← 返回任务列表

Opus 5.5 vs GPT-6 Sol and Luna

281 段 · 1 位说话人 · 原片 30:47
M1
M10:00

本周前半段,我们看到了不是一个、两个、三个,而是四个重要的模型发布,其中包括本周二的两个——OpenAI的两个模型和Anthropic的一个。

The first half of this week has seen not one, not two, not three, but four major model releases, including two on Tuesday this week—a pair from OpenAI and one from Anthropic.

对于OpenAI来说,GPT-6、Sol和Luna继续他们的使命,在智能堆栈的每一层都构建成本高效的模型。

For OpenAI, GPT-6, Sol, and Luna continue their quest to build cost-efficient models at every level of the intelligence stack.

而对于Anthropic,早期迹象表明Opus 5.5是回归辉煌,至少是自Opus 4.6以来公司未曾见过的早期用户赞誉。

And for Anthropic, early indications suggest that Opus 5.5 is a return to glory, or at least early adopter acclaim that the company has not seen since Opus 4.6.

《AI每日简报》是一个每日播客和视频节目,关注人工智能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,在我们开始之前,快速通知一下。

All right, friends, quick announcements before we dive in.
M1
M10:54

现在,和大型新模型发布时通常一样,这期节目会是纯主节目,老实说,即使这样我们也很难把内容全部讲完。

Now, as is normal when it comes to big new model releases, this will be a main-only episode, and honestly, we're gonna have a hard time getting it all in even with that.

那我们就开始吧。

So let's dive in.

欢迎回到《AI每日简报》。

Welcome back to the AI Daily Brief.

不可否认,《AI每日简报》最棒的日子,当然也是最有趣、最激动人心、最有活力、啊呀,我等不及做完这一期因为我可以去做别的事情的那种节目,就是当我们拿到新模型的时候。

It's fairly undeniable that the best days on the AI Daily Brief, certainly the most fun, exciting, dynamic, ooh boy, I can't wait to be done with this episode because now I get to go do things types of episodes are ones where we get new models.

而昨天我们得到了绝对稀有的惊喜。

And yesterday we got the absolute rarest of treats.

老实说我不记得以前有没有发生过这种情况,那就是当下两个最重要的实验室,OpenAI和Anthropic,在同一天发布了新模型。

A thing which I can't frankly remember ever happening before, which is the two most important labs of the moment, OpenAI and Anthropic, both releasing new models on the same day.

当然,我们之前也有过模型发布很接近的情况。

Certainly we've had models released close together before.

事实上,我们通常看到的模式是像SpaceX AI昨天那样,提前发布他们的模型,以免被他们知道会获得更多关注的模型所淹没。

In fact, usually the pattern that we see is what SpaceX AI did yesterday, releasing their model in advance so as not to be drowned out by models that they knew would get more attention than them.

同一天发布的挑战在于,它不可避免地让人们不再问这个特定模型有什么价值以及它如何融入我的工具组合,而是问这两个哪个更好,以及这说明了哪个实验室处于领先地位?

The challenge of the same-day release is that it inevitably gets people to ask not what's valuable about this particular model and where it's going to fit in my rotation, but instead, which of these is better and what does it say about which lab is in the lead?

所以今天我们打算回顾一下发布了什么、基准测试的情况、初步反应、一些特定用例的例子、对竞争格局的影响、这对整个前沿竞赛意味着什么,以及根据社区的评价,谁赢了这一天。

So today we are going to go through what was released, where things stand on the benchmarks, the first reactions, examples around some particular use cases, the impact on the competitive landscape, what it says about the whole pacing the frontier thing, and in the community's estimation, who won the day.

我们拿到的第一个模型是Claude Opus 5.5,老实说,自从Opus级别的模型成为主角已经有一段时间了。

The first model we got was Claude Opus 5.5, and frankly, it's been some time since an Opus-class model was the big show.

现在最大的原因当然是Mythos和Fable的推出,但即使在那之前,人们对Opus 4.6如此喜爱,以至于4.7和4.8对很多人来说,如果不是退步,那也顶多是增量改进,而且在某些方面甚至没有领先。

Now the biggest reason for that is of course the introduction of Mythos and then Fable, but even before that, people had so much love for Opus 4.6 that 4.7 and 4.8 were for many, if not a regression, certainly very incremental at best, and not even incrementally ahead in certain cases.

事实上,到了Opus 5,人们根本就不喜欢它。

And in fact, when it came to Opus 5, people just genuinely did not like the thing at all.

现在快进到这次讨论的落脚点,AI内容创作者Peter Yang给出了完美的总结,他用了一个马的插画梗——一开始是一匹美丽完整的马,当然代表Opus 4.6,然后到Opus 4.7和Opus 4.8变成更粗糙、更简单、更幼稚的线条画,再到Opus 5几乎是涂鸦,最后到了Opus 5.5,又是一幅完美细节的马的正面插画。

Now, jumping ahead to where this conversation is landing, the perfect encapsulation comes from AI content creator Peter Yang, who used a meme of an illustration of a horse— beautiful and complete at the beginning, representing Opus 4.6 of course, to a much scratchier, more simplistic and childlike line drawing for Opus 4.7 and Opus 4.8, culminating in near scribbles for Opus 5, to finally, once again, the beautiful front side of a horse in perfect illustrated detail for Opus 5.5.

换句话说,大家真的很喜欢这个模型。

The people, in other words, are really liking this model.

但Anthropic是怎么宣传它的呢?

But how did Anthropic pitch it?

在公告中,他们表示Opus 5.5在大多数任务上的表现达到Claude Fable 5.1的水平,但运行成本降低了40%,不仅比Fable 5.1低,而且比Opus 5也低。

In their announcement post, they said that Opus 5.5 performs at the level of Claude Fable 5.1 for most tasks, but costs 40% less to run, not even just than Fable 5.1, but than Opus 5.

在他们的公告串中,他们指出这是从Opus 5的一次重大升级,而且老实说,至少从基准测试来看,它也比Fable 5.1有所提升。

In their announcement thread, they point out that it is a major step up from Opus 5, but also frankly, at least when it comes to the benchmarks, it's also a step up from Fable 5.1.

在Terminal Bench 4.0上,模型从Fable 5.1的55.8%跃升至Opus 5.5的66.4%。

On Terminal Bench 4.0, the model jumped from Fable 5.1's 55.8% to Opus 5.5's 66.4%.

Cursor Bench、Frontier Code v1.1和Humanity's Last Exam也都出现了提升,不仅相对于Opus 5,而且相对于Fable 5.1也是如此。

Cursor Bench, Frontier Code v1.1, and Humanity's Last Exam also all saw jumps, not only from Opus 5, but from Fable 5.1 as well.

事实上,没有一个基准测试是Opus 5.5没有领先于其他Anthropic模型(Fable 5.1和Opus 5)的,只有两个基准测试——衡量业务流程的Automation Bench和衡量智能体科学研究的Terminal Bench Science——Opus 5.5也没有超过GPT-6 Astra。

In fact, there wasn't a single benchmark that Opus 5.5 wasn't ahead of the other Anthropic models, Fable 5.1 and Opus 5, and only two Automation Bench, which measures business workflows, and Terminal Bench Science, which measures agentic scientific research, where Opus 5.5 wasn't ahead of GPT-6 Astra as well.

重要的是,这不仅是性能的提升,也是定价上的优势。

Importantly, this is not just a performance gain, but also a pricing gain as well.

Claude官方账号写道,因为Opus 5.5需要的计算资源比Opus 5少,所以它的定价也相应降低了。

The Claude account wrote that because Opus 5.5 requires less compute to serve than Opus 5, its pricing is consequently down.

相比Opus 5每百万输入token 5美元和每百万输出token 25美元,Opus 5.5是每百万输入4美元,每百万输出20美元。

Compared to Opus 5's $5 per million input tokens and $25 per million output, Opus 5.5 is $4 per million input and $20 per million output.

不过,由于额外的效率提升,他们表示成本节约实际上能达到约40%,而不仅仅是成本数字上反映的20%。

However, because of additional efficiencies, they said that the cost gains will actually be about 40% as opposed to just the 20% reflected in the cost.

虽然它没有被定位为以速度为主的模型,但他们确实指出,因为它的许多默认努力设置比其他模型在更高设置下能产生更好的结果,所以Opus 5.5生成输出的速度比其他模型快得多,包括比Opus 5平均快30%。

Although it's not pitched as a model focused on speed, they do note that because many of its default effort settings deliver better results than other models running at higher settings, that Opus 5.5 generates output much faster than those other models, including an average of 30% faster than Opus 5.

锦上添花的是,他们把Pro、Max和Team套餐的5小时使用限制都提高了,还给其他订阅用户增加了一次速率限制重置,大家随时可以用。

As a cherry on top, they increased 5-hour usage limits on all of their Pro, Max, and Team plans, and threw in the rest of the subscription users a rate limit reset, which can be used whenever people need.

在安全方面,他们指出这是他们在发布《Pacing the Frontier》报告后推出的第一个模型,而且Opus 5.5在发布前经过了外部评估者的测试,包括Frontier Design和Meter。

On the safety front, they note that this is their first model released in the wake of the Pacing the Frontier note, and that Opus 5.5 was tested before release by external evaluators, including Frontier Design and Meter.

总的来说,他们认为Opus 5.5基本上是他们迄今为止发布的与人类价值观最一致的模型。

And overall, they argued that Opus 5.5 is basically the best aligned model they've released yet.

在Anthropic负责对齐工作的Sam Baumann写道:我们认为Opus 5.5比之前的模型安全得多,发布它——更有可能的是——会降低与对齐不良相关的风险。

Sam Baumann, who works on alignment at Anthropic, wrote, We think Opus 5.5 is sufficiently safer than its predecessors that releasing it, more likely than not, reduces risks related to misalignment.

单单Opus 5.5就足够做一整期节目了,但OpenAI决心不让Claude独占风头,于是在美国东部时间周二下午早些时候,他们发布了自己的模型系列:GPT-6 Sol和GPT-6 Luna。

Now, Opus 5.5 alone would be more than enough for a complete episode, but OpenAI was determined not to let Claude have all the fun, and in the early afternoon Eastern time on Tuesday, they dropped their own set of models, GPT-6 Sol and GPT-6 Luna.

如果说Opus 5.5的发布稍微侧重于成本和效率,那么GPT-6 Sol和GPT-6 Luna的发布则完全聚焦在成本和效率上。

And if the Opus 5.5 announcement was a little bit focused on cost and efficiency, the GPT-6 Sol and GPT-6 Luna announcement was all about cost and efficiency.

他们在X上的发布帖写道:GPT-6 Sol和Luna建立在GPT-6 Astra的进步之上,将其许多优势融入到更快、更实惠的模型中,以支持大规模工作。

Their announcement post on X reads, GPT-6 Sol and Luna build on the advances behind GPT-6 Astra, bringing much of its strengths into faster and more affordable models to support work at scale.

我们还让缓存和推理变得更高效,并且把这些节省直接传递给你们。

We've also made caching and inference more efficient, and we're passing the savings directly to you.

与GPT-5.6的促销定价相比,Sol和Luna的API价格降低了50%。

50% Lower API prices for Sol and Luna compared with GPT-5.6 promotional pricing.

换句话说,他们不只是说这个定价比Astra低;他们甚至是在上一代的基础上进一步削减成本。

In other words, they're not just talking about this pricing being lower than Astra; they're cutting costs from the previous generation even before that.

有意思的是,他们显然也在把降低成本视为使用工具的新方式。

And what's interesting is that they're clearly presenting cost as enabling new ways to use the tool as well.

正如他们所说:更高的使用限制和更低的成本,让你有更多灵活性和迭代空间。

As they put it, higher usage limits and lower cost give you more flexibility and room to iterate.

也就是说,他们认为人们利用这些成本节约的方式之一,就是在使用模型时更加注重迭代。

In other words, one of the ways that they think that people should cash in on those cost savings is to be more iterative in how they use the model.

现在回溯几代模型。

Going back a couple of generations of models at this point.

Codex加GPT的使用方式,与Fable加Claude的使用方式之间,一个微妙的区别是:人们倾向于用Fable和Claude做更放手、长期运行的任务,而在Codex里用GPT模型做需要更多交互和迭代的事情。

One of the subtle divides between Codex and GPT usage and Claude and Fable usage, on the other hand, has been an impulse to use Fable and Claude for more hands-off, long-running tasks and the GPT models inside Codex for things where you need more interaction and iteration.

而新的Sol和Luna模型似乎也延续了这一思路。

And these new Sol and Luna models seem to build on that as well.

和Opus 5.5的发布一样,OpenAI也提到了对齐方面的进步。

Like the Opus 5.5 announcement, OpenAI also points to advancements in alignment.

值得注意的是,当你查看OpenAI页面上的基准测试时,你会发现Anthropic还在使用传统的图表——显示基准名称和一组模型的一组百分比分数——而OpenAI已经完全放弃了那种方式,只展示以性能为Y轴、成本为X轴的图表。

And what's notable when you get into the benchmarks on the OpenAI page is that where Anthropic is still using the traditional charts, which show a benchmark name and then a set of percentage scores across a set of models, OpenAI has entirely abandoned that, only showing benchmarks on graphs that map performance on the y-axis versus cost on the x-axis.

Sam Altman在推文中说:GPT-6、Sol和Luna在智能、对齐、工作输出、编码、计算机使用等方面都比5/6系列的前代模型有显著改进。

Sam Altman tweeted, GPT-6, Sol, and Luna are big improvements on intelligence, alignment, work output, coding, computer use, and more over their 5/6 family predecessors.

而且每个token的价格降低了一半,每个任务的价格甚至更低。

They are also half the price per token and even less per task.

接着他又跟进并进一步强调了这种思维方式对现在的OpenAI有多重要,尤其是按任务定价的比较。

He then went and followed up and really reinforced just how important this way of thinking is to OpenAI now, especially compared by per-task pricing.

Sam说:这才是真正应该看重的指标,我不认为市场上有什么能比得上。

Which is the metric that should matter, Sam says, I don't think there is anything competitive anywhere in the market.

我们希望人们能够大量使用AI。

We want people to be able to use tons of AI.

这对于能够探索眼前的这场新复兴非常重要。

It's important to being able to explore this new renaissance in front of us.

为了说明他们所说的“大量”是什么量级,在发布帖中他们指出,按API价格计算,OpenAI的研究人员中位数每天使用价值600美元的token,而第90百分位的研究人员每天使用价值7000美元的token。

Putting some magnitude around what they mean by tons, in the announcement post they point out that valued at API prices, the median researcher at OpenAI uses $600 in tokens per day, while the 90th percentile of researchers use $7,000 of tokens per day.

他们几乎是轻描淡写地写道:随着编码代理承担更长、更复杂的任务,持续使用的成本变得更重要。

Almost understating it, they write, as coding agents take on longer and more demanding tasks, the cost of sustained use matters more.

结论是什么?

The upshot?

GPT-6 Sol和Luna,他们写道:将强大的编码性能与更低的API价格相结合,让开发者有更多迭代空间,也让团队更有信心在Codex上承担更雄心勃勃的任务。

GPT-6 Sol and Luna, they write, combine strong coding performance with lower API prices, giving developers more room to iterate and teams the confidence to be more ambitious about what they ask Codex to take on.

但外部基准测试怎么样呢?

But what about external benchmarks?

就算是对基准测试比较在意的人,也都知道内部测试数据得打个折扣来看。新模型一出来,大家最先去看的往往是像 Artificial Analysis 这样的独立来源。

Even for those folks who care about benchmarks, everyone assumes you have to take any internal measures with a grain of salt, and one of the first places that people look when a new model comes out is independent sources like Artificial Analysis.

在他们的新 V3 指数上,GPT-5.6 Luna 和 GPT-6 Luna 总分都是 37,基本持平;而 GPT-6 Sol 则略胜 GPT-5.6 Sol。

On their new V3 index, GPT-5.6 Luna and GPT-6 Luna basically score the same at 37 overall, while GPT-6 Sol slightly beats GPT-5.6 Sol.

不过话说回来,这两个模型达到这个分数,单位智能的成本更低。

That said, both of those models get those scores at a lower cost per intelligence.

实际上,Artificial Analysis 的第一句话就是:GPT-6 Sol 和 Luna 把成本效率边界往前推了,成本比 GPT-5.6 Sol 和 Luna 降低了一半。

Indeed, Artificial Analysis's first sentence is, GPT-6 Sol and Luna push the cost-efficiency frontier by halving cost relative to GPT-5.6 Sol and Luna.

另一个亮点是,Artificial Analysis 发现这两个模型在幻觉方面都有明显减少。

In another positive highlight, Artificial Analysis found that both models saw a significant reduction in hallucination.

Zapier 也用 GPT-6 Sol 测试了一堆真实工作流程。

Zapier also tested GPT-6 Sol on a slate of real workflows.

他们发现 6 Sol 得分 33.2%,而 GPT-5.6 Sol 是 28.77%,而且成本大约是 5.6 的一半。

They found that 6 Sol scored 33.2% compared to GPT-5.6 Sol's 28.77%, and did so at roughly half the price of 5.6.

他们的结论很简单:所有在用 5.6 的工作流,都该升级了。

Their summation was pretty simple: upgrade any of your workflows that are using 5.6.

更便宜还更好——这种组合可不多见。

It's cheaper and better— a rare combo.

不过 Opus 5.5 就完全是另一回事了。

Opus 5.5, however, was a different kettle of fish entirely.

Opus 5.5 之所以引人注目,是因为它把其他所有模型都甩在身后,在智能指数上整整高出 5 分,拿下了新榜首。

Opus 5.5 was notable because it absolutely thumped everything else to get the new top score on the Intelligence Index by 5 whole points.

之前 Claude Fable 5.1 和 GPT-6 Astro 以 53 分并列第一,但 Claude Opus 5.5 一下子跳到了 58 分。

Claude Fable 5.1 and GPT-6 Astro were previously holding it down at the top with scores of 53, but Claude Opus 5.5 jumped all the way to 58.

而且虽然在成本和效率提升上可能不如 GPT-6 Sol 那么显著,但 Artificial Analysis 提到,这次性能提升同时还降价了 20%。

And while its cost and efficiency gains might not have been as much as GPT-6 Sol, Artificial Analysis did note that this increase in performance also came with a 20% price cut.

说实话,虽然我一直尽量保持分析客观中立,但看看大家的初期反应,毫无疑问,这些模型里最火的就是 Opus 5.5。

And honestly, while I've been trying to keep this analysis largely aligned and equivalent so far, watching people's first responses, there is absolutely no doubt which model among these dominated the discourse.

Opus 5.5 成了地球上最强的模型,而且比 Astra 和 Fable 还便宜——这我可真没想到。

Opus 5.5 being the absolute best model on the planet and being cheaper than Astra and Fable was not on my bingo card.

我用得越多,就越喜欢 Opus 5.5。

The more I use, the more I fall in love with Opus 5.5.

它特别好,速度快了好多,而且啰嗦少了很多。

It's so good and so much faster and so much less verbose.

这正是我想要的。

It's everything I could have asked for.

就好像我心爱的 Opus 4.6 回来了,但更强了,改名叫 5.5。

It's as if my beloved Opus 4.6 came back but better and rebranded as 5.5.

后面我们会再聊聊这种喜爱,不过大家具体都在哪些领域测试 Opus 5.5 呢?

Now we'll get more into that love, but what were some of the specific areas that people were testing Opus 5.5 on?

Aaron Levy 和 Brocks 团队把它用在——引用原话——‘各种处理非结构化数据的复杂企业知识工作’上,不仅性能有提升,效率也大幅提高。

Aaron Levy and the team at Brocks brought it into, quote, a variety of complex enterprise knowledge work tasks dealing with unstructured data, and saw not just improvements in performance, but major efficiency gains as well.

跟 Opus 5 相比,他们发现 token 用量减少了 63%,啰嗦程度降低了 42%,速度提升了 30%。

Versus Opus 5, they found 63% fewer tokens used, 42% less verbosity, and 30% faster.

在金融服务任务中,任务准确率提升了 39%。

On financial services tasks, they found an increase of 39% task accuracy.

在云成本分析和技术用例中,任务准确率提升了 65%。

On cloud cost analysis and technology use cases, they found an increase of 65% task accuracy.

而在消费品和临床诊断相关用例中,任务准确率分别提升了 17% 和 15%。

And in use cases related to consumer products and clinical diagnostics, they found an increase of 17% and 15% task accuracy respectively.

如果你在企业里负责 AI 这块,你肯定明白现在的差距不是能力,而是执行。

If you're leading AI inside an enterprise, you already know that the gap right now isn't capability, but execution.
M1
M111:41

哪些管用,哪些不管用,以及从试点到真正规模化落地到底需要什么。

What's working, what's not, and what it actually takes to move from pilots to real scaled impact.

涉及战略、数据就绪度、治理、人才和业务价值。

Across strategy, data readiness, governance, workforce, and value.

当然,主持人是我,Nathaniel Whittemore。

And of course, it's co-hosted by me, Nathaniel Whittemore.
M1
M112:03

这就是为什么大多数遗留系统现代化项目会失败的原因。

Here's why most legacy modernization projects fail.

做这个工作的AI无法理解大规模的代码库。

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

它只能看到一小段上下文,检查一下语法,却错过了分散在全球应用生态系统中多年积累的决策。

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

Blitzy用解决一切问题的方式来处理这个。

Blitzy solves this the way it solves everything.

在任何迁移开始之前,就把你的代码作为基础,Blitzy的智能体把整个遗留系统逆向工程成一个持久的知识图谱。

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

包括每一个依赖、每一个约束、每一条原本只存在于某个工程师脑子里的隐性知识。

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

基于这种理解,Blitzy自主执行语言迁移、框架升级和单体到微服务的转型,全程端到端验证。

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

Blitzy的一个客户用16周将一个价值1000万美元的单体保险系统现代化,而如果用编程智能体,基线是137周。

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

这就是9倍的压缩。

That's 9x compression.

在加速你的路线图的同时清除技术债务。

Retire technical debt while accelerating your roadmap.
M1
M112:55

这里有个残酷的事实。

Here's a harsh truth.

你的公司可能正花着几千甚至几百万美元在AI工具上,而这些工具被严重地未充分利用。

Your company is probably spending thousands or millions of dollars on AI tools that are being massively underutilized.

一半的公司有AI工具,但只有12%的公司用它们来创造商业价值。

Half of companies have AI tools, but only 12% use them for business value.

大多数员工还在用AI来总结会议记录。

Most employees are still using AI to summarize meeting notes.

如果你是在公司负责推广AI的人,你需要Section。

If you're the one responsible for AI adoption at your company, you need Section.

Section是一个平台,帮助你管理整个组织的AI转型。

Section is a platform that helps you manage AI transformation across your entire organization.

它指导员工使用真实案例,追踪谁在用AI产生商业影响,并且精确地告诉你AI在哪些地方创造价值,哪些地方没有。

It coaches employees on real use cases, tracks who's using AI for business impact, and shows you exactly where AI is and isn't creating value.

结果呢?

The result?

你从部署工具变成驱动可衡量的AI价值。

You go from rolling out tools to driving measurable AI value.

你的员工从只会总结会议记录,变成解决真正的业务问题。

Your employees move from meeting summaries to solving actual business problems.

而且你能证明投资回报率。

And you can prove the ROI.

别再猜测你的AI投资是否有效了。

Stop guessing if your AI investment is working.
M1
M114:23

说到社交媒体上的演示,和之前几个模型的情况一样,大量的帖子都是非常高视觉性的内容,比如图形、3D渲染和游戏设计,这些在网上看着挺有意思,虽然不一定能很好地反映大多数人实际会用到这个模型的场景。

Now, when it came to the social media demos, as has been the case with the last several models, a huge amount of the posts were highly visual things like graphics, 3D renderings, and game design, which are interesting to look at on the internet, if not necessarily super reflective of the type of use cases that most people are going to actually have for this model.

不过,考虑到3D图形、视觉效果和Blender在GPT-6 Astra发布中占了很大比重,很多Opus 5.5的演示基本上就是在这些用例上碾压那个模型。

Still, given how much 3D graphics and visuals and Blender were a big part of the story of the GPT-6 Astra release, a lot of the demos of Opus 5.5 basically have it mogging that model on exactly that type of use.

而且,鉴于我大概一两周前(或者不管什么时候我们得到GPT-6 Astra)的分析,重点在于那些能力如何让我们以更广阔的视角思考AI带来的机会,那么值得注意的是,Opus 5.5至少在初步看来,在这些能力集上也是又一次飞跃。

And given that my analysis just a week or two ago, or whenever it was that we got GPT-6 Astra, was about how much those sort of capabilities asked us to think in more expansive terms about the opportunities that AI opened up, it's worth noting that Opus 5.5 at least initially seems to be another jump in those capability sets as well.

Anthropic团队的Alex Albert展示了Opus 5.5只用一个提示就用Blender制作了黏土动画,Chase Lean做了一个交互式珊瑚礁壁纸,还有一些人展示了这些模型足够强大,可以用生成的代码而不是图像模型来创造真实的世界视觉表现。

Alex Albert from the Anthropic team showed Opus 5.5 using Blender to make claymations with a single prompt, Chase Lean made an interactive coral reef wallpaper, and some people showed how the models were powerful enough to use generated code rather than image models to create real visual representations of the world.

Peter Yang分享了一段看起来像是金门大桥的视频,但指出这完全是由Opus 5.5用代码生成的。

Peter Yang shares what looks like a video of the Golden Gate Bridge, but notes that it's actually entirely generated using code by Opus 5.5.

他补充说,根据我的测试,我认为Opus在构建3D场景方面和Astra一样好。

He adds, from my testing, I think Opus is just as good at building 3D scenes as Astra.

Alex Albert也做了类似的事情,只用Opus 5.5代码在Blender中可视化了1906年地震前旧金山市场街的历史准确面貌,没有使用图像生成。

Alex Albert again did something similar, visualizing a historically accurate San Francisco Market Street in 1906 pre-earthquake using only Opus 5.5 code in Blender, no image generation.

Dan Wood说,Opus 5.5在“这类3D生成世界里绝对碾压其他所有模型”,并补充说这是在空间感知方面“惊人的飞跃”。

Dan Wood says that Opus 5.5, quote, absolutely frame-mogs every other model on these sort of 3D-generated worlds, adding that it is a, quote, astounding leap in spatial awareness.

在一篇爆火的帖子中,Anthropic的Jake Eaton分享了几幅看起来像是画作的东西,并说,过去几个月我一直在让我们的模型画画。

In a wildly viral post, Jake Eaton from Anthropic shared a number of different what appear to be paintings and said, for the past few months I've been asking our models to paint.

Opus 5.5 非常擅长模仿不同的风格。

Opus 5.5 is very skilled at emulating different styles.

这里的每一张图都是由 Python 程序逐像素生成的。

Every image here is a Python program generated pixel by pixel.

没有用到任何图像模型,也没有现成的艺术软件。

There is no image model and no off-the-shelf art software.

实际上是大约七千五百行代码,用标准库来模仿不同的笔触风格。

Instead, it's about 7,500 lines of code using standard libraries to emulate different brush styles.

这些智能体完全不参考任何图片,仅仅依靠它们对每位画家的了解来工作。

The agents don't use any pictures as reference, instead working only from what they know about each painter.

当然,也正是在这里,你可以看到这些能力——虽然在社交媒体上看着挺酷,但对大多数人来说未必那么相关——开始变得稍微有点实际意义了。

And certainly this is where you can see these capabilities, which might otherwise be cool to look at on social media, but not necessarily all that relevant for most of us, maybe start to become a little bit more relevant.

如果 Opus 5.5 的高级编程能力足够强,就像 TAC 的这个帖子展示的那样,能一次性生成一段三十秒的营销动画,那么瞬间就会打开一堆你之前可能没想到的用例。

If the advanced coding capabilities of Opus 5.5 are enough that it can, as in the case of this post by TAC, generate a one-shot 30-second-long marketing animation, well, certainly all of a sudden a bunch of use cases open up that you might not have thought of before.

Claude Code 团队的 Tariq 让 Opus 对他的个人网站做了好几次重新设计,然后用所有迭代版本做了一个预告片。

Tariq from the Claude Code team asked Opus to do a bunch of redesigns on his personal website and then make a trailer with all of its iterations.

乍一看,它做得相当不错。

And at first glance, it seems to have done a great job.

顺便说一句,这也是我在 Opus 5.5 上试过的一个用例。

For what it's worth, this is one use case that I tried on Opus 5.5 as well.

它对我那个 AI Daily Brief 网站的分析和评价,我确实很满意。

And I was definitely pleased with its analysis and review of my AI Daily Brief website.

分析清晰,思路简洁,而且非常实用。

It was clear, had good concision of thought, and was very practical.

不过我得说,对于那些只想分出绝对好或绝对差的人来说,它并不一定比 Fable 5.1 的分析更好,只是风格不同而已。

Although I will say for those who are just trying to live in the world of strictly betters or strictly worsts, It wasn't that it was necessarily strictly better than Fable 5.1's analysis, it was just different.

但在一个方面,几乎所有人都一致认为 Opus 5.5 确实更好了,那就是写作。

And yet one area where almost everyone seems to be in agreement that Opus 5.5 is indeed strictly better is writing.

Anthropic 的 Sholto Douglas 转发了他们的公告,并写道:重要消息,我们修复了写作。

Anthropic's Sholto Douglas reposted their announcement and wrote, Also important news, we fixed the writing.

那么这意味着什么呢?

So what does that mean?

嗯,一部分原因,正如 Theo 用全大写说的那样,他们去掉了破折号。

Well, part of it is, as Theo put it in all caps, they removed the em dashes.

但 Anthropic 官方账号实际说的是:Opus 5.5 的交流更自然,解决了我们在 Opus 5 上收到的最常见的反馈。

But what the Anthropic account actually said about this is, Opus 5.5 communicates more naturally, addressing some of the most common feedback we heard on Opus 5.

它会把最重要的信息放在前面,并且遵循你给出的写作规则,这样长时间的使用过程就更清晰了。

It puts the most important information up front and follows the writing rules you give it, which makes long sessions easier to follow.

Every 团队——他们拥有测试 AI 写作的最佳基准和流程之一——写道:Opus 5.5 产生了我们见过的 Anthropic 或 OpenAI 模型中最易读的文章。

The team at Every, who have one of the best benchmarks and processes for testing AI writing, wrote, Opus 5.5 produces the most readable prose we've seen from an Anthropic or OpenAI model.

事实上,预示着一个在后面对话中会反复出现的观点:他们写道,虽然 Opus 5 曾让他们的作家把 Claude 模型踢出了房间,但 5.5 —— 引用原话 —— “让我们想把 Claude 请回房间”。

In fact, presaging a point that would come up a lot more later in the conversation, They wrote that while Opus 5 had made their writers kick Claude models to the curb, 5.5, quote, makes us want Claude back in the room.

他们接着说:它能很好地回应反馈,基于你给出的素材进行构建,并且会解释自己的选择,而 Opus 5 不会。

They continue, it responds well to feedback, builds on the material you give it, and explains its choices where Opus 5 wouldn't.

重要的是,这不仅是写作输出的改进,更是写作过程本身的改进。

And importantly, it turns out that this isn't just an improvement on the output of writing, but an improvement on the actual process itself.

他们写道:和 Opus 5.5 一起写作是一种享受。

They write that Opus 5.5 is a pleasure to write with.

它接受反馈,不会反抗,并且会在素材基础上进一步完善,而不是仅仅把素材整理得更整洁。

It takes feedback without a fight and builds on material instead of handing it back tidier.

事实上,团队写道:最大的升级其实并不完全在于写作,而在于这个模型的感觉和氛围。

In fact, the team writes, the biggest upgrade is not even exactly about the writing, it's about the feel and vibe of the model.

正如 Every 团队所说:Anthropic 修复了 Opus 的个性。

As the Every team puts it, Anthropic fixed Opus's personality.

它不再是个顽固的小家伙了。

It's not an obstinate little turd anymore.

这是很多人的共同体验。

And this is an experience that many have had.

AI 构建者 McKay Wrigley 写道:Opus 5.5 等于我们迫切想要找回的 Opus 4.6 的个性,加上 Fable 5.1 的智慧和品味。

AI builder McKay Wrigley writes, Opus 5.5 equals the personality of Opus 4.6 that we all desperately wanted back and the intelligence and taste of Fable 5.1.

而Anthropic当然知道这一点。

And certainly Anthropic knew it.

Anthropic的Nat McAleese发帖说,Opus 5.5比Opus 5好太多太多了。

Nat McAleese from Anthropic posted, Opus 5.5 is way, way, way better than Opus 5.

抱歉,那个模型不怎么样。

Sorry about that model.

请试试这个。

Please try this one.

那么,大家用Opus 5.5有没有遇到什么问题?

So have people had any issues yet with Opus 5.5?

有一个问题特别突出,对某些使用场景来说,基本上就是行不通。

There is one that stands out and that for certain use cases will be effectively a non-starter.

X上的Chief of Staff写道,我把所有使用额度都烧光了,用来对比Opus 5.5和Fable 5.1在复杂开放式法律工作中的表现。

Chief of Staff on X wrote, I burned all my usage reset benchmarking Opus 5.5 against Fable 5.1 for complex open-ended legal work.

也就是起草给客户和法院的文件、合同修订、法律和案件相关的研究。

I.e., drafting client and court-ready documents, draft and contract redlining, legal and matter-related research.

据我所知,Opus 5.5要差得多,因为安全拒绝增加了,而且我猜你们会称之为努力预算分配不当。

As far as I can tell, Opus 5.5 is much worse, due to increased safety rejections and what I'd guess you would call poor effort budgeting.

Opus 5.5看起来不错,但作为一家生命科学商业化公司,我有点担心这一点:因为Opus 5.5在生物学和网络安全方面与Claude Mythos 5.1相当,所以我们部署它时采用了与Claude Fable 5.1类似的安全措施。

Opus 5.5 looks good, but as a life science commercialization company, I'm a bit concerned by this: Because Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, We're deploying it with safeguards similar to those on Claude Fable 5.1.

由于这些限制,我们在使用Fable完成某些任务时遇到了问题。

We've faced issues using Fable for some tasks because of these guardrails.

我们申请了Anthropic的生命科学验证计划,但尚未获批。

We've applied for Anthropic's Life Science Verification Program but haven't yet been approved.

之前Fable因为生物安全限制拒绝请求时,我们会用Opus,但现在看来Opus也行不通了。

We had used Opus when Fable refused a request because of Fable's biology guardrails, but now it seems like this won't be possible anymore with Opus either.

虽然不是普遍现象,但我确实看到不少人抱怨这些过于严格的安全措施,不幸的是,如果这正好影响了你的使用场景,那基本上这个模型就没用了。

It's not everywhere, but I certainly have seen a number of people complaining about these overzealous safeguards, which is unfortunately the type of thing where if that hits your use case, it basically makes the model null and void.

那么,在这一大波发布之后,大家有什么重要的收获?

So what are some of the big takeaways for folks after this huge slate of releases?

AI领域真是疯狂的一天。

What an insane day in AI.

前沿模型的价格大幅下降,Opus 5.5降价了,现在GPT-6 Sol和Luna的token价格又降低了50%。

The frontier models just became substantially cheaper with the Opus 5.5 price cuts and now with GPT-6 Sol and Luna dropping token prices by 50%.

AI中每任务成本在同等条件下的下降速度是历史上任何其他技术都无法比拟的,而且每次AI成本下降,可以部署agent的使用场景就会大幅增加。

The rate at which the cost per task on a like-for-like basis drops in AI is unlike any other type of technology in history, and every time the cost of AI drops, the use cases you can deploy agents against dramatically increase.

这就是Jayvon悖论在agent上的体现。

This is Jayvon's paradox applied to agents.

Aaron继续预测说,这些改进将直接导致AI在经济中更广泛的普及,因为我们可以用agent来处理所有数据、扫描代码以发现安全问题、读取所有日志数据来做决策、让agent运行自动化工作流等等。

Making a bit of a prediction, Aaron continues, these improvements will directly lead to broader diffusion of AI in the economy, as we can use agents to process all of our data, scan our code for security issues, read through all log data to make decisions, have agents run automated workflows, and much more.

token的成本直接关系到这些使用场景能否大规模开放。

The cost of tokens is directly correlated to these use cases being opened up at scale.

而且还有一些研究数据支持这一点?

And putting some research heft behind that?

Epic AI Research昨天也发布了数据,认为AI变便宜的速度比历史上任何其他变革性技术都要快。

Epic AI Research yesterday also dropped data arguing that AI is getting cheaper more quickly than any other transformative tech in history.

在给定性能水平下,自2023年以来,成本每季度下降约47%。

At a given level of performance, cost has fallen around 47% per quarter since 2023.

他们指出,这比DNA测序快4倍,比计算能力快6倍,比锂电池快18倍,至少到1973年,比电力的成本下降快54倍。

They point out that that's 4x faster than DNA sequencing, 6x faster than compute, 18x faster than lithium batteries, and at least up to 1973, 54 times faster than electricity.

现在,如果要进行正面比较,你会发现人们对这些模型中哪个总体上最有价值争论不休。

Now when it comes to a head-to-head analysis, you can find people arguing all sides when it comes to which of these models are the most valuable in aggregate.

To Your Taxes写道,更便宜、更好的SOL和LUNA。

To Your Taxes writes, cheaper, better SOL and LUNA.

我猜它们是原生训练来当Astra的手下的。

I assume they're natively trained for being Astra's henchmen.

这大大加强了OpenAI的价值主张。

This makes OpenAI's value proposition considerably stronger.

我怀疑Opus 5在成本和多agent方面的优势实际上被Astra+SOL+LUNA抵消了。

I suspect that Opus 5 wins on cost and multi-agent are effectively negated with Astra+ SOL+ LUNA.

还有很多人只是乐于享受每个模型各自带来的独特价值。

And there are also many folks just happy to enjoy the unique values that each of these different models provided on their own terms.

在回答谁赢了这个问题时,Chubby写道,OpenAI和Anthropic都给了我们很多让人兴奋的东西。

In answering the question of whether there was a winner, Chubby writes, both OpenAI and Anthropic gave us plenty to be excited about.

如果你对OpenAI的策略有任何疑问,Sam Altman的态度再明确不过了。

And if you have any question about the strategy that OpenAI is pursuing, Sam Altman could not have made it any more clear.

他写道,我们希望OpenAI API在每个价位都能提供最好的模型,并且在每种模态——文本、代码、图像、视频等等——都是最好的。

He wrote, We want the OpenAI API to feature the best model at every price point and to be the best at every modality— text, code, image, video, etc.

然后我们希望大家能想出好主意,去实现它们,成为开心的用户。

And then we all want you to come up with great ideas and build them and get to be happy users.

再说一遍,记住他在另一篇帖子里写过,我们希望人们能大量使用AI。

Again, remember that he had written in a separate post, We want people to be able to use tons of AI.

这对于探索眼前这场新文艺复兴来说很重要。

It is important to being able to explore this new renaissance in front of us.

我认为从很多方面来看,OpenAI这一系列的发布可以说是他们很明显已经追求了一段时间的策略的延续,那就是在成本和效率上非常激进地投入,并构建一个模型系列,适应人们从单一模型转向复杂模型架构的时代——这些架构实际上是根据任务匹配合适类型的智能。

I think in many ways you can view this set of releases from OpenAI as a continuation of the strategy that they've clearly been pursuing for a while now, which is be really aggressive about focusing on cost and efficiencies and building a model slate for an era in which people are moving away from a single model towards complex model architectures that actually match tasks with the right type of intelligence.

而Claude这边,感觉像是为了重新夺回一些势头。

Claude, on the other hand, feels like it was out to reclaim some momentum.

如果真是这样,他们做得相当成功。

And if that is the case, they've done so pretty successfully.

Every是这么说的:Sol 就像一次S级的iPhone发布。

Here's how Every put it: Sol feels like an S-class iPhone release.

它给你带来Astra的大部分能力,但价格只有五分之一左右。

It will give you much of Astra's power at about a fifth of the price.

Opus 5.5 才是更大的惊喜。

Opus 5.5 is the bigger surprise.

5.5 让我们团队里好几个人都动了心,想从Codex换回来。

5.5 Is tempting a few people on our team to switch back from Codex.

如果你已经在用Claude生态系统,你会喜欢这个模型;如果你是Codex用户,也值得一看,特别是针对高端编程任务。

You'll love this model if you're already in the Claude ecosystem, and if you're a Codex user, it's worth a look, especially for your top-end coding tasks.

从某种程度上说,这只不过是我之前提到过的那个分歧的延续,而且我们已经看了好几个模型迭代了。

Now in some ways, this is just a continuation of the divide which I mentioned before and which we've seen for several model iterations now.

Every的Dan在那篇文章中总结说,如果你想要一个快速、实惠的日常驱动模型,用来阅读、写作和完成任务,你会喜欢Sol 6。

Dan from Every concludes that article, you'll like Sol 6 if you want a fast, affordable daily driver for reading, writing, and getting things done.

这是我一直在用的模型。

It's the model I keep reaching for.

如果你愿意为更强大的性能——在雄心勃勃的编程和视觉项目上——付出更多,你会喜欢Opus 5.5。

You'll like Opus 5.5 if you're willing to pay more for stronger performance on ambitious coding and visual projects.

在我们测试中,它的最佳表现超过了Sol,足以把我们团队的一些人拉回Claude。

Its best work surpassed Sol in our tests, enough to pull some of our team back towards Claude.

把人们拉回Claude这个说法,确实是很多不同的人都在表达的一种感受。

And the pulled people back to Claude is a real take that I'm seeing from a number of different people.

Yuchen Jin写道,Claude回来了!

Yuchen Jin writes, Claude is back!

是时候重新打开Claude code了,我已经忽略它一个月了。

Time for me to open Claude code again after ignoring it for a month.

而且随着时间的推移,我看到人们对Opus 5.5的评价越来越正面。

And if anything, the more time that goes on, the more positive I'm seeing people get on Opus 5.5.

今天早上,前投资人、现AI构建者Jeffrey Emanuel又发言了,天哪,Opus 5.5 真是令人叹为观止。

This morning, former investor turned AI builder Jeffrey Emanuel chimed back in, my god, Opus 5.5 is breathtaking.

它就像处理不费吹灰之力一样,轻松搞定极其棘手的难题。

It's just grinding through incredibly tricky stuff like it's nothing.

它找出了Fable和Astra有时几周都没能发现的bug和问题,展现出我从未见过的自主性和决心。

Finding bugs and problems that eluded Fable and Astra for weeks in some cases, and showing a level of agency and resolve I haven't seen before.

至于我个人,我打算再多花点时间用用这些模型,然后会在本周晚些时候出一期更偏向实用指南的节目,告诉大家如何充分利用这两个模型。

Now for me personally, I'm going to save a bit of my personal analysis for a little bit more time with the models, which will come together in more of a how-to episode for getting the most out of both of these models later in the week.

话虽如此,我觉得有几个有趣的观察值得在最后提一下。

That said, I do think that there are a few interesting observations that are worth closing on.

顺便说一句,我现在讲解时用的这张图是由GPT-6 SOL生成的,它加了不少我原本没有的文字,但我还是留着它,或许可以作为一个模型怪癖的例子吧。

By the way, this visual that I'm talking over was created by GPT-6 SOL, Which added quite a bit of text that I didn't have in there, but I left it perhaps as an example of particular model quirks.

观察第一点,即使是那些切换模型很果断、嘴上说只关心性能的早期用户,其实也有自己的个性偏好。

Observation number 1 is that even early adopters who are more ruthless about switching and who profess to really only care about performance definitely have personality preferences.

Opus 5.5 之所以让人兴奋,很大程度上不是因为它做到了AI以前做不到的事,而是因为用起来特别爽。

So much of the excitement around Opus 5.5 is not about, wow, it does this thing which AI has never been able to do before, but about, oh my gosh, it is such a pleasure to use this.

这种体验好长一段时间都没有过了。

In a way that it hasn't been for some time.

我其实觉得这一点非常重要。

I actually think that this is extremely important.

之前ChatGPT-4o被下架时引起那么大争议,大家容易把它当成只有那些狂热的普通用户才在意的事情——他们迷恋一个讨好型的模型。

When there was such a big dust-up around ChatGPT-4o being deprecated, there was a temptation to treat it like a phenomenon exclusively for crazy normies who had gotten obsessed with a sycophantic model.

但随着时间的推移,虽然AI对我们有些人来说是个工具,但它明显不是那种普通的工具。

But as time has gone on, for as much as AI might be a tool to some of us, it's clearly a very different type of tool.

这个工具至少是近似于有性格的,而不同版本的性格用起来爽不爽,实际上会影响到我们能从中获得多少价值。

It's a tool that at least approximates having a personality, and different versions of that tool's personalities are more or less enjoyable to use, actually has implications for how much we can get out of them.

简单来说,在大模型这里,性格就是用户体验,我们得把它当作用户体验来对待。

In short, when it comes to LLMs, personality is UX, and we should make sure we're treating it as such.

观察第二点。

Observation number 2.

虽然如果单看话题度和兴奋度,Opus 5.5肯定是赢家,但大模型的竞争现在早就不只是比一件事了。

While it is definitely the case that if you were just looking to declare one winner from the standpoint of buzz and excitement, it would be Opus 5.5, the battle of LLMs is not really about one thing anymore.

首先得搞清楚你需要哪些选项组成一个堆栈,然后是为堆栈里的每个位置展开竞争。

It's about one, figuring out what the stack of options needs to be, and then 2, battling for each slot in that stack.

换句话说,即使是最喜欢Opus 5.5的人,也不会说你得在所有场景都用它。

In other words, even the people who loved Opus 5.5 the most aren't saying you have to use this for absolutely every single use case.

而且我甚至觉得,虽然乍一看Opus 5.5和GPT-6 Sol是直接竞争对手,但两家开发者的思路其实有点不同。

And in many ways, I'd even argue that although at first glance, Opus 5.5 and GPT-6 Sol seem to be direct competitors, the way that their respective builders are thinking about them is a little bit different.

现阶段重要的是个人和公司能理解自己的完整需求,以及不同模型和工具组合如何最好地满足这些需求——在性能、能力、性价比和速度这些方面。

What matters at this stage is individuals and companies being able to understand their complete set of needs and how different models and harnesses combine to best fill those needs with respect to performance, capability, cost efficiency, and speed.

第三点观察是,我们现在肯定处于一个无法完全把模型和工具分开的时代。

A third observation is that we're definitely in the era where you can't really separate models from harnesses anymore, at least not fully.

比如很多人已经完全投入Codex生态了,就算Opus 5.5好很多,对他们来说也没区别。

For a lot of folks who are fully invested, for example, in the Codex ecosystem now, it doesn't matter if Opus 5.5 is much better.

只要OpenAI的选项足够接近,他们就不会完全迁移出那个工具。

As long as their OpenAI options are close enough, they're not going to fully shift out of that harness.

Prime Intellect的Will Brown写道:我每周都在换桌面agent,已经失控了。

Will Brown from Prime Intellect wrote, I keep switching desktop agents every week and it's getting out of control.

Codex没有Fable 5.1,Claude没有Astra,AMP两者都有但现在没有Opus 5.5了,所以我换回了Claude。

Codex didn't have Fable 5.1, Claude didn't have Astra, AMP had them both but now doesn't have Opus 5.5 so I'm back to Claude.

自己搭建又太麻烦,维护不起来。

Self-build is annoying to maintain.

我们这期没怎么聊工具,但大概一周后,真正的检验就是:有没有人因为模型变化而真的改变了自己的整体使用习惯?还是大家测试了一圈,找到自己喜欢的,然后回到之前用的生态——主要受工具影响,因为你的上下文、工具、规则都存那里。

We didn't talk much about harnesses in this episode, but give it about a week and the real proof in the pudding will be whether anyone has actually switched their overall behavior because of any of these model changes, or whether everyone tested things, found out what they liked, and just went back to the ecosystem that they were already using before, dictated in large part by the harness where they store their context, their tools, their rules, etc.

第四点观察其实不关Anthropic或OpenAI的事,而是随着这些模型发布,大家热议的另一个话题是——我认为第一个真正受欢迎的AI产品,用户根本不知道也不在乎底层模型是什么。

A 4th observation is actually not so much about either Anthropic or OpenAI, but about the fact that as these models were being released, the other big buzzy thing that people were talking about was, I would argue, the first AI product that is seeing a lot of traction where the people using it genuinely don't know or don't care about the model underneath.

我说的当然是Muse,从Meta首席AI官Alexander Wang的帖子就能看出来,Meta团队对它的接受度和反响非常兴奋。

I'm talking of course about Muse, which you can tell just from the posting from Meta's Chief AI Officer Alexander Wang, the team over there at Meta is very excited about the uptake and reception for.

如果模型真的开始被抽象掉,这会怎么改变我们讨论新模型的方式呢?

How will it change this sort of new model conversation if the models actually start to get abstracted away?

这件事我们讨论了很久,但到目前为止还没真正实现。

That's something we've talked about for a long time, but which so far hasn't really played out in practice.

但谁知道呢,也许我们正处在AI的产品时代而不是模型时代的开始,模型真的会融入系统消失不见。

But who knows, maybe we're at the beginning of the product rather than the model era of AI in which the models really will disappear into the system.

今天的第五点也是最后一点观察是,这一切对所谓的'前沿节奏'说明了什么。

The 5th and final observation for today is what this all says about pacing the frontier.

X上有个经典段子能拿到很多转发和点赞:人们两眼无神地盯着两天内四个主要模型发布,然后问——这就是前沿节奏?

One of the standard jokes that could get you a bunch of reposts and likes on X was people staring with bleary eyes at 4 major model releases in 2 days, and asking some version of, this is pacing the frontier?

但我认为Theo说得对,他写道:我的辛辣观点是,这就是前沿节奏该有的样子。

But I think Theo had the right of this when he wrote: Hot take: this is what pacing looks like.

今天的这些发布没有一个达到了Astra或Fable那个级别。

None of today's releases were Astra or Fable tier.

这是有意为之的。

This is intentional.

节奏控制的目的不是停止迭代和改进。

The point of pacing isn't to stop iteration and improvement.

目标是防止更大模型的发展失控。

The goal is to prevent the development of bigger models from spiraling out of control.

Opus和Sol级别的模型是目前我们关注的很好的方向。

Opus and Sol-class models are a great place for our focus to go right now.

有很多取得真正成果的机会,而且风险没那么大。

Lots of opportunity for real wins without as much risk.

我认为这是对的。

And I think that that's correct.

如果这就是我们在前沿时代放缓节奏所得到的——换句话说,就是那些能解决前代模型特定问题、并以显著的成本和效率提升带来渐进式性能或价值增益的模型——那这似乎是个值得花点时间的好方向。

If this is what we get from the pacing the frontier era, in other words, models that solve specific problems with previous iterations of those models and deliver incremental performance or value gains at significant cost and efficiency gains, that seems like a pretty good place to spend some time.

而且这并不是一个会导致各类应用放缓的方向。

And not one where somehow all sorts of adoption is going to slow down.

Claude Code团队的Tariq明确指出了这是实验室的一个关键方向。

Tariq from the Claude Code team is clearly pointing at this as a key direction for the labs.

他写道,正确使用模型能力的方式不是向生产环境推送10倍更多的功能。

He wrote, the right way to use model capabilities is not to ship 10x more features to prod.

而是花更多时间了解用户、做实验、构建原型、学习你不理解的东西,这样你才能交付真正有效的东西。

It's to spend more time understanding your users, trying experiments, building prototypes, learning about things you don't understand so that you can ship things that actually work.

事实上,我认为你可以说,因为自ChatGPT之后生成式AI的整个历史中,能力提升如此迅猛、如此快速,实验室们其实从来不需要从产品的角度去思考。

In fact, I think you could argue that because there have been such radical capability increases at such a fast clip for the entire history of generative AI post-ChatGPT, the labs have never really had to think in product terms.

整个时期里原始能力的跃升如此显著,以至于他们只要把最新东西甩给我们,我们就会照单全收。

The raw capabilities jumps have been so pronounced throughout the entire period that they can just splatter the latest thing at us and we're gonna eat it up.

但到了某个点,愿意这么做的市场人群饱和了,那种方法能覆盖的商业用例比例也饱和了,你就必须开始从人和系统的实际情况去考虑。

But at some point, the market of people who are willing to do that saturates, the percentage of business use cases that that approach can get to saturates, and you actually have to think in terms of human and system realities.

不过,我完全不认为这意味着所有资源会突然都流向完善现有类别的模型,而不是尝试构建更大的东西。

Now, I don't at all think that somehow this means that all of a sudden all the resources are going to go to perfecting existing classes of models rather than trying to build bigger things.

但这确实表明,无休止地追求更大模型并不是实验室唯一最重要的事。

But it certainly shows that the endless pursuit of bigger models is not the only be-all and end-all for the labs.

总之,各位,今天是个相当了不起的日子。

Anyways guys, it is a pretty phenomenal day.

如果你还没试过,我强烈建议你去花点时间体验一下这些新模型。

If you haven't yet, I would highly encourage you to go out and spend some time with these new models.

那么,这期AI Daily Brief就到这里。

For now, that is going to do it for this edition of the AI Daily Brief.

一如既往感谢你的收听或收看。

I appreciate you listening or watching as always.

下次见,peace。

Until next time, peace.
已剔除 5 处广告(点击展开查看)
M1
M10:38广告 · 已剔除

First of all, thank you to today's sponsors, KPMG, Blitzy, Section, 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
M111:29广告 · 已剔除

That's why KPMG's You Can with AI is back with a new season featuring conversations with leaders like Surojit Chatterjee of EMA, Mei Habib of Ryder, McKesson CIO Ellery Fisher, and others focused on practical execution.

M1
M111:52广告 · 已剔除

Go listen and subscribe at www.kpmg.us/AIPodcasts.

That's www.kpmg.us/AIPodcasts.

M1
M112:49广告 · 已剔除

See how at blitzy.com.

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

M1
M113:37广告 · 已剔除

Check out Section at sectionai.com.

That's sectionai.com.

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.