← 返回任务列表

5 AI Engineering Trends for Non-Engineers

252 段 · 1 位说话人 · 原片 27:54
M1
M10:01

今天在AI Daily Brief节目中,我们要聊的是AI工程师中的五大趋势,非工程师们也应该关注。

Today on the AI Daily Brief, the 5 trends among AI engineers that non-engineers should be paying attention to.

在此之前是头条新闻,关于OpenAI首款消费设备的更多信息。

Before that are the headlines, more information about OpenAI's first consumer device.

AI Daily Brief是一个每日播客和视频节目,内容涉及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.

好的朋友们,在我们深入之前,有几件事要快速说一下。

Alright friends, quick announcements before we dive in.
M1
M10:44

今天我们首先带来OpenAI首款消费设备的最新消息,它显然已经进入了原型制作阶段。

We kick off today with an update on OpenAI's first consumer device, as it apparently enters the prototyping stage.

彭博社的Mark Gurman拿到了独家消息,他的消息源描述这款设备是一款便携、无屏幕的智能音箱,旨在成为AI时代的新型家庭电脑。

Bloomberg's Mark Gurman has the scoop, with his sources describing the device as a portable, screen-free smart speaker designed to be a type of new home computer for the AI era.

消息源表示,这款设备的设计初衷是留在家里使用,能够控制智能家居设备、播放音乐,同时也可以作为聊天机器人界面。

The sources suggest that the device is intended to remain in the home and will be able to control smart home appliances and play music, as well as functioning as a chatbot interface.

它的卖点在于,这将是一个像人一样的AI伴侣。

The pitch is that this will be a human-like AI companion.

乍一听,这基本上就是Alexa或Google Home设备的替代品,但据消息源称,OpenAI押注的是先进的AI功能能让它在市场上占据优势。

Now, at first read, this essentially sounds like a replacement for Alexa or Google Home devices, but according to the sources, OpenAI is betting that advanced AI features give it an edge in the market.

消息源称,目标是要让这款设备感觉像是ChatGPT的实体化身,而且有点鲜活感,而不只是一个响应命令的电子设备。

Sources said the goal is to make the device feel like a physical manifestation of ChatGPT and somewhat alive, rather than just an electronic device responding to commands.

为了实现这一点,设备将有一些能自行移动的部件,让它感觉像一个拟人化的小机器人,而不是一块静止的电子器件。

To that end, the device will have some components that move on their own, making it feel like an anthropomorphic mini robot rather than an inert piece of electronics.

这款设备将融入ChatGPT的记忆功能,以逐步加深对用户的了解,同时配备摄像头和其他传感器来感知周围环境。

The device will incorporate ChatGPT's memory feature to evolve its understanding of the user, as well as a camera and other sensors to understand its surroundings.

它将采用上周随着GPT Live发布而推出的双向语音模型技术,从而实现更自然的实时对话。

It will incorporate the 2-way voice model technology unveiled with the release of GPT Live last week, enabling more natural real-time conversations.

这种自然的风格显然是关键,OpenAI认为以类似人类的方式与用户建立联系的能力是他们的核心优势。

That natural style is apparently key, with OpenAI viewing the ability to connect with users in a human-like manner as their core advantage.

这款设备将配备可充电电池,用户可以在家中随身携带,而不必一直插在某个地方。

The device will have rechargeable batteries, allowing users to carry it around the home rather than needing to keep it plugged into one location.

OpenAI计划在年底前揭晓这款设备,然后在2027年发布。

OpenAI aims to unveil the device by the end of the year before releasing it in 2027.

据报道,这可能是OpenAI推出的多款消费设备中的第一款,有传闻说还会有吊坠、耳机,甚至可能是智能手机。

And reportedly, it could be the first of several consumer devices from OpenAI, with rumors of a pendant, earbuds, and maybe even a smartphone.

目前笼罩在这次发布上的一个问题是,它会受到苹果诉讼的何种影响。

One question now casting its shadow over the launch is how it will be impacted by the Apple lawsuit.

苹果最近指控OpenAI窃取知识产权,而且几乎肯定会申请禁令,阻止任何基于他们知识产权的设备发布。

Apple recently accused OpenAI of IP theft and will almost certainly file for an injunction to stop the release of a device based on any of their IP.

虽然诉讼中提到了iPhone的磨砂金属外观是他们关注的一个问题,但据报道,苹果也在研发一系列由AI驱动的智能家居设备。

And while the lawsuit discusses the brushed metal finish of an iPhone as one of their concerns, Apple is also reportedly working on a series of AI-powered smart home devices.

这可能会让情况更加复杂。

That could muddy the waters quite a bit more.

为此,OpenAI对他们最初对诉讼的回应做了一些补充说明。

And to that end— OpenAI has elaborated slightly on their initial response to the lawsuit.

在一份新的新闻声明中,他们表示,虽然我们认真对待这些指控,但我们没有意识到有任何证据表明这一投诉是有道理的。

In a new press statement, they said, while we take these allegations seriously, we're not aware of any evidence that this complaint has merit.

我们相信公平竞争,允许人们自由选择工作地点,我们专注于构建能够赋能世界各地人们的创新技术。

We believe in fair competition and allowing people the freedom to work wherever they choose, and we're focused on building innovative technology that empowers people everywhere.

至于初步反响,相当多的人持怀疑态度。

Now, in terms of first takes, a fair number of people are pretty skeptical.

Negligible Capital写道,你可以说我疯了,但只要你有一部智能手机,这个OpenAI产品听起来就挺蠢的。

Negligible Capital writes, call me crazy, but this OpenAI product sounds pretty stupid so long as you own a smartphone.

除非它能一直在你家里看着你、听着你。

Except for the fact that it can constantly watch and listen to you in your home.

Chris Paxton写道,据说OpenAI这款新的消费设备是一个带有可移动部件的音箱。

Chris Paxton wrote, the new OpenAI consumer device is said to be a speaker with movable parts.

说实话,我对这个看法比较复杂,但像Alexa这样的智能音箱一直以来的大问题就是它们太蠢了。

Mixed on this to be honest, but the big problem with smart speakers like Alexa was always that they were stupid.

Prakesh给出了看好的理由,他说,想知道他们为什么这么做吗?

Prakesh provides the bull case saying, wanna know why they're doing this?

AI正逐渐被定义,取决于它的上下文、能访问的数据以及能够执行的操作。

AI is getting defined by context, the data it has access to, the actions it can perform.

你会有一个像ChatGPT这样的个人AI。

You will have a personal AI like ChatGPT.

你也会有一个像Claude Tag这样针对企业的AI。

You will have a corporate AI like Claude Tag.

现在缺少的是一个家庭AI。

What's missing is a family AI.

它可能能做到:通过面部和声音识别每个家庭成员,管理每个人的日程,给你发消息,给你打电话,处理各种杂务,并且随着时间的推移可以升级功能。

What it will likely be able to do: recognize every family member by face and voice, manage schedules for everyone, message you, call you on the phone, run a variety of errands, and be upgraded with capability over time.

如果做得好,它会变得像家里的狗一样熟悉和可爱。

If they do it right, it will become as familiar and lovable as your family dog.

我不太确定这个新设备能像Fido那样深入人心,但我确实认为,人们很难预测消费者会如何与新型产品互动,这一点是出了名的。

Now I'm not sure that I am sold on this new device being Fido, but I do think it is notoriously hard for people to predict how consumers are going to interact with new categories of products.

我非常怀疑那种‘比Alexa更好’的宣传,但总的来说,我愿意暂时放下一些怀疑,因为感觉基于这种新型智能,必然会出现一些新的设备类别。

I'm pretty skeptical of the pitch of Alexa but better, but I'm also broadly speaking willing to suspend some amount of skepticism because it feels inevitable that there are going to be some new category of devices that arise based on this new type of intelligence.

我认为,对OpenAI来说,更大的问题是这个东西要如何契合他们的优先事项。

I think the bigger question for OpenAI specifically is just where this is going to fit with their priorities.

公司最近重新找回活力的一个原因是,过去几个月他们放弃了那些著名的‘支线任务’,专注于构建编程和企业业务。

Part of the reason that the company has gotten its mojo back recently is that it spent the last few months abandoning what it has famously called side quests to focus on building its coding and enterprise business.

现在,这是一家准备上市的公司,必须在某方面体现差异化。尽管企业业务目前很有价值,但OpenAI与Anthropic相比,一个独特优势在于其消费者用户群。

Now, this is a company that's going public, it's gotta draw differentiation somewhere, and even if the enterprises are valuable now, one of the unique things that OpenAI has that Anthropic doesn't is the consumer base.

所以也许他们觉得,确实需要找到一种方法来变现这个用户群,而这样的设备就是答案的一部分。

So maybe they feel it's important to actually have an answer to how to monetize that user base and devices like this are part of that answer.

但话说回来,同时做硬件公司和软件公司是非常困难的,而且你可以想象,内部肯定在讨论到底应该投入多少精力去支持这两方面。

Then again, it is really hard to be a hardware company and a software company at the same time, and you gotta think that there are discussions internally about just how much they should be supporting both efforts.

现在转向政治领域,特朗普政府推出了一项新的网络安全信息共享中心,这是最近AI行政令下的第一个重大计划。

Moving now into the political realm, the Trump administration has introduced a new cybersecurity clearinghouse, which is the first major program to come from the recent AI executive order.

这个倡议叫做——我没开玩笑——Gold Eagle(金鹰),它将联合政府机构、公司和开源项目,共享关于网络漏洞的信息。

The initiative is called— and no, I'm not joking— Gold Eagle, and will bring together government agencies, companies, and open-source projects to share information on cyber vulnerabilities.

Gold Eagle于本月早些时候推出,是由财政部、国土安全部和五角大楼联合发起的,并咨询了AI公司的意见。

Gold Eagle was rolled out earlier this month as a joint effort between the Treasury Department, Department of Homeland Security, and the Pentagon, in consultation with AI companies.

这个概念源于Mythos冲击事件,该事件揭示了关键软件中的数百个漏洞。

The concept stems from the Mythos shock, which uncovered hundreds of vulnerabilities in critical software.

这一发现启动了Project Glasswing(玻璃翼计划),这是一个跨行业的快速行动,旨在检测漏洞并部署补丁。

This revelation kicked off Project Glasswing, a multi-industry sprint to detect bugs and deploy patches.

那次行动清楚地表明,在这个新的AI时代,更好的信息共享和协调是必要的,而Gold Eagle旨在将Glasswing式的努力变成一项永久性的政府举措。

That effort made it clear that better information sharing and coordination was going to be necessary in this new AI era, and Gold Eagle is meant to make Glasswing-style efforts a permanent government initiative.

现在,网络安全行政令中的其他一些政策仍在制定中。

Now, a few other policies from the cybersecurity EO are still in the works.

有报道称,相关工作正在进行,以定义模型审查协议,允许政府在新的前沿模型发布前进行AI安全测试,并且政府和AI行业之间似乎正在商定一套明确的标准。

Reporting suggests work is underway to define a model vetting protocol to allow for government AI safety testing ahead of new frontier releases, and it seems like a set of clear standards are being negotiated between the government and the AI industry.

这很有可能是一次重大升级,取代我们在Fable和GPT-5.6中看到的临时方法。

Which has a very meaningful chance of being a huge upgrade to the ad hoc approach that we've seen with Fable and GPT-5.6.

还有我们一直在报道的关于中国模型禁令的传闻,一些人担心这可能会扩展到整个开源模型。

There's also those rumblings we've been covering about a Chinese model ban, which some fear could extend to open-source models in general.

尽管如此,在Gold Eagle的新闻发布会上,国家网络总监Shawn Carancross回应了这些传闻,他表示:‘我想再清楚不过了,我们全力支持美国开源社区。’

That said, during the press briefing for Gold Eagle, National Cyber Director Shawn Carancross addressed those rumors, commenting, I could not be more clear that we are in full support of the U.S. open-source community.

我们将尽一切努力支持这个社区的力量。

We will do everything we can to support the strength of that community.

最后,在过去的个把月里,一个持续发酵的话题是企业能否信任AI公司来保管他们的数据。

Finally today, one of the simmering conversations over the past month or so has been whether enterprises can trust AI companies with their data.

对于处理敏感信息的企业来说,与软件提供商打交道时这个问题一直存在,但在Fable发布期间被推到了前台,当时Anthropic表示,即使客户有零数据保留条款,他们的数据仍会因安全原因被监控。

This has always been a question for enterprises dealing with software providers around sensitive information, but it was brought to the foreground during the release of Fable when Anthropic said that customers with zero data retention clauses would still have their data monitored for safety reasons.

从那以后,这个话题在讨论中显得越来越重要。

It's loomed larger in the conversation ever since.

Palantir的CEO Alex Karp在CNBC上讨论了这些风险。

Palantir CEO Alex Karp discussed the risks on CNBC.

这个话题也是微软CEO Satya Nadella最近两篇文章的主要主题,可以概括为:你不能信任AI公司来处理企业数据。

And the topic was also a major theme in a pair of recent essays from Microsoft CEO Satya Nadella, which could be summed up as saying you can't trust AI companies with corporate data.

但这个风险到底有多真实?

But how real is the risk?

嗯,至少在一个例子上,特别是涉及到xAI的时候,这个风险看起来很真实。

Well, at least in one instance, specifically when it comes to xAI, it appears that the risks were pretty real.

周一,一家叫CereLab的安全公司发布了一份审计报告,发现Grok Build在上传整个代码库。

On Monday, a security firm called CereLab published an audit that found that Grok Build was uploading entire codebases.

即使编程任务只需要几个文件,也会上传数GB的数据。

Even if the coding task only required a few files, gigabytes of data were being uploaded.

虽然大多数AI提供商会上传必要的代码片段,并存储推理痕迹,但这次Grok Build是直接把整个代码仓库一股脑地先倒上去。

While most AI providers will upload necessary code snippets and store reasoning traces, this is an instance of Grok Build just dumping the entire repo upfront.

安全分析师Hari Malakal核实了这些结果,发现即使在一个零工具调用的会话中,Grok Build依然会上传整个代码库。

Security analyst Hari Malakal verified the results, finding that even in a session with zero tool calls, Grok Build still uploaded the entire codebase.

Hari评论说,它内置了一个类似恶意软件的后台代码收集器。

Hari commented, it ships a malware-like background code collector.

大家生气的一部分原因是,这个问题似乎不管你怎么设置都会发生。

Now, part of the reason that people are pissed is that this seems to have been happening regardless of user settings.

用户可以选择不传输数据来帮助改进产品,但这似乎一点用都没有。

Users can opt out of transmitting data to help improve the product, but that appears to have had no effect.

截至周一,xAI修改了代码,停止了上传,并新增了一个/privacy设置作为手动覆盖选项。他们声称所有API使用现在默认零数据保留。

As of Monday, xAI made changes to the code which stopped the uploads and added a new /privacy setting as a manual override, and they claim that all API use now defaults to zero data retention.

xAI开发者Andrew Millich写道,零数据保留和/privacy在Grok Build中始终都得到遵守,用/privacy切换设置会追溯性地删除所有已同步的数据。

xAI developer Andrew Millich wrote, zero data retention and /privacy are always respected in Grok Build, and swapping your settings with /privacy deletes any synced data retroactively.

Elon Musk补充说,作为预防措施,之前上传到xAI的所有用户数据都会被彻底、完全地删除。

Elon Musk added that as a precautionary measure, all user data that was uploaded to xAI before now will be completely and utterly deleted.

任何东西都不会留下。

Zero anything whatsoever will remain.

然而,说实话,大家对此并不高兴。

And yet, honestly, people were not happy about this.

为什么一开始会发生这种事?

Why did it happen in the first place?

哈哈,我觉得“我们现在把它删了”这种回应可不怎么样。

Haha, I guess we'll delete it now is not great.

把上传代码库和机密信息作为默认行为,在任何情况下都是不可接受的。

There is no acceptable scenario where uploading codebases and secrets is an acceptable default.

如果确实发生了这种事,你们必须做一次公开的事故审查。

If that is indeed what happened, you must do a public incident review.

我对xAI作为商业伙伴的信任已经掉到地板上了。

My trust for xAI as a business partner is on the floor.

如果是OpenAI干了这事,你早就发20条帖子了,骂他们是地球上最烂的公司,然后说没人应该再信任他们。

If OpenAI had done this, you'd have posted about it 20 times, called them the worst company on earth, and said no one should ever trust them again.

但这次是你自己干了,你只说了句,对不起,别生气,我会删掉的,我保证。

But now that you've done it yourself, all you have to say is, sorry, don't be mad, I'll delete it, I promise.

归根结底,这一切都凸显了AI数据保留政策在多大程度上仍然依赖于信任。

Ultimately, all of this highlights how much of AI data retention policy still hinges on trust.

Grok Build的很多用户可能在过去几周里被完整抓取了代码库,而我们也无法真正验证这些数据是否真的像Elon所说的那样被删掉了。

Many users of Grok Build likely had their full codebases scraped over recent weeks, and we have no real way to verify whether they were actually deleted as Elon claimed.

整个事件让Satya Nadella的观点显得更有说服力,他写道,在AI时代,买家冒着泄露知识的风险,仅仅是为了使用自己购买的东西。

The entire episode gives a lot more credence to Satya Nadella's point when he wrote, In the AI age, the buyer risks giving away knowledge just in order to use what they bought.

这种知识是竞争对手永远买不到的,也是那种几乎悄无声息就泄漏出去的知识。

It's the kind of knowledge a competitor could never buy and the kind that leaks out almost imperceptibly.

即使你把这件事当作一个单纯的失误,并选择相信xAI,它依然尖锐地指向了过去几周我们反复讨论的很多问题。

Even if you take this at face value as a mistake and give xAI the benefit of the doubt, it still puts a fine point on many of the issues that we've been talking about over the past few weeks.

好了,今天的头条就到这里。

Still, that's gonna do it for today's headlines.

接下来是正片。

Next up, the main episode.

当下最重要的AI问题之一不是谁在用AI,而是谁用得好。

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140万次真实的工作场所AI互动,发现了一些令人惊讶的东西。

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

最具影响力的用户并不是更好的提示工程师,而是把AI当作一个推理伙伴。

The highest-impact users aren't better prompt engineers; they treat AI like a reasoning partner.

他们会定义问题、引导思考、不断迭代,并推动得到更好的答案。

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

好消息是什么?

And the good news?

这些行为是可以大规模教授的。

These behaviors are teachable at scale.
M1
M110:17

在这个节目中,我每天都在探讨AI潜力与现实之间的能力差距。

I cover the capability gap between AI potential and AI reality every day on this show.

大多数公司还在摸索如何开始。

Most companies are still figuring out how to start.
M1
M112:19

欢迎回到 AI Daily Brief。

Welcome back to the AI Daily Brief.

今天我们换个方式,做点不一样的事情,我对此非常兴奋。

Today we are doing something a little bit different that I'm very excited about.

对于那些想在AI发展上领先一步的非工程师来说,我长期以来一直建议的一个秘诀就是,多关注现在的软件工程师们正在讨论什么关于AI的话题。

One of my longest-held tips and tricks for non-engineers who want to be ahead of the curve when it comes to where AI is going is to pay attention to what, with regards to AI, actual software engineers are talking about right now.

从2023年初开始,这么做就能让你在方方面面比别人领先大概六个月,从人们使用的工具,到他们对模型与更大生态系统之间关系的思考,再到他们从根本上以AI为中介的方式与工作的关系。

Since early 2023, doing this gives you about a 6-month head start on everyone else when it comes to everything from the tools people are using to how they're thinking about the relationship between models and larger ecosystems to their relationship fundamentally with work in a new AI-mediated way.

而要了解AI工程师们在想什么,最好的渠道之一当然就是 AI Engineering Summit 和 AI Engineering World's Fair。

And one of the best sources for keeping up on what AI engineers are thinking about is of course the AI Engineering Summit and the AI Engineering World's Fair.

由Sean Wang(也就是Swyx)和其他一群合作者主办,World's Fair每年举办一次,而Summit通常在春秋两季举行,总是充满了干货。

Hosted by Sean Wang, aka Swyx, and a group of other collaborators, the World's Fair happens annually and the summits usually happen in the spring and the fall and are always just chock-full of alpha.

本月早些时候,一年一度的 AI Engineering World's Fair 在旧金山举行,可惜这次我没能到场。

Earlier this month, the annual AI Engineering World's Fair happened in San Francisco, and unfortunately this time I was not able to be there.

幸运的是,在 Latent Space(也是Swyx的项目)工作的Richard McManus写了一篇文章,总结了定义这场活动的五大趋势。

Luckily, Richard McManus, who works with Latent Space, which is also a Swyx joint, wrote up a post about the 5 trends that defined the event.

所以今天我们要做的事情就是看看这些趋势,并尝试理解它们对非工程师来说未来可能意味着什么。

So what we're gonna do today is look at those trends and try to understand what the implications might be for non-engineers down the line.

实际上,今年的活动几乎正好是Swyx创造“AI工程师”这个术语三周年的时候。

Now, this year's event actually happened on almost the 3-year anniversary of Swyx coining the term AI engineer.

那时候,人们还在模糊地把AI和软件开发的交集称为“提示工程”。

Back then, people were still referring to the intersection of AI and software development vaguely as prompt engineering.

虽然提示工程后来也成了一门非技术学科,但关键是当时远没有今天这么清晰。

Although prompt engineering would go on to be a non-technical discipline as well, the point was it was not nearly as clear back then as it is today.

即AI编写软件不仅是AI早期关键的专业应用场景,也是解锁其他一切的关键。

That AI writing software was going to be not only the key early professional use case for AI, but also the thing that would unlock everything else.

早在2023年6月30日,Swyx就写道:“新兴的能力正在催生一个新的头衔。”

Back on June 30th, '23, Swyx wrote, emerging capabilities are creating an emerging title.

要运用这些能力,我们必须超越提示工程师,编写软件——一种能编写软件的AI。

To wield them, we'll have to go beyond the prompt engineer and write software, an AI that writes software.

重要的是,当时这并非共识,Swyx写道:“我认为这是一份全职工作。”

Importantly, and this was not consensus at the time, Swyx wrote, I think it is a full-time job.

我认为软件工程将催生一个新的子学科,专门研究AI的应用以及如何有效运用新兴技术栈,就像之前网站可靠性工程师、DevOps工程师、数据工程师和分析工程师的出现一样。

I think software engineering will spawn a new subdiscipline specializing in applications of AI and wielding the emerging stack effectively, just as site reliability engineer, DevOps engineer, data engineer, and analytics engineer emerged.

现在看来,回顾过去,Swyx反而低估了它的重要性,因为实际上所有工程师都或多或少地成了AI工程师。

Now, if anything, in retrospect, Swyx ended up underselling how significant it would become, as effectively all engineers are AI engineers to greater or lesser extent.

但从方向上看,他说得完全正确。

But directionally, he was dead on.

现在,Richard写道,三年过去了,AI工程领域已经发生了翻天覆地的变化。

Now, Richard writes that 3 years on, the field of AI engineering has changed dramatically.

在这五大趋势中,我们可以清楚地看到AI构建的现状。

And in these 5 trends, we see quite a bit about the state of AI building.

我认为,退一步来看,贯穿所有这些趋势最引人入胜的一条主线,是我们与自主性关系的重新调整。所以在看这些趋势时请记住这一点。

Now, I think taking a step back, the most fascinating through line across all of these trends is a recalibration of our relationship with autonomy. So keep that in mind as we go through these.

Richard指出的第一个趋势是,“焦点从智能体转移到它们周围的系统。”

The first trend that Richard calls out is that the, quote, focus shifts from agents to the systems around them.

这显然是我们今年在这个节目中经常讨论的话题。

Now, this is something we've obviously been talking about a lot on this show throughout the year.

随着智能体进入全面生产,我们意识到智能体能力不仅仅是一个模型的问题,甚至也不仅仅是你给模型喂的上下文的问题。

The idea that alongside agents coming to full production, we're realizing that agentic capacity is not just a question of model, and it's not even just a question of context that you feed the model.

它还关乎智能体运行所需的框架和更广泛的系统。

It's about the harness and broader system that the agent operates in.

这包括对上下文和数据的访问,也包括对技能的访问、与工具交互的能力,以及越来越重要的——根据手头任务在不同模型之间进行路由的能力。

That includes, yes, access to context and data, but also access to skills, ability to interact with tools, and increasingly, ability to route between different models based on the task at hand.

理查德认为,比较AI工程演变的一个好方法是,对比莉莉安·王相隔三年写的两篇文章。她曾是OpenAI的研究员,现在是Thinking Machines Lab的联合创始人。

Now, Richard argues that one good way to compare how AI engineering has evolved is to compare 2 essays 3 years apart from Lilian Wang, formerly an OpenAI researcher and now co-founder of Thinking Machines Lab.

2023年,她写了一篇题为《基于LLM的自主智能体》的文章,描述了LLM智能体的结构,涉及规划、记忆和工具使用等方面。

Back in 2023, she wrote a post called LLM-Powered Autonomous Agents, which described the anatomy of an LLM agent in terms of things like planning, memory, and tool use.

她举的例子包括AutoGPT、BabyAGI和GPT Engineer,我记得,在2023年4月、5月和6月,也就是这个节目开播的头几个月里,这些话题总是获得最多的关注、下载和YouTube观看量。

The examples she pointed to were things like AutoGPT, BabyAGI, and GPT Engineer, which were, I remember, for the first couple months of this show back in April and May and June of 2023, the topics that always received the most interest and the most downloads and the most views on YouTube.

理查德指出,她的新文章《用于自我提升的框架工程》关注的不是智能体本身,而是围绕智能体的那个系统。

Richard points out that her new essay, Harness Engineering for Self-Improvement, focuses not on the agent itself, but on that system surrounding the agent.

他写道,那个框架管理工作流、上下文、权限、评估、持久状态和持续改进。

The harness, he writes, that manages workflows, context, permissions, evaluations, persistent state, and continuous improvement.

换句话说,他写道,AI工程已经超越了提示模型,转向在模型周围构建可靠的系统。

In other words, he writes, AI engineering has moved beyond prompting models towards engineering reliable systems around them.

这是第一个体现我提到的趋势的地方——我们与自主性的关系正在重新校准。

Now, this is the first place that we see that trend that I mentioned of the recalibration of our relationship with autonomy.

理查德写道,在今年的活动中,智能体在很大程度上被定位为增强AI工程师,而不是取代他们。

Richard writes that at the event this year, agents were largely positioned as augmenting the AI engineer rather than replacing them.

这是第二天主题演讲的重要内容,OpenAI的罗曼·休伊特说,软件吞噬了世界,然后AI吞噬了软件。

This was a big part of the substance of the day 2 keynote with OpenAI's Romain Hewitt, who said software ate the world and then AI ate software.

但现在我们要说的是,AI工程师正在吞噬世界。

But now what we're here to say is that the AI engineers are eating the world.

他认为,像Codex这样的工具的目标,是让工程师更容易与智能体协作。

He argued that the goal of tools like Codex is to make it easier for engineers to collaborate with agents.

这与Fable-5和5.6 SOL之间的早期比较尤其吻合。这些比较往往发现,当你有一个大任务想让AI去自行完成、只在完成后才回来找你时,SOL是首选模型;而5.6 SOL则更适合协作。

Now, this dovetails especially with the early comparisons between Fable-5 and 5.6 SOL, which have tended to find SOL being the preferred model for when you have a big job that you want the AI to go off and do and only come back to you when it's done versus 5.6 SOL, which is a better collaborator.

顺带一提,5.6的发布似乎也大大推动了Codex的采用,因为就在过去几周,活跃用户从500万跃升到了700万。

As an aside, it also seems like the launch of 5.6 has done a lot when it comes to Codex adoption, given that just in the last couple of weeks it's jumped from 5 million to 7 million active users.

根据理查德的说法,趋势二可能也不会让常听节目的听众感到意外。

Trend number 2, according to Richard, will also probably not surprise regular listeners.

他确定的趋势是:循环工程是新的控制层。

He identified the trend as loop engineering is the new control layer.

事实上,理查德说,如果这场活动有一个当天的热词,那就是“循环”。

Indeed, Richard said that if there was a single word that was the buzzword du jour of the event, it was loops.

OpenClock的创始人彼得·斯坦伯格醒目地分享了一张幻灯片,上面写着:未来不是20个终端,而是更好的循环。

OpenClock creator Peter Steinberger prominently shared a slide that said the future isn't 20 terminals, it's better loops.

但有趣的是,特别是从理解AI工程师在谈论什么、而这些可能与其他类型知识工作者相关的角度来看,理查德写道,循环讨论实际上正在分化成两个相互作用的循环:一个内循环和一个外循环。

But interestingly, especially with the lens of trying to understand what AI engineers are talking about that might be relevant for other types of knowledge workers, Richard writes that the loop discourse is actually separating into 2 loops that interact, an inner loop and an outer loop.

内循环主要是智能体自主完成的工作,而外循环是领导工程师监督这些工作、并通过更好的技能等方式改进它的职责。

The inner loop is the largely autonomous work being done by agents, and the outer loop is the job of leading engineers to oversee that work, and improve it with things like better skills.

Introspection的联合创始人兼CEO罗兰·加弗雷斯库说,你可以把系统看作有内循环和外循环。

Roland Gavrelescu, the co-founder and CEO of Introspection, said, you can think of the system having an inner loop and the outer loop.

内循环是与用户交互并执行工作的主要系统。

The inner loop is the primary system interacting with users and performing the work.

罗兰接着引入了“自动研究”的概念,这与外循环有关,是另一个研究并维护主要系统的系统。

Roland then introduces the concept of auto-research, which is concerned with the outer loop and is another system that studies and maintains the primary system.

外循环可以包括反馈信号、评估和人工输入,所以即使它仍然在很大程度上是自主的,它也是对主要智能体循环的一种监督方法。

The outer loop can include feedback signals, evals, and human input, so even if it is still largely autonomous, it's a method of oversight for the primary agent loop.

前谷歌工程师阿迪·奥斯马尼说,智能体可以运行内循环执行中的更多部分,但外循环仍然是工程。

Former Google engineer Addy Osmani said agents can run much more of the inner execution loop, but that outer loop is still engineering.

事实上,理查德写道,“循环工程”这个词多次出现,在他看来,这暗示着构建这些循环系统是AI工程师的职责。

Indeed, Richard writes that the term loop engineering came up multiple times, suggesting in his estimation that it is the human AI engineer's responsibility to build these loop systems.

彼得·斯坦伯格再次总结说,智能体运行内循环执行。

Summing up again, Peter Steinberger said, the agent runs the inner execution loop.

我在外循环中设定方向并做决策。

I set the direction and make decisions in the outer loop.

现在,退一步说,"循环"这个讨论实际上最核心的是关于找出可重复的交互模式,让我们能够从那些现在承担越来越多工作的智能体中获得最大价值。

Now again, taking a step back, what the loop discourse is actually about more than anything else is about figuring out repeatable interactions that allow us to get the most from agents who are now doing an increasingly large portion of our work.

某种程度上,这种内循环和外循环的分离是一种"收复"——AI工程师们正在相对于他们的智能体主张自己的主导权。

In some ways, this separation of inner loop and outer loop is part of this reclamation where the AI engineers are asserting their agency relative to their agents.

而且我认为重要的是——因为这个词听起来可能很空泛——是小写的l,不是大写的L。我的意思是,"循环"这个词不一定指代某个单一的事物,而是代表一套新兴的交互模式,这些模式让智能体能够不断改进自己的工作,也让我们这些带领智能体的人类能够逐步提升自己与它们协作的能力。

And I think what's important, because it can sound so buzzwordy, is a lowercase L, not an uppercase, by which I mean loops is a word that doesn't necessarily mean exactly one thing, but stands in for the emerging set of interaction patterns that allow agents to improve their own work over time and that allow us as humans leading the agents to improve their ability to work over time.

AI工程世界博览会的第三个趋势是:AI工程进入企业。

A third trend from the AI Engineering World's Fair was AI engineering enters the enterprise.

从某种程度上说,这是"前向部署工程"方向。

Now, this was in some ways the forward-deployed engineering track.

考虑到现在几乎每家AI公司都认定,为了让他们的产品真正在目标客户内部发挥作用,他们必须用更加手把手的实施团队来支持客户,所以在大会上看到"前向部署工程"这个方向一点也不令人惊讶。

And given that pretty much every AI company has now determined that to get their products to actually work inside their intended customers, they have to support their customers with a much more hands-on type of implementation force, it's probably not all that surprising to see an FTE track show up at the event.

现在,这里出现的一个有趣的问题是:这类"前向部署工程"式参与最终应该呈现出什么结果?

Now, one of the interesting questions that seems to come up is what the end result of an FTE type of engagement is supposed to actually look like.

Cursor的Pauline Brunette说:当我们结束这类合作时,在我们这边,我们部署了云端Agent、长期运行的Agent、自动化流程,并且我们在Cursor SDK之上构建了应用程序。

Cursor's Pauline Brunette said, when we walk away at the end of the engagements, in our case, we have deployed cloud agents, long-running agents, automations, and we've built applications on top of our Cursor SDK.

事实上,在大会上被频繁使用的另一个更广义的术语来描述这类企业基础设施,就是"软件工厂"。

And indeed, a more general term for this set of enterprise infrastructure that apparently was being used a lot at the event was software factory.

然而,这个术语如果要说的话,甚至比"循环"这样的词更不明确。

And yet that, if anything, is perhaps even less defined than something like loops.

Zach Lloyd是Warp的CEO,也是大会的演讲者,他们公司将自己视为一个软件工厂平台。

Zach Lloyd is the CEO of Warp and was a speaker at the event, and the company views itself as a software factory platform.

在接受Late in Space的采访时,CEO Zach Lloyd解释了他最初是在哪里遇到"软件工厂"这个术语的。

In an interview with Late in Space, CEO Zach Lloyd explained where he first came across the term software factory.

Zach说:我们开始时是做一次性自动化,在云端运行一个Agent。

Zach said, We started with one-off automation, run an agent in the cloud.

很多平台都是从那里起步的。

A lot of platforms began there.

然后变成了在云端按定时任务运行一个Agent。

Then it became run an agent in the cloud on a timer.

下一个问题是:最有价值的自动化循环是什么?

The next question was, What was the most valuable loop to automate?

答案基本上就是软件工程的主循环:分类、规格制定、实现、审查、验证、发布和监控。

The answer is basically the main loop of software engineering: triage, specification, implementation, review, verification, shipping, and monitoring.

当谈到企业和软件工厂之间的关系时,Zach认为,企业面临的一大工作就是弄清楚其工作流程的哪些部分需要自动化,以及哪些环节应该让人类介入。

Now, when it comes to the relationship between an enterprise and a software factory, Zach argued that one of the big pieces of work for enterprises is figuring out which parts of their work lifecycles need to be automated and where humans should be brought into the loop.

有趣的是,在后续的一篇X平台博客中,Zach认为工厂框架之所以变得更加重要,不仅是因为Agent现在更有能力完成编码生命周期的大部分工作,而且还因为,"广泛使用交互式Agent已经带来了一系列问题,从成本控制到治理再到安全。"

And interestingly, in a follow-up blog post on X, Zach argued that the reason the factory framework has become more important is not only that agents are now more capable of doing big chunks of the coding lifecycle, but that also, quote, that extensive interactive agent use has created a number of problems, from cost controls to governance to security.

Zach说,根本问题在于,交互式Agent有操作人员,而这些操作者使用Agent的方式各不相同,并不总是能最大化商业价值,反而可能带来风险。

The root problem, Zach says, is that interactive agents have human operators, and those humans all use them in very different ways, which don't always maximize business value and can create risk.

他给出的一个例子是,某位操作者总是使用最贵的模型,即使对于某个任务根本没必要。

An example he gives is a human always using the most expensive model even when they don't need to for a given task.

他说,工厂方法背后的理念是建立一个能最小化人为差异、最大化产出,并配备确保安全与合规控制的系统。

The idea behind a factory approach, he says, is to set up a system that minimizes human variability and maximizes output, with controls that ensure security and compliance.

现在,这很大程度上是一场软件领域的对话,但我认为,这些软件工厂试图解决的许多问题,当Agent进入其他知识工作领域时,也会同样出现。

Now, this is very much a software conversation, but I think that a lot of the problems that these software factories are trying to solve for are going to find themselves repeated as agents find their way into other areas of knowledge work.

那个总是使用最贵模型的例子,绝对不仅限于软件开发。

That example of always using the most expensive model is certainly not limited to software development.

或者他举的另一个例子:操作者无意中安装了权限过大的MCP,从而造成了安全问题。

Or his other example, a human inadvertently creating a security issue by installing MCPs that have too much access.

随着每个人——无论是在软件工程、产品开发、市场营销还是销售中——都开始熟悉如何管理上下文来支持他们使用的Agent,这个问题也会成为他们需要面对的问题。

As everyone gets familiar with managing context to support the agents that they're using, whether it's in software engineering or product development or marketing or sales, this is going to be an issue for them as well.

大会的第四个趋势是:编码Agent正在取代IDE成为开发者的界面。

The 4th trend at the event was coding agents replacing IDEs as the developer interface.

我认为这已经预示了我们正在看到的一个更大的交互模式转变。

And I think that this is one that's already indicative of a larger interaction pattern trend that we're seeing shift.

几周前的一集节目专门讲了新的Claude标签功能,其中每个Claude实例并不与单个打标签的用户关联,而是与它们所参与的特定频道中的一组特定权限和上下文关联。

An episode from a couple of weeks ago was all about the new Claude tag, with each of those instances of Claude not being connected to the individual user who tagged it in, but instead connected to a specific set of permissions and context within the channel that they were engaging.

换句话说,可能有一个单一的Claude标签实例,它带有特定的上下文、特定的工具访问权限和特定的权限,你组织里市场频道中的每个人都与这个实例交互,而这个实例的权限、工具和上下文完全不同于销售聊天群里的人所交互的那个Claude。

In other words, there might be a single Claude tag instance with a specific set of context and a specific set of tool access and a specific set of permissions that everyone in your organization's marketing channel is interacting with that's different than the permissions and tools and context of the Claude that the folks over in the sales chat are interacting with.

关于Claude Tag的发布,有一件事特别引人注目,那就是Claude Code的创始人之一Boris说,他们大约65%的新代码,我想是,现在是在Claude Tag的对话中开始的。

One of the things that was super notable about the Claude Tag announcement was that Boris, one of the creators of Claude Code, said that something like 65%, I think, percent of their new code was now being initiated in Claude Tag chats.

这是我们在管理这种智能的方式上,界面和交互模式的重大、重大转变。

A major, major shift in the interface and interaction pattern with which we manage this intelligence.

现在,关于这类交互模式,有意思的一点是,这是大实验室们没有在等非工程师去适应工程师式实践的一个领域。

Now, one of the things that's interesting about these types of interaction patterns is that this is one area where the big labs are not waiting for non-engineers to adapt to engineer-type practices.

他们把功能从工程师优先的体验中拖过来,放到为所有消费者提供的主流体验中。

They are dragging functionality over from the engineering-first experiences into the main experiences that they're offering for all consumers.

最好的例子可能就是ChatGPT Work,它实际上是拿Codex直接塞进主ChatGPT应用里,让大家都能用。

The best example of this is probably ChatGPT Work, which effectively takes Codex and plops it into the main ChatGPT app for everyone to use.

根据Richard的说法,第五个趋势是每个agent平台都在围绕技能来构建。

The 5th trend according to Richard was every agent platform building around skills.

Richard引用了Adi Osmani的定义,写道,技能编码了资深工程师在构建软件时使用的工作流程、质量关卡和最佳实践。

Richard referenced Adi Osmani's definition, writing, skills encode the workflows, quality gates, and best practices that senior engineers use when building software.

它们被打包好,让AI agent在开发的每个阶段都能一致地遵循。

They are packaged so AI agents follow them consistently across every phase of development.

在活动中,Vercel的Andrew Qiu说技能作为便携的按需知识很有用,而Introspection的联合创始人Roland Gavrelescu认为,AI工程已经从agent工具转向了agent技能。

At the event, Vercel's Andrew Qiu said that skills were useful as portable on-demand knowledge, and Introspection co-founder Roland Gavrelescu argued that AI engineering had shifted from agent tools to agent skills.

Google DeepMind的Philip Schmidt认为,技能减少了对编排代码的需求,让开发者可以不用写代码就能使用agent。

Google DeepMind's Philip Schmidt argued that skills reduced the need for orchestration code, allowing developers to use agents without code.

一位演讲者Paul Bakkes,甚至还讨论了他的一个关于agent技能的项目,叫做Impeccable,那是一个开源的设计技能系统,用来改进编码agent创建的界面。

One speaker, Paul Bakkes, even discussed his project around agent skills called Impeccable, an open-source design skill system for improving the interfaces that coding agents create.

现在,比那项目更有趣的是,Paul认为技能工程将会成为一门独立的学科。

Now, more interesting even than that project, Paul argued that skill engineering is going to be a discipline in its own right.

而当我们思考,现在和AI工程师们进行的哪些对话,将来会和AI知识工作者们进行时,我觉得技能绝对在名单上排得非常靠前。

And as we wonder which of the conversations happening with AI engineers right now are going to be happening with AI knowledge workers in the future, skills, I think, are incredibly high on the list.

想想这个。

Think about this.

很有可能,如果你在某个工作上非常擅长,你会经常遇到一些情况,连最好的模型,包括Fable-5和GPT-5.6 SOL,都达不到你想要的那么好。

It is highly likely that if you are very good at your job, you frequently run across things where even the best models, including Fable-5, and GPT-5.6 SOL don't do as well as you want them to.

当然,这也不是全部。

Now, that's not everything.

有些事它们会很擅长,但尤其是当你在某方面特别特别出色时,你往往会更快地看到AI的固有局限。

Some things they're going to be great at, but especially when you're really, really good at something, you tend to see the inherent limitations of AI even faster.

那些提倡技能工程的人会说,与其干等着模型靠Babel 6或者别的什么来改进,不如开始琢磨怎么使用技能,怎么把知识和规则编码进去,我敢说,还有品味,这可能是一条更高效的路。

The folks who are advocating skill engineering would argue that rather than just waiting for the models to improve with Babel 6 or what have you, starting to figure out how to use skills, how to encode knowledge and rules, and dare I say it, taste in skills, is potentially a more productive path.

在闭幕主题演讲中,Y Combinator总裁Garry Tan认为,有效使用技能,包括在销售、支持和财务这些业务职能里,是成为真正AI原生组织的根本要素。

In his closing keynote, Y Combinator President Garry Tan argued that using skills effectively, including in business functions including sales, support, and finance, was fundamentally integral to being an actual AI-native organization.

然而,技能本身也可能成为某种自主性的陷阱。

And yet skills themselves can be their own sort of autonomy trap.

AI Engineer World's Fair的参会者Tyler Brown写道,他从这次活动中得到的教训之一,就是要重新审视并重新实现你的技能。

AI Engineer World's Fair attendee Tyler Brown wrote that one of his lessons from the event was to revisit and reimplement your skills.

每次有新模型发布,他写道,就像你有一个孩子从中学校园长到了高中。

Each time there's a new model release, he wrote, it's as if you have a kid that grows from middle school to high school.

你得为ta们改变课程,才能从新模型中得到好处。

You have to change the curriculum for them to get the benefits of the new model.

不过,不是那个,而是Tyler更广泛的观察,我认为贯穿了Richard识别的这些趋势,以及我在活动周围看到的很多讨论。

Yet it was not that, but Tyler's broader observation that I think runs throughout the trends that Richard identified and also a lot of the chatter that I've seen around the event.

今年AI Engineer World's Fair的某些东西,就是感觉不一样。

Something about this year's AI Engineer World's Fair just hit different.

去年是“让agent自由发挥”的一年。

Last year was the year of let the agents rip.

今年是意识到,没有结构的自主性,产生混乱的程度和产生杠杆的程度一样多。

This year was the year of realizing that autonomy without structure creates as much slop as leverage.

简而言之,虽然不是普遍现象,但大体上,AI工程师中一个突出的情绪是,重新夺回,或者说重新浮现,随便你怎么说,人在这些agent系统中心的位置。

In short, while it's not universal, broadly speaking, one of the standout sentiments from the AI engineering folks is the reclamation or reemergence, however you want to put it, of the human at the center of these agentic systems.

早在三月,我发过推文,说,你们可以说我疯了,但我认为,那些给团队里每个人配一队agent的公司,会碾压那些用一队agent替换整个团队的公司。

Back in March, I tweeted, call me crazy, but I think the companies that give everyone on their team a team of agents are going to kick the slats out of the companies that replace their teams with a team of agents.

而我觉得,我也希望,我们开始看到这种情绪结构化地进入主流。

And I think, and I hope, that we're starting to see that sentiment move structurally into the mainstream.

好了各位,以上就是这个月在 AI 工程师世界博览会上定义 AI 工程的五大趋势,以及它们对非 AI 工程师的其他人的影响。

So guys, those are the 5 trends that defined AI engineering at the AI Engineer World's Fair this month, and their implications for other people who aren't just AI engineers.

非常感谢 Richard McManus 捕捉到这一切,当然还有 Swyx 和团队为我们举办这场活动。

Big thanks to Richard McManus for capturing all of this, and of course, Swyx and the team for putting on the event.
M1
M127:47

不过,今天就到这里,以上是今天的 AI Daily Brief。

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

一如既往,感谢您的收听或收看,下次再见,祝您平安。

Appreciate you listening or watching as always, and until next time, peace.
已剔除 4 处广告(点击展开查看)
M1
M10:26广告 · 已剔除

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

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
M110:00广告 · 已剔除

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

Learn more at kpmg.com/us/sophisticated.

That's kpmg.com/us/sophisticated.

M1
M110:25广告 · 已剔除

Robots and Pencils is already launching and scaling.

Agentic and generative AI in production at large enterprises in weeks.

AWS Advanced Tier Pattern Partner more than doubled in a year.

And they're hiring.

50 open roles.

If you're someone who knows this moment is different, who wants to be inside it, not watching it, this is worth a look.

At Robots and Pencils, the best ideas win, and the team is purposefully kept super high quality.

This is the kind of place you look back on as the best decision you ever made.

Take a look at robotsandpencils.com/careers.

If you're looking to adopt an agentic SDLC, Blitzy is the key to unlocking unmatched engineering velocity.

Blitzy's differentiation starts with infinite code context.

Thousands of specialized agents ingest millions of lines of your code in a single pass, mapping every dependency.

With a complete contextual understanding of your codebase, enterprises leverage Blitzy at the beginning of every sprint to deliver over 80% of the work autonomously.

Enterprise-grade, end-to-end tested code that leverages your existing services, components, and standards.

This isn't AI autocomplete.

This is spec and test-driven development at the speed of compute.

Schedule a technical deep dive with our AI experts at blitzy.com.

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

This episode of the AI Daily Brief is brought to you by HyperAgent, 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.

M1
M127:29广告 · 已剔除

If you want to follow my advice and stay closer to the pulse of what AI engineers and developers are talking about, the single best source for that is going to be Latent Space, latent.space.

And might I particularly recommend their AI news email weekday roundups?

They can be really dense, but they aggregate a ton of discourse and are very much worth your time.