← 返回任务列表

How to Get the Most Out of Fable 5 and GPT-5.6 Sol

273 段 · 1 位说话人 · 原片 26:22
M1
M10:00

今天在AI每日简报中,我们来聊聊如何最大化利用前沿模型。

Today on the AI Daily Brief, how to get the most out of frontier models.

AI每日简报是一个每日更新的播客和视频节目,专注于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,在深入之前先快速说几个通知。

Alright friends, quick announcements before we dive in.
M1
M10:32

另外,关于今天这期节目有个小提示。

And a quick note about today's episode.

这期是提前录制的。

This was recorded in advance.

实际上,我是在周四下午录的,当时大家都在为Gemini 3抓狂。

In fact, I am recording it on Thursday afternoon as everyone freaks out about Gemini 3.

所以如果Dario和Sam——你懂的——因为那个让纳斯达克市值暴跌的威胁而失去理智,发布了新版Fable和GPT,那你就知道我为什么现在不谈这个了。

So in the meantime, you know, if Dario and Sam have lost their minds and released new versions of Fable and GPT in response to the threat that's tearing value off the Nasdaq, you'll know why I am not talking about it right now.

周一我稍微出趟差。

Just a quick little bit of travel on Monday.

周二我会回来,发一期常规节目。

I will be back on Tuesday with a normal episode.

不过,不管模型界的风云如何变幻,今天的节目重点还是关于如何最大化利用最先进、最新的模型。

But still, regardless of what is going on in the wide world of models out there, today's episode is all about how to get the most out of the most advanced and newest models.

好,废话不多说,我们直接开始。

So without any further ado, let's dive in.

现在,Fable 5和GPT-5.6这批新模型已经推出几个星期了。

It has now been a couple of weeks with this new class of models in Fable 5 and GPT-5.6.

奇怪的是,这些模型实际存在的时间其实比几个星期还要长一些。

Now weirdly, these models have actually been around a little longer than a couple of weeks.

GPT-5.6的早期访问期特别长,而Fable 5则来了几天就又没了。

There was a particularly long early access period for GPT-5.6, and Fable 5 was here for a couple days before going away.

但现在,基本上所有人都已经用了一段时间这些模型了。

But at this point now, pretty much everyone has now had these models for some time.

实际上,我们对这些模型的访问权限还在不断延长和重置。

In fact, our access to them keeps getting extended and reset.

与此同时,人们也开始分享他们最大化利用这些模型的技巧和诀窍。

And along with that, people have started to publish their tips and tricks for getting the most out of them.

新模型总是需要新的交互方式来充分发挥它们的潜力。

Now, it is always the case that new models demand new ways of interacting with those models to get the most out of them.

但这类信息恰恰是基准测试无法体现的,只能通过反复试错来积累经验。

But this is exactly the sort of information that can't really be captured in anything like benchmarks and just has to go be experienced through trial and error.

你会看到,在5.6 SOL和Fable 5之间有很多共同点,我认为这不仅指出了充分利用这些模型的新方法,还揭示了一些未来会越来越常见的交互模式。

As you will see, there are a number of common threads that cut across both 5.6 SOL and Fable 5 that suggest, I think, not just some new ways to get the most out of these models, but for some new patterns of interaction that are going to become increasingly common from here on out.

我们先从一些官方来源和评论说起。

Now we're going to start with some official sources and commentary.

Codex团队成员Eric Provencher写道,很多人用5.6 SOL时,提示词和当年用5.5一模一样。

Codex team member Eric Provencher wrote, with 5.6 SOL, a lot of people are still prompting the model exactly as they did 5.5.

需要指出的是,5.6 SOL比之前的模型更坚韧、更彻底。

It's important to note that 5.6 SOL is a lot more tenacious and thorough than previous models.

Eric在官方的learn.chatgpt.com网站上发布了一份提示词指南,虽然这份指南并非专门针对如何从5.6而不是5.5中获得更多价值,但其中几点特别值得一提。

Eric published a prompting guide on the official learn.chatgpt.com site, and while in this case it is not presented as a way to get more out of 5.6 as opposed to 5.5, there are a few things that are specifically worth noting.

其中一条建议就源于这种坚韧性,那就是设定边界。

One piece of guidance that comes from that tenacity is around setting boundaries.

指南写道,边界是ChatGPT需要的一些指令,以避免产生额外工作或执行你不希望的操作。

Boundaries, writes the guide, are the few instructions ChatGPT needs to avoid creating extra work or taking an action you didn't intend.

当修改错误的细节会导致结果不可用时,或者当你希望在影响他人之前先审核某些内容时,就加上边界。

Add one when changing the wrong detail would make the result unusable, or when you want to review something before it affects other people.

例如:保持已批准的日期和预算数字不变。

Examples: keep the approved dates and budget figures unchanged.

仅使用提供的来源。

Use only the supplied sources.

如果信息缺失,则标记出来,不要凭空猜测。

Flag missing information instead of guessing.

把建议控制在给定的预算范围内。

Keep recommendations within the stated budget.

把消息准备成草稿。

Prepare the message as a draft.

别发出去。

Don't send it.

你可以看到,在每种情况下,一个更执着的模型——用Eric的话说——可能会比用户觉得舒服的程度多发挥一点主动性,然后真的去做一些事情,结果对提示者想达成的目标来说并不理想。

You can see how in each of these cases, a more tenacious model, to use Eric's word, might assert a little bit more agency than the user would be comfortable with and actually go off and do something that ended up not being optimal for whatever the prompter was trying to achieve.

这些边界的一个例子是明确说明你希望它停在哪里,以及你不希望它采取哪些行动,比如不要发消息,只是准备成草稿。

One example of these boundaries was being explicit about where you wanted it to stop and actions you didn't want it to take, i.e., don't send the message, just prepare it as a draft.

另一个关于批准日期和预算数字的例子是限制和聚焦执着模型应用注意力的范围。

Another example around the approved dates and budget figures was to limit and focus where the tenacious model was applying its attention.

至于只使用提供来源的指令,一个能访问整个互联网的执着模型,如果没被告知不要这么做,可能会跑偏,陷入一些严重的兔子洞。

When it comes to the injunction to use only the supplied sources, a tenacious model that has access to the entire internet could go off and get lost in some serious rabbit holes if not told not to do that.

现在,显然设定边界不是什么新鲜事,但关键在于,模型越强大,这些边界就越重要,而且如果没设定边界,可能带来的现实后果就越严重。

Now, obviously setting boundaries is nothing new, but the point here is that the more powerful the model, the more significant those boundaries become, and the more real-world consequences there can be if those boundaries aren't set.

在后果光谱的低端,就是消耗比你实际需要的多得多的token,在这个越来越注重成本的时代,这也不是小事。

On the lower end of the spectrum of consequence, there's just burning through way more tokens than you actually needed to, which in an increasingly cost-conscious era isn't nothing.

但当然还有更大的后果,比如把还没批准的消息发给重要的客户、合作伙伴或同事。

But then of course there's much bigger consequences, like sending a message that hasn't been approved yet to a critical customer, partner, or colleague.

Eric的另一个提示技巧——这也不是什么新东西,但肯定符合5.6 SOL这种模型的特点,即你实际上要跟它互动,而不是只交给它一个任务然后让它跑去做自己的事——就是通过后续消息与模型迭代来改进结果。

Another prompting tip from Eric, which again isn't different, but certainly comports with the idea of 5.6 SOL being a model where you actually want to interact with it as opposed to just give it a task and let it go off and do its thing, is to iterate with the model to improve the results with follow-up messages.

一旦你看到第一版,你就可以跟进一些东西,比如让开头更直接,保留证据,把这里或那里调整一下位置。

Once you see the first version, you can follow up with things like keep the opening more direct, keep the evidence, move this or that part around.

有趣的是,迭代越来越不会完全以回合制的方式发生,即你必须停下来等一个输出,然后才能引导模型下一步去哪里。

And interestingly, increasingly iteration isn't going to happen in a completely turn-based way where you have to stop and wait for an output before you can steer where the model goes next.

随着ChatGPT和Codex整合在一起,一些在Codex中常见的引导行为也可以进入主应用体验。

As ChatGPT and Codex come together, some of the sort of steering behaviors that are normal in Codex can come to that main app experience as well.

指南里写道,当Codex已经在工作时,你可以在当前运行还没完成时再发一条消息。

The guide writes, when Codex is already working, you can send another message without waiting for the current run to finish.

Steer会把消息添加到当前运行中。

Steer adds the message to the current run.

用它来改变方向、添加缺失的细节或分享新的信息。

Use it to change direction, add a missing detail, or share new information.

Queue会保存消息用于下一次运行。

Queue saves the message for the next run.

把它当作一个要等当前工作完成后再进行的后续操作。

Use it as a follow-up that should wait until the current work finishes.

这种行为减少了与AI协作的延迟,随着模型变得更强大,这种延迟的重要性会越来越高。

This sort of behavior reduces the latency of collaborating with AI in ways that become important, the more powerful the models get.

现在,最近Codex和ChatGPT的整合,以及ChatGPT拆分为chat和work,带来的一个结果是不同体验类型需要不同的提示技巧。

Now, one of the things that follows from the recent integration of Codex and ChatGPT and the split of ChatGPT into chat and work is different prompting tips for different types of experiences.

在这篇帖子里,Eric和ChatGPT团队明确区分了针对chat而不是work的提示最佳实践和示例。

In this post, Eric and the team at ChatGPT explicitly separate prompting best practices and examples for chat as opposed to for work.

很多关于work的提示都带有成本和效率意识。

A lot of the tips around work have this element of cost and efficiency consciousness embedded within them.

有一节叫“高效使用Work”,里面说,work适合处理耗时或重复的任务,或者适合处理可以重复使用的完整文件。

One section is called Use Work Efficiently and says, work is useful for time-consuming or recurring tasks or for finished files you can reuse.

一个消耗更多credits的任务仍然是值得的,如果它能节省时间、提高质量或帮助你做出重要决定的话。

A task that uses more credits can still be worthwhile if it saves time, improves quality, or helps you make an important decision.

而且他们的很多建议还是关于管理Eric提到的那个执着特性的。

And a lot of their recommendations are again about managing that tenacity that Eric was talking about.

他们建议从一个你可以审查的结果开始,并在任务开始做你不再需要的工作时,缩小范围或停止任务。

They suggest starting with only one result that you can review and doing things like narrowing or stopping the task if it starts doing work you no longer need.

当然这里还有很多内容,但我最后特别想提的一个建议是使用语音听写——如果你常待在这个领域,你会经常听到这个建议。

Now there is of course a lot more in here, but the last one that I wanted to note specifically was the suggestion, which you'll hear a lot if you stick around these parts, to use voice dictation.

ChatGPT的一个很棒的地方是,即使你电脑上没装像WhisperFlow这样的工具,ChatGPT的听写功能也绝对是同类中最好的。

One thing that's great about ChatGPT is that even if you don't have something like WhisperFlow set up on your computer, ChatGPT's dictation is absolutely best in class.

所以你可以使用应用内自带的原生听写工具。

And so you can use the native dictation tool that's embedded right within the app.

我之前说过,在AI的背景下,语音听写之所以这么有价值,原因之一就是——上下文。

As I've said before, one of the reasons that voice dictation is so valuable in the context of AI is, well, context.

在这个AI需要更多信息才能做好工作的世界里,啰嗦的人将继承地球。

In a world where AI needs more information to do its job well, the ramblers shall inherit the earth.

对着电脑说上几分钟,哪怕完全是毫无结构的意识流,往往也比一份超精确、表达清晰但未必包含所有上下文的打字笔记,更能给AI提供它需要的有效信息。

And talking at your computer for a bunch of minutes, even if it's a totally unstructured stream of consciousness, is often going to be more effective in giving the AI what it needs to do well than a hyper-precise and clearly articulated set of typed notes that don't necessarily include all of that context.

现在,这不是5.6发布时唯一的新模型指南。

Now, this is not the only new model guide that came out around 5.6.

在OpenAI开发者网站上,他们也为GPT-5.6列出了一套最佳实践。

Over on the OpenAI developer site, they also have a set of best practices for GPT-5.6.

酷的是,你实际上可以点击回看,看到之前模型的最佳实践,一路追溯到GPT-4.1。

And what's cool is you can actually click back and see best practices for previous models as well, going all the way back to GPT-4.1.

AI内容创作者Ollie Lehmann觉得这份文档特别有价值,并总结了他发现的一些最佳技巧。

AI content creator Ollie Lehmann found the document really valuable and summed up some of the best tricks he found.

一,他写道,从你的旧提示中删除指令。

One, he writes, delete instructions from your old prompts.

OpenAI的规则是,每条指令只陈述一次。

OpenAI's rule is to state each instruction exactly once.

他们发现,移除重复指令能让分数提高10%到15%,同时减少最高达66%的Token消耗。

They found that removing repeated instructions raised scores by 10 to 15% while cutting tokens by up to 66%.

你为旧模型写的那些冗长规则列表,现在反而让GPT-5.6回答更差,还花你更多钱。

The giant rule lists you wrote for older models now make GPT-5.6's answers worse and cost you more.

二,让计算量匹配任务。

2, Match the compute to the job.

现在有两个独立的调节钮:一是模型大小,比如Sol用于最难的难题,Terra用于日常业务工作,Luna用于便宜快速的任务;二是它思考的难度。

There are 2 separate dials now, which model size, i.e., Sol for the hardest problems, Terra for everyday business work, and Luna for cheap fast tasks, and how hard it thinks.

六个努力级别,从无到最大。

6 Effort levels from none to max.

OpenAI关于思考旋钮的建议是:从你在上一个模型上用的设置开始,然后测试低一级。

OpenAI's advice on the thinking dial: start at whatever setting you used on the last model, then test one level lower.

新一代通常需要更少。

The new generation usually needs less.

把最大级别留给你真正最难的难题。

Save max for your genuinely hardest problems.

这是那种听起来极其简单,但做起来几乎情感上都很难的建议。

This is one of those pieces of advice that sounds incredibly simple, but is almost emotionally hard to do.

对我们很多人来说,面对几乎每个问题,诱惑就是把设置调到最大,因为如果其他条件都一样,你难道不希望在每一个问题上都获得最多的智能吗?

For so many of us, the temptation for just about every problem is to dial up the settings all the way to max because all things held equal, wouldn't you want the most intelligence on every single problem?

即使不考虑成本,也越来越明显,那不一定是最优行为。

Even outside of costs, it is increasingly clear that that is not necessarily the optimal behavior.

所以,这正是OpenAI试图给出一些离散指导的原因,告诉你应该怎么做。

And so that's why OpenAI is trying to give some discrete guidance around what you should do instead.

另一个关于撤销先前指令的技巧是,检查你之前给出的关于简洁性的规则。

Another tip in the area of undoing previous instructions is to check older rules you gave about brevity.

Ollie写道,GPT-5.6默认的回答比5.5更短,所以你为旧模型添加的简洁规则现在可能剪掉太多了。

Ollie writes, GPT-5.6 defaults to shorter answers than 5.5, so brevity rules you added for older models can now cut too much.

当你确实想要短答案时,告诉它保留哪些信息,可以省略哪些细节,而不是笼统地说“保持简短”。

When you do want something short, tell it which information to keep and which detail it can drop instead of a blanket keep it brief.

关于语气,具体的指令胜过抽象的指令。

On tone, concrete instructions beat abstract instructions.

例如,OpenAI建议“友好”和“有同理心”这类词太抽象了,相反,Ollie写道,要明确写出你想要的写作行为——要有多直接,开头怎么写,跳过什么。

For example, OpenAI suggests that terms like friendly and empathetic are going to be too abstract, and instead Ollie writes, spell out the actual writing behavior you want— how direct to be, what to open with, what to skip.

比如像这样:“在第一行说出客户的问题,给出编号步骤的解决方案,跳过道歉段落”,这样每次能得到相同的语气。

Something like, name the customer's problem in your first line, give the fix as numbered steps, skip the apology paragraph, gets you the same tone every single time.

话虽如此,如果这些技巧都非常清晰、实用,是你马上就能用的东西,那还有另一个我开始在许多讨论中看到的主题,那就是我们对待这些最先进模型的野心水平。

And yet, if these tips are all very clear and practical, and stuff that you can go use right away, there is another theme that I'm starting to see across a lot of the discourse, which is about the level of ambition we're bringing to these most advanced models.

Intuit的AI UX产品经理Christine Zhu写了一篇帖子,名叫《你对Claude还不够有野心》。

Christine Zhu, an AI UX PM at Intuit, wrote a post called You're Not Ambitious Enough with Claude.

Christine写道,我日常工作中最大的生产力和能力提升,发生在我超越自动化琐碎工作,开始让Claude去做更高杠杆的工作时。

Christine writes, the biggest productivity and capacity unlock in my daily work happened when I went beyond automating busywork to asking Claude to do more high-leverage work.

几个月来,她写道,我一直用Claude来清理多巴胺积压——那些我因渴望获得即时快感和虚假进展感而过于急切完成的小任务。

For months, she wrote, I was using Claude to clear the dopamine backlog, the queue of little tasks that I respond to too eagerly because completing gave a dopamine hit, an illusion of progress.

这些自动化帮我省下了很多时间。

These automations freed up a lot of my time.

有了这些新腾出的时间和Fable,我开始用它做更高杠杆的工作。

With that newfound time and Fable, I started pushing it for higher-leverage work.

那些我之前不放心交给Claude的硬任务。

Hard tasks I didn't trust Claude with before.

这让我的生产力、工作容量和心流状态都有了最大的飞跃。

This made the biggest leap in my productivity, capacity, and sense of flow.

Christine引用了Stripe产品负责人Shreyas Doshi的一个概念,他写道:产品工作有三个层级——第一层是执行层,第二层是影响层,第三层是观感层。

Christine calls upon a concept from Stripe product leader Shreyas Doshi, who wrote, there are 3 levels to product work: 1, the execution level; 2, the impact level; and 3, the optics level.

Christine指出,引用原话:“很多人用Claude来自动化琐事、观感工作和执行工作,而把影响工作留给自己。”

Christine notes that, quote, many of us use Claude to automate busywork and optics and execution and leave the impact work to ourselves.

我们告诉自己,判断力才是人类不可被替代的部分。

We tell ourselves the story that judgment is the defensible human thing.

但一旦给对了上下文,Claude做苦活累活可以比你更快更好。

But when given the right context, Claude can do the hard work better and faster than you.

降低开始的启动能量本身就是一个重大突破。

Lowering the activation energy of starting is a major unlock in itself.

实际上,Christine建议针对这三类工作分别用不同的方式来使用Fable 5。

In fact, Christine suggests using Fable 5 in different ways for each of these 3 categories of work.

比如观感工作,她说,就是要在合适的场合向合适的人展示团队的进展。

Optics work, for example, she says, is about making your team's progress visible to the right people in the right forums.

对此,她建议用Claude搭配Fable的自动驾驶模式来替代生产工作。

For this, she suggests using Claude with Fable as autopilot to replace production.

她举例说:我在Cowork上设置了一个定时任务,每周运行两次。

The example she gives: I have a scheduled task on Cowork that runs twice a week.

它会从关键来源拉取上下文,并将每个工作流的最新更新写入一个Google Sheet模板。

It pulls context from key sources and writes the latest updates for each work stream into a Google Sheet template.

这个Google Sheet状态板对团队可见,他们可以自己找到答案,不用再来问我。

This G Sheet status board is visible for the teams to get their questions answered without asking me.

我还有一个技能,能把这些更新按不同层级打包,呈现给合适的受众和场合。

I also have a skill that packages these updates at different levels for the right audience and forums.

她继续说:观感工作是应该被无情自动化的那一层,因为它能清空多巴胺积压,让你保持心流状态。

Optics work, she continues, is the layer to automate ruthlessly because it takes care of the dopamine backlog so you can stay in flow.

事实上,她表示:如果这些流程很快被产品化集成到Claude本身,我一点也不会惊讶。

In fact, she suggests, I wouldn't be surprised if these flows are productized soon in Claude itself.

在下一层,对于执行工作,她建议不要把Claude当自动化工具,而是当副驾驶。

At the next level, for execution work, instead of Claude as Automator, she suggests Claude as Copilot.

Christine接着说:每周一,我会运行一个上下文倾倒技能,它读取我过去一周在Slack、日历、笔记和代码仓库中的活动,给我一份效率分析和需要关注的模式,推荐应该优先做什么,还会给出具体的工作建议帮我启动。

Every Monday, Christine continues, I run the context dump skill, which reads my past week across Slack, calendar, notes, and repo activity, gives me an analysis of my efficiency and patterns to watch, recommends what I should prioritize, and offers specific work to help me get started on.

这读起来不像一份总结,更像一个教练,他注意到我的个人习惯,能给出真正有帮助的观察和建议。

It reads less like a summary and more like a coach who notices my personal patterns and actually makes helpful observations and recommendations.

我还会运行一个规划技能,它会检查Jira状态和路线图,告诉我团队目前跟踪的进展以及下一步该安排什么。

I also run a planning skill that checks Jira status and the roadmap and shows me how the team is tracking and what to line up next.

为了贴近客户,我每月运行一个定时任务,分析我们的客户支持渠道,提炼出最热门的主题、趋势,并推荐我可以优先解决的最重要的请求或问题。

For staying close to the customer, I run a scheduled task monthly that analyzes our customer support channel to give me the top themes, the trends, and recommends the top requests or problems I can prioritize.

总的来说,她说,要对你的需求更有野心。

Overall, she says, get more ambitious with the asks.

别把Claude只当作提炼摘要的工具。

Don't treat Claude as just for distilling a summary.

让它帮你做判断。

Ask for judgment.

最后,对于影响工作,她建议把Claude当作一个练习伙伴,帮你启动最困难的任务。

Lastly, for impact work, she suggests using Claude as sparring partner to get the hardest work started.

她说,这一层是最高杠杆的工作,但也是最难启动的,比如思考一个新产品的赌注、为高风险的演讲准备叙事、用业务单元战略来验证策略等等。

This level, she says, is the high-leverage work that is the hardest to start, e.g., thinking through a new product bet, preparing narrative for a high-stakes presentation, testing strategy against business unit strategies, etc.

针对这一点,她建议值得花时间让Claude上手,并为它设置好正确的上下文。

Now around this, she suggests it's worthwhile to onboard Claude and set it up with the right context.

这是我们反复讨论过的一个话题,创建一个个人上下文档案,包含任何AI或模型真正成为有用协作者所需的全部信息。

This is something we've talked about a lot, creating a personal context portfolio that includes all the information that any given AI or model needs to be able to actually be a useful collaborator.

总体来看,最让我印象深刻的是,Christine似乎第一次真正感受到Fable为她解锁了这类高影响力工作。

And overall, the biggest thing is that it feels to me like Christine is really feeling like Fable unlocks this sort of impact work for her for the first time.

她写道,Fable在影响力工作方面明显优于之前的模型。

She writes, Fable is noticeably better than previous models for impact work.

我特别喜欢它的沉稳感。

I especially enjoy its calmness.

它听起来更简洁。

It sounds more concise.

感觉更像是在和一位聪明的同事对话,而不是读一堆充满不必要离题的长篇大论。

It feels more like a conversation with a smart colleague than reading walls of text with unnecessary digressions.

这对保持思路连贯有很大帮助。

This makes a big difference in keeping a train of thought going.

当下最重要的AI问题之一,不是谁在用AI,而是谁把AI用得好。

One of the most important AI questions right now isn't who's using AI, it's who's using it well.

KPMG和德克萨斯大学奥斯汀分校刚刚分析了140万次真实的职场AI交互,结果发现了令人意外的事情。

KPMG and the University of Texas at Austin just analyzed 1.4 million real workplace AI interactions and found something surprising.

影响力最大的用户并不是更擅长写提示词的人。

The highest impact users aren't better prompt engineers.

他们把AI当作推理伙伴来对待。

They treat AI like a reasoning partner.

他们界定问题,引导思路,反复迭代,推动获得更好的答案。

They frame problems, guide thinking, iterate, and push for better answers.

好消息是什么呢?

And the good news?

这些行为是可以大规模传授的。

These behaviors are teachable at scale.

如果你正试图从“能用AI”迈向“真正用好AI”,KPMG关于高级AI协作的研究值得你花时间看看。

If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time.
M1
M113:38

随着2022年AI代码生成的出现,NVIDIA首席发明家和哈佛工程师Sid Pareshi采取了逆向立场:推理时计算和智能体编排,而不是预训练,才是解锁企业中高质量AI驱动软件开发的关键。

With the emergence of AI code generation in 2022, NVIDIA Master Inventor and Harvard engineer Sid Pareshi took a contrarian stance: inference-time compute and agent orchestration, not pre-training, would be the key to unlocking high-quality AI-driven software development in the enterprise.

他认为真正的突破不在于AI生成代码的速度有多快,而在于它推理的深度如何,这才能构建企业级应用。

He believed the real breakthrough wasn't in how fast AI could generate code, but in how deeply it could reason to build enterprise-grade applications.

当整个世界都聚焦于“副驾驶”模式时,他构建了从根本上不同的东西。

While the rest of the world focused on copilots, he architected something fundamentally different.

Blitzy,首个利用数千个智能体、专为企业级代码库打造的自主软件开发平台。

Blitzy, the first autonomous software development platform leveraging thousands of agents that is purpose-built for enterprise-scale codebases.

财富500强领导者正在通过Blitzy解锁5倍的工程效率,将数月的工作量在几天内完成。

Fortune 500 leaders are unlocking 5x engineering velocity and delivering months of engineering work in a matter of days with Blitzy.
M1
M115:49

现在,谈到把Fable引入更高影响力和更高层级的工作,Claude Code团队的Tariq最近写了一篇文章,题为《Fable实地指南:发现你的未知》。

Now, along the idea of moving Fable into more high-impact and higher-order types of work, Tariq from the Claude Code team recently wrote a post called A Field Guide to Fable: Finding Your Unknowns.

Tariq写道,和Claude Fable 5一起工作不断让我复习一个老道理:地图不是疆域。

Tariq writes, working with Claude Fable 5 keeps reteaching me an old lesson: The map is not the territory.

地图,即对待完成工作的表征,是我的提示词、技能和上下文。

The map, a representation of the work to be done, is my prompts and skills and context.

这是我给Claude的东西。

It's what I give Claude.

疆域是工作真正需要发生的地方。

The territory is where the work needs to happen.

代码库、现实世界、其真实的约束条件。

The codebase, the real world, its actual constraints.

地图和疆域之间的差异,就是我所谓的未知。

The difference between the map and the territory is what I call unknowns.

当Claude遇到一个未知时,它需要基于对我想法的最佳猜测来做决策。

When Claude runs into an unknown, it needs to make a decision based on its best guess of what I want.

工作越多,Claude可能遇到的未知就越多。

The more work being done, the more unknowns Claude might run into.

Fable是第一个让我觉得,工作质量受制于我能多好地澄清其未知的模型。

Fable is the first model where I find the quality of the work is bottlenecked by my ability to clarify its unknowns.

塔里克这篇文章的另一个重点是,用他的话说,光提前规划是不够的。

And the rest of the thrust of Tariq's post is to point out that, in his words, planning ahead isn't enough.

相反,他说,用Fable工作是一个迭代的过程,在实施前、实施中、实施后不断发现自己的未知。

And instead that, in his words, working with Fable is an iterative process of discovering my unknowns before, during, and after implementation.

那么有哪些例子呢?

So what are examples of this?

塔里克说,他开始用Fable时,会把它拆成四个方面。

Tariq says that when he starts engaging with Fable, he breaks it down in 4 ways.

这四个类别是:已知的已知,基本上就是提示词里的内容,也就是说,我告诉代理我想要什么?

The 4 categories are known knowns, or essentially what is in his prompt, i.e., what do I tell the agent that I want?

已知的未知,是他还没搞清楚,但意识到自己没搞清楚的。

The known unknowns are what he hasn't figured out yet but is aware that he hasn't.

未知的已知,是那些太明显、他永远不会写下来,但看到就会认出来的东西。

The unknown knowns are what are so obvious he'd never write it down but would recognize it if he saw it.

而未知的未知,是他完全没考虑到的。

And the unknown unknowns are what hasn't he considered at all.

塔里克认为,减少和规划未知是代理编程的技能,并且建议有办法可以改进这一点。

Tariq argues that reducing and planning for unknowns is the skill of agentic coding and suggests that there are ways to improve upon it.

关键在于帮助Claude来帮助你。

And key to that is helping Claude help you.

塔里克写道,指导Claude需要一种微妙的平衡。

Tariq writes, instructing Claude is a delicate balance.

如果你太具体,Claude会遵循你的指示,即使转向可能更合适。

If you're too specific, Claude will follow your instructions even when a pivot may be more appropriate.

如果你太模糊,Claude往往会根据行业最佳实践做出选择和假设,可能不适合你的任务。

If you're too vague, Claude will often make choices and assumptions based on industry best practices that may not be a fit for your task.

重要的是,当你不考虑自己的未知时,你会两头都失败。

Importantly, when you don't account for your unknowns, you fail both ways.

你不知道路什么时候会充满障碍,也不知道路什么时候会畅通,但你仍然希望Claude能转向。

You don't know when the path will be filled with obstacles, and you don't know when the path will be clear, but you still want Claude to veer.

Claude可以帮助你更快地发现自己的未知。

Claude can help you discover your unknowns faster.

他说,最重要的是给Claude关于你起点的背景信息。

The most important part, he says, is to give Claude context about your starting point.

比如,告诉它你在思考过程的哪个阶段,分享你对这个问题的经验,让它像思考伙伴一样和你合作。

For example, tell it where you are in your thought process, disclose your experience with the problem, and let it work with you like a thought partner.

几个具体的想法包括盲点检查之类的。

A couple specific ideas are things like a blind spot pass.

他举的一个例子是:我不知道什么是调色,但我需要给这个视频调色。

An example he gives: I don't know what color grading is, but I need to grade this video.

你能教我说,让我理解自己对调色的未知的未知,这样我就能更好地提问吗?

Can you teach me to understand my unknown unknowns about color grading so that I can prompt better?

另一个建议是头脑风暴和原型设计。

Another suggestion is to brainstorm and prototype.

塔里克写道,当我在一个有很多未知的已知的领域工作时,涉及一些我只有看到才能定义的标准,我喜欢让Claude和我一起头脑风暴和做原型。

Tariq writes, when I'm working in an area with a lot of unknown knowns involving criteria I only know to define when I see it, I like to ask Claude to brainstorm and prototype with me.

他继续说,在原型阶段早期识别和表达未知的已知非常有价值,因为在实施阶段才发现它们,成本相对较高。

It's extremely valuable, he continues, to identify and verbalize unknown knowns early during prototyping because finding them out during implementation can be relatively expensive.

功能或规格上的小改动,可能导致代码实现上的大幅差异,而且让代理回滚之前的改动可能更困难。

Small changes in a feature or spec can cause drastically different implementations in code and can be more difficult for your agent to revert previous changes.

我觉得这可能是你们很多人已经在直觉上做的事情。

Now, this is something I think a lot of you guys are probably doing intuitively.

他给的一个示例提示是:我想要一个数据仪表盘,但我不懂视觉审美,也不知道能做什么。

An example prompt he gives is, I want a dashboard for this data, but I have no visual taste and I don't know what's possible.

给我做一个HTML页面,包含四种完全不同的设计方向,让我来反馈。

Make me an HTML page with 4 wildly different design directions so I can react to them.

塔里克还给了很多其他处理未知的想法,很多最终归结为一个元过程,就是邀请Fable作为过程的共同创造者,来充分利用它的能力。

Now, Tariq gives a bunch of other ideas about how to work with unknowns, and a lot of it ultimately comes to a meta process where you invite Fable in as a co-creator of the process to get the most out of what it can do.

丹尼尔·迈斯纳认为,即使超越像Fable-5这样的单一模型,每当有新的一流模型发布时,有一组提示词可以重跑,帮助你重新调整经验,充分利用那个新模型。

Now, Daniel Meissner thinks that even beyond a single model like Fable-5, that there are a set of prompts to rerun every time a new state-of-the-art model drops that can help you reorient your experience to take most advantage of that new model.

他称这些为战术性元提示词,每次智能水平有显著提升时都值得重新运行一遍。

He calls these tactical meta prompts worth rerunning every time there's a significant jump in intelligence.

第一类是关于框架优化的,全都和改善你的AI框架有关,也就是围绕模型的那套资源和信息,目的是从最新版本中榨取最大价值。

The first category is around harness optimization and are all related to improving your AI harness, i.e., the set of resources and information around your model to get the most out of the newest version.

他举的一个例子是自我模型审计。

One example he provides is the self-model audit.

读一遍我的框架所相信的关于我的一切——身份、目标、声音和偏好——然后找出它在哪些地方建模的是一个过时、理想化或完全错误的我。

Read everything my harness believes about me— identity, goals, voice, and preferences— and find where it's modeling a version of me that's stale, aspirational, or just wrong.

把我的文件里说的我是什么样,和我最近的言行及工作实际展现出来的我做个对比。

Compare what my files say I am against what my recent behavior and work actually revealed.

标出系统在为“过去的我”优化而非“现在的我”的每一处地方,并提出具体的修改建议来弥合这个差距。

Flag every place the system is optimizing for who I said I was instead of who I am now, and propose the specific edits that close the gap.

现在,从某种程度上说,这其实就是上下文清理。

Now, to some extent, this is just context hygiene.

而Daniel真正在做的是,利用新模型发布这个时机来进行这种上下文清理,确保你围绕模型放置的所有信息,比如你的agents.md文件,确实反映你带给新模型的那个人和那些项目。

And what Daniel is really doing here is using the moment of a new model release to engage in that sort of context hygiene and make sure all of the information that you've surrounded your model with, things like your agents.md file, are actually reflective of the person and projects that you're bringing to that new model.

Daniel也喜欢通过搞大动作来测试新模型。

Daniel's also a fan of testing new models by going really, really big.

他有一整类提示词,他称之为整体生活和工作优化。

He has a whole category of prompts that he calls overall life and work optimization.

也就是说,大规模范围的提示词,基本上帮你理清优先级、项目,并制定一个计划来最大化接下来的关键几年。

I.e., massive scope prompts that basically help you sort out your priorities, your projects, and come up with a plan for maximizing the next crucial few years.

其中一个他称为“大图景”的提示词是这样:看看我在网上写的各种项目、我们在The Harness里做的所有活动,以及你通过网络搜索了解到的关于我的一切,然后深入分析。

One he calls Big Picture is a prompt like, take a look at all my various projects in writing online, all the activities we've been doing in The Harness, and everything you know about me from web search and analyze it deeply.

然后看看我的领域、AI领域、社会以及整体未来正在发生什么,告诉我什么能帮我实现日本概念中的“ikigai”。

Then look at what's going on in my field, in AI, and in society, and in the future in general, and tell me what solves the Japanese concept of ikigai for me.

我应该做什么工作,既能让我有成就感,又能赚钱——是自己单干、和合作者一起,还是为公司打工?

What should I be working on that gives me fulfillment but is also lucrative, doing it for myself, with collaborators, or working for a corporation?

如果你有具体建议就给我具体的,在生成输出之前也可以随便采访我以获取更多背景信息。

Make concrete recommendations if you have them, and feel free to interview me for more context before creating the output.

现在,我不认为这一定是所有人的菜,但我确实觉得这类提示词是个很好的方式,哪怕只是为了感受一下模型的“氛围温度”,看看它怎么应对这类问题。

Now, I don't think that this is necessarily going to be everyone's cup of tea, but I do think that this sort of prompt is a great way, even if it's just for the sake of taking the vibe temperature of a model, to understand how it engages with these sorts of questions.

Daniel的另一个建议,也开始越来越多地出现,就是再次以循环的思维来考虑问题。

Another suggestion from Daniel, which is starting to come up more and more frequently, is to once again think in terms of loops.

Matt Schumer其实也把这个包括在了他的Fable 5技巧里,他给的建议叫做“循环到达标为止”,特别适用于创意任务。

Matt Schumer actually included this in his Fable 5 tips as well, with a recommendation he called loop it until it hits the bar, especially for creative tasks.

对于这个,你得回到Matt之前的建议,给Fable一个真正的完成标准。

Now for this, you have to go back to Matt's previous recommendation to give Fable a real bar for done.

Matt写道:如果你告诉Fable做出高质量的东西,它会在它自己认为“够好”的地方停下来,而这通常比你的标准低。

Matt writes, if you tell Fable to make something high quality, it stops at its own idea of good enough, which is usually lower than yours.

所以我不用形容词。

So I don't use adjectives.

我给它一个它能自己对照检查的标准,而且我把这个标准设得很高。

I give it a bar it can check itself against, and I make that bar hard.

有时候我自己写测试,比如像“陌生人看不出我们的渲染图和真实照片的区别”这种具体的东西。

Sometimes I write the test myself, something concrete like a stranger can't tell our render from the real photo.

其他时候我甚至不知道怎么衡量我想要的东西,所以我也把这个难题交给Fable。

Other times I don't even know how to measure the thing I want, so I hand that problem to Fable too.

Matt接着说,一旦有了标准,我就让Fable对着它循环,然后放手让它跑。

Now Matt continues, once there's a bar, I put Fable on a loop against it and let it go.

它构建、自查、找出最大的差距、弥合它,然后再来一轮。

It builds, checks itself, finds the biggest gap, closes it, and goes again.

Matt说他经常用/loop命令来做这个,特别是在创意工作上,因为总有具体的东西可以持续对照衡量,直到真正达到目标。

Matt says that he uses the /loop command for this constantly, especially on creative work, where there's always something concrete to keep measuring against until it's actually there.

重要的是,Matt写道,循环的全部意义在于Fable永远不能自己决定它完成了。

Importantly, Matt writes, the whole point of the loop is that Fable never gets to decide it's finished.

总是还有下一个差距。

There's always a next gap.

它在我说完成的时候停下,或者当它真的找不到任何需要修的东西时停下——而如果你设置得当,这种情况很少见。

It stops when I say it's done or when it genuinely can't find anything left to fix, which is rare if you've set this up right.

以后我们肯定会专门做一期节目来聊这些循环策略,但今天我想先分享一些来自 Claude Devs 最近一篇文章的笔记,讲的是怎么开始用循环。他们把循环拆成几类,我觉得这些分类对理解核心概念挺有帮助的。

Now at some point we'll do a show entirely about some of these loop strategies, but I did want to share just a couple notes from a Claude Devs post recently about getting started with loops where they break loops apart into a set of different categories that I think are useful in helping understand the concept at core.

第一种是基于回合的循环。

Turn-based loops are the first type.

这种循环由用户提示触发,当 Claude 判断任务已完成或需要更多信息时就会停止。

They are triggered by a user prompt and they stop when Claude judges that it has completed the task or needs additional context.

这种循环最适合那种比较短的任务,不是那种常规流程或固定计划的。

Turn-based loops are best for shorter tasks that are not part of a regular process or schedule.

他们写道,你发的每一条提示都会启动一个手动循环,由你来指挥每一回合。

Every prompt you send, they write, starts a manual loop with you directing each turn.

Claude 会收集上下文、采取行动、检查自己的工作、需要的话重复这些,然后给出回应。

Claude gathers context, takes actions, checks its work, repeats if needed, and responds.

比如说,他们提到,让 Claude 创建一个点赞按钮。

For example, they say, ask Claude to create a like button.

它会读你的代码,做修改,跑测试,然后交回一个它认为能用的结果。

It reads your code, makes the edit, runs the tests, and hands back something it believes works.

然后你手动检查结果,再写下一个提示。

You then manually check the work and write the next prompt.

你可以改进验证步骤,方法就是把你的手动检查步骤写进一个 skill.md 文件,这样 Claude 就能更完整地自己检查更多工作了。

You can improve the verification step by encoding your manual steps as a skill.md so Claude can check more of its own work end-to-end.

对比一下基于目标的循环,这种循环由手动提示实时触发,停止条件是目标达成或者达到最大回合数。

Contrast that with a goal-based loop, which is triggered by a manual prompt in real time, where the stop criteria is the goal being achieved or the maximum number of turns being reached.

他们写道,这种循环最适合有可验证退出标准的任务,而且当单回合不够用时特别有用。

This, they say, is best used for tasks that have verifiable exit criteria and is useful when a single turn is not enough.

他们还说,当智能体能够反复迭代时,表现会更好。

Agents do better, they write, when they can iterate.

而且当你明确了成功标准,Claude 就不用自己决定什么算足够好了,也不会提前结束循环。

And when you define the success criteria, they say, Claude doesn't have to make a determination on what is good enough and end the loop early.

每次 Claude 想停下来,评估模型就会检查你的条件,然后让它继续工作,直到目标达成或达到你设定的回合数。

Each time Claude tries to stop, an evaluator model checks your condition and sends it back to work until the goal is met or a number of turns you define is reached.

基于时间的循环是由特定的时间间隔触发的,当用户取消它或工作完成时可以停止,这种循环很适合处理重复性任务。

A time-based loop is triggered by a specific time interval that can be stopped when the user cancels it or when the work completes, and this is going to be useful for things like recurring work.

而主动式循环,则是通过事件或日程安排触发的,不需要有人实时在场去启动它们。

Proactive loops, on the other hand, are those that are triggered by an event or schedule and don't require a human in real time to set them going.

停止条件同样是某个特定目标达成,整个例程会自动运行,直到你把它关掉。

The stop criteria is once again when a specific goal is met, with the routine running itself until you turn it off.

接下来几个月我们会探索很多内容,核心就是把这种基于循环的交互方式,从编码领域扩展到更有创意和知识型的工作中。

Now, a lot of what we'll be exploring over the next few months is how to bring this sort of loop-based interaction outside of the realm of coding and into more creative and knowledge work.

虽然我们在循环这个话题上没深入太多,但在这期节目结尾提到它,是因为所有这些建议的大主题和核心要点有两层。

Even though we're not getting deeper than that when it comes to loops, the reason it's worth mentioning them at the end of this particular episode is that the big theme and takeaway of all of these recommendations is two-part.

最简单的层面,它们都在提醒我们,每当拿到一个新模型,尤其是当这些新模型代表智能水平巨大飞跃时,我们都需要花时间、有时是费劲地逐项测试,找出我们过去提示和交互老模型的方式,哪些不再能充分发挥新模型的潜力,甚至可能对它有害。

At a simplest level, they're all reminders that every time we get a new model, especially when those new models represent a big jump in intelligence, we need to do the hard, sometimes time-consuming work of going through and trying everything and figuring out where the ways that we prompted and interacted with the old models either no longer get the most out of the new models or actively harm them.

但第二点,从长远来看可能更重要,就是我们需要找出真正能解锁这种智能飞跃所带来的差异化能力的新技巧。

But secondly, and perhaps in the long run even more importantly, we need to figure out the new techniques which actually unlock the differentiated capabilities of that jump in intelligence.

很多时候,这些不只是提示技巧,而是全新的交互模式,就像这些循环代表的那样。

And many times those things aren't just going to be prompting tips, but totally new interaction patterns like these loops represent.

如果要从所有这些内容里提炼一个核心要点,那就是要提升自己的雄心,几乎要假设最新模型没有局限,这样通过尝试你能想到的最大、最具挑战性的事情,你才能真正发现那些限制,一旦它们无可避免地暴露出来。

If there is one takeaway from all of these pieces, it is to ratchet up one's ambition and almost to assume no limits on what the newest model can do so that by trying the biggest, most challenging things you can think of, you can actually figure out what those limits are once they inevitably reveal themselves.

这项工作第一层永远是个体层面的,但我觉得对组织来说也有类似的道理。

The first layer of this work is always going to happen on an individual level, but I think that there is an analogy for organizations here as well.

尤其是在商业或工作环境中,很容易默认用 AI 来做那些我们已经在用 AI 做的事情。

It is so easy, especially in a business or work context, to default to using AI for the things that we're already using AI for.

那些我们已经找到价值的地方,我们已经在做的工作,然后只是把 AI 看作一种更快、更便宜或稍微好一点的方式,但归根结底还是做同样的工作。

The places where we've already found value, the type of work that we already do, and to view AI as just a way to do it faster or cheaper or maybe a little bit better, but ultimately to do that same work.

真正的解锁,当然是与工作建立一种新关系,甚至开启以前不可能的新类别的工作。

The unlock, of course, is a new relationship with work and even unlocking new categories of work that weren't possible before.

弄清楚怎么做没那么简单,但同时也非常令人兴奋。

Figuring out how to do that is a lot less simple, but it's also really exciting.

希望今天你听到的这些小技巧,能为你提供工具,让你跨越整个工作光谱去运用它们。

Hopefully some of the tips you've heard today give you tools you can use across that full spectrum of work.

希望再过五分钟,等我们迎来下一次模型飞跃时,这些内容还能派上用场。

And hopefully they're still useful about 5 minutes from now when we get the next model leap.

今天的人工智能日报就到这里了。

For now, that's going to do it for today's AI Daily Brief.

一如既往,感谢您的收听或收看。

Appreciate you listening or watching as always.

下次再见,Peace。

And until next time, peace.
已剔除 3 处广告(点击展开查看)
M1
M10:17广告 · 已剔除

First of all, thank you to today's sponsors, KPMG, Blitzy, Retool, and Airtable.

To get an ad-free version of the show, go to patreon.com/AIDailyBrief, or you can subscribe on Apple Podcasts, To learn more about sponsoring the show, send us a note at [email protected].

M1
M113:28广告 · 已剔除

Learn more at kpmg.com/us/sophisticated.

That's kpmg.com/us/sophisticated.

M1
M114:15广告 · 已剔除

Transform the way you develop software.

Discover how at blitzy.com.

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

This episode is brought to you by Retool.

Generating a working app now takes about 5 minutes thanks to AI.

Getting it safely into production— with auth, permissions, audit logs, security reviews— that part still takes time.

That's the gap Retool closes.

Build apps however you want: natively in Retool, with Claude Code, Codex, or any coding agent, or by importing React you've already built.

It all deploys into Retool and picks up governance automatically.

Security lives in the platform, not in whatever the AI wrote.

So your team ships at AI speed without the shadow IT and the endless reviews that stall out most vibe-coded projects.

It's why teams at Amazon, Stripe, and Brex build on Retool.

And new enterprise customers who sign by September 30th get up to $10,000 in AI credits per year.

Start building at retool.com/AIDaily.

This episode of the AI Daily Brief is brought to you by HyperAgent, where you run fleets of agents your team can manage together.

New users get $1,000 in inference.

Forget local agents and chat workflows waiting on your laptop to be prompted.

HyperAgent deploys always-on agents in the cloud, doing real work across the tools your team already uses.

Marketing's agent turns competitor moves into landing pages.

Sales' agent enriches leads, drafts emails, and updates the CRM.

Ops' agent chases the paperwork and tracks the budget.

Every agent has access to shared context and follows your rules about scope and approvals.

It's time you add agents that feel like teammates.

Hire yours at HyperAgent, built by the team at Airtable.

Claim your $1,000 in inference at hyperagent.com/AIDailyBrief.