← 返回任务列表

Why GPT-6 Astra Is So Significant and So Confounding

269 段 · 1 位说话人 · 原片 29:10
M1
M10:00

GPT-6 Astra 来了,如果你觉得大家的第一印象有点奇怪,那你不孤单。

GPT-6 Astra is here, and if you feel like people's first impressions are a little strange, you're not alone.

这应该算是 OpenAI 推出的一个超级先进的模型。

This is supposed to be this crazy advanced model from OpenAI.

它是一个全新的预训练基础。

It's a totally new pre-training base.

它太强大了,所以他们不得不暂时保密了一段时间。

It was so powerful they had to keep it under wraps for a little while.

它本应是用来回应 Anthropic 的 Fable 和 Mythos 的。

It's supposed to answer Anthropic's Fable and Mythos.

当然,社交媒体上出现的一些东西,比如那些超厉害的一次性生成游戏、3D 模型之类的,确实让人印象深刻,至少视觉上很震撼。

And certainly some of the stuff that's showing up on social, these incredible one-shot games and 3D models and things like that, are really impressive, at least visually.

但另一方面,在一些更基础的领域,人们并不觉得它有多大的提升。

But then in some other more basic areas, people aren't necessarily finding it all that much better.

那么,Astra 的故事到底是什么呢?

So what is the story of Astra?

挑战在于,GPT-6 Astra 并不是一个效率型 AI 模型。

The challenge is that GPT-6 Astra is not an efficiency AI model.

换句话说,它的重点不是让你把现在做的事情做得更好。

In other words, it is not about doing what you currently do better.

它从头到尾都是一个机会型 AI 模型,会挑战你换个角度思考自己能做什么。

It is through and through an opportunity AI model that is going to challenge you to think differently about what you can do.

而且,它不仅带来了新的能力集成——比如计算机使用能力——也带来了一种新的默认交互模式。在这种模式下,我们不再需要坐在电脑前点来点去、打字输入,而是可以轻松地跟电脑说话,让它替我们完成以前需要手动操作的事情。

And not only is it bringing a new capability set with its computer use capabilities, it's also bringing along a new default interaction pattern where increasingly we will not be sitting there clicking around and typing to our computer, but instead we'll be ambiently talking to it as it does all the things that we used to do.

总之,它带来的变化非常巨大,但很难用一个周末的测试就全部概括出来。

In short, the changes that it represents are immense, but not easy to package up in a weekend of testing.

所以,我们一起来尝试找到看待 Astra 的正确方式,也许能帮你弄清楚该从哪里开始。

So let's try to figure out together the right way to look at Astra and maybe help you figure out where to begin.

《AI 每日简报》是一个每日更新的播客和视频节目,关注 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,在我们深入之前,先说几个快速通知。

Alright friends, quick announcements before we dive in.
M1
M11:45

在进入这期重磅节目之前,还有两件小事要说。

2 Other quick things before we get into this very big episode.

首先,我们新一期的《高管代理领导力项目》和《高管进阶项目》就要开课了,这是最后的上车机会。

First is a last call for our new cohorts of our Executive Agent Leadership Program and our Executive Catch-Up Program.

这两个项目都会在本周启动,不过你现在加入也还来得及。

Both of those are kicking off this week, although you are not too late to join right now.

相关信息可以在 aidailybrief.ai 的顶部找到链接。

You can find the information to that linked off the top of aidailybrief.ai.

当然,你也能在顶部看到新推出的《团队多人 AI 冲刺》项目的链接。

And of course, you can also find linked at the top the new Multiplayer AI Sprint for Teams.

这是 AIDB 继《AIDB 新年》和《爪营》、《AgentOS》和《夏日冒险》之后推出的最新免费项目。

This is the latest AIDB free program following AIDB New Year's and Claw Camp, AgentOS and the Summer Adventure.

而这次是第一个专门为团队设计、让大家一起使用代理的项目,而不仅仅是个人。

And this one is the first one that's designed specifically not just for individuals, but for teams to come together to use agents together.

我相信,从单机 AI 到多人 AI 的转变,将是今年秋天定义最成功、最会利用 AI 的公司的重要趋势之一。

Which I believe this shift from single-player to multiplayer AI is going to be one of the biggest trends defining the best, most successful AI-using companies this fall.

你可以在主网站或 multiplayerai.ai 找到这个冲刺项目,再次强调,完全免费。

You can find the sprint, which is again completely free, linked from the main website or at multiplayerai.ai.

欢迎回到《AI 每日简报》。

Welcome back to the AI Daily Brief.

今天我们要讨论的是很长一段时间以来——至少是 2026 年全年——最重磅、也最令人困惑的模型发布之一。

Today we are talking about one of the most significant and confounding model releases for a very long time, certainly for all of 2026.

你看,我刚计划好劳工节周末和我的生日旅行,就知道肯定会有一波重要模型发布,正好赶上我起飞的时候,OpenAI 果然没让我失望。

Now inevitably, as soon as I planned a trip for Labor Day weekend and my birthday, you just knew that we were going to get a bunch of important model releases right as I was on my way wheels up, and OpenAI did not disappoint.

Astra 这个模型我们已经等了一段时间了。

Now Astra is one we have been waiting for for some time.

它比之前的 GPT-5.6 系列要更直接地回应 Anthropic 的 Claude 系列模型。

It is intended much more than the GPT-5.6 class ever was to be an answer to Anthropic's Claude class models.

不像从5.5到5.6那种升级,Astra从一开始就是一次根本不一样的训练。虽然这个模型名义上早就准备好了,但Astra属于新一代模型,而这一代模型的运行模式是,由于安全和安保方面的考量,发布前需要推迟一段时间。

Unlike the jump between something like 5.5 and 5.6, Astra started as a fundamentally different training run, and while the model has nominally been ready for some time now, Astra is of the new generation of models which now operate in the paradigm of having to be delayed somewhat before their release based on safety and security considerations.

事实上,在发布公告中Sam Altman说:“我们花了额外的时间来确保能达到这个能力级别所需的安全和对齐标准,但我们相信你会发现这等待是值得的。”

Indeed, in his announcement, Sam Altman said, it took us some extra time to ensure that we could meet the safety and alignment standards required for this capability level, but we think you will find it worth the wait.

不过,这次发布过程有点混乱。

Now, the rollout was a little bit messy.

模型是在周四公布的,但只有加入了以网络安全为核心的Daybreak项目的合作伙伴才能使用。

The model was announced on Thursday, but availability was limited to partners enrolled in the cybersecurity-focused Daybreak program.

包括我在内的很多人都非常着急,想赶紧用上这个期待了那么久的东西。

And a lot of people, myself included, were very antsy to get our hands on the thing that had been so long anticipated.

OpenAI的Thibault承诺,对于那些没能用上Astra的日常付费订阅用户,会提供一次保留的重置机会。

Thibault from OpenAI promised that for everyday paid subscribers didn't have access to Astra, they would give them one banked reset.

Altman也承诺,会尽快向API和ChatGPT订阅用户全面开放。

And Altman promised to begin the broad rollout to their API and ChatGPT subscribers as soon as possible.

到了周五晚上,大家就都能用上新模型了。

And by Friday night, the new models were in everyone's hands.

Astra的发布公告视频在X上彻底爆火。

The announcement video for Astra went absolutely mega viral on X.

截至周二早上录制的时候,这个视频已经被观看了超过1.32亿次,被收藏超过10万次。

At the time of recording on Tuesday morning, it has been viewed more than 132 million times and saved more than 100,000 times.

这个视频大概3分钟,内容就是人们在房间里坐着或来回踱步,墙上投着一块大屏幕,对着一个打开的笔记本电脑说话,看着Astra帮他们做事。

The video is 3 minutes of people sitting in or pacing around a room with a big projection of a screen on the wall, talking to an open laptop and watching Astra do things for them.

用例分布在商务、个人效率还有娱乐几个方面。

The use cases are split between business, personal productivity, and recreation.

有一个人在eBay上架商品,一个人在审核合同,一个人在制作演示文稿,还有一个人在做一个游戏。

One person is listing an item on eBay, one person is doing a contract review, one person is building a presentation, one person is making a game.

视频还展示了这些人摆脱了浏览器窗口、鼠标这些烦人的老东西,同时处理多件事。

They also show those same people untethered from old pesky things like browser windows and mouses doing multiple things at once.

他们在前台处理主要的效率任务,在后台做一些行政或简单的任务,比如点餐或者预订网球场。

While in the foreground they are working on their main productivity task, in the background they're working on some administrative or simple task like ordering food or booking a tennis court.

而且光看那个视频,你就能感觉到Astra这个模型确实会有点不一样。

And just from that video alone, you could tell that there was kind of going to be something different about this Astra model.

首批上手感受很快就出来了,其中一些结果可以说是意料之中。

The first takes quickly started rolling in, and some parts were kind of expected.

Every的Dan Shipper总结说,相比5.6,这是一次巨大的升级,但同时有一些恼人的习惯,让它没法在最高水准上和Fable一较高下。

Dan Shipper from Every summed up, it's a big upgrade from 5.6 so well with some frustrating habits that keep it from matching Fable at the top end.

Dan说他试过的最好的写作模型就是它,非常容易引导,产出也很少垃圾内容。

Dan said that it was the best writing model he'd tried, being very easy to steer and producing very little slop.

他发现它非常擅长使用电脑,这显然是这次模型发布的一个重点。Dan说,Astra可以——引述原话——“连续几个小时使用各种复杂应用来完成工作。”

He found that it was incredibly good at computer use, clearly a major priority for this model launch, with Dan saying that Astra can, quote, go for hours at a time using complicated apps to get work done.

他的这番话也呼应了整个周末X上都在刷屏的话题——顺便说一句,如果你只是在听音频,这可能是你转而去看视频版的好理由。Dan还指出,Astra在制作3D游戏和可视化内容方面也非常厉害。

Harkening to something that would be all over X all weekend, and which, by the way, is a good reason to potentially switch to the video version of this if you're just listening, Dan also pointed out that Astra was incredibly good at making 3D games and visualizations.

但他也说,Astra有把事情过度复杂化的倾向。

But he said, Astra also had a tendency to overcomplicate things.

比如,他想要一个简单的界面,结果上面到处是多余的标签、按钮和功能。

For example, when he asked for a simple interface, he got extra labels, extra buttons, and features all over the place.

他说,原话,“它还是无法像Fable那样直观地理解你的提示词,做出令人满意又不过度复杂的东西。”

He suggested that it, quote, doesn't quite match Fable's ability to intuitively understand your prompt and do something delightful without overcomplicating.

那么,一方面,OpenAI发布的新模型——即使是为了让他们超过Anthropic和Fable——在某些方面让一些忠实用户更喜欢Fable,这一点其实并不太令人惊讶。

Now, on the one hand, it's not all that surprising that a new release from OpenAI, even one that is meant to catapult them ahead of Anthropic and Fable, had aspects where people who were dedicated users preferred Fable more.

与此同时,看了很多其他人对Astra的初步实验之后,再去读Every的那篇氛围评测,感觉非常像是在用一套并不完全合适的旧标准来判断一个新东西。

At the same time, reading the Every vibe check after seeing a lot of other people's first experiments with it felt very much like trying to judge a new thing with old criteria that didn't quite fit.

它的能力组合与人们用来评估它的工具和测试之间,几乎有些格格不入。

There was something almost discordant about the capability set and the tools and tests people were using to evaluate it.

当然,一个更戏剧化的例子是Artificial Analysis的智能指数分数公布的时候。

Certainly an even more dramatic example of this came when the Artificial Analysis Intelligence Index scores came out.

最初,Astra在这个智能指数上得了61分。

Initially, Astra scored a 61 on the Intelligence Index.

这跟GPT-5.6 SOL的得分一样,但整整比Fable 5.1低了5分。

That gave it the same score as GPT-5.6 SOL, and left it a full 5 points behind Fable 5.1.

这也意味着,同样在上周末发布的Meta Muse Spark——我们还没机会深入聊它,但会聊的——比GPT-6 Astra高出了一点。

It also meant that Meta Muse Spark, which was also released at the end of last week and which we haven't had a chance to dig into here yet, but we will, was one point ahead of GPT-6 Astra.

对一些人来说,这正好印证了我刚才说的观点:我们用来评判旧模型的测试,实际上并不能真正反映它们可能产生的影响。

And for some, this was evidence of exactly the idea that I just mentioned, that the tests that we're using to judge old models don't really reflect what their impact is likely to be.

具体到人工分析部分,这个指数比较偏向于那些奖励记忆力、事实背诵的基准测试。

Now, when it comes to artificial analysis specifically, the index is fairly skewed towards benchmarks that reward, for example, memorization of facts.

它们对高级编程的测试较少,关于高级计算机使用的测试就更少了。

They have fewer tests that require advanced coding and even fewer still that are about advanced computer use.

而且,别以为这只是那些在智能指数上得分不高的人面临的问题——AI团队其实在上周末就赶着更新了指数,发布了新版本4.2。

And lest one think that this is just a problem for the people who don't get a great score on the Intelligence Index, the AI team actually rushed out their update to the index over the weekend, announcing their new index version 4.2.

这个更新通过AI Briefcase测试加强了对智能体任务的评估,同时对许多现有测试的结构和权重做了调整。

The update includes more emphasis on agentic tasks with their AI Briefcase test, plus a different structure and weighting on many of the existing tests.

在新版4.2中,GPT-6 Astra仍然落后于Fable 5.1,但领先于其他所有模型。

In that new version 4.2 iteration, GPT-6 Astra was still behind Fable 5.1, but was ahead of everything else.

大家好。

Hello everyone.
M1
M110:41

那么,这到底是怎么回事呢?

So what is going on here?

稍后,我们会看看OpenAI选择重点宣传的那些基准测试,以及人们在周末分享的使用Astra做的那些独特而不同的事情。

Well, in just a minute, we're gonna look at the benchmarks that OpenAI chose to highlight and some of the unique and different things that people share that they were doing with Astra over the weekend.

但先别埋没重点:GPT-6 Astra不是一个效率型AI模型。

But to not bury the lead, GPT-6 Astra is not an efficiency AI model.

这是一个机会型AI模型。

This is an opportunity AI model.

换句话说,GPT-6 Astra不是让你把现有工作做得更好,而是拓展你能做的事情。

In other words, GPT-6 Astra is not about doing what you currently do better.

这让Astra变得极其有趣,可能极具价值,但也充满挑战。

It's about expanding what you can do.

今天这期节目,部分内容就是回顾一些类似的机会型AI模型发布的案例,看看我们当时是否以及如何发现这些模型的不同之处,以及花了多长时间。

That makes Astra extremely interesting, potentially extremely valuable, but also challenging.

不过,我们先回到其他基准测试,具体来说是OpenAI选择重点宣传的那些。

And part of what we're gonna do in today's episode is look at some comparable past examples of this type of opportunity AI model release to see if and how we figured out what made those models different and how long it took.

即便你对实验室的基准测试持保留态度,他们选择把重点放在哪里,也能在一定程度上告诉我们,OpenAI团队如何看待Astra相对于其他模型的独特差异化价值。

But first, let's go back to some other benchmarks, specifically to OpenAI and what they chose to highlight.

现在已经成为惯例的是,OpenAI展示基准测试时,不再仅仅列出对比性能数字,而是用图表来绘制性能与成本的关系,把效率直接纳入核心分析。

Even if one takes lab benchmarks with a grain of salt, where they choose to put their focus can still tell us something about how the OpenAI team sees the unique and differentiated value of Astra relative to other models.

在编程基准测试TerminalBench 4.0上,该模型得分57.6%,超过了Fable 5.1的55.8%,也比GPT-5 6 SOL的37.3%有大幅提升。

As has become the norm, OpenAI now shows benchmarks not just as a list of comparative performance numbers, but on charts that can map performance versus cost, building efficiency right there into the core analysis.

在DeepSWE上,Astra得分74.1%,略高于Fable 5.1的73.7%。

On coding benchmark TerminalBench 4.0, the model scored 57.6%, which beat Fable 5.1 at 55.8% and represented a large jump from GPT-5 6 SOL at 37.3%.

有趣的是,在这两个基准测试中,Astra在中等或较高设置下取得了最佳成绩,而将努力水平调到最高时,成绩反而略有下降。

On DeepSWE, Astra scored 74.1%, which was slightly higher than Fable 5.1 at 73.7%.

这说明更高的努力程度可能导致过度思考、偏离方向,并且重要的是,OpenAI选择强调Astra不仅得分更高,而且成本远低于Fable 5.1之类的模型。

Now interestingly, for both of those benchmarks, Astra achieved its best score on high or extra settings, and actually declined a little bit with effort set to max.

在衡量计算机使用的Automation Bench上,分数更是惊人:Astra得分41.1%,远高于Fable 5.1的31.4%和GPT-5.6 SOL的18.1%。

This suggests that higher effort could mean overthinking and getting sidetracked, and importantly, OpenAI chose to highlight that not only did Astra score higher, but they did it a lot more cheaply than, for example, Fable 5.1.

说实话,要是只能选一件事让OpenAI特别想让你了解Astra,那就是它在使用计算机方面带来的变革性差异。

Now on Automation Bench, which measures computer use, the scores were off the chart.

在发布公告中,他们明确将其称为世界上最好的计算机使用模型,声称Astra在计算机使用的速度、准确性和安全性方面开辟了新前沿。

Astra scored 41.1%, which was much higher than Fable 5.1 at 31.4% and GPT-5.6 SOL at 18.1%.

OpenAI还强调了在科学和数学方面能力的巨大提升。

And honestly, if you had to pick just one thing that OpenAI really wanted you to know about Astra, it is how transformatively different at using computers it is.

在TerminalBench Science上,Astra得分64.6%,超过了Fable 5.1的52.6%,并且完全碾压了GPT-5.6 SOL的22.4%。

In the announcement post, they call it explicitly the world's best computer use model, claiming that Astra marks a new frontier in the speed, accuracy, and safety of computer use.

在Frontier Math Tier 4上,Astra取得了近乎完美的97.6%,比Fable 5.1的90.2%又是一次飞跃。

OpenAI also emphasized a big increase in capabilities for science and math.

在 TerminalBench 科学测试中,Astra 得了 64.6%,超过了 Fable 5.1 的 52.6%,而且把 GPT-5.6 SOL 的 22.4% 远远甩在后面。

On TerminalBench Science, Astra scored 64.6%, beating Fable 5.1 at 52.6%, and absolutely demolishing GPT-5.6 SOL's score of 22.4%.

在 Frontier Math 第四级测试里,Astra 拿到了近乎满分的 97.6%,相比 Fable 5.1 的 90.2% 又上了一个台阶。

And on Frontier Math Tier 4, Astra achieved a near-perfect score of 97.6%, another jump from Fable 5.1's score of 90.2%.

在网络安全领域,先进的计算机使用似乎也带来了能力上的重大突破。

In cybersecurity, advanced computer use seems to have led to a big breakthrough in capabilities as well.

Astra在Exploit Bench上所有努力级别都获得了100%的分数,说明它几乎不费吹灰之力就能构建和执行漏洞利用。

Astra scored 100% on Exploit Bench across all effort levels, suggesting it can build and execute exploits with little difficulty.

使用一个针对近期披露漏洞的内部基准,Astra得分39%,而GPT-5.6 SOL只有5.5%。

Using an internal benchmark of recently disclosed vulnerabilities, Astra scored 39% compared to just 5.5% for GPT-5.6 SOL.

除了基准测试,GPT-6 Astra的发布公告还强调了一些使用案例,很明显OpenAI不只是把它宣传成一种更好的撰写文档和做研究的方式。

Now, in addition to benchmarks, the GPT-6 Astra announcement post also highlighted some use cases, and it's very clear that OpenAI is not just pitching this as a better way to write documents and do research.

他们提到了游戏开发、电路板印刷、汽车变速箱设计,还有很多其他完全超出这类模型发布公告通常关注的传统用例范围的东西。

They point to game development, circuit board printing, car transmission design, and again, many other things that are very much outside the classic set of use cases that these sort of model announcement posts focus on.

这在上周末X平台上走红的各类实验中也得到了体现。

And this was reflected in the types of experiments that were going on viral on X over the weekend.

OpenAI的Thomas Ricard写道,Astra非常擅长3D建模,然后展示了一个逐步讲解,演示了他是如何为他们的博客文章构建一个演示房屋漫游的。

OpenAI's Thomas Ricard wrote, Astra is very good at 3D modeling, and then showed a walkthrough of how he built a demo house walkthrough for their blog post.

Arita的Peter Gostev在Blender中使用Astra重现了苹果一段著名的发布会视频,内容是Craig Federighi做一些奇怪的跑酷动作。

Arita's Peter Gostev used Astra in Blender to recreate a famous Apple announcement video of Craig Federighi doing some weird parkour moves.

Blender这个3D创建程序在Astra相关的帖子里到处都是。

Now Blender, which is a 3D creation program, was all over Astra-related posts.

Alex Oliver展示了它的能力和持久性,在Blender里构建了一只逼真的蝙蝠,说它一直在做直到我的token用完。

Alex Oliver showed its capability and persistence by building a photorealistic bat in Blender, saying it just kept going until I ran out of tokens.

Yunfan Yi模仿了OpenAI博客上的房屋游览用例,写道,给它一个Zillow房源链接。

Yunfan Yi modeled the house tour use case that had been on the OpenAI blog post, writing, give it a Zillow listing.

它可以根据房源照片对房子进行3D建模,并创建一个很酷的宣传视频。

It can 3D model the house based on the listing photos and create a cool promotional video.

这个视频是一次性生成的,还有一些细节错误,但我相信如果让它进一步打磨,效果会好得多。

This video was created in one shot, and there are still some wrong details, but I'm sure it would be much better if I asked it to polish further.

Duncan Trussell构建了一个看起来像老式VHS录像带的恐怖版Backrooms,分享了一个看起来像老VHS的东西,但实际上不是AI生成的视频,只是完全在Blender里建的。

Duncan Trussell built a creepy VHS-looking version of Backrooms, sharing something that looks like an old VHS, but actually wasn't AI-generated video, just was all built in Blender.

Duncan接着说,我想知道它在3D角色绑定和动画方面表现如何。

Duncan went on, I wonder how it is at rigging 3D characters and animating them.

我接下来试试看。

I'll try that next.

他又补充道,是的,我们完蛋了。

He followed up, yeah, we're cooked.

它做到了,一点问题都没有。

It did it, no problem.

绑定了一个3D角色,把它放进了Backrooms,让它走来走去。

Rigged a 3D character, put it in the Backrooms, and made it walk around.

如果你从来没摆弄过Blender,这可能看起来没什么大不了的,但绑定真的很烦人。

If you've never messed with Blender, this might not seem like a big deal, but rigging sucks.

我之前用其他版本的GPT试过这个,从来没真正成功过。

I've tried this with other versions of GPT, and it never really works.

这个版本一次提示就完美搞定了。

This did it perfectly in one prompt.

我用了GPT-6 Astra创建了一个3D网站,把一辆特斯拉Model X拆解成334个建模零件。

I used GPT-6 Astra to create a 3D website that pulls apart a Tesla Model X into 334 modeled pieces.

Google Applied AI的产品经理Baron Roth写道,Astra帮我赢了一场和未婚妻的争论——关于洗碗后眼镜正着放干得快还是倒着放干得快——通过模拟水分子和附近空气湿度的物理原理。

Google Applied AI PM Baron Roth wrote, Astra helped me win an argument I had with my fiancée on whether glasses dry faster right side up or upside down after dishwashing by simulating the physics of the water molecules and nearby air humidity.

很多人都想知道这是不是一个值得推荐的用例,Michael Briggs回应说,科学一小步,沙发一大步。

Now, many wonder whether this is an advisable use case or not, with Michael Briggs responding, One small step for science, one giant leap towards sleeping on the couch.

和你听到的差不多,很多人也通过构建游戏来测试Astra。

As you might expect from what you've heard so far, lots of people also tested Astra by building games.

Ethan Mollick写道,我让GPT-6 Astra把Zork——那款经典的1977年文字冒险游戏——变成一个完整的3D动作冒险游戏。

Ethan Mollick wrote, I asked GPT-6 Astra to turn Zork, the classic 1977 text adventure, into a full 3D action-adventure game.

它保留了原版的剧情和谜题,增加了战斗场景,并且直接在Three.js里构建了所有角色和环境。

It kept the original plot and puzzles, added fight scenes, and built all of the characters and environment directly in Three.js.

90年代中期我超喜欢玩的一款游戏是Doom。

One of the games I absolutely loved playing in the mid-'90s was Doom.

30多年后的今天,我用Astra只花了几小时就创作了自己的一款Doom风格游戏。

Now, more than 30 years later, I've created my own Doom-inspired game in just a few hours with Astra.

我还在不断添加新功能、新敌人和新关卡,玩起来特别爽,而且几乎没什么我能发现的bug,哪怕是最初版本也一样。

And I keep adding new features, enemies, and levels that are so enjoyable to play and have almost no bugs I can notice, even in first versions.

做这个游戏让我想起了早期游戏时代的很多美好回忆,过程特别开心。

Building it brought back so many great memories from those early days of gaming and was so much fun.

太不可思议了,以前需要传奇团队和超强工程才能做的东西,现在像我这样没有软件工程经验的人,几个小时就能重新做出来、试验、扩展。

It's beyond remarkable to think that something that once required a legendary team and extraordinary engineering can now be recreated, experimented with, and expanded in hours by someone like me, who has no software engineering expertise.

Theo做了一个叫Fish Slop的游戏,他写道:GPT-6 Astra在Blender和三维推理方面是世界级的。

Theo built a game called Fish Slop, writing, GPT-6 Astra is world-class at Blender and 3-dimensional reasoning.

这是个一次性生成的游戏,全部在浏览器里运行。

This was a one-shot game it created all running in browser.

大家以为这肯定要消耗大量token,Theo回应说:生成Fish Slop花不到30美元。

Now when people assumed that this must be just absolutely eating through tokens, Theo responded, Fish Slop would have been under $30 to generate.

在200美元的订阅计划下,这点钱根本不算什么。

Wouldn't have even noticed the dent on the $200 sub.

创业者Anshu也发现了类似的情况,分享了一个他们做的游戏,还说:兄弟,GPT-6 Astra简直就是3D游戏的涡轮AGI机器神。

Entrepreneur Anshu found something similar, sharing a game that they created and adding, Dude, GPT-6 Astra is some kind of turbo AGI machine god for 3D games.

它一次性就在45分钟内搞定了,只用了我的配额里不到几个百分点。

It one-shot this in 45 minutes for hardly a couple percentage of my quota.

AI Battle账号做了一个索尼克的游戏,发现用最高设置,Astra在Pro X5账户上花了53分钟,用掉了周配额的4%,而中等设置下只用了25分钟,耗用了周配额的1%。

The AI Battle account built Sonic and found that on max settings, Astra was able to build it in 53 minutes using 4% of weekly usage on a Pro X5 account, while Astra on medium took 25 minutes and used 1% of the weekly usage.

连Sam Altman都对这种用法发表了看法,他说:和其他能力比这显然微不足道,但是Astra能让我做出任何我想象的有趣小游戏,几分钟后就能玩到,这太酷了。

Even Sam Altman weighed in on this use case saying, it is obviously trivial relative to everything else, but the fact that Astra can make me whatever fun little game I can imagine and I can be playing it a few minutes later is so cool.

还有人把这种3D能力用在了学习方面。

Others took this 3D capability in a learning direction.

昨天我做了一个细胞模型。

Yesterday I built a cell model.

今天它给我展示了Ozempic(一种药)的工作原理。

Today it showed me how Ozempic works.

我让它解释机制,一小时后,就出了这个。

I asked it to explain the mechanism, and 1 hour later, this.

Andrew分享的视频是一个交互式3D模型,精确展示了Ozempic的作用。

The video that Andrew shares is an interactive 3D model that shows exactly what Ozempic does.

产品负责人把Astra和Higgsfield结合起来,做了一个看足球时的叠加层,可以实时追踪球员、映射传球、跟踪控球情况。

Head of product combined Astra and Higgsfield to build an overlay for watching soccer that tracks players, maps passes, and follows possession all in real time.

Scenario公司的Emmanuel说,GPT-6 Astra帮他做出了他小时候希望拥有的东西。

And Emmanuel from Scenario said that GPT-6 Astra helped him build something that he wished he'd had as a kid.

只要输入一张图片或几句话的想法,就能得到一个可以完全拼装的乐高套装,100%根据你的提示或图片定制,用的是可以在线订购的官方乐高零件。

Input just an image or an idea in a few words and get a fully buildable Lego set, 100% customized to your prompt or image using official Lego parts you can order online.

我让GPT-6 Astra帮我理解自己的脚踝疼痛。

I asked GPT-6 Astra to help me understand my own ankle pain.

它给了我一个完整的交互式3D图集——骨骼、韧带、肌腱、真实的运动轴、用于跖屈和内翻的滑块,以及每个韧带正在做什么的实时读数。

It gave me a full interactive 3D atlas— bones, ligaments, tendons, real motion axes, sliders for plantar flexion and inversion, and a live readout of what each ligament is doing.

就一次对话。

One session.

Ethan Mollick写道:我不认为这是目标,但Astra在像Blender这样视觉效果好的3D工作上非常出色,这让它在社交媒体关注的战争中比Fable更有感知优势。

Ethan Mollick writes, I don't think this was the goal, but the fact that Astra is incredibly good at visually pleasing 3D work like Blender gives it a perception edge over Fable in the war for social media attention.

其他类型的输出更难判断,但视觉上的东西一目了然。

It's harder to judge other kinds of outputs, but the visual stuff comes through.

当然,Ethan说得对。

And of course Ethan is right.

这些东西确实很酷。

This stuff is really cool.

但对普通人来说,这有多重要呢?

But how much does it matter to the average person?

如果你知道这种3D设计是Astra擅长的事情,你就能想到自己有什么可以用它的地方吗?

If you knew that this sort of 3D design was what Astra was really good at, would you even have something that you could think of to use it for?

的确,当人们分享一些更通用的用例时,并不全是阳光和彩虹。

And indeed, when people were sharing some of their more generic use cases, it wasn't all sunshine and rainbows.

A16z的Martin Casado写道:新模型很厉害,但似乎编码方面已经饱和了。

A16z's Martin Casado wrote, the new models are amazing, but it seems coding has saturated.

在电脑操作上确实有明显的进步,但在我做的编码工作上,我没感觉到有什么质的提升。

There's clearly a meaningful step in computer use, but I don't notice a meaningful step in coding for the work I'm doing.

可能是我技术不行。

Maybe a skill issue.

但其他人也有类似的反馈。

But others report something similar.

Armin Ronacher 写道,我不知道 Astra 是从哪儿学的 Python,但只要它离开常规代码稍远一点,就会写出很奇怪的 Python 垃圾代码。

Armin Ronacher wrote, I don't know where Astra learned Python programming, but when it's one step removed from normal code, it writes weird Python slop.

它写的单元测试简直惨不忍睹。

The unit tests it writes are absolutely horrific.

Stripe 的 Chris Puckett 写道,对 Astra 的感受:它在处理一些着色器动画作品时,效果简直惊艳到我了。

Stripe's Chris Puckett wrote, Astra thoughts: absolutely blew me away with taking some shader animation work to an incredible level.

但在前端设计上,完全是一塌糊涂,可以说是最糟糕的垃圾代码了。

Absolutely crapped the bed, the sloppiest slop to ever slop on frontend design.

GPT-6 Astra 有一个很大的问题:前端和UI设计。

GPT-6 Astra has one pretty big problem: frontend and UI design.

Astra 在空间推理、3D、电脑操作、数学、智能体、编码方面都极其出色。

Astra is absurdly good at spatial reasoning, 3D, computer use, math, agents, coding.

但要它做一个漂亮的网站,Claude 做出来的仍然明显更好看。

But ask it to make a beautiful website and Claude can still look noticeably better.

这可能是我依然会继续用 Claude 的最大原因了。

This might genuinely be the biggest reason I'd still use Claude.

当然,这不是每个人的体验都一样。

Which is not to say that's been everyone's experience.

来自 How I AI 播客的 Claire 在她的长篇评测里说,她一直在尝试构建一个架构很复杂的产品情报应用,用来跨平台聚合和分析数据。

Claire from the How I AI podcast described in her long review how she had been trying to build an architecturally complex product intelligence app that aggregates and analyzes data across platforms.

她说 Fable 在架构方面做了些疯狂的事情。

She said Fable did insane things with the architecture.

5.6 版本在洞察质量上卡住了。

5.6 Just got stuck on quality of insights.

而 Astra 一次就搞定了。

Astra one-shotted it.

她说,这就是我使用这个模型一直以来的体验。

This, she says, has been my experience with this entire model.

这些非常具体的任务一直很难完成。

It's been so hard to do these very specific tasks.

我反复尝试了半年,然后 Astra 一次就搞定了。

I've tried them over and over again for 6 months, then Astra one-shotted it.

然而,从 Claire 或 Allie K. Miller 这样更全面的评测中可以看出,除了这些3D能力之外,Astra 真正具有变革性的地方似乎是在电脑操作领域。

And yet still, what comes through in the much more full reviews, like the one from Claire, or like Allie K. Miller's review, is that in addition to these 3D capabilities, the really transformative thing around Astra seems to be in the realm of computer use.

Claire 自称是电脑操作的狂热用户,她说 Astra 把她的电脑操作水平提升到了一个新高度。

Claire, describing herself as a computer use maxer, said that Astra has taken her computer use to the next level.

她举了一些例子,比如管理复杂的网页UI,或者在CRM系统中自动执行商机分配流程,并说这类工作如果全靠手动输入和思考,会花很长时间。

She described things like managing complex web UI or automating lead routing workflow in CRM and said this kind of thing, just typing and thinking, takes forever.

在 Astra 之前,我觉得电脑操作没法很好地处理这些节点。

Before Astra, I didn't feel like computer use could navigate these nodes as well.

但现在,她说,我基本上可以完全不管电脑了。

Now, however, she said, I'm just hands-off my computer all the time now.

她还指出,这种与电脑互动的方式,在思维模式上代表了一种根本性的转变。

And pointed out how different a mental paradigm this type of interacting with computers really represents.

同意,电脑操作和浏览器操作简直太棒了。

Agreed, writing, computer use and browser use are just incredible.

我确信,任何在电脑上进行的稳定工作流程,至少可以部分由AI完成。

I'm confident that pretty much any stable workflow done on a computer can be at least partially done by AI.

真的非常非常出色。

It's just so, so good.

想象一下,如果你是个变态老板,连续7天录下了你的屏幕。

Imagine you were a creepy boss and recorded your screen for 7 straight days.

你会让AI做那些你自己一直在做的事呢?

What would you have AI do that you've been doing?

你能挑战自己一天不用鼠标吗?

Can you challenge yourself to go mouse-free for a day?

现在,还有更多内容要讲,我们才刚刚开始揭开这个模型的神秘面纱,但我想以这个想法来结束——为什么我同时把 Astra 描述为意义重大又令人困惑。

Now, there's way more to get into, and we are just barely beginning to pull back the curtain on this model, but I wanna close out on this idea of why I'm describing Astra as both significant and confounding all at the same time.

当你想到高级AI的能力时,像GPT-3.5和后来的GPT-4这样的LLM,直接击中了一大批人日常做的核心事情:查答案、做研究、写东西。

When you think about capabilities of advanced AI, LLMs like GPT-3.5 and then 4 hit pretty directly down the line of an extremely core set of things that lots and lots of people do, looking up answers to questions, researching, writing things.

这些任务和查询构成了我们日常体验的很大一部分,尤其是在工作场景中。

These sort of tasks and queries make up a big part of our daily experience, especially in the work world.

而且,基本上就是给机器人发短信这种交互模式,大部分人当时已经很习惯了。

And what's more, the interaction pattern of effectively just sending text messages to a bot was something that most people were at that point really comfortable with.

但是,即使LLM那些核心的写作和研究用途吸引了大量用户,我们也看到了一些超出人们日常习惯的重大跳跃。

But even as those core writing and research type use cases of LLMs have attracted legions of users, we've had a few major jumps outside of what people normally do as well.

第一个就是图像生成。

The first was image generation.

这个有点模糊,因为显然在Midjourney、Stable Diffusion以及后来的Nano Banana和GPT Image出现之前,人们就已经在和图像打交道了。

Now, this is a little bit borderline because obviously even before things like Midjourney and Stable Diffusion and later Nano Banana and GPT Image, people were interacting with images.

但即便如此,在图像生成变得非常容易之前,和图像打交道——或者至少是创作图像——仍然是一个更专业的领域。

But still, before it was really easy to generate them, interacting with images, or at least creating images, remained a more specialized domain.

如果以2022年11月那个大型生成式AI发布的时间点作为参考,ChatGPT比同月发布的Midjourney第四版要普及得更快也就不奇怪了。

If you look at the big generative AI releases from November 2022 as a proxy, it's not surprising that ChatGPT was so quickly so much more ubiquitous than Midjourney's version 4, which also came out that month.

因为当时有更多人、更多已经成熟的用例适用于ChatGPT能做的事情,至少相对于Midjourney 4的能力来说是这样。

There are simply more people and more established use cases for the things that ChatGPT could do than at that time, at least for the things that Midjourney 4 could do.

不过随着时间的推移,随着AI能力范围的变化,人们使用生成式图像的方式也在增加和扩展,使得它也成为了这项技术的主流用途之一。

Over time though, as the AI capability set has changed, the way that people use generative images has increased and expanded and made it a major mainstream use of this technology as well.

我认为第二次重大跳跃是高级AI编程的引入,这超出了普通人的——姑且称之为知识工作者——日常范围。

The second time I think we had a major jump outside of what the average person, call it knowledge worker, does was with the introduction of advanced AI coding.

有意思的是,对很多人——也就是AI早期采用者中的很大一部分——来说,编程本身就是他们日常工作的主要内容。

Now, what makes this interesting is that obviously for some people, a significant number and class of AI early adopters, coding was in fact the major substance of what they do in a day-to-day way.

但对于我们这些迅速变成“氛围型编程者”的人来说,情况就有点不一样了。

But for the rest of us who had quickly become vibe coders, it was a little bit different.

在那之前,我们没办法通过编写代码来构建东西、解决自己的问题、完成工作,直到一系列模型和工具让我们不需要了解底层编程语言就能做到这一点。

We hadn't had the ability to interact with code as a way to build things and solve our own problems and accomplish our work up until a set of models and tools made it viable to do so without having to know the underlying coding languages.

从很多方面来看,你可以把2026年的漫长历史追溯到2025年11月底,当时我们得到了Claude Opus 4.5,紧接着就是GPT-5.2,人们在那之后的一个月左右意识到,这两者共同代表了AI编程能力的一次彻底飞跃,让高质量工作变得可行。

In many ways, you can chart the long history of 2026 back to the end of November 2025, when we got Claude Opus 4.5, which was followed quickly after by GPT-5.2, and which together people realized over the next month or so represented a total step change in their ability to do significant quality work with AI coding.

现在,距离那时已经过了10个月,很多像我这样之前不会编程的人已经以各种方式把AI编程整合到了自己的工作流程中。

At this point now, 10 months on from that, many of us previously non-coders have integrated AI coding into our workstreams in various ways.

但这仍然是一个进行中的过程。

But that's still a work in progress.

它还没有完全渗透到整个知识工作领域,尽管对我们这些已经全面拥抱它的人来说,感觉某种程度的普及是不可避免的。

It hasn't yet translated all the way across the entire span of knowledge work, even though for those of us who have fully embraced it, it feels like some version of that is inevitable.

所有这些3D建模和设计的东西,感觉就是这种模式的第三个例子。

All of this 3D modeling and design stuff feels like a third example of that pattern.

换句话说,AI的新进展解锁了一系列能力,让那些之前无法接触这些能力的人现在有了机会去做。

In other words, a set of capabilities unlocked by some new advance in AI that gives people who previously had no way to interact with that capability set the opportunity to now do so.

问题是,对于那些目前工作通常不涉及3D建模或设计的人来说,哪些用例会变得正常化?

The question will be, what, if any, use cases will become normalized for that for people whose work at this moment doesn't normally include 3D modeling or design?

这些用例会一直停留在新奇阶段,还是有一些会像AI和智能体编程一样,跃升为更广泛的知识工作者常规使用的东西?

Will these use cases remain entirely novelty, or will some make the jump, like AI and agentic coding has made the jump, to just a thing that now a much broader set of knowledge workers actually use in a regular way?

我在Astra身上看到的另一个变化模式,体现在它塑造的交互方式上。

The other pattern of change that I see with Astra comes in the interaction patterns that it models.

我的意思是,有时重大的转变不仅仅在于能力本身,还在于我们如何通过交互来利用这些能力。

What I mean by that is that sometimes the big shift isn't just in capabilities, but in how we interact to leverage those capabilities.

当Nano Banana(也就是Gemini 2.5 Flash Image)最初在2025年8月发布时,它之所以引人注目,并不只是因为它在生成逼真图像或特定风格图像方面有了巨大提升。

When Nano Banana, aka Gemini 2.5 Flash Image, was originally released back in August of 2025, it wasn't notable specifically because it was massively better at making photorealistic images or images in particular styles.

它之所以如此强大,是因为你不再需要为了修改一个细节而重新生成整张图像,而是可以直接提示它只修改某个特定部分。

What made it so powerful was that instead of having to regenerate an entire image from scratch to change one detail, You could instead prompt it to fix just one specific part.

结果就是,交互模式上的这么一个小小改变,解锁了大量用例。

And it turns out that that one small change in the interaction pattern unlocks a tremendous number of use cases.

整个2026年,编程领域也见证了交互模式的一系列快速转变。

Throughout 2026, coding has also been witness to a set of rapid shifts in the interaction pattern as well.

现在,其中一些显然也是能力上的变化,但今年夏天AI和智能体编程的关键词是什么呢?

Now, some of this obviously has been capability changes as well, but what has been the big watchword for AI and agentic coding this summer?

一直都是关于循环。

It's been all about loops.

也就是说,从想着给编程智能体一个提示去完成某件事,转向建立一个循环结构,让它能够自我管理,持续朝着目标推进,能够评估自己的进展,并一直继续下去,直到真正实现那个目标。

And shifting from thinking of prompting a coding agent to do a thing to instead setting up a loop structure in which it can manage itself and continuously progress towards a goal in a way that it can evaluate its own progress and continue on until it actually achieves that goal.

这是一种完全不同的交互模式,它将改变我们能利用AI做什么,以及我们能从中获得多少价值。而我们目前正处于尝试将其从最初的编程领域扩展到其他知识工作领域的过程中。

That is a totally different type of interaction pattern that will change what we can do with and how much we can get out of AI, and that we are in the middle of trying to expand outside of just the initial domain of coding to other areas of knowledge work as well.

而且我觉得,即使3D建模和设计比起图像生成或编程,作为所有知识工作的默认部分来说仍然更小众,OpenAI至少提出了一个非常有力的观点:计算机使用能力正在改变,并且将改变我们对与计算机交互的思考方式。

And it feels to me like even if 3D modeling and design stays more niche than either image generation or coding as a default part of all knowledge work, OpenAI at least is making a very strong argument that computer use capabilities are changing and will change how we think about interacting with computers.

我觉得这绝非偶然——整个三分钟的发布视频,已经被观看了一亿三千万次,里面都是人们完全不用手,坐在椅子上、来回踱步,用语言和外部处理的方式完成各种不同类型的工作,全靠语音,而AI则管理着界面。

I don't think it's an accident at all that the entire 3-minute launch video that has been seen 130 million times is people completely hands-free sitting in a chair, pacing around, verbalizing and externally processing their way through a variety of different types of work, all with their voice as the AI manages the interface.

这种交互模式长期以来一直是科幻小说的一部分。

Now, this sort of interaction pattern has long been a part of sci-fi.

大家都提到托尼·斯塔克的贾维斯,这不是没有原因的。随着语音模式的发展,我们已经朝着这个方向迈出了初步的、小小的步伐。

There's a reason that everyone references Tony Stark's Jarvis, and we've had baby steps towards this with developments in voice mode.

今年我发布的任何免费训练项目,基本上做过的人都会听到我不断唠叨,强调要围绕语音模式来锚定你的交互方式。

Anyone who's done pretty much any of the free training programs that I've released this year has heard me squawk endlessly about anchoring the way that you interact by using voice mode.

但很明显,我们现在仍然处于早期阶段。

But it's pretty clear that we're still early days.

而我认为,这一切的最终结论是,Astra并不是一个我们短期内应该指望能完全理解或充分发挥其优势的模型。

And I think the upshot of all of this is that Astra is not a model that I think we should expect to fully understand or take anywhere near full advantage of in the immediate term.

虽然关于这是否就是AGI会有很多争论,比如OpenAI的Greg Brockman说他认为GPT-6 Astra就代表了AGI,但接下来几个月的真正工作将是弄清楚Astra不是在哪里取代GPT-5.6 Sol或Fable 5来处理我们日常做的事情,而是以更宏大、更核心的方式,为我们解锁了哪些新的机会。

While there's gonna be lots of debates around whether this is now AGI, as for example, OpenAI's Greg Brockman says that he believes GPT-6 Astra represents, the real work of the next few months is going to be to figure out not where Astra replaces GPT-5.6 Sol, or Fable 5 for the everyday sort of things we do, but in a much bigger and more core way, what new opportunities it unlocks for us.

一个值得关注的地方是OpenAI员工自己透露的一些信息。

One place to keep an eye on that is what's coming out from OpenAI staffers themselves.

例如,负责产品业务的Thibault说,当Astra还没有普遍可用时,它可能是我们最大的竞争优势。

Thibault, who leads product over there, said, for example, Astra was probably our biggest competitive advantage while it wasn't generally available.

自从我们拥有了它,我们的生产力提升了很多,以至于我们把一些计划提前了六个月,将在Dev Day发布,而不是明年年中。

Since we've had it, our productivity jumped so much that we shifted some of our plans 6 months ahead and we'll ship them at Dev Day instead of mid-next year.

那只是因为更好的智能体解码,还是完全是别的什么?

Was it just better agent decoding, or was it something else entirely?

这就是我们面临的问题,我们需要花更多的时间去弄清楚。

That is the question that faces us, and we're gonna have to take a lot more time to figure it out.

不过今天就到这里吧,这是今天的《AI每日简报》。虽然我们已经聊得非常久了,但我觉得我们只是刚刚触及皮毛。

For now though, that is gonna do it for today's AI Daily Brief, one where even though we've gone wildly long, I feel like we are just barely scratching the surface.

一如既往地感谢你们的收听或收看,下次再见,平安。

Appreciate you listening or watching as always, and until next time, peace.
已剔除 2 处广告(点击展开查看)
M1
M11:29广告 · 已剔除

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

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
M17:59广告 · 已剔除

One big change around AI is we've shifted our thinking from how we rank our pages to how do we become the source that AI trusts enough to answer with.

At KPMG, they're seeing this firsthand.

AI-generated results now surface answers directly, often without a single click.

That's why they are increasingly focused on generative engine optimization, or GEO, structuring content so AI systems can retrieve it, understand it, and cite it as trusted authority.

This is not just an SEO evolution, but a visibility mandate.

And indeed, the GEO mandate from KPMG is simple.

If AI is shaping decisions, your expertise needs to show up inside the answer.

Read all about it at kpmg.com/us/geo.

Again, that is kpmg.com/us/geo.

Blitzy deeply understands your codebase before it writes code.

Here's the first place that pays off: security in the age of AI.

Vulnerabilities don't live in isolation—they live buried inside millions of lines of interconnected code, where patching one thing quietly breaks three others.

That's why surface-level scans fail.

Blitzy starts from its knowledge graph of your entire application, identifies and surfaces CVEs across the full estate, proactively recommends patches, and can execute the PR.

Each fix is grounded in how your systems connect and validate so nothing new breaks, and the knowledge graph dynamically updates, keeping you ahead of an ever-accelerating threat landscape.

One Blitzy customer resolved 21 active CVEs across 6 core microservices in 4 days.

Zero compile errors, every validation scan clean, months of planned work fixed in less than a week.

Security remediation grounded in real architectural context at the speed of compute.

Harden your codebase at blitzy.com.

That's blitzy.com.

The best teams don't have a single star carrying everyone else.

They know their own strengths and each other's weaknesses and play to both.

That's the team Robots and Pencils has built on purpose.

Nobody there is grinding through busywork to pad a headcount number.

People come for the hard problems, and they stay because everyone around them is leveling up at the same time.

In a market full of companies that are just trying to hire fast, that's worth a look.

Check out robotsandpencils.com/careers.

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

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 agents turn competitor moves into landing pages.

Sales agents enrich leads, draft emails, and update 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 had agents that feel like teammates.

Hire yours at HyperAgent.

Get $100 in credits at hyperagent.com/AIDailyBrief.