新企业之战：谁拥有模型

今天《AI每日简报》继续聚焦模型动态，昨天推出的一款新型开源权重模型，企业们值得重点关注。

Today on the AI Daily Brief, the month of model continues, and businesses might wanna pay attention to this new open weight model introduced yesterday.

在那之前，先看几条头条：苹果正在物色芯片收购目标，微软与OpenAI、Anthropic展开竞争，Cursor也在更深地布局模型领域。

Before that in the headlines, Apple hunting for a chip acquisition, Microsoft competing with OpenAI and Anthropic, and Cursor also getting deeper into the model game.

《AI每日简报》是一档每天更新的播客和视频节目，聚焦AI领域最重要的新闻与讨论。

AI daily brief is a daily podcast and video about the most important news and discussions in AI.

好了，各位。

Alright, friends.

在正式开始之前，先说几件小事。

Quick announcements before we dive in.

首先，感谢今天的赞助商：普华永道、Rackspace、Blitzy 和 Airtable。

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

想收听无广告版本的节目，请访问 patreon.com/aidailybrief，或者在Apple播客上订阅。

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

如果想了解节目赞助事宜，欢迎发送邮件至 sponsors@AIdailybrief.AI。

And to learn more about sponsoring the show, send us a note at sponsors@AIdailybrief.AI.

今天这期节目有个特别有意思的地方，新闻部分和主内容其实主题高度一致，因为各位，我们正处在一个模型主导的时代。

Today's one of those days where the headlines part of the episode actually has a pretty coherent theme with the main episode because, friends, we are in our model era.

最近有报道提到，Cursor在五月开过一次全员会议，CEO迈克尔·特鲁尔阐述了公司未来的发展方向。

The information recently reported on an all hands meeting that Cursor held back in May where CEO Michael Trull laid out the future direction for the company.

这场会议是在SpaceX交易公布几周后召开的，当时那笔交易是计算资源与训练方面的合作，附带收购选项，但即便那时，大家也普遍觉得，等SpaceX上市后，这笔收购基本板上钉钉。

The meeting took place a couple of weeks after the SpaceX deal was announced, which at the time was a compute and training partnership with an acquisition option, although even then it seemed pretty likely that the deal would be finalized after the SpaceX IPO.

特鲁尔告诉员工，Cursor的目标是成为一流的模型研发公司，而不仅仅局限于AI编程工具。

Trull told staff that Cursor would aim to become a top tier model developer in their own right, not exclusively limited to AI coding.

短期内，他计划在今年底前推出一款顶尖水平的模型，到2027年，他希望积累显著的算力优势，真正推动技术前沿的发展。

Indeed, in the short term, he aimed to produce a state of the art model by the end of the year, and by 2027, he wanted to accrue a significant compute advantage and actually push the frontier forward.

特鲁尔还提到，员工中确实有人对被SpaceX收购这件事感到不安。

Truell also apparently acknowledged some unease among staff about the SpaceX acquisition.

他承认Cursor正处于快速变革和高速增长的阶段，并承诺会随着交易进展，不断给大家更清晰的说明。

He described Cursor as being in the midst of rapid change and significant growth and promised to provide more clarity around the deal as it developed.

他还解释说，SpaceX看中的是Cursor的品牌影响力、现有的企业客户关系，以及更成熟的市场推广团队。

Truell also explained that SpaceX wanted to leverage Cursor's brand, existing enterprise relationships, and larger go to market team.

此后，我们看到了首个由SpaceX与Cursor合作训练的AI模型Grok 4.5发布。

Now subsequent to this, we've seen the release of the first SpaceX AI model trained in partnership with Cursor in Grok 4.5.

撇开最近关于数据留存的争议不谈，这个模型本身反响不错，表现也很有竞争力，尤其是在我们进入更先进的模型架构时，各大公司都在把顶尖模型和性能稍弱但更实惠的模型搭配使用。

And holding aside recent controversy around data retention, the actual model itself has been well received and seems competitive, especially as we get into more advanced model architectures where companies are pairing state of the art and frontier with slightly less performant but slightly more affordable models.

幕后消息说，Cursor还在开发一款对标Claude Cowork的产品，这将是他们首次大规模拓展到编程之外的领域。

Behind the scenes, Cursor is also rumored to be working a competitor to Claude Cowork, which would be their first big expansion beyond coding.

很明显，马斯克和SpaceX的AI团队对这次整合有着宏大的计划。

Now it's very clear that Elon and SpaceX AI have big plans for this integration.

但对一直关注的人来说，这更多是顺势而为，而不是颠覆性转变——毕竟早在去年年底，Cursor就已经决定，模型研发这场游戏，他们必须亲自上场。

And for those who've been paying attention, this is more evolution than revolution given that Cursor had already decided as early as the end of last year that the model development game was going to have to be a game they played.

现在我们终于确认了，外界看到的所有信号，确实反映了公司内部的真实想法。作为消费者，我觉得这意味着我们会有更多优质的选择，我很期待他们接下来的举动。

Still, we have now confirmation that all of the signals we're seeing from outside are what they're talking about inside, and I think as consumers, you gotta think that this just means more quality options for us, so I'm excited to see what they do.

再说回SpaceX的市场表现，公司股价首次跌破了IPO价格。

Now speaking of SpaceX on the market side, the company's stock dipped below its IPO price for the first time.

周三，SpaceX股价跌至133美元，首次跌破了135美元的IPO价。

On Wednesday, SpaceX traded down to a $133, breaking through the $135 IPO price for the first time.

不过当天收盘时略有回升，收于135.27美元，但股价相比历史高点已下跌33%，埃隆·马斯克也失去了万亿富翁的身份。

It did recover very slightly to close the day at $135.27 but the stock is now down 33% from its all time high and Elon Musk has lost his trillionaire status.

当然，马斯克似乎并不太在意，反而把注意力放在了下一次星舰发射上。

Musk, of course, doesn't seem all that concerned focusing instead on the next Starship launch.

与此同时，不少看空者指出，内部人士的禁售期还要一个月才开始。

Meanwhile, there are plenty of bears pointing out that insider lockups won't even start for another month.

这件事更重要的影响，可能不在SpaceX本身，而在于它对其他准备上市的AI公司意味着什么。

The more significant implications of this than anything happening in the SpaceX world may be actually around what it means for other AI companies looking to IPO.

我们上一次听说的是，OpenAI越来越倾向于把IPO推迟到明年。

Last we had heard, OpenAI were increasingly inclined to wait until next year for their IPO plans.

截至六月底，有顾问 reportedly 告诉萨姆·阿尔特曼，他们认为实现他设定的万亿美元市值目标可能性不大。

As of late June, advisors were reportedly telling Sam Altman that they thought that they were unlikely to achieve his target of a trillion dollar market cap.

另一方面，Anthropic似乎仍计划在今年秋季进行IPO。

Anthropic, on the other hand, seemed to still be on track for an IPO in the fall.

彭博社报道称，Anthropic已经任命了投资银行主导此次上市，并将在未来几周与投资者会面。

Bloomberg reports that Anthropic has appointed investment banks to lead the offering and are setting meetings with investors over the coming weeks.

Anthropic原本计划最早在九月或十月上市，这个目标仍然有可能实现。

Anthropic had been targeting an IPO as early as September or October, which could still be possible.

公司还新增了数十亿美元的循环信贷额度，明显是在为进军公开市场做准备。

The company has also added a couple billion dollars in a revolving credit facility, sending all the signals that they are, in fact, gearing up for a run at the public markets.

不过接下来几周至关重要，Anthropic将与公众投资者会面，试探市场反应，看看现在是否适合上市。

Still, it feels like the next few weeks are going to be an extremely important period as Anthropic will be meeting with public investors to test the waters and see whether an IPO is viable at this time.

再回到模型领域，英伟达又发布了一款新的开源模型，这次是想填补一个空白领域。

Moving back to models, NVIDIA has released another new open source model, this time attempting to fill a missing niche.

这款名为Cosmos 3 Edge的模型是一款轻量级物理AI模型，参数量仅有40亿。

The model, called Cosmos three Edge, is a small physical AI model weighing in at just 4,000,000,000 parameters.

它专为部署在机器人等边缘设备上设计，比如英伟达的Jensen硬件——这周刚更新了产品线。

It's designed to be run on edge devices that are physically installed in robots such as NVIDIA's Jensen hardware, which also saw a product refresh this week.

这款模型采用双重架构，既能当世界模型用，也能像普通大语言模型那样处理视觉与语言任务。

The model takes a dual approach to robotics, able to function as both a world model as well as a vision language model similar to standard LLM architecture.

上个月，英伟达发布Nemotron 3 Ultra的时机还挺巧。

Last month, NVIDIA's release of Nemotron three Ultra was pretty timely.

这个模型是他们开源模型系列中更大、更强的版本，旨在满足对强大且低成本的美国本土模型的需求，以应对不断攀升的令牌开销。

That model was a larger and more capable version of their open model family, addressing the demand for powerful and cheap US based models to deal with spiraling token budgets.

这次，NVIDIA推出的新物理AI模型，恰逢新一轮机器人设计纷纷亮相。

This time, NVIDIA's new model for physical AI arrives as another wave of robot designs are unveiled.

上周，One X发布了他们Neo人形机器人的新版本，手部动作已经超越了人类水平。

Last week, One X previewed the new version of their Neo Humanoid with superhuman hand movements.

NVIDIA本身也在这个领域积极建立大量合作，包括与丰田的深化合作。

NVIDIA themselves are striking up a ton of partnerships across the space, including an expanded partnership with Toyota.

继近年来共同研发自动驾驶汽车后，双方的合作现在将扩展到机器人、智慧城市和自动化工厂等领域。

After working together on self driving cars in recent years, the partnership will now include robots, smart cities, automated factories.

稍微岔开一下话题，我一直关注着物理AI领域的动态。

Now as a slight detour, one thing that I do keep a bit of an eye on is everything going on with physical AI.

目前，我们在这里的讨论显然主要集中在生成式大语言模型上，而且特别偏向于与工作相关的内容。

Right now, obviously, the main focus of most of our discussions here is in the generative LLM world with a particular bias towards what matters at work.

但你得明白，未来几年，我们会越来越多地关注AI在物理世界中的应用。

But you have to think that over the next few years, we're gonna have a lot more context to focus on AI in the physical world.

有一个特别有意思的例子，是一家刚上线的公司，叫Chip。

One really interesting example of that of a company that just launched is called Chip.

Chip 算是近几年来，我们看到的第一个对家用车型的全新设计。

Chip is honestly one of the first new takes we've seen on a family vehicle for some time.

过去几年，高尔夫球车那种小型车确实越来越流行，但Chip直接把这事儿推到了新高度。

Now golf cart style smaller vehicles have been getting increasingly popular over the last couple years, but Chip kinda takes things to the next level.

它的理念是把人工智能和自动驾驶结合起来，让你的车能更高效地变成一个替你办事的机器人。

The idea is that it combines AI and autonomy to more efficiently turn your car into a robot that can do things on your behalf.

公司刚发布的首发视频里展示了几个场景：比如一位父亲在足球赛现场，派Chip去买东西、买零食；一对夫妇决定从酒吧打车回家，就让Chip自己去停车——很明显，他们想重新定义一辆车能做什么。

The company just went live in the initial launch video, shows examples like a father at a soccer game sending Chip to run errands and pick up snacks, to go park itself after the couple decides that they're gonna take an Uber home from the bar, and clearly the idea is to rethink what a vehicle can actually do.

你可以去他们刚上线的网站 chipmotors.com 看看，这其实预示着未来几年，AI将如何深度融入物理世界。

You can check it out at the just launched chipmotors.com, but as a herald of the type of ways we're gonna see AI integrate itself into the physical world in the coming years.

说到芯片，不过是指另一种芯片了，苹果似乎正在物色收购目标，想自研AI服务器芯片。

Now speaking of chip, but a very different type of chip, Apple is apparently shopping for acquisitions to build AI server chips.

据消息显示，苹果在收购一家芯片公司这件事上，已经进展得相当深入了。

According to the information, Apple is quite far along in an effort to buy up a chipmaker.

据报道，他们已经与银行讨论过为潜在收购提供资金，并接触了一些半导体初创公司，试探他们的意向。

They've reportedly spoken with bankers about funding possible deals and have approached semiconductor startups to gauge interest.

苹果一向避开大规模收购，更倾向于自主研发能力。

Now Apple has famously shied away from big acquisitions, preferring to develop capacity internally.

事实上，他们有史以来最大的一笔收购是2014年收购Beats，花了30亿美元。

In fact, their biggest acquisition ever is the Beats deal in 2014, which cost just $3,000,000,000.

现在有种说法是，AI时代可能正在改变他们的想法，有消息推测，他们最近的行动源于一个认识：苹果目前缺乏支持AI产品所需的技术。

There is a sense that the AI era might be changing their tune on that, and the information speculates that their most recent push grew out of a realization that Apple doesn't have the technology they need to support AI product.

M系列芯片在专业消费市场表现非常出色，Mac mini 一上市就被抢购一空。

The M series chips have been a massive hit when it comes to local inference in the prosumer market, I.

但苹果在运行AI服务器时，依然依赖相同的M2 Ultra芯片，而随着苹果推出AI版Siri，这些芯片可能撑不住了。

E, selling out Mac minis, but Apple relies on the same M two Ultra chips to run their AI servers, and they might not be up to the task as Apple rolls out AI Siri.

苹果不仅请谷歌帮忙开发Siri背后的模型，还把服务器算力外包给了谷歌云，用的是英伟达的芯片。

Not only did Apple contract with Google to develop the models that power Siri, they're also outsourcing server capacity to Google Cloud running on NVIDIA chips.

苹果原本在研发一款服务器芯片，代号叫Baltra，原计划今年推出，但据传这个项目已经推迟了。

Apple did have a solution in development, a server grade chip codenamed Baltra, which was expected to ship this year, but that project has reportedly been delayed.

如果苹果能找到合适的收购目标，这笔交易可能会非常大，但即便他们愿意出高价，真正靠谱的芯片公司也寥寥无几。

Now this could end up being a very large deal if Apple can find a suitable acquisition target, but even if the company is willing to pay up, the selection of proven chipmakers is pretty slim.

英伟达刚花了两千亿美元把Grok买下来，而Cerberus在公开市场的估值也达到了四千亿。

NVIDIA just paid $20,000,000,000 to take Grok off the board, and Cerberus is currently valued at $40,000,000,000 in public markets.

苹果从2014年起就和博通合作开发服务器芯片，但博通估值高达一万八千亿，这已经不是收购，而是合并了——不仅不太像苹果的风格，还会引来巨大的反垄断审查。

Apple has been working with Broadcom on server chips since 2014, but that company's $1,800,000,000,000 valuation, which would mean a merger rather than an acquisition, which not only would be not particularly Apple y but would also come with a huge amount of antitrust scrutiny.

之前有传闻说Tenstorrent收到了博通的收购要约，不过后来被否认了；但这家公司其实挺合适的，毕竟那位传奇芯片设计师、首席执行官吉姆·凯勒，当年就在苹果工作过。

Earlier, there were rumors of Tenstorrent fielding acquisition offers from Broadcom although that was later denied, and that could be a natural fit given that legendary chip designer and CEO Jim Keller worked at Apple in the late two thousands.

目前还没有任何确凿的消息。

At this stage, there's no solid rumors.

我们只知道，苹果正在积极寻找机会，想重新杀回AI战场，这对苹果粉丝来说绝对是好消息。

All we know is that Apple is in the market to buy their way back into the AI race, which will be music to Apple fans' ears.

今天最后一个故事，其实是为接下来的主节目做铺垫。

Now our last story today is actually kind of a setup to the main episode.

微软正在培训销售团队，直接跟OpenAI和Anthropic竞争，重点突出那些大型模型实验室的短板。

Microsoft is preparing their sales teams to compete directly with OpenAI and Anthropic, training them to emphasize the drawbacks of the major model labs.

据彭博社报道，微软高层在周二的全员会议上公布了这一计划。

According to Bloomberg reporting, Microsoft leadership laid out the plan in a department wide meeting on Tuesday.

这一计划被定位为从本月开始的新财年战略。

It was framed as a new strategy for the fiscal year that began this month.

销售团队被要求重点强调微软自研MAI模型在效率和成本上的优势，同时突出微软垂直整合的AI架构作为核心卖点。

Sales staff were instructed to push the efficiency and cost effectiveness of Microsoft's in house MAI models compared to their rivals, and Microsoft's vertically integrated AI stack was also highlighted as a key selling point.

目前的主打方向是对比Claude，微软正逐渐淡化对整体模型质量的讨论。

Selling against Claude appears to be the initial focus, and Microsoft is increasingly sidestepping overall model quality.

有高管 reportedly 告诉销售团队，谈到在微软Office套件中的表现时，Claude‘速度更慢、准确度更低，而且缺乏必要的安全集成’。

Instead, one executive reportedly told sales staff to say that when it comes to performance in Microsoft's Office Suite specifically, Claude is, quote, slower and less accurate and lack the proper security integrations.

微软可以指出，客户已经在使用Copilot中的MAI模型，但可能根本没察觉到区别。

Microsoft can point to the fact that customers have already begun using the MAI models in Copilot and might not have noticed the difference.

本月初，彭博社曾报道，微软已开始将部分Copilot功能转向MAI，以降低对OpenAI和Anthropic的依赖成本。

Earlier this month, Bloomberg reported that Microsoft had begun switching some Copilot functions over to MAI as a cost cutting measure compared to OpenAI and Anthropic.

这一新的销售策略，正好呼应了萨提亚·纳德拉近期关于AI时代企业数据风险的表态。

This new sales pitch aligns directly with Satya Nadella's recent commentary on the risk to corporate data in the AI era.

这条帖子，还有他几周前发的那篇，其言外之意其实很明显：企业不该把数据交给OpenAI或Anthropic，因为这两家公司都有动机去开发自己的竞品服务。

The not so subtle subtext of this post and one that he published a couple of weeks ago was that companies shouldn't trust OpenAI or Anthropic with their data because both of those companies have an incentive to build competing services.

但在我看来，这次转向也部分是为了先让微软自己的模型站稳脚跟，之后再以定制化微调作为增值服务来推动客户升级。

But it also feels to me like this pivot is in part about giving Microsoft a foot in the door with their own models before pushing customer specific fine tuning as an upsell.

上个月，微软CEO穆斯塔法·苏莱曼公布了这一长期计划，推出了微软前沿微调技术。

Last month, Microsoft CEO, Mustafa Soleiman, laid out this long term plan with the introduction of Microsoft Frontier tuning.

正如我们将在正片中看到的，企业自主微调模型这个想法，今年下半年将被多家实验室列为重要主题。

Now as we'll see from the main episode, this idea of enterprises tuning their own models is something that a number of labs are going to be pushing as a major theme for the second half of this year.

企业到底会多大程度接受这种做法，现在还不好说，但肯定会上升为行业热议的话题。

What sort of enterprise uptake of that there is remains to be seen, but it is certainly going to be in the conversation.

有意思的是，第一个宣布这类微调服务的并不是微软，而是去年年底Thinking Machines实验室推出的Tinker API。

And interestingly enough, it was not Microsoft that was the first company to announce this sort of fine tuning, but Thinking Machines Lab back at the end of last year with their Tinker API.

今天，Tinker刚刚发布了重大更新，这正是我们正片要重点聊的内容，好了，新闻部分就到这里，我们直接进入正题。

Well, Tinker just got a big new update today, and that is going to be the subject of our main episode, so with that, we'll close the headlines and head on over to the main.

现在AI领域最重要的问题，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.

普华永道和德克萨斯大学奥斯汀分校刚刚分析了140万次真实职场中的AI互动，发现了一个令人意外的现象：最有影响力的人并不是擅长写提示词的人，而是把AI当成一个思维伙伴来合作。

KPMG and the University of Texas at Austin just analyzed 1,400,000 real workplace AI interactions and found something surprising: the highest impact users aren't better prompt engineers they treat AI like a reasoning partner.

他们会明确问题、引导思考、反复迭代，并不断追求更好的答案。

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

好消息是？

And the good news?

这些行为是可以大规模培养的。

These behaviors are teachable at scale.

如果你正想从单纯使用AI转向真正提升能力，普华永道关于高级AI协作的研究非常值得一看。

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

了解更多，请访问 kpmg.com/us/sophisticated。

Learn more at kpmg.com/us/sophisticated.

详情请见 kpmg.com/us/sophisticated。

That's kpmg.com/us/sophisticated.

目前企业级AI领域一个更有趣的趋势是，讨论焦点正迅速转向基础设施和运营层面。

One of the more interesting shifts in enterprise AI right now is how quickly the conversation is moving towards infrastructure and operations.

随着人工智能深入核心工作流程、受监管的数据环境和自主系统，企业需要一套具备可操作性和明确责任机制的受控基础设施与推理系统，从一开始就确保稳定运行。

As AI moves into core workflows, regulated data environments, and agentic systems, enterprises need governed infrastructure and inference that can operate reliably day to day, with clear operational accountability built in from the start.

当这些系统不断扩展时，运营模式本身正逐渐成为人工智能战略的核心部分。

As those systems scale, the operating model increasingly becomes part of the AI strategy itself.

Rackspace Technology 是全栈企业级人工智能的运营方，覆盖私有云、混合云和边缘环境中的智能代理与底层设施。

Rackspace Technology is the operator of the full enterprise AI stack, from agents to infrastructure across private cloud, hybrid cloud, and edge environments.

Rackspace 为那些对数据主权、合规性与系统可用性有严格要求的企业，构建并运维受控的人工智能基础设施、推理系统和生产级AI应用。

Rackspace builds and operates governed AI infrastructure, inference, and production AI systems for organizations where sovereignty, compliance, and uptime are non negotiable.

他们的现场工程师在部署后仍长期驻场，协助将人工智能真正融入日常运营并保障稳定运行。

Their forward deployed engineers stay embedded beyond deployment to help operationalize and run AI in live environments.

想了解更多企业级人工智能的落地场景与成果扩展方式？请访问 rackspace.com。

To learn more about where enterprise AI runs and outcomes scale, go to rackspace.com.

想把企业软件开发速度提升五倍吗？

Want to accelerate enterprise software development velocity by 5x?

你需要 Blitzy——唯一专为大型企业代码库设计的自主软件开发平台。

You need Blitzy, the only autonomous software development platform built for enterprise codebases.

你的工程师定义项目，无论是新功能、代码重构，还是全新开发。

Your engineers define the project, a new feature, refactor, or greenfield build.

Blitzy 的智能代理会先扫描并梳理你的整个代码库，然后平台生成一份专属的执行计划，供你的团队审阅和确认。

Blitzy agents first ingest and map your entire codebase, then the platform generates a bespoke agent action plan for your team to review and approve.

一旦获得批准，Blitzy 就会自主生成数十万行经过验证、端到端测试过的代码。

Once approved, Blitzy gets to work autonomously generating hundreds of thousands of lines of validated, end to end tested code.

单次运行就能完成超过百分之八十的工作量。

More than 80% of the work completed in a single run.

Blitzy 不只是在写代码，它是在以计算的速度开发软件。

Blitzy is not generating code, it's developing software at the speed of compute.

你的工程师负责审阅、优化，然后上线。

Your engineers review, refine, and ship.

这就是财富500强企业如何把耗时数月的项目压缩到一个冲刺周期内完成的方式。

This is how Fortune 500 companies are compressing multi month projects into a single sprint.

将工程效率提升五倍。

Accelerating engineering velocity by 5x.

前往 blitzy.com 亲自体验 Blitzy。

Experience Blitzy firsthand at blitzy.com.

就是 blitzy.com。

That's blitzy.com.

本集《AI每日简报》由 Hyperagent 赞助播出，您可以在这里管理团队协作的智能代理群组。

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

新用户可获得 1000 美元的推理算力。

New users get $1,000 in inference.

别再依赖本地代理或等着在电脑上手动触发的聊天流程了。

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

Hyperagent 将智能代理部署在云端，7x24 小时运行，无缝对接您团队正在使用的各类工具。

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

市场团队的代理能将竞争对手的动态自动生成落地页。

Marketing's agent turns competitor moves into landing pages.

销售团队的代理会完善客户信息、撰写邮件，并同步更新 CRM 系统。

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

运维代理负责处理文书工作并跟踪预算。

Ops agent chases the paperwork and tracks the budget.

每个代理都能访问共享信息，并遵循你设定的范围和审批规则。

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

是时候引入那些像同事一样的代理了。

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

去 Hyperagent 上雇佣你的代理吧，这是 Airtable 团队打造的产品。

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

前往 hyperagent.com/aideallybrief 领取你的 1000 美元推理额度。

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

欢迎回到 AI 每日简报。

Welcome back to the AI Daily Brief.

今天，模型月继续进行。

Today, the month of models continues.

对于普通观察者来说，我们接下来要聊的这个模型，初看可能有点平淡无奇。

And for the casual observer, this first one we're going to talk about could at first glance be a little underwhelming.

不过在这期节目中，我们会看到，这其实属于一系列更重要的趋势的一部分，尤其是对企业来说，非常值得关注。

What we'll see though throughout this episode is that this is actually part of a much more significant set of trends, and I think particularly for enterprises, is really worth paying attention to.

Thinking Machines Lab 发布了他们的首个大型语言模型。

Thinking Machines Lab has released their first large language model.

这个模型叫 Inkling，采用专家混合架构，总参数达9750亿，其中活跃参数为410亿。

Called Inkling, the model is built on a mixture of experts architecture with nine seventy five billion total parameters and 41,000,000,000 active parameters.

它支持百万级上下文窗口，并能原生处理文本、图像和音频的多模态推理。

It supports a million token context window and reasons natively across text, images, and audio.

TML 是前 OpenAI 首席技术官米拉·莫拉迪离开公司后创立的衍生实验室，他们表示，这是他们计划发布的一系列模型中的第一个，还包括一个名为 Inkling Small 的120亿参数模型，该模型已于周三开放预览。

Now TML, which is of course the spin off lab built by former OpenAI CTO Mira Moradi after she left that company, says that this is the first in a family of models they plan to release, including a 12,000,000,000 parameter model called Inkling Small, which was also released in preview on Wednesday.

作为参考，9750亿参数的 Inkling 比 DeepSeek v4 Pro 的1600亿参数要小不少，但略大于 zAI 的 GLM 5.2 的750亿参数。

Now for context, at 975,000,000,000 parameters, Inkling is quite a bit smaller than the 1,600,000,000 parameters of DeepSeek v four Pro, but a little larger than the seven fifty billion parameters of zAI's GLM 5.2.

重要的是，TML 从零开始预训练了 Inkling，除了少量基于 Kimikay 2.5 蒸馏的初始数据外，其余微调数据全部来自 TML 自己收集的人工生成与合成数据。

Importantly, TML pre trained Inkling from scratch, and aside from a small bootstrap distilled from Kimikay 2.5, the rest of the fine tuning run was based on a set of human created and synthetic data collected by TML.

他们写道：我们训练 Inkling 成为一个广泛均衡的基础模型，在多个领域表现强劲，且具备足够的灵活性以适应不同需求。

They wrote: We trained Inkling to be a broad, balanced foundation model, strong across many domains, flexible enough to adapt.

Inkling 目前并不是最强的通用模型，无论是开源还是闭源的。

Inkling is not the strongest overall model available today, open or closed.

真正让它在开源模型中脱颖而出的，是它在定制化、多模态能力、高效推理，以及可在 Tinker 上微调等方面的综合优势。

Instead, a combination of qualities make it a good open weight space for customization, multimodal capabilities, efficient thinking, and availability on Tinker for fine tuning.

而正是这种可获取性，正如我们接下来会看到的，才是最关键的。

And indeed it is that availability which, as we'll see, is the most important part.

不过，要衡量它的实际表现，更准确的说法是：它有一定竞争力，但在各类主流基准测试中远未达到顶尖水平。

Still, to get a sense of where this ranks when it comes to performance, a fair way to put it would be to call it somewhat competitive but certainly not state of the art across a range of different benchmarks.

在 Humanity 的最后一次测评中，它在不使用工具的情况下得分为 29.7%，使用工具后提升到 46%。

On Humanity's last exam, which measures knowledge and web search, it scores 29.7% without tools and 46% with.

这个成绩比 NVIDIA 的 Nemotron 3 Ultra 略好，和 Kimikay 2.5 差不多，但比 Kimi K 2.6、GLM 5.2 和 GPT 5.6 SOL 低大约 10 个百分点，更远落后于 Fable five。

This is ahead of Nemotron three Ultra from NVIDIA, roughly in line with Kimikay 2.5, but around 10 points behind Kimi K 2.6, GLM 5.2, and GPT 5.6 SOL, and around 20 points behind Fable five.

在编程能力上，Inkling 在 SweetBench Pro 上得 54.3%，在 TerminalBench 2.1 上得 63.8%，虽然超过 Kimikay 2.5，但略逊于 Kimi K 2.6，和 GLM 5.2、5.6 SOL 以及 Fable five 相比仍有明显差距。

On coding, Inkling scored 54.3% on SweetBench Pro and 63.8% on TerminalBench 2.1, again putting it ahead of Neutron in k 2.5, but a little behind k 2.6 and at a significant shortfall compared to GLM 5.2, 5.6 SOL, and Fable five.

在 GDPVal 测试中——这个测试评估的是真实场景任务和工具调用能力——它得了 12.38 分，超越了所有主流开源模型，除了 GLM 5.2 和 DeepSeek v4 Pro，但和 5.6 以及 Fable five 相比仍有很大距离。

On GDPVal, which measures real world tasks and is a strong test of tool call execution, it scored twelve thirty eight, placing it ahead of all major open models other than GLM 5.2 and DeepSeek v four Pro, but a long way behind 5.6 and Fable five.

TML 还特别提到了 Inkling 在音频和视觉基准测试中的表现，这些数据通常不会对外公开。

TML also chose to highlight Inkling scores on some of the audio and vision benchmarks, which typically aren't published for frontier models.

在音频方面，它明显领先于大多数开源模型，堪比 Quen 3.5 Omni Plus，接近 Gemini 3.1 Pro，这点值得注意，因为后者以多模态能力见长。

On audio, it seems well ahead of most open models, a match for Quen 3.5 Omni plus and pretty close to Gemini 3.1 Pro, which is notable for its multimodal trading.

在视觉方面，它的表现和 Kimi 2.5 差不多，略逊于 Kimi 2.6 和 Gemini 3.1 Pro。

And on vision, it's roughly a match for Kimi 2.5 and only slightly behind k 2.6 and Gemini 3.1 Pro.

人工评估机构给 Inkling 的智能指数打出了 41 分，排在第十九位，落后于 Kimi 2.6、DeepSeek v4 Pro 和 MiniMax v2.5 Pro，但领先于 DeepSeek v4 Flash 和 Nemotron 3 Ultra。

Artificial analysis gave Inkling a score of 41 on their intelligence index, ranking nineteenth behind Kimi 2.6, DeepSeek v4 Pro, and MiniMax v2.5 Pro, and ahead of DeepSeek v4 Flash and Nemotron three Ultra.

它比 5.6 SOL 低了 18 分，比 Fable five 低了 19 分。

It's 18 points behind 5.6 SOL and 19 behind Fable five.

AA 也指出，Inkling 在令牌效率上表现非常出色，平均每个任务使用的令牌数只有 5.2、2.6 和 DeepSeek v4 Pro 的三分之二左右。

AA did note very strong token efficiency compared to open weights leaders, averaging around two thirds as many tokens per task as five two, 2.6, and DeepSeek v four Pro.

不过，常听节目的朋友都知道，这些基准测试通常并不能完整反映真实表现。

And yet, as regular listeners well know, benchmarks don't usually tell much of a story.

遗憾的是，虽然目前关于它的第一手体验还不多，但现有的反馈都不太理想。

Unfortunately, although there aren't all that many first impressions, the ones that we have aren't great.

伊桑·马利克教授说，很高兴看到一个新的开源模型，但在我测试中，Inkling的表现还很粗糙。

Professor Ethan Malik wrote, happy to see a new open weights model, but so far, Inkling is pretty rough in my tests.

跟国内顶尖的开源模型比，差距还很大。

Not close to frontier Chinese open weights models.

举个例子，它连LEM测试都过不了，而这个测试自从DeepSeek C1和Sonnet 3.5以来，所有前沿模型都通过了。

As one example, here it is failing to pass the LEM test, done by every Frontier model since Deepsea Car one and Sonnet 3.5.

伊桑接着问：真有人在用Inkling吗？

Ethan followed up, are people actually trying Inkling?

我试了好几次，哪怕调到最高配置，也稳定不下来，稍微简单点的请求，它的推理就乱成一团。

I can't seem to get it to work solidly on any of my tests even on extra high, and the chain of thought goes crazy at even simple requests.

是我哪里没弄对吗？

Am I missing something?

Dreams ASI的评价更尖锐，直接说：Thinking Machines该改名叫Lobotomized Machines了。

Dreams ASI was even more harsh, writing: Thinking Machines should rename to Lobotomized Machines.

第一次用Inkling的感觉，就像一个准确率更低的GPT 5.2。

First impression of Inkling feels like a GPT 5.2 that's less accurate.

企业级限制多得离谱，动不动就拒绝，好像默认用户都是在试图注入恶意提示。

Corporate guardrails galore, plenty of refusals, and treating the user as by default an adversarial prompt injector.

Lissonscaling01 写道：基准测试表现并不理想。

Lissonscaling01 writes: Benchmarks don't look that great.

这 basically 就是另一个 Kimi K 2.6，在 GLM 5.2 的测试里输给了所有闭源模型。

It's basically another Kimi K 2.6 and loses to all closed models in GLM 5.2.

这让人感觉他们是在赶着发布，想在 Kimmy K3 和 DeepSeek v4 正式上线前抢占先机。

This makes it more feel like they are rushing this release before Kimmy K3 and DeepSeek v4 GA.

但这并不是唯一的解读。

But this is not the only interpretation.

张云峰写道：几年后回头看，这个模型可能会变得很重要。

Yunfeng Zhang writes: This is going to be an important model when we look back in a few years.

这个世界需要更多公司推出开源大语言模型。

The world needs open LLMs from as many players as possible.

以 TLM 的人才、经验、资金和算力，要不了多久就能追上 Opus，甚至接近 Fable 五的水平。

Given TLMs talent, know how funding, compute, they will soon reach Opus if not Fable five level.

所以，云帆的意思是，这应该是TML推出的第一个模型，而不是最后一个，应该被视为未来更好产品的前奏。

So basically, Yunfan's argument is that this should be treated as the first, not last model that TML is going to put out and seen as a precursor for better things to come.

这个观点当然合理，但还不是最乐观的解读。

Certainly a reasonable interpretation, but not even the most positive thing out there.

Tortaxis写道：我喜欢Inkling在基准测试中表现平平。

Tortaxis writes: I like that Inkling is mediocre on benchmarks.

这说明他们在蒸馏过程中并没有过度偷工减料，他们的独立数据管道最终会显现优势，而不仅仅是靠分数说话。

This suggests that they haven't been cutting corners too much with distillation, so their independent data pipeline will shine through and it's not reducible to scores.

到第二、第三个版本更新时，他们就会成为一个不可忽视的玩家。

By the second, third update, they become a meaningful player.

这和刚才那条评论类似，都是说它是个未来可期的前奏，但也点出了一点：他们明显没有像很多中国模型那样一味追求基准分数。

So this is similar to that last comment in the sense that it's a precursor for better things to come, but also a point out that they're clearly not benchmark maxing in the same way that many Chinese models do.

不过，也有人从完全不同的角度看待这件事，他们看重的不只是未来潜力，更是当下就已经存在的价值。

Some people viewed this in fundamentally different terms, though, where the benefits weren't just about the future, but about aspects of what we have right now.

Burton Shaw写道：Inkling最厉害的地方，是它的发布方式。

Burton Shaw wrote: The sickest thing about Inkling is the manner of their release.

他们推出了一款架构和应用场景都很有新意的模型，明确告诉我们它能做什么、不能做什么，还提供了用于部署和后续训练的工具，用不用全由我们自己决定。

They brought out a model with genuinely interesting architecture and applications, told us what it does and does not do, and gave us tools to serve and post train it, which we can choose to use or not.

开源领域开始出现更多元的声音，各自对如何构建AI有着独到且合理的见解。

The open scene is beginning to get diverse voices with their own valid takes on how AI should be built.

不过杰克·莫里斯走得更远，他认为人们低估了这件事的重要性。

Jack Morris, though, goes farther, arguing that people are underestimating what a big deal this is.

杰克写道：这是唯一一个没有通过蒸馏OpenAI或Anthropic模型来训练的开源权重模型。

Jack writes: This is the only open weight model that's trained without distilling from OpenAI or Anthropic.

KIMI是通过蒸馏来的。

KIMI distills.

GLM也是蒸馏出来的。

GLM distills.

QUEN同样是蒸馏的产物。

QUEN distills.

Nemotron是蒸馏的，KIMI和DeepSeq也算在内。

Nemotron distills, Kimi and DeepSeq which counts.

本质上，这是一个完全不同的技术栈，首个纯粹开源的代码生成模型。

Basically, a fully different tech stack, the first pure open frontier coding model.

不过社区笔记指出，TML 自己也提到，Inkling 虽然是从零开始预训练的，但还是用了一点来自开源模型（包括 K 2.5）的合成数据做初步引导，同时他们也提到像 LAMA 3.1 这样的模型同样是从头训练，没有依赖 OpenAI 或 Anthropic 的蒸馏。

Now Community Note did point out that TML itself did say in their notes that while Inkling had been pre trained from scratch, it did use a small bootstrap on synthetic data from open models, including k 2.5, as well as pointing out that other models like LAMA 3.1 were also trained from scratch without OpenAI and Anthropic Distillation.

但 Jack 想强调的关键是：在我们正走向的这个世界里，企业开始越来越关注开源权重模型，而这款模型是美国产的，且主要不是从封闭模型蒸馏来的——这一点很快会变得非常重要。

But the significant point that Jack is getting at is that in the world that we're heading to, with the unspoken piece of this being that enterprises are starting to pay more attention to open weights models, the fact that this is an American model that is not primarily distilled from one of the closed models could become more significant very, very soon.

与此同时，开源专家 Nathan Lambert 指出，这显然是 TML 的 Tinker 平台故事中的一部分。

Meanwhile, open source expert Nathan Lambert pointed out that clearly this is one part of the story alongside TML's Tinker platform.

Nathan 写道，一旦 Tinker 火起来，Inkling 的出现几乎是必然的。

Nathan wrote, Inkling was somewhat inevitable once Tinker took off.

TML 整个技术栈中整合了微调服务，这是目前最出色的开源模型商业案例之一。

Integration of post training services across TML's entire stack, and it's one of the best open model business stories to date.

Jeffrey Emanuel 也探讨了这一点，他写道：这是一个相当聪明且有差异化的商业策略，这很合理，因为要和顶级实验室正面硬刚实在太难了。

Jeffrey Emanuel explored this dimension of it as well, writing: This is a pretty smart and differentiated business strategy, which makes sense because it's hard to compete head to head with the biggest labs at their own game.

所以你瞄准的是他们的软肋——大公司对泄露核心优势越来越严重的焦虑。

So you focus on their weakness, which is the emerging competitive paranoia among big companies about leaking alpha.

要做到这一点，企业需要使用开源权重模型，这样他们才能在自己的基础设施上运行，而不会泄露任何数据。

To do that, you need open weight models so that the companies can run it on their own infrastructure without leaking anything.

但开源权重模型的难题在于如何盈利，因为推理服务很容易变成价格战，变得同质化，而且资本投入巨大。

But the problem with open weight is obviously how to monetize it because inference becomes a race to the bottom and commoditized and capital heavy.

TML的解决方案似乎是通过微调来盈利——根据企业的具体需求，利用企业自身的内部数据定制模型，同时确保这些数据带来的学习成果和优势只属于该企业独有。

TML's solution seems to be to monetize the fine tuning process of customizing their model for the particular problems a company wants to solve using that company's own internal data in a way that keeps the learning and benefits of that data exclusive to that one company.

很聪明。

Smart.

你完全可以相信，这正是TML推广这款产品时的核心卖点。

And you better believe that this is the thrust of TML's marketing for this.

他们发布文章的最后一部分叫‘定制Inkling’，开头就写道：许多现实世界的问题，即便是最好的通用模型也难以很好解决，而差距正是通过利用组织专业知识进行微调来填补的。

The last section of their launch post is called Customizing Inkling, and it starts, many real world problems aren't solved well by even the best generalist models, with the gap being closed by fine tuning that utilizes an organization's specialized knowledge.

我们的Tinker客户的经验也印证了这一点。

The experience of our Tinker customers points in the same direction.

我们在大规模强化学习后的训练结果表明，Inkling能够快速从微调中学习。

Our post training and results of reinforcement learning at scale suggest that Inkling is capable of rapidly learning from fine tuning.

一旦你把它看作是专为Tinker设计的基础模型，很多事情就说得通了。

And once you view it as the model that was designed to be the base model for Tinker, it starts to make a lot more sense.

在这篇博客的另一部分，他们提到，Inkling的设计目标是通用性。

In another part of that blog post, they wrote, Inkling is designed to be broad.

我们训练它覆盖代理行为、推理、编程、指令遵循、事实准确性、视觉和音频任务，而不是只专注某一个领域。

We train it across agentic, reasoning, coding, instruction following, factuality, vision, and audio task rather than narrowly optimizing for one domain.

这种广泛的能力对定制化和实际应用至关重要。

That breadth matters for customization and real world use.

不同用户需要的是能适应各种工作流程的模型，而不仅仅是能在基准测试中表现优异。

Different users need models that can adapt to very different workflows, not just excel on the benchmarks.

如果你一直关注微软，就会发现这正是他们正在推动的方向。

Now if you're watching Microsoft closely, this is clearly the direction they're pushing as well.

萨提亚·纳德拉最近在X上的博文都在强调，企业必须掌握自己的学习基础设施，而微软新推出的MAI模型系列中，重点产品之一就是他们的Microsoft Frontier Tuning。

Satya Nadella's recent blog posts on X have been all about how companies need to own their own learning infrastructure, and one of the big product pushes aligning with the new MAI models that Microsoft has released is their Microsoft Frontier Tuning product.

但即便如此，使用MAI的Microsoft Frontier Tuning仍然要求企业把数据交给微软来处理。

And yet Microsoft Frontier Tuning using MAI still requires enterprises to trust Microsoft with their data.

现在微软凭借过去几十年建立的信任，占据了有利位置，对许多公司来说，信任他们比信任OpenAI或Anthropic要自然得多，但这和Inkling这样的开源权重模型相比，仍然有明显不同。

Now Microsoft is in a good position to cash in on the trust that they've built over the past decades, making it less of a leap and a jump to trust them as opposed to an OpenAI or an Anthropic for many companies, but that still does look very different to an open weights version of that such as the one provided by Inkling.

不同公司对这些权衡会有不同的考量。

Now different companies are gonna have different senses of the trade offs.

我敢说，肯定会有公司尝试这种后训练或微调的方式，但他们其实并不太在意模型是开源还是闭源。

There will be, I guarantee, companies who do experiment with that sort of post training or fine tuning approach, but who don't ultimately care that much about it being open versus closed.

换句话说，他们更看重成本优势，而不是数据安全，或者即便存在数据安全问题，也愿意把微软当作合作伙伴。

In other words, they want the cost benefits more than the data security benefits or are comfortable with Microsoft as the partner regardless of the data security issues.

但另一些公司可能会更青睐那种更独立、更少受大公司控制的模型，比如Inkling这样的开源权重模型所代表的方向。

But others, you have to think, might jump at a model that has even more independence and sovereignty from a big player the way that an open weights model like Inkling might represent.

当法律团队和风控部门介入后，这款模型作为一款未经过压缩的美国本土模型，其独特性就更加突出了。

And when you start to pile the lawyers and the risk group on top of things, the fact that it's an American model that is not distilled becomes pretty unique relative to the other options out there.

丹尼尔·卡普兰认为，这可能标志着一个全新服务类别的开端。

Daniel Kaplan thinks that this could be the beginning of an entire new subdivision of a services category.

他写道：‘欢迎来到前向部署的微调师时代。’

He writes, enter the forward deployed fine tuner.

Thinking Machines 的两个 GA 产品，正好完美契合美国市场推出新基础模型的需求。

Thinking Machines two GA products equal the perfect go to market for a new foundation model in The US.

想要一个定制的本地部署模型吗？

Want a custom on prem model?

有自己的 AI 研发团队吗？

Got an AI research team?

用 Inkling 加上 Tinker 就行。

Inkling plus Tinker.

全程自助服务，当然你也可以付我们一笔合理的费用，获得额外的专业支持、优先体验新模型等等。

Full self serve, and you're welcome to pay us a moderate amount of money for additional expertise, early access to new models, etcetera.

想要定制模型，但没有 AI 研发团队？

Want your own custom model but no AI research team?

觉得 Tinker 有点难懂？

Tinker seem confusing?

没关系。

No problem.

我们会派一个高级微调专家过来，费用不菲。

We'll send over a forward deployed fine tuner for a lot of money.

适合大型企业、想保护核心优势的AI初创公司、寻找低成本算力的AI应用公司，以及从巨头实验室出走的新人创办的初创团队。

Scales to big companies and AI startups who want to protect their alpha, AI app layer companies in search of cheaper tokens, and new startups founded by runaways from the behemoth labs.

我觉得这里面逻辑非常清晰，当微软和这家热门实验室都在讲同样的故事时，企业自然会留意。

I think there is very clearly a logic here, and when both Microsoft and this buzzy lab are telling the same story, enterprises will pay attention.

但与此同时，这项技术背后的基础理念并非没有争议。

At the same time, the fundamental technology underneath this is not without controversy.

事实上，在某些方面，它甚至和人们普遍认为的通用模型与微调模型之间的关系背道而驰。

In fact, in some ways, it runs counter to what people's broad perceptions have been about how generalist models compare to fine tuned versions.

这指的是所谓的‘痛苦教训’——一种过于简化的观点，认为人类的学习方式未必适合机器学习，真正有效的策略其实是给足数据和算力，让模型自己摸索，而不是过度依赖人类的特殊经验。

Referring to the bitter lesson, which is a theory that oversimplistically says that human ways of learning aren't necessarily going to be the best ways of machine learning, and that in general, the best strategies for machine learning are going to be giving it as much data and compute as possible and letting it do its thing rather than overly relying on unique human expertise.

西蒙·史密斯写道：所以，Thinking Machines 其实是在挑战‘痛苦教训’这个观点？

Simon Smith writes: So thinking machines is basically a bet against the bitter lesson?

感觉确实如此，但我个人并不完全认同。

That's how it feels and I'm personally not convinced.

根据我微调模型的经验，这比大多数人想的要麻烦得多。

From my experience fine tuning models, it's way more effort than people think.

这种投入是持续不断的——要应对新出现的数据边界情况和模型更新，模型可能会丧失原有能力，或者突然冒出意想不到的问题，最终一个带点上下文的大通用模型出现，直接碾压你的辛苦成果，而具备这种能力的模型成本又会断崖式下跌。

That effort is ongoing to address new data edge cases and model updates, models can lose capabilities or have unexpected issues introduced, and ultimately a big general model with a bit of context comes along and beats your hard work, and then the cost of models with that capability drops off a cliff.

当人们谈论令牌、微调和运行自有模型时，往往只算微调本身的每令牌成本，却忽略了持续收集和整理数据、开发维护数据与训练流水线、部署管理服务器基础设施等全部开销。

When people talk about tokens and fine tuning and running their own models, they often only factor the cost per token of their fine tune, not the fully loaded cost of continuously collecting and curating data, developing and maintaining data and training pipelines, deploying and administering server infrastructure, and so forth.

不过，时间会给出答案。

Anyway, time will tell.

我依然相信那个‘苦涩的教训’。

My money's still on the bitter lesson.

我觉得这是个非常关键的观点，现在特别需要明白的一点是：意识到数据主权和令牌成本确实存在问题是没错的，但这并不意味着市场急着推出的解决方案——比如微软的Frontier Tuning或者Tinker——最终就一定是正确的路径。

I think it's a super important point, and one of the really important things to keep in mind right now is that the recognition that there is a problem and a challenge with both data sovereignty and the cost of tokens does not mean that the immediate answers that the market is jumping in to provide, I e, Microsoft Frontier Tuning or Tinker, is ultimately going to be the right answer.

短短几句话里，西蒙·史密斯就点出了所有这些潜在挑战，它们可能最终证明，这种微调方式未必能真正解决这些问题。

In just a few sentences, Simon Smith pointed out all of those different challenges that may end up showing that this sort of fine tuning isn't necessarily the right way to solve those problems.

但反过来讲，如果你现在是一家TML公司，你至少可以安心一点，因为你的方案采用开放权重，正好同时回应了令牌成本和数据主权这两大难题。

The flip side, though, is that if you are a TML right now, you have to feel good that your solution, because it is open weights, is addressing both sides of that equation, both the token cost side as well as the data sovereignty side.

换句话说，像微软这样的闭源模型为基础的微调方案，可能处于中间地带，即便它服务的是那些真正关心数据主权和成本的细分市场。

In other words, it might be that a fine tuning solution based on a closed model like Microsoft is two in between, and that even if it's a niche market that requires both the data sovereignty and the cost side for the people who actually care about it.

我们还不确定，但很明显，这将是一个正在兴起的趋势，值得密切关注。

We don't know, but it's very clear that this is going to be an emerging trend and one to keep an eye on.

前白宫人工智能顾问、16Z投资人的希拉姆·克里什南写道：开源模型和工具现在正迎来属于自己的时刻。

Former White House AI advisor and a 16 z investor, Shriram Krishnan, wrote, it is clear open source models and harnesses are having a moment.

这其中有几个关键因素在起作用。

There's a few factors at work.

第一，现在很明显，你完全可以追上接近顶尖的性能，而且还能清楚地追溯训练过程。

One, it is now obvious that you can catch up to near state of the art performance and do so with a clear training lineage.

看看TML今天推出的Inkling就知道了。

See TML's Inkling launch today.

第二，美国和海外已经有好几支资金充足、人才出众的团队正在开发开源权重模型。

Two, there are several well funded, talented teams building open weight models now in The US and abroad.

再加上Grok、Cursor、MuSpark等一大批接近顶尖水平的模型纷纷涌现，我们可以预见，在编程和智能体应用领域，未来将形成一个多元化的模型生态。

Along with the explosion of other near state of the art models like Grok and Cursor, MuSpark, it's clear that we are going to have a diverse ecosystem of models at least on coding and agentic use.

第三，越来越多的组织希望掌控自己数据的使用方式，愿意为此放弃一些前沿模型的访问权限。

Three, organizations are increasingly looking for control over how their data is used and are willing to trade off some access to frontier level tokens for this control.

企业和国家越来越担心，前沿实验室未来可能会成为自己的竞争对手，因此不希望自己的数据帮对方打造未来的对手。

Organizations and countries are increasingly nervous about the frontier labs potentially competing with them down the road and don't want their data to enable a future competitor.

第四，开源就像一个调节旋钮。

Four, open source is a slider.

你可以自己搭建开源框架，使用自己的评估标准和业务场景，自由选择最适合的模型。

You could bring your own open harness, your evals, your business context, and are free to pick and choose your model of choice.

第五，企业现在已不再纠结于如何让员工用上模型，而是开始对模型费用飙升却看不到明确收益感到不安。

Five, companies have now actively shifted from how do we get our people to use tokens to being uncomfortable with their token costs ballooning without a clear line to revenue.

第六，从地缘政治角度看，各国会把开源模型视为一种方式，把前沿级别的能力引入那些原本无法接触的受控环境。

Sixth, geopolitically, countries will be weighing open weight models as a way to get frontier level tokens inside controlled environments that may not be otherwise possible.

所有这些都为我们带来了更多选择，我认为这才是关键所在。

All of this leads to more choice for all of us, and I think this is the key point.

几个月前，你要是觉得企业都在纠结该选OpenAI、Anthropic，还是两个都选，那也不算想错，那时他们的选择范围就只有这些。

A few months ago, one would be forgiven for thinking that the way things were headed, pretty much every enterprise was just asking the question of whether it should sign up with OpenAI or Anthropic or both, and that was the extent of their decision.

现在我们讨论的不仅仅是模型的选择更多了，还包括模型运行的框架、复杂的架构——可能涉及多个模型，以及像微调这样的全新方法。

Now we are talking about so much more choice, not only in terms of models, but in terms of the harnesses in which model lives and complex model architectures which could involve multiple models and entirely new approaches to models such as these fine tuning approaches.

感觉现在正处在一个充满新尝试和新探索的转折点上。

It feels very much like a moment where we're going to see a lot of new shots on goal and new explorations emerge.

但就像往常一样，当大量新事物涌现时，并不是每一个都能站得住脚。

Now as is always the case, when a whole bunch of new flowers bloom, not all of them will work.

也许在未来六到十二个月内，行业巨头们会看清哪些方向真正被采纳，然后把最好的部分收拢过来，最终企业又回到只在少数几种预设方案中做选择的局面。

It may be that the juggernauts in the space figure out where there's adoption and uptake over the next six to twelve months, take in the best parts, and we're back to most enterprises simply choosing between a very small handful of predescribed solutions.

但很难想象，我们即将进入的这个实验阶段，不会对这些解决方案的形态产生深远且确定的影响；也很难想象，在这个蓬勃发展的高竞争环境中，小型团队和更灵活的企业没有大量机会可以抓住。

But it's very hard to imagine that, a, this experimental period we're moving into won't have a significant and deterministic impact on the shape of those solutions, and b, that there isn't a ton of room for alpha, especially among small teams and more nimble enterprises that can take advantage of this flourishing and highly competitive period.

这并不意味着所有企业都得立刻去试用 Tinker、微软的前沿微调，或者像赶时髦一样马上切换到 Open Router，或者跑去折腾 GLM 5.2。

That doesn't mean that you at enterprises all need to rush out and sign up for Tinker or Microsoft Frontier tuning or anything like that in the same way that you don't need to immediately switch all your activity to Open Router or go start hacking on GLM 5.2.

但现在已经很清楚了：如果你负责企业的采购决策，却对这些新方案连试一试、关注一下都不做，那你真的会错过很多机会。

But at this point, it's pretty clear that if you are in charge of any sort of enterprise buying decisions and you're not at least playing around with and experimenting with or closely paying attention to all these different solutions, you're missing quite a bit of opportunity.

我会持续关注这些动态的发展，不过今天就先说到这里，这就是本期《AI每日简报》的全部内容。

I will continue to keep track of these things as they develop, but for now, that is gonna do it for today's AI Daily Brief.

感谢你收听或观看，我们下次再会，再见。

Appreciate you listening or watching as always, and until next time, peace.