configuration, then you can try and run that in a— using this method.

但对于运营类任务,我猜它受整体配置的影响最大,这类任务要特别小心,不能脱离实际配置去判断,因为它高度依赖你整个系统。

But for operations, I would guess it's going to be most influenced by the overall setup, and that's the one to be a little bit more careful to judge without the setup because it's highly dependent on the entire system that you have.
M3
M345:21

太棒了。

Awesome.

然后有人问到开源模型,问我们在测试的时候有没有天然注意到什么,显然你那边有 KIMI。

So there were some questions about open source models and if we notice inherently anytime whenever we're doing our testing, obviously you had KIMI up there.

你注意到什么了吗?

Are you noticing anything?

还有一些关于 Mistral 的问题。

So there were some questions about Mistral.

你有没有发现很大的差异?还是说你通常很乐意把开源模型和前沿模型都测一遍,然后直接做比较?

Do you notice any big differences or do you find that you're generally happy to test through the open source and the frontier models and just kind of compare?
M2
M245:41

这又回到了最终标准的问题,因为如果你出于安全或公司政策原因,特别偏向开源,或者反过来特别偏向闭源,那你本身就带了偏见。

So it goes back to the final criteria because if you are very biased towards open source or the other way around for security or company policy reasons, you are biased again.

也就是说你可能会完全不纳入开源,或者主要纳入开源。

That either don't include or primarily include open source.

如果成本是个关键因素,那你可能会更倾向于那些便宜、同时结果也够好的模型。

If cost is a critical consideration, perhaps you will be leaning much more heavier towards models that are cheap while giving you good enough results.

而且任务本身也影响很大,因为有些任务只是门槛级别的要求,对吧?

And also the task itself has a huge implication because some of the tasks are table stakes, right?

比如写邮件草稿,或者基础研究,这些都属于门槛级任务。

For example, email drafting or basic research, those are table stakes.

对这些任务,你可能会发现任何像样的模型——比如 KIMI 及以上,或者 Mistral、GLM 及以上——你都看不出明显差别,这时候开源模型就已经足够好了。

And for them, you will probably see that any decent model, let's call it KIMI and above or Mistral and GLM and above, you'll not be able to see any noticeable difference, and then open source will be as good as it gets.

但对于更进阶或更复杂的任务,有时候你仍然会看到商业模型和开源模型之间的差距,不过这又回到了我一开始提出的那个观点。

For the more aggressive things or more sophisticated stuff, sometimes you will still see the gap between commercial and open source, but it goes back also to the objection that I raised at the beginning.

如果你大部分用例都追求“够用就好”,那么到目前为止,开源模型在大多数知识型工作任务上已经够用了,尤其是当它运行在一个很好的 harness 里。

If you are primarily working with good enough for most of your use cases, then as of now, open source is good enough for most knowledge work tasks, especially if the harness in which it's working is a good one.

很好。

Great.
M3
M346:43

另一个问题是关于如何给出指令和提示词。

And another question that came up was regarding trying to figure out how to give instructions and prompts.

这让我想起上次我们做的关于循环的线上研讨会,当时谈到什么算是好的内容,比如“如果不符合就拒绝”,因为有一些评论提到,在构建时人们注意到你有一些非常具体的指令。

And I'm reminded of our last webinar that we did about loops in terms of what does good like, reject if, because there were some comments just in terms of whenever you were building and people noticing that you had some very specific instructions.

在怎么做这方面,你有什么建议吗?

Do you have any pointers in terms of how to do that?

或者我们有没有什么资源可以帮助大家做到这一点?

Or perhaps we have resources that can help people do that?
M2
M247:04

所以提示词本身能帮你理清任务,而且好的提示词能起到很大的作用,尤其是在一个智能模型、一个了解你的工具里。

So the prompt itself that will help you figure out the tasks and the prompt will go a long way to do that and do that in a smart model, in a tool that knows you.

一般来说,它需要是可衡量的。

In general, it needs to be measurable.

我们主要讨论的是智能体工具。

We're talking primarily about agentic tools.

智能体工具的目标。

Agentic tools goals.

“写一篇好的博客文章”是不可衡量的。

Saying create a good blog post is not measurable.

“写一篇能让高管和普通员工都感兴趣的博客文章”就好一些;而“写一篇不包含以下内容、字数限制在多少、并包含以下部分的博客文章”就更好了,因为它高度可衡量且高度可实现。

Saying create a blog post that will be engaging for executives as well as ICs is better, and saying create a blog post that does not include the following and is limited by this number of words and has the following parts is even better because it's highly measurable and highly attainable.
M3
M347:38

太好了。

Great.

另外,有一个关于成本的问题,我觉得 OpenWriter 能做出这些改变很好。

And in terms of, there was one question about, I think, cost, and it's great that OpenWriter makes those changes.

在考虑切换时,关于成本方面你还有什么其他想法吗?

Do you have anything else to think about whenever you're considering switching as it comes to costs?

即使有了 OpenRadar,你能看到每个任务的成本,这很好,但关于切换的成本你还有其他想法吗?

Even having, because it's nice with OpenRadar, you're getting that cost per task, but do you have any other thoughts around the cost of switching?
M2
M247:58

是的。

Yeah.

要记住,如果你有一个慷慨的订阅方案,比如 Pro 版或 Teams 版,有些方案实际上是在隐藏或补贴更高的实际 token 成本。

So remember that if you have one of the generous subscriptions, like the Pro tier or the Teams tier, some of them are basically hiding or subsidizing a much higher actual token cost.

所以仅仅评估成本本身并不一定是全貌。

So just evaluating cost per se is not necessarily the full picture.

因为,比如你每月花200美元用 Claude 或 OpenAI,你可能从 Astra 或 Fable 那里得到了超出你应得份额的所需资源,那么成本对你来说可能就不像那些按实际消耗付费的人那么重要。

Because if, for example, with your $200 a month Claude or OpenAI, you're getting more than your fair share of what you need from Astra or Fable, then perhaps the cost does not matter to you as much as to someone who actually pays per consumption.

所以这有点取决于你的订阅方案,而且我认为一个不错的预测是,未来可能不会一直这么慷慨,到那时我们都会关心具体成本,而不仅仅是大致范围。

So it's a little bit conditional to your subscription, and I think a good prediction to make is that it might not stay as generous in the future, and then we all will care about the specific cost and not just the ballpark.

但如果你即便用了慷慨的订阅方案,也经常遇到速率限制,那么成本对你就很重要,你就应该更侧重于优化成本。

But if you are a person that also with the generous subscription hit the rate limit very frequently, then cost matters to you, and then you should optimize more for cost.

所以是视情况而定的。

So conditional.
M3
M348:47

太棒了。

Amazing.

NLW,这就是我能够大致整理的所有问题了。

NLW, that was all the questions that I was able to kind of catalog.

我不知道你还有没有其他问题。

I don't know if you have any others.
M1
M148:52

没有了,我觉得这非常有帮助。

No, this— I think this is super helpful.

我认为这是那种比我们常见的操作指南类主题更需要个人深思的领域,对吧?

I think this is one of those areas where I think even more than our normal how-to type sessions, this one requires definitionally a lot of personal consideration, right?

比如,如果其中有你觉得没用的部分,全部的意义就在于让它变得有用。

Like, this is— if there are pieces of this that you don't find useful, the whole point is to make it useful.

所以拒绝它,试试别的。

So reject, try something else.

然后再分享回来,这样其他人也能了解情况。

And then share it back so other people get a sense of where things are.

我超级凭感觉,因为我什么都用,而且我经常用多个不同的模型跑同一个东西。

I'm super feel-based because I'm using absolutely everything, and I'll run the same thing through multiple different models all the time.

所以我的做法会和别人不太一样。

So I'm— my approach to this is going to be different than some other folks.

不过Nufar,还是要谢谢你一如既往地把这个搞出来,希望大家玩得开心。

But, Nufar, thanks as always for getting this out there, and I hope everyone had a good time.
M3
M349:33

嗯。

Yep.
M2
M249:33

等我们定下来下一期内容后,记得来参加哦。

And join us for the next one once we decide what it is.
M3
M349:36

干杯。

Cheers.
M2
M249:37

拜拜。

Bye.
已剔除 5 处广告(点击展开查看)
M1
M11:05广告 · 已剔除

First of all, thank you to today's sponsors, KPMG, Robots and Pencils, Harbor, and Blitzy.

To get an ad-free version of the show, go to patreon.com/AIDailyBrief, or you can subscribe on Apple Podcasts.

And to learn more about sponsoring the show, send us a note at [email protected].

M1
M122:29广告 · 已剔除

Learn more about what separates AI Amplifiers from everyone else at kpmg.com/us/aiamplifiers.

M1
M123:04广告 · 已剔除

Head to robotsandpencils.com.

M1
M123:22广告 · 已剔除

Harbor Capital Advisors' AI Lab Ecosystem ETF suite gives investors a way to gain exposure to the AI ecosystem they believe is best positioned for success.

Search Harbor AI Lab Ecosystem ETFs wherever you invest, or follow @HarborCapital on X to learn more.

Visit harborcapital.com for a prospectus containing investment objectives, risks, fees, expenses, and other important information.

Read and consider it carefully before investing.

Risks include principal loss and artificial intelligence-related risks.

Harbor ETFs are distributed by Foresight Fund Services, LLC.

Harbor is not affiliated with AI Daily Brief, and the funds are not affiliated with, sponsored by, or endorsed by any AI lab.

This is a paid advertisement and not personalized investment advice.

Investing involves risk, including possible loss of principal.

Here's why most legacy modernization projects fail.

The AI doing the work can't understand codebases at scale.

It sees a small slice of context, examines syntax, and misses years of decisions distributed across the global application ecosystem.

Blitzy solves this the way it solves everything.

Grounded in your code before any migration begins, Blitzy's agents reverse engineer the entire legacy system into a persistent knowledge graph.

Every dependency, every constraint, every piece of tribal knowledge that used to live in one engineer's head.

From that understanding, Blitzy autonomously executes language migrations, framework upgrades, and monolith-to-microservices transformations, all validated end-to-end.

One Blitzy customer modernized a $10 million monolithic insurance stack in 16 weeks against a 137-week baseline with coding agents.

That's 9x compression.

Retire technical debt while accelerating your roadmap.

See how at blitzy.com.

That's B-L-I-T-Z-Y dot com.

M2
M241:37广告 · 已剔除

If you also want to go deeper into anything that you've seen or to become a master in building, we do have our two existing courses for Executive Agent Leadership and Executive Catch-Up courses that are starting over the next two weeks.

So we'll be happy to have you join us.