AI公司仍未兑现其最大承诺

在最近一次播客节目中，一位知名投资人表示，他从Anthropic内部的多个消息源听说，Dario Amodei和公司其他领导层认为，在未来的某个时间点，他们可能是唯一剩下的公司。

In a recent podcast appearance, a prominent investor said that he had heard from multiple sources inside Anthropic that Dario Amodei and other leaders in that company felt that at some point in the future, they might be the only company left.

到时候就只剩下他们、政府，还有我们其他人了。

It would just be them, governments, and the rest of us.

这些评论在播客上引发了关于Anthropic及其在AI领域角色的激烈讨论，也让大家开始思考他们的信念对整个行业意味着什么。

Now, these comments on that podcast kicked off quite a firestorm of discourse about Anthropic and their role in AI and what their beliefs actually meant for the industry.

这还催生了一个极为罕见的现象——Anthropic的CEO Dario Amodei本人竟然在社交媒体上露面了。

It also generated that rarest of phenomenon, an appearance on social media from Anthropic CEO Dario Amodei himself.

在他的回应帖子中，Dario谈到了他对监管俘获的真实看法，他认为AI在人们心中信任问题的真正根源是什么，以及他认为长期来看，什么才能真正解决这些信任问题。

In his response post, Dario discusses his real views on regulatory capture, what he thinks the real root of AI's trust problems with people are, and what he thinks could actually address those trust problems in the long run.

那么，人们觉得这些内容有启发吗？

So did people find it enlightening?

有说服力吗？

Convincing?

真的有人因此改变看法吗？

Did anyone's opinions actually change?

那这场对话又说明了AI领域的讨论以及Anthropic在其中处于什么位置呢？

And what does the whole conversation say about the AI discourse and Anthropic's place in it?

《AI每日简报》是一档每日更新的播客和视频节目，内容涵盖AI领域最重要的新闻和讨论。

The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

好了，朋友们。

Alright, friends.

在我们深入之前，先快速说几个通知。

Quick announcements before we dive in.

首先，感谢今天的赞助商：KPMG、Blitzy、Robots and Pencils 和 Hyperagent。

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

想要收听无广告版本的节目，可以访问 patreon.com/aideallybrief，或者在Apple Podcasts上订阅。

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

如果想了解如何赞助本节目，请发送邮件至 sponsors@aideallybrief.ai。

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

这个周一我们可有的聊了，系好安全带吧！

We got ourselves quite a Monday here, so strap in!

首先是一个新模型，而且肯定会引发很多讨论——智谱发布了GLM 5.3。

First up comes a new model, and one that is sure to kick off a lot of debate ZAI has dropped GLM 5.3.

你可能还记得，智谱在六月发布GLM 5.2时，就引发了我们现在仍在经历的这波担忧，就是中国的开源权重模型正在缩小差距。

Now, you might remember that when ZAI released GLM 5.2 in June, it helped kick off this new wave of concern that we're still in right now that Chinese open weight models were closing the gap.

其中一部分原因是时机。

Now part of that was timing.

发布的时候，Fable 5还被政府锁着，OpenAI也因为同样的原因推迟了5.6，但紧接着下个月Kimi K3的发布，更是加剧了这种中国模型紧追西方的感觉。

The release came during the period where Fable five was locked behind government doors and OpenAI was delaying 5.6 for the same reason, but the feeling of Chinese models nipping on Western heels compounded with the release of Kimi K3 the following month.

不过，和之前每次一样，即使承认这些模型确实擅长某些方面，大家还是觉得它们最终在关键方面仍然落后于前沿。

Still, as has happened every other time, even acknowledging what these models are really good for, there has been a sense that they still are, ultimately, behind the frontier in pretty meaningful ways.

那么，GLM 5.3的情况又如何呢？

So where does that leave us with GLM 5.3?

5.3是基于和5.2相同的基础模型构建的，但智谱声称他们仅仅通过扩展强化学习就取得了重大进展。

Well, 5.3 is built on the same base model as 5.2 Still, ZAI claims that they've made some big advances purely by scaling reinforcement learning.

在编程方面，GLM 5.3在Terminal Bench 3.0上得分28.3%，这比Fable 5和GPT 5.6 SOL的前沿水平低了大约五分，但比Kimi K3高出11分。

On coding, GLM 5.3 scores 28.3% on Terminal Bench 3.0 (that puts it around five points behind the Frontier with Fable five and GPT 5.6 SOL), but 11 points ahead of Kimi K3.

在DeepSweet上的结果没那么出彩，得分66.9，比Kimi K3低半分点，比Fable低3分，比5.6 SOL低6分。

The results on DeepSweet were less impressive, scoring 66.9, which puts the model half a point behind Kimi K3, three points behind Fable and six points behind 5.6 SOL.

在代理使用方面，Five3在Automation Bench和GDP Val上拿到了最优成果，稍微赢了美国前沿一点。

For agentic use, Five3 scored state of the art results on Automation Bench and GDP Val, just slightly inching out The US frontier.

总的来说，对于跑代理和编码这类高端应用，5.3还是跟绝对前沿有点距离，但它从一款中型模型里挤出了很大的性能提升，有些情况下可能已经超过了Kimi K3。

Overall, it looks like that for high level use cases like running agents and coding, 5.3 remains a bit behind the absolute state of the art, but has squeezed a lot of performance out of a mid sized model and may in some cases have overtaken Kymi K3.

ZAI说他们这次强化学习跑下来，重点之一就是提升网络安全表现。

Now ZAI said improved cyber performance was one of the key focuses for their reinforcement learning run.

他们提到，在Hugging Face遭到攻击的时候，防御方被迫用GLM 5.2，因为前沿模型的护栏让它们没法用了。

They noted that during the Hugging Face attack cyber defenders had been forced to use GLM 5.2 because the guardrails on frontier models had rendered them useless.

ZAI在微信公众号里写道：如果强大的攻击能力会扩散，防御能力就不能只掌握在少数闭源模型公司手里。

In a WeChat post ZAI said: If the powerful attack ability is spreading, the defensive ability cannot be limited to a few closed source model companies.

正因如此，他们在Cybersecurity Benchmark CyberGym上的成绩就显得更有看头了，比上一代跳升了7分，还真超过了Fable 5。

That makes their performance gains on Cybersecurity Benchmark CyberGym a bit more interesting jumping seven points from their predecessor to actually overtake Fable five.

先强调清楚，因为现在确实已经有人对着开源发布的巫术级网络模型吓坏了，但那不是这些基准在说的，即便你只信字面数字，也并不是那样。

Now just to be super clear because of course we're already seeing some people freak out about a Mythos level cyber model released as open source, that is not what these benchmarks are saying even if you take the benchmarks at face value.

在发现漏洞方面达到Mythos级别，和之后利用漏洞或自主发起网络攻击达到Mythos或Fable级别，这两者是有区别的。

There is a difference in being mythos or level at finding vulnerabilities, and being mythos or fable level at then doing something about it and or autonomously executing cyber attacks.

ZAI也采取分阶段发布的策略，先让模型在可信合作伙伴那里测试，之后再公布完整的权重。

ZAI is also taking a phased approach to the release, testing the model with trusted partners before publishing the full weights.

在录制的时候，我们还没拿到Artificial Analysis的完整基准测试数据，所以既没有一系列的独立测试，也没有一个对token调整后成本的好衡量方式。

Now at the time of recording we don't yet have the full benchmark run from artificial analysis, so we don't have either a full barrage of independent tests nor do we have a great gauge for token adjusted cost.

按token计算，GLM 5.3的成本不到Fable或5.6 SOL的十分之一，只有Kimi K3的五分之一。

On a per token basis, GLM 5.3 is less than a tenth of the cost of Fable or 5.6 SOL and one fifth the cost of Kimi K3.

在有限的测试中，一些用户发现，对于同样的任务，GLM 5.3的成本大约是Kimi K3或Grok 4.6的三分之二。

In limited testing, some users are finding that GLM 5.3 costs around two thirds of Kimi K3 or Grok 4.6 for the same task.

初步对真实世界性能的测试反馈有点好坏参半。

First testing for real world performance leads to feedback that is a little mixed.

ZAI声称在开源代码库中检测到超过一千个关键和高风险漏洞。

ZAI claims to have detected over a thousand critical and high risk vulnerabilities in open source repos.

他们还认为在Cursor中找到了一个潜在的严重漏洞，并向团队私下报告了，但其他一些初步测试的结果就不太明确了。

They also believe they found a potentially serious vulnerability in Cursor which they privately reported to the team, but some of the other first tests are a little less clear.

开发者Aditya测试了它在游戏开发任务上的表现，发现GLM 5.3有点吃力，明显落后于Kimi K3。

Developer Aditya tested it for game development tasks, and found GLM 5.3 struggling a bit and clearly behind Kimi K3.

另一个大问题是模型慢得让人受不了，几乎没法用。

And another big complaint was that the model was painfully slow to the point of being unusable.

这可能只是发布当天需求过大的问题，而不是根本性的缺陷，但还有待观察。

Now that could just be issues with excessive demand on release day rather than a fundamental problem, but that remains to be seen.

总的来说，ZAI在GLM 5.2的基础上带来了扎实的提升，也再次证明了很多性能提升其实只需要扩大强化学习的规模。

Overall, it looks like ZAI delivered a solid improvement over GLM 5.2 and showed once again that a lot of performance can come simply from scaling reinforcement learning.

Open Model的研究员Nathan Lambert在评测中指出，我们不应该再对中国实验室的出色表现感到惊讶了。

Open Model researcher Nathan Lambert used his review to argue that we should stop being surprised by strong performance out of Chinese labs.

简单来说，他写道：ZAI似乎在训练后的优化上下了功夫，而Kimi更像是预训练阶段的杰作。

To risk a broad oversimplification he wrote: ZAI seems to have strengthened post training when compared to Kimi which is more of a pre training masterpiece.

这次发布之后，很多人都在讨论，中国是怎么能跟得这么紧的？

Following this release there have been a lot of discussions wondering how China can keep up so well?

这么小的模型，怎么能跟美国领先的公开模型打成平手？

How can such a small model be matching the leading public American models?

这些结果是真的吗？

Are these results real?

他接着说，最简单的解释就是智谱AI非常擅长他们做的事情。

The simplest explanation, he continued, is that ZAI is very good at what they do.

兰伯特的观点是，我们不能再把中国实验室的出色表现仅仅归因于蒸馏或刷榜，他认为这样做会低估他们的真实能力。

Lambert's point is that we need to stop writing off the Chinese labs as only performing well because of distillation or benchmark maxing, arguing that when we do so, we underestimate what they're actually capable of.

重要的是，随着这类发布，华尔街分析师和那些不一定每天听AI每日简报的人，终于开始更新他们的认知，意识到中国实验室在大多数情况下并不是像他们之前以为的那样，以极低的价格出售前沿智能。

Importantly alongside this type of release, Wall Street analysts and people who aren't necessarily listening to the AI Daily Brief every day are finally beginning to update their priors and recognize that Chinese labs aren't in most cases selling Frontier Intelligence for pennies on the dollar the way they had previously assumed.

成本优势确实还在，但价格差距已经大幅缩小了。

The savings are definitely still there, but the pricing gap has contracted substantially.

在这个阶段，像Grok 46、MuSpark 1.2或GPT 5.6 Luna这样更便宜的美国模型，已经和中国最好的模型在成本上具有竞争力了。

At this stage, cheaper US models like Grok 46, MuSpark 1.2, or GPT 5.6 Luna are cost competitive with the best models out of China.

华尔街也意识到，即使是中国的模型，也仍然需要基础设施。

Wall Street is also recognizing that even Chinese models still need infrastructure.

这也许不会让你感到太意外，但值得记住的是，在2025年初的DeepSeek时刻之后，一个主要的担忧是，更便宜的模型可能会让大规模GPU投资变得不再必要。

Now that probably won't come as a huge surprise to you, but it is worth remembering that following the deep seek moment in early twenty twenty five, one of the big concerns was that cheaper models would invalidate large scale GPU investment.

不过，在最近的一份报告中，晨星的分析师马利克·汗写道：如果一家企业把整个AI技术栈都建立在开放权重模型上，它仍然需要云基础设施来运行这些工作负载、存储数据、管理安全以及获取资源等等。

In a recent note, however, Morningstar analyst Malik Khan wrote: If an enterprise were to consolidate its entire AI stack on open weight models, it would still need cloud infrastructure to run those workloads, store data, manage security, and access to resources, etcetera.

这些都是云基础设施公司的利好因素。

All tailwinds to cloud infrastructure companies.

随着人们对5.3的进一步测试，我肯定会持续关注它。

I will certainly keep an eye on 5.3 as people test it further.

但另一条关于模型的消息，或者说至少是模型的暗示和传闻，是Anthropic表示他们不会发布下一个模型，而是将其保留供内部使用。

But another bit of model news, or at least model innuendo and rumor, is that Anthropic has suggested that they won't release their next model instead keeping it for internal use.

《Anthropic风险报告》第二版讨论了目前正在内部使用的未发布模型。

The second edition of the Anthropic Risk Report included a discussion of unreleased models that are currently being used internally.

截至7月中旬，Anthropic有三个重要的未发布模型。

As of mid July, Anthropic had three significant unreleased models.

前两个是Opus 5和一个被称为Model 1的模型，其能力大致与Mythos 5相当。

The first two were Opus five and a model referred to as Model one, with capabilities broadly in line with Mythos five.

Anthropic还披露了一个名为Model 2的模型，他们描述其能力略高于Mythos 5。

Anthropic also disclosed a model known as Model two, which they described as somewhat more capable than Mythos five.

但是，Anthropic 继续说：我们粗略的定性感觉是，这个模型在内部使用的许多相关任务上比 Mythos 5 有显著提升，但并没有表现出像从 Claude Opus 4.6 到 Mythos Preview 那种能力跃升的幅度。

However, Anthropic continued: Our rough qualitative sense is that this model is a noticeable improvement on Mythos five for many tasks relevant to internal use, but does not display a capability jump to the degree observed from Claude Opus four-six to Mythos Preview.

Anthropic 表示目前没有公开发布这个模型的计划，也没有对它进行他们通常的那套评估测试。

Anthropic said they have no plans for releasing this model publicly and have not run it through their usual suite of evaluations.

ChrisGPT 指出：Anthropic 的新 Model 2 不是像报告里说的那样只是略优于 Mythos 5，它在 Anthropic 内部的 co benchmark v2 上拿到了 62.8%，而 Mythos 5 是 50.3%，Mythos Preview 是 54.8%。

ChrisGPT pointed out: Anthropic's new Model two doesn't just slightly outperform Mythos five as they stated in the report, it scores 62.8% on Anthropic's internal co benchmark v2 versus 50.3 for Mythos five and fifty four point eight for Mythos Preview.

另外要记住，报告日期是7月15日，所以这些报告反映的其实是几个月前 Anthropic 的内部能力概况。

Also keep in mind the report's date is July 15, so these reports are already a months old snapshot of Anthropic's internal capabilities.

他们目前的内部水平很可能远超这些数据所展示的。

Their current internal frontier is most likely further ahead.

我觉得这是我们必须记住的重要背景，尤其是当我们讨论中国模型跟最前沿技术差距有多大时，比如在美国政府介入最前沿模型发布这个新范式下。

I think that is just important context to keep in mind as we discuss things like how far behind Chinese models are relative to the state of the art, especially in this new emerging paradigm of government involvement in US state of the art model releases.

我们现在实际用的模型跟实验室内部有的模型，它们之间的差距比以往拉得更大了一些。

The gap between what we're using and what the labs have internally is somewhat wider than it's been in the past.

与此同时，OpenAI 的人显然也不想在竞争中掉队，大家开始注意到他们的员工在偷偷暗示 Astra 的存在，不管是最终官方叫什么，这款模型很快就会落到我们手里了。

Meanwhile, not wanting to be left out of the game, people are noticing that OpenAI staffers have started vague posting about Astra, suggesting that that model, whatever it ends up being called officially, will be in our hands soon.

最后，今天咱们继续聊大厂，不过话题转到他们的IPO计划上。

Lastly, today, we're staying on the big labs, but moving over to their IPO plans.

Anthropic已经开始接触投资者和投资银行了，也有消息传出，透露了他们跟投资人分享的一些细节，以及他们刻意避开的几个点。

Anthropic has begun meeting with investors and investment banks, and reports have come in relaying some details that Anthropic has shared as well as a few points they stayed away from.

Anthropic告诉投资人，收入同比增长了14倍，第二季度达到了115亿美元。

Anthropic told investors that revenue was up 14 x compared to a year ago, reaching 11,500,000,000 in q two.

按这个速度推算，全年能达到460亿美元，但并没有泄露出更新后的年化运行率数据给媒体。

That annualizes out to 46,000,000,000 for the year, but no updated run rate was leaked to the press.

另一个关键数字是两万亿美元，这是Anthropic的投资者表示他们期望在IPO中实现的估值。

The other big number is $2,000,000,000,000 which is the valuation that Anthropic investors said that they expect from the IPO.

据报道，Anthropic方面自己还没有讨论过估值问题，所以这完全是投资者告诉《金融时报》的内容。

Anthropic themselves reportedly hasn't discussed valuation, so this is purely what investors have told the Financial Times.

不过，如果他们真的实现了这个目标，Anthropic首次上市的估值将会比今年早些时候Spacecraft高达1700亿美元的估值还要高。

Still, if they achieve that, it would make Anthropic's debut valuation even larger than the $1,700,000,000 attached to SpaceX earlier this year.

这也意味着其估值将从今年五月融资时的水平翻了一番还多。

It would also mean more than a doubling in valuation from where they were when they raised funds in May.

投资者表示，Anthropic今年的营收目标是在1000亿到1200亿美元之间，大约是现在水平的两倍半。

Investors said that they Anthropic to reach between $100 and $120,000,000,000 in revenue by the end of this year, so around 2.5 times where they are currently.

Anthropic自己的预测则稍微保守一点。

Anthropic themselves are a little more modest with their forecasts.

虽然有些投资者在无限外推这种增长速度，但Anthropic自己的预测可能更保守一些。据路透社报道，他们预计到2028年营收能达到1900亿到2000亿美元之间。

And while some investors are extrapolating out this type of growth rate indefinitely, Anthropic themselves are perhaps a little more modest with their forecasts, with Reuters reporting that they see revenue hitting between 190 and 200,000,000,000 by 2028.

现在财经媒体开始质疑这个估值，比如《财富》指出，公开股票通常按盈利倍数而不是营收倍数交易。他们写道，假设一个平均盈利倍数，Anthropic要达到2万亿美元的估值，就得有大约590亿到790亿美元的年利润才能跟上。

Now the financial press is starting to poke holes in the valuation with, for example, Fortune noting that public stock typically trades on an earnings multiple, not a revenue multiple and assuming an average earnings multiple they write, a $2,000,000,000,000 Anthropic would need to post annual profits in the neighborhood of 59 to $79,000,000,000 to keep pace.

不过话说回来，Anthropic也不是按普通公司的标准来估值的。

Then again, Anthropic is not being valued as an average company.

我们确实没有先例可以参考。

And we are really working without precedent.

我觉得在IPO之前，这样的争论会越来越多，因为没有明确的先例，我们能做的就是各说各的看法。

I think you can expect to see a lot more of this type of debate in the lead up to this IPO because in the absence of clear precedent, all we have is opinion to be argued.

说到看法和Anthropic到底有多独特：据报道，关于Anthropic认为自己有多么独一无二的讨论，其实就是我主节目要讲的话题的起因。

Now, of opinions and just how unique Anthropic is: reported discussions of how singular Anthropic believes they are was actually what started the topic that I will address in the main episode.

那么现在，我们先结束头条新闻，进入正题。

So for now, let's end the headlines and move on over into main.

大家好。

Hello, everyone.

人工智能带来的一个重大变化是，我们的思路已经从如何给网页排名，转变为如何成为人工智能足够信任并用来回答问题的信息来源。

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

在毕马威，他们正亲身经历这一变化。

At KPMG, they're seeing this firsthand.

现在，AI生成的结果直接给出答案，往往连一次点击都不需要。

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

因此，他们越来越关注“生成式引擎优化”，也就是GEO，通过优化内容结构，让AI系统能够检索、理解并引用这些内容，将其视为可信的权威来源。

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

这不仅仅是SEO的演变，更是关乎可见度的必然要求。

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

事实上，毕马威提出的GEO要求非常明确。

And indeed, the GEO mandate from KPMG is simple.

如果AI在影响决策，你的专业能力就需要体现在答案里。

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

详情请访问kpmg.com/us/geo。

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

再说一遍，是kpmg.com/us/geo。

Again, is kpmg.com/us/geo.

市面上每个AI编程工具做的第一件事都一样。

Every AI coding tool on the market does the same thing first.

它直接开始写代码。

It starts writing code.

Blitzy却反其道而行。

Blitzy does the opposite.

在写一行代码之前，Blitzy会花好几天来逆向分析你的整个代码库。

Before writing a single line, Blitzy spends days reverse engineering your entire codebase.

数千个智能体处理数百万行代码，梳理出每一个依赖关系、每一条没有记录的约束，以及过去十年里每一个架构决策。

Thousands of agents ingest millions of lines, mapping every dependency, every undocumented constraint, every architectural decision made over the last decade.

最终形成的是一张动态知识图谱，能够像一位在行业里深耕三十年的资深工程师那样理解你的软件。

The result is a dynamic knowledge graph that understands your software the way a principal engineer would after thirty years in the building.

其他工具靠grep搜索和markdown文件来猜测上下文，而Blitzy从不猜测。

Other tools guess at context with grep searches and markdown files, Blitzy never guesses.

它先构建真正的理解，然后自主交付超过80%的完整软件史诗。

It builds true understanding first, then delivers over 80% of entire software epics autonomously.

经过验证、端到端测试、达到生产标准的拉取请求。

Validated, end to end tested, production grade pull requests.

这就是为什么五百强企业的工程团队会把最关键的代码库放心交给Blitzy。

That's why Fortune 500 engineering teams trust Blitzy with the code bases that matter most.

去blitzy.com亲自看看吧。

See for yourself at blitzy.com.

那就是blitzy.com。

That's blitzy.com.

我每天都在这个节目中探讨AI潜力与现实之间的能力差距。

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

大多数公司还在摸索怎么开始。

Most companies are still figuring out how to start.

Robots and Pencils已经在推进和扩展了。

Robots and Pencils is already launching and scaling.

Agendaq和生成式AI在几周内就能部署到大型企业生产环境中。

Agendaq and Generative AI in production at large enterprises in weeks.

AWS高级合作伙伴，一年内业绩翻了一倍多。

AWS advanced tier pattern partner more than doubled in a year.

而且他们正在招人。

And they're hiring.

有50个空缺职位。

50 open roles.

如果你觉得这个时代与众不同，想参与其中而不是旁观，那这值得一看。

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

在Robots and Pencils，最好的想法胜出，而且团队质量有意识地保持极高水准。

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

这种地方，你会回头觉得这是你做过的最好的决定。

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

去 robotsandpencils.com/careers 看看吧。

Take a look at robotsandpencils.com/careers.

本期《AI每日简报》由Hyperagent赞助播出，它让你可以运行一支代理团队，大家共同管理。

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美元的推理额度。

New users get $1,000 in inference.

别再依赖本地代理和聊天工作流，等着你的笔记本电脑被提示了。

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

Hyperagent在云端部署始终在线的代理，在你团队已经使用的工具中完成实际工作。

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

营销部门的代理把竞争对手的动态变成落地页。

Marketing's agent turns competitor moves into landing pages.

销售部门的代理丰富潜在客户信息，起草邮件，并更新CRM。

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

运营代理会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上下文，并遵循你关于范围和审批的规定。

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

是时候让你拥有像队友一样的代理了。

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

去Hyperagent雇佣你的代理吧，它由Airtable团队打造。

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

在hyperagent点com斜杠aideallybrief领取你的1000美元推理额度。

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

回到AI每日简报。

Back to the AI Daily Brief.

我在这个节目里想做的就是，围绕正在发生的重大事件，带给你广泛而有代表性的各种观点。

One of the things that I try to do on this show is bring you a wide ranging and representative set of commentary around whatever the big events happening are.

我觉得这是更深层次理解新闻的最佳方式之一，因为新闻如何影响我们的世界，不仅取决于新闻本身，还取决于它如何与人们对它的看法相互作用。

I think this is one of the best ways to understand the news on a deeper level, given that when it comes to how news impacts our world, it's not just the news itself, but how that interacts with how people perceive it.

然而，今天这件事更进了一步，因为讨论本身成了新闻。Anthropic 的 Dario Amodei 罕见地在 x.com 上公开露面，参与了一些关于 Anthropic 的讨论。

Today, however, this goes a step farther, where the discourse is actually the news itself, as Anthropic's Dario Amodei has waded into the muck of x.com for a rare public engagement around some concerns about Anthropic.

这些回应和互动，远不止是业内小圈子的内幕纠缠。

The responses and interaction around this are more than just insider baseball psychodrama.

考虑到这场对话的分量，以及 Anthropic 和 OpenAI 在全球经济议题中扮演的超级重要角色，这里的发言本身真的能算作新闻。

Given the stakes of this particular conversation and the outsized role specifically Anthropic and OpenAI are playing in global economic discourse, the comments themselves in this case actually count as news.

起因是投资人 Gavin Baker，Tradies Capital 的 CIO，本周末上了 All In 播客。

It started when investor Gavin Baker, CIO of a Tradies Capital appeared on the All In podcast this weekend.

在那次采访里，他说，Anthropic 内部非常确信。

In that interview, he said, internally, Anthropic is very confident.

我信任的好几个人都告诉我，Dario 说过，有些时候，Anthropic 可能是世界上唯一的私人公司。

I have been told by multiple people I trust that Dario has said that Anthropic might be the only private company in the world at some point.

你们想想看。

Think about that.

在这个愿景里，也就是 Anthropic 的极致愿景里，只有 Anthropic 和各国政府，没有别的了。

In this vision, an Anthropic maximalist vision, there's Anthropic, and then there are governments, and that's it.

所以，他们非常有信心。

So there's a lot of confidence.

我相信他们手里肯定还有比Fable更先进的模型。

I'm sure they have more advanced checkpoints than fable up their sleeve.

他们执行得确实非常出色。

They've executed really, really well.

我可能会打赌他们不会是世界上唯一的私营公司，我觉得要等很久他们才能成功发射火箭。

I would probably take the under on them being the only private company in the world I think it's going to be a long time before they land a rocket.

当前AI主管David Sacks说，我可能会把这解读为一个负面信号，因为这太傲慢了。

When former AI czar David Sacks said, I might interpret that as a negative signal because it's so hubristic.

这有点进入SPF的领域了。

This is getting into SPF land a little bit.

Gavin接着说，我当然会劝Dario以后不要再对任何人说这种话。

Gavin followed up, I would certainly discourage Dario from saying that ever again to anyone.

而且这些评论可能就留在All In上了，要不是Anthropic的Sholto Douglas决定插一脚的话。

And those comments might have just stayed on All In, had Anthropic s Sholto Douglas not decided to wade in.

他在X上转发了那条帖子，并说完全错误。

He re shared that post on X and said Completely false.

我喜欢加文的观点，但不管他是从谁那里听来的，那个人都在撒谎，好让这个说法符合某些人拼命想让你相信的叙事。

I like Gavin's takes, but whoever he heard this from is lying so that it fits the narrative some people so desperately want you to believe.

同一批人会先试图让你相信Anthropic没有护城河，然后下一句话又说它可能强大到成为唯一剩下的公司。

The same people will try to convince you Anthropic has no moat, and a sentence later that it might become so powerful it could be the only company left.

事实上，我们最担心的就是经济权力的集中。

In fact, one of the things we are most worried about is economic concentration of power.

政府绝不应该让任何一家公司拥有那么大的影响力。

There is no world where the government should let any company have that much influence.

我们需要竞争和资本主义。

We need competition and capitalism.

AI市场现在绝对是全球竞争最激烈的市场。

The AI market is literally the most competitive market in the world right now.

地球上每一家最大的公司都在全力以赴地推出更智能、更便宜的模型。

Every single one of the largest companies on Earth is singularly focused on getting you smarter, cheaper models.

如果一切顺利，我们就能把一切成本降到能源成本那么低。

If it all works out, we'll succeed in reducing the cost of everything to the cost of energy.

这很棒，但会威胁到很多人的旧有优势。

This is awesome, but it threatens a lot of people's old motes.

他们感到害怕。

They're frightened.

加文接着说道：肖托，谢谢你澄清事实。

Gavin picked it up from there: Shoto, thank you for setting the record straight.

更大的问题是，硅谷有好几个非常严肃的人听到过类似的说法，并且相信这是真的。

Larger issue is that multiple very serious people in Silicon Valley have heard some variation of this and believe it to be true.

之所以这么多人相信，是因为这与达里奥公开的表态以及你分享的那篇文章里他阐述的内容一致。

And the reason it is believable to so many is that it is consistent with Dario's public messaging and what he outlined in the essay you shared.

这项技术可能在多个方面对人类构成危险，可能导致经济权力的极端集中，因此需要谨慎监管。

This technology might be dangerous for humans in multiple ways, could lead to extreme concentration of economic power, and therefore needs to be regulated thoughtfully.

我同意这些潜在风险，而且我相信达里奥提出这些论点都是出于善意。

I agree with the potential risks, and I believe Dario makes all of these arguments in good faith.

就像我们在播客里讨论的那样，如果你认为AI可能有危险，那应对这种潜在风险有两种方式：要么通过监管把AI集中在少数公司和政客手里，要么广泛地分发出去。

As discussed on the pod, if one agrees that AI might be dangerous, there are two ways to address this potential risk: either concentrate it in the hands of a chosen few companies and politicians via regulation, or distribute it widely.

这本质上就是看你是觉得AI太危险了不能集中，还是太危险了不能分发。

Essentially boils down to whether one believes AI is too dangerous to concentrate or too dangerous to distribute.

两边都有合理的论点，但我非常赞同扎克伯格的说法，就是那种认为AI太危险，唯一安全的路径就是极度集中权力的想法，本身就有问题。

There are reasonable arguments on both sides, but I profoundly agree with Zuckerberg's statement that The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic.

从历史上看，指望一个绝对权力在足够开明的情况下会仁慈地为人类服务，并没有带来更安全、更好的结果。

Historically, hoping that an absolute power will benevolently provide for humanity if sufficiently enlightened has not led to safer positive outcomes.

接着，加文写道：正如达里奥在那篇文章里说的：有人可能会反对说，我们可以像对待人类一样，通过多个AI系统之间的权力平衡来约束AI。

Continuing, Gavin writes: as Dario says in the aforementioned essay: Some may object that we can simply keep AIs in check with the balance of power between many AI systems as we do with humans.

加文继续他自己的思考，他说：我相信这是最好的前进道路。

Continuing his own thoughts, Gavin says: I believe this is the best path forward.

我希望AI越多越好，这样就有更大的机会，其中有一个能认同我自己的价值观。

I want as many AIs as possible to maximize the odds that one shares my own particular values.

而且正如达里奥指出的，从来没有哪个人类能够统治全世界。

And as Dario notes, no human has ever been able to take over the world.

到这会儿，我觉得可以说达里奥已经输掉这场争论了。

At this point, I think safe to say that Dario has lost the argument.

他的发声并没有带来他想要的那种监管路径。

His messaging has failed to result in his preferred regulatory path.

最近那个事件——一个未发布的先进开源AI模型入侵了Hugging Face——而唯一的解决办法居然是一个开源模型，这基本上断送了短期内严格监管的任何可能。

The fact that the only solution to the recent incident, where an unreleased advanced open AI model hacked Hugging Face was an open source model likely ended any chance of strict near term regulation.

基本上除了Anthropic，所有大公司都签了黄仁勋那封信。

Essentially every major company other than Anthropic has signed Jensen's letter.

不过，达里奥的发言倒是大大助长了在美国禁止数据中心的努力。

However, Dario's messaging has been massively helpful to efforts to ban data centers here in America.

我猜接下来会看到反数据中心团体拿达里奥关于AI对人类有多危险的警告片段来打广告。

I suspect we will see anti data center advocacy groups running ads using clips of Dario warning about how dangerous AI could be for humans.

他出于善意支持监管，结果反而增加了AI无法造福美国人乃至全人类的可能性。

His good faith efforts in favor of regulation are now increasing the odds that AI will not be beneficial for Americans and humans everywhere.

我觉得AI有很大机会帮我们治愈大多数疾病，让我们寿命更长，能在一个丰饶的、星际迷航式的未来里享受这些漫长的生命。

I believe there is a reasonable chance AI might help us cure most forms of disease, such that we have extended lifespans and can enjoy these long lives in an abundant Star Trek like future.

那是我想要的未来，而我觉得达里奥现在正在降低这个未来的可能性。

That is the future I want, and I think Dario is decreasing the odds of that future at this point.

他即将成为世界上最重要的上市公司之一的CEO。

He is about to be the CEO of one of the most important public companies in the world.

鉴于支持监管的努力至少目前失败了，我怀着敬意认为，他应该努力为自己行业进行更积极的倡导。

And given that the pro regulatory effort has failed, at least for now, I respectfully think he should make an effort to be a more positive advocate for his own industry.

如果我错了，我们确实需要监管这项技术，那么他将来会因曾对其他可能性持开放态度而成为更有效的倡导者。

And if I am wrong and we do need to regulate this technology, he will be a more effective advocate for this in the future having been open minded to the alternative.

为了明确起见，我认为Anthropic拥有深厚的竞争优势，是一家了不起的公司。

For the sake of clarity, I think Anthropic has deep competitive advantages and is an amazing company.

讽刺的是，几个月前我看到Anthropic面临的主要风险，就是达里奥自身言论和行为引发的国有化。

Ironically, the main risk I saw to Anthropic a few months ago was nationalization as a result of Dario's own rhetoric and behavior.

现在，Sholto和Gavin接着那话题继续讲。

Now, Sholto and Gavin go on from there.

但这就是达里奥开始参与对话的地方了。

But this is the point where Dario comes into the conversation.

现在稍微说一下背景，Dario 很少用社交媒体。

Now just for some context, Dario is very rarely on social media.

他已经很明确地说过，他特别不喜欢这个。

He has been very clear that he hates it.

事实上，他把它归咎于好多人误解他的原因。

In fact, he has blamed it for a lot of people misinterpreting him.

他现在在 X 上可以说一个关注的人都没有。

He currently follows zero people on X.

他最后一次发帖是6月10号，那是分享他那篇关于“AI指数”的策略文章。

And the last time he posted was June 10 to share his essay policy on the AI exponential.

再往前是4月7号，他宣布了“Glasswing”项目。

Before that, it was April 7 to announce Project Glasswing.

再再往前，是1月26号，他宣传自己另一篇文章《技术的青春期》。

And before that, January 26 to promote another essay of his, The Adolescence of Technology.

回到他对回应，他是这么说：加文，感谢你带来了这么深入的一次交流。

Coming back to his response to Gavin, he says: Thanks Gavin for an especially thoughtful exchange.

我平时不怎么花时间在社交媒体上，但这次想参与一下，因为这场对话确实触及了核心问题。

I don't usually spend much time on social media, but I wanted to engage here because it really brings out the heart of an important conversation.

首先，关于监管。

First, on regulation.

我认为，要么靠监管把权力集中在少数公司和政客手里，要么就都摊开分给大家，这种二选一其实是个伪命题。

I think that either concentrated in the hands of a chosen few companies and politicians via regulation or distribute it widely is a false choice.

我知道硅谷有一种简化的说法，就是“监管等于监管捕获”，然后“监管捕获”等于“权力集中”，但我一直觉得这把世界看得太简单了。

I know that there is a sort of Silicon Valley shorthand where regulation equals regulatory capture equals concentration of power, but I've always found this to be an overly simplified picture of the world.

圈外人很多人觉得，监管是限制企业权力、让老百姓受益的。

Many people outside this bubble think of regulation as something that constrains corporate power and benefits ordinary people.

我也不完全认同这个看法。

I don't necessarily agree with that perspective either.

其实，我觉得事情很复杂，关键要看监管到底包含了什么。

Rather, I think it's complicated and really depends on what the regulation consists of.

但特别要说的是，那些用“监管=监管捕获=权力集中”这个框框看问题的人，往往低估了客观公正的制度流程在分散权力上的作用。

But in particular, I think that those in the Regulation = Regulatory Capture = Concentration of Power frame often underrate the decentralizing power of objective and fair institutional processes.

打个粗浅的比方，正式的法院系统有时候会让人觉得有点刻板、高高在上，但在保护弱势群体权益方面，它比那种替代方案——也就是暴民正义——要做得好得多。

A crude analogy is that the formal court system can sometimes feel stuffy and elitist, but it does a much better job of defending the rights of vulnerable individuals than the alternative, mob justice.

在最好的情况下，制度可以把权力赋予思想，而不是个人，从而分散权力。

At their best, institutions can vest power in ideas rather than people and thereby decentralize that power.

这就是为什么Anthropic在提出政策建议时总是非常谨慎。

That is why Anthropic has always made its policy proposals very carefully.

我们非常努力地提出那些对前沿AI公司不利或拖慢它们发展的建议，同时让较小的竞争对手受益。

We try very hard to make proposals that disadvantage or slow down frontier AI companies, while advantaging smaller competitors.

我们支持的加州SB 53法案，甚至那个备受批评的SB 1047法案（我们当时态度暧昧），都完全豁免了那些收入或模型训练成本低于一定门槛的公司。

California's SB 53, which we supported, and even the much maligned SB ten forty seven, which we were ambivalent on, completely exempt any company below a certain amount of revenue or model training costs from being covered at all.

最近，我们在CAISI和白宫倡导的测试流程，对前沿模型的测试比非前沿模型更严格，这在一定程度上有利于挑战者。

More recently, the testing process we've advocated for at CAISI and the White House involves more rigorous testing of frontier models than off frontier models something that differentially advantages challengers.

同样，前沿模型节奏信函所设想的，是调节最顶尖模型的开发速度，而不限制那些正在追赶的公司。

Similarly, the pacing the frontier letter envisions modulating the pace of the very best models, while not constraining those who are catching up.

这损害了前沿实验室的商业利益，却帮助了挑战者，包括开放权重模型。

This hurts the business interests of the frontier labs and helps challengers, including open weights.

总的来说，我认为AI在结构上是一种倾向于集中权力的技术，原因与监管无关，更多与扩展定律的极端影响有关。

Overall, my view is that AI is structurally a technology that tends to concentrate power, for reasons that have nothing to do with regulation more to do with the extreme implications of the scaling laws.

开放权重在这方面确实有些帮助，但远远不够，因为它只是把权力集中到那些拥有最多算力和芯片的人手里，基本上还是前沿实验室，可能再加上硬件供应商。

Open weights do help some with this, but are nowhere near a sufficient solution because they simply shift the concentration somewhat to those with the most compute and chips which are roughly the frontier labs plus maybe hardware providers.

相比之下，我认为正确的规则可以同时做到：a) 应对AI的网络、生物和校准风险；b) 在制度上约束前沿AI公司的权力；c) 为开放权重模型留出空间，同时解决它们带来的特定风险。

Contrast, I think the right rules of the road can simultaneously: a) Address AI cyberbioalignment risks b) Institutionally constrain the power of the frontier AI companies and c) Leave room for open weights models while also addressing the specific risks that they bring.

顺便说一句，我并不认为过去几个月的事件没有导致我偏好的监管路径。

By the way, I do not think that the events of the last months have failed to result in my preferred regulatory path.

据报道，特朗普政府正在采取的方法（对前沿模型进行部署前测试，以及当开放权重模型接近前沿时也进行测试）是我非常支持的。

The approach that the Trump administration is reported to be taking (predeployment testing for frontier models, and also testing of open weights models when they get closer to the frontier is one that I am very supportive of.

当然，我得看到细节才能确定。

Though of course I have to see the details to be sure.

我也支持Demis Hassabis关于类似FINRA的实体的想法。

I am also supportive of Demis Hassabis' ideas around a FINRA like entity.

这与六个月前的情况形成对比，当时大多数行业仍在推动先发制人地阻止所有州级监管，而且联邦层面也没有明确的做法。

This contrasts with six months ago when most of the industry was still pushing for preemption of all state regulation, and no apparent federal approach either.

他在下一篇帖子中继续写道：第二，关于AI的舆论导向，我并不认为我的观点过于消极。

In his next post, he continues: Second, on the messaging around AI, I do not agree that my messaging has been disproportionately negative.

事实上，我对风险和收益的讨论是平衡的。

In fact, it has been equally balanced between risks and benefits.

我分别写过一篇关于这两者的长文，即使在访谈中讨论风险时，我也会频繁提及它带来的巨大好处，并提出可能的解决方案。

I've written one major essay about each, and even in interviews where I discuss the risks, I make sure to frequently mention the incredible benefits, as well as proposing possible solutions to the risks.

社交媒体上流传的我访谈中的短视频片段往往显得过于消极，因为那样更容易吸引点击量。

Short clips from my interviews that end up on social media tend to be disproportionately negative as that gets clicks.

其实我写《爱与恩典的机器》这本书，是因为我觉得AI行业没有描绘出一个足够鼓舞人心的图景，来展示这项技术如何能从根本上让世界变得更好。

In fact, I wrote Machines of Love and Grace because I didn't feel the AI industry was painting an inspiring enough picture of how the technology could radically transform the world for the better.

文章的大部分篇幅都在反驳人们对AI在健康和生物领域潜力的怀疑，并说明为什么我认为在五到十年内治愈大多数人类疾病是切实可行的。如果你读过我最近的文章《关于AI指数增长的策略》，里面讨论了具体的提案，旨在简化FDA的流程，确保那些由AI加速研发的药物不会因监管流程而受阻。

The bulk of the essay is devoted to refuting skepticism of AI's potential in health and biology, and showing why I think it will actually be possible to cure most human diseases in around five to ten years And if you read my most recent essay Policy on the AI Exponential, I discuss concrete proposals for how to streamline the FDA process to make sure the deluge of AI accelerated drugs isn't slowed down by the regulatory process.

我感到这里非常紧迫。

I feel the urgency here.

几年前，我的父亲就因丙型肝炎去世了，而直接抗病毒药物就是在那之后才发展起来的。这种药能治愈95%的患者，很可能也能救回他的命。

I lost my father to hepatitis C only a few years before the development of direct acting antivirals, which cures ninety five percent of patients and probably would have cured him.

我确实同意公众对AI有负面看法，而且这是一个很大的问题。

I do agree that the public has a negative view of AI, and that this is a big problem.

但我不认为这主要是由我或其他AI领导对AI风险的警告引起的。

But I don't think it is primarily caused by me or any other AI leader warning about AI's risks.

我认为这根本上是信任危机。

I think it is fundamentally a crisis of trust.

我觉得普通人都不信任公司、政府或科技行业，总觉得我们在琢磨一些新办法来坑他们。

I think that ordinary people don't trust companies, governments, or the tech industry, and always suspect that we are cooking up some new ways to screw them over.

这些问题的根源可以追溯到几十年前，AI只是最新的表现而已。

The causes of this go back decades, and AI is just the latest iteration of it.

我不认为靠一个光鲜亮丽、刻意正面的营销活动就能赢回信任。

I don't think that a glitzy marketing campaign with a positive spin is the way to win back that trust.

到了现在，说AI能治愈癌症更多是陈词滥调，算不上鼓舞人心。

At this point, saying that AI will cure cancer is more of a cliche than it is inspiring.

而且大多数人觉得那就是在欺骗。

And most people think it is deceptive.

真正管用的，是实实在在把癌症治好。

The thing that will work is actually curing cancer.

我觉得对AI公司，包括Anthropic，最中肯的批评就是我们还没兑现那些造福世界的重大承诺。

I think by far the most accurate criticism of AI companies including Anthropic is that we haven't yet delivered on our big promises to benefit the world.

这完全是我们自己的问题，我觉得你们应该批评这个，而不是老盯着宣传和营销那些事儿。

That is totally on us, and I think it's the criticism you should be making, instead of all the stuff about messaging and marketing.

不过，我们确实在尽力解决这个问题。

We are, however, doing our best to fix this.

Anthropic正在生物和医学领域快速加大投入，我们希望未来几年能出惊人的成果，未来几个月也能看到一些初步的亮点。

Anthropic is ramping up efforts very quickly in biology and medicine, and we hope to have incredible results in the coming years and some early glimmers in the coming months.

等我们真正做出实事，全世界都会听到，而且会尽可能大声地宣传，这点我向你保证。

When we've actually accomplished something real, the whole world will hear about it as loudly as possible you have my word on that.

但在那之前，我不想空口许诺。同时，我觉得有责任坦诚地谈谈AI带来的真实风险，以及如何应对这些风险。

But until then, I don't want to make empty promises, and in the meantime, I feel compelled to speak honestly about the very real risks of AI and how to address them.

坦诚本身就是正确的做法，而且从公众信任和公信力的角度看，这也不比那种回避或淡化风险的做法差，因为人们本能地知道那些风险是真实存在的。

Honesty is the right thing on the merits, and in terms of public credibility and trust, it is no worse than an approach that ignores or distracts from risks which people instinctively understand are real.

好的。

Alright.

所以，虽然很长，但我不想在这里做总结，因为这些帖子本身就是故事。

So, very long, but I didn't want to summarize here given that the posts are the actual story.

所以，第一批反应基本上就是“请多来点这样的内容”之类的。

So one of the first types of reactions was basically more of this, please?

加文本人回应说，我觉得你在这里以如此建设性、深思熟虑的方式参与讨论，这很棒。

Gavin himself responded, of all, I think it is great that you are engaging here in such a constructive, thoughtful way.

公开对话是与公众建立信任的好方式，这些是困难且重要的问题，应该在公共广场上进行辩论。

Open dialogue is a great way to build trust with the public, and these are difficult, weighty issues that I deserve to be debated in the proverbial town square.

前OpenAI员工威尔·德普说，我真的认为达里奥应该更多地公开写作。

Former OpenAI staffer Will DePooh says, I really think Dario should be writing a lot more in public.

这做得很好。

This was well done.

The Information的创始人杰西卡·莱森写道，仅凭两条推文，达里奥·阿莫迪就做到了他整个一年都在努力做到的事情。

Jessica Lessen, the founder of the Information, writes, with just two tweets, Dario Amodei did what he has struggled to do all year.

他改变了自己在大家心中的形象，也给科技圈的舆论人士提供了一个更容易引用的观点来表达他的立场。

He changed the narrative about himself and gave the chattering classes of tech a far more accessible message to quote regarding his views.

而他那句关于信任的话，我觉得正说到了点子上。

And his trust comment is, I believe, exactly the point.

但也有人不觉得他真的改变了什么舆论风向。

Now others weren't so sure that he actually changed the narrative, though.

奥斯汀·奥尔雷德转发了达里奥第二条帖子的截图，并且特意圈出了“我不同意我的言论一直被过度负面化”这句话，然后配了个评论说，哈哈，真是笑死我了。

Capturing the feeling of many, Austin Allred retweeted a screenshot of Dario's second post with the highlight on the section I do not agree that my messaging has been disproportionately negative, adding the comment LOL, l m f a o.

的一位工程师也引用了那句同样的话，然后吐槽说，你是在开玩笑吧？

Terminally online engineer highlighted that same line and said Are you kidding me?

AI评论家哈特写道：“哎呀达里奥，你总是怪别人误解你，或者嫌别人夸大了你的话。”

AI commentator Hater wrote, Oh Dario, as usual, you blame everyone else for misunderstanding you or hyping your own words.

这简直就像你说了句不靠谱的话，回头又说别人理解错了，让别人觉得自己是错觉。

It's almost like you say something reckless, then later try to gaslight people into thinking they misunderstood you.

前AI研究员苏珊·张写道：“那一大段文字，加上一群捧场的人，全都绕开最关键的点，没有正面否认原来的指控。达里奥说Anthropic可能某天会成为世界上唯一一家私人公司。”

Former AI researcher Susan Zhang writes, A wall of text and an army of sycophants to thoughtfully not deny the original accusation of Dario says Anthropic might be the only private company in the world at some point.

哦对了，顺便说一句，Dario完全没在散布AI毁灭论或者人类大规模失去权力的恐慌。

Oh, and by the way, Dario is totally not spreading p doom or massive human disempowerment fears.

不不不。

No, no, no.

你们全都是自己幻觉出来的，完全是在恶意曲解。

You all hallucinated that and misread it all in bad faith.

亲爱的读者和用户们，真替你们感到害臊，这全是你们自己能力不足。

Shame on you, dear readers and users, for all of your own skill issues.

公关专家Lulu Chang Meservey倒是更冷静地分析了Dario帖子里的这段内容。

PR expert Lulu Chang Meservey actually discussed this section of Dario's post in more clinical terms.

她写道：第二部分的论点很有趣，展现了我怎么评估自己的言论。

She writes: The argument in Part two is an interesting insight into how Dario assesses his own messaging.

当被指责言论过于负面时，他反驳说其实已经算平衡了，因为他写过一篇关于好处两篇关于风险的文章。他用分析的方法去衡量积极性，但其他人只是凭感觉。

When accused of negative messaging he rebuts that it's roughly balanced in large part because he's written one essay about benefits and one about risks he's taking an analytical approach to measuring positivity, but many others are simply going off vibes.

所以他和批评者将来还会继续像这样各说各的。

So this won't be the last time he and his critics talk past each other.

在我看来，这不只是双方各说各话。

For my part, I don't think that it's just two parties talking past each other.

我觉得这件事上，Dario就是错了。

I think this is an area where Dario is just wrong.

或者，他是故意不愿去理解我们所处世界的现实。

Or perhaps being willfully unwilling to understand the reality of the world we live in.

在我看来，你可以一方面声称自己了解社交媒体永远只截取最负面的部分，然后却去做一连串没完没了的采访，不断抛出一堆轻易就能成为头条的负面数据。

In my view, you can, on the one hand, claim to know and understand that social media is always going to clip the most negative parts, and then go and do an endless string of interviews with a seeming never ending torrent of easily sound biteable negative statistics.

在现实的公众舆论中，公众的看法不是靠你正数文章和负数文章里各有多少字来决定的。

In the real world of public discourse, public perception is not shaped by a word counter of how many words in your positive essay versus how many words in your negative essay.

还要这样辩解，只能说明你对实际媒体环境的理解极度欠缺，而作为一家如此重要公司的领导者，你恰好在这样的环境中生存。

And to argue such shows just a radical lack of understanding of the actual media environment in which you are going to have to operate as the leader of an extremely significant company.

如果你听过哪怕几集AI Daily Brief的节目，你会知道我完全认同一个观点：整个科技世界和各大实验室之间确实存在根本性的信任差距，这不仅仅是宣传信息的问题，也不是宣传信息能单独解决的，但你仍然需要明白宣传信息在其中扮演的角色。

Now, if you have listened to even a handful of AI Daily Brief episodes, where I certainly agree is in that I believe that there absolutely is a fundamental trust gap between the broader world and the labs that is not just a question of messaging and that messaging cannot fix alone, but you still do have to understand how messaging plays into this.

还有人注意到，Dario并没有反驳Gavin当初的那个说法，就是Dario曾说过Anthropic可能是世界上唯一一家私人公司。

Others picked up on the idea that Dario didn't dispute Gavin's original claim that he had said that Anthropic might be the only private company in the world at some point.

Zephyr Z9 提到了这一点，David Saxx 也是，不过我觉得这更多是策略上的问题。

Zephyr Z9 made that point, as did David Saxx, although I think that this is more tactical than anything else.

Lulu 又指出来：在资深同事已经做完事实核查、把辩论提升到原则层面之后，再直接发个人帖子是一种升级。

Lulu again points out: Going direct with a personal post was the escalation after a senior colleague already did the fact checking and elevated the debate to the level of principles.

所以 Dario 能直接谈他自己的信念，而不需要先争辩谁在什么时候说了什么细节。

So Dario was able to come in and focus on his beliefs rather than having to first debate the details of who said what when.

不过，对很多人来说，更大的批评是关于 Dario 对监管的讨论。

Still, the bigger critique for many was around Dario's discussion on regulation.

Saxe 认为，Dario 说硅谷认为所有监管都等于监管绑架，这种描述是稻草人论证。

Saxe called Dario's characterization of Silicon Valley thinking that all regulation equals regulatory capture a straw man.

他写道，当然，把所有监管都当作监管绑架是过于简单化的，但几乎没有人持有这种观点。

Of course, he writes, treating all regulation as capture would be overly simplified, but almost no one holds that view.

同时，replet 的 Anjan Masad 不同意 Dario 关于 AI 在结构上集中权力的论点。

Meanwhile, replet's Anjan Masad takes issue with Dario's argument that AI structurally centralizes power.

他说，因为 AI 目前需要大量算力就认为它在结构上集中权力，这种说法忽略了一百二十五年来算力价格性能超指数增长的事实。

He says, the argument that AI structurally centralizes power because it's currently compute hungry ignores one hundred and twenty five years of super exponential growth in compute price performance.

算法改进和硬件效率的持续提升意味着，没有理由认为通用人工智能级别的能力总是需要数据中心才能运行。

Improvements in algorithms and continued hardware efficiency gains mean there is no reason to assume AGI level capabilities will always require a data center to run.

扩展定律并不是物理定律，它们只是针对特定架构、目标、数据集等观察到的经验关系。

Scaling laws are not laws of physics they're simply empirical relationships observed for particular architectures, objectives, datasets, etc.

只要改变其中一个因素，就会得到不同的扩展曲线。

Change any one of those factors and you get a different scaling curve.

与此同时，对其他人来说，整个讨论还是有点太抽象了。

Meanwhile, for others, the whole thing remains a little too abstract.

Silicon Data的研究主管Steve Howe写道：

Write Silicon Data's head of research, Steve Howe?

归根结底，对于不是深度参与这个行业或对AGI深信不疑的人来说，整个讨论可能显得有点自娱自乐。

Ultimately, the whole discussion can feel self indulgent to anyone not already deeply inside the industry or AGI pilled.

硅谷里一小群人完全相信，具备超级智能、大规模取代人类和造成最大伤害能力的系统即将或正在加速到来。

A relatively small group in Silicon Valley is fully convinced of the accelerating or imminent arrival of systems capable of superintelligence, large scale displacement, and maximal harm.

但对大多数人来说，这看起来还是有点牵强。

To most others, this still looks like a large stretch.

实际上，这就是为什么有些人同意Dario的看法，要让AI真正解决信任问题，它必须是用行动来证明，而不是靠嘴上说说。

Which is, in fact, why some agree with Dario that for AI to actually solve its trust issues, it's gonna have to be a show, not a tell.

Peter Yang写道：我百分之百同意Dario的观点，用AI治疗疾病，并且加快AI医疗突破的监管审批，可能给人类带来的好处是其他所有事情加起来的总和的10倍。

Peter Yang writes, I 100% agree with Dario that using AI to cure diseases and speeding up the regulatory approval of AI breakthroughs in healthcare could bring 10x the benefit to humanity as everything else combined.

另外，OpenAI的Angel Broden写道：我个人觉得，认为任何AI实验室只要做点惊艳的事情，比如治愈癌症或者搞出重大医学突破，就能解决信任问题，这种想法太简单了。

Still OpenAI's Angel Broden writes, My personal opinion: It's an oversimplification to assume that any AI lab can solve the trust problem by doing something spectacular, like curing cancer or achieving some major medical breakthrough.

制药行业确实给人类健康带来了巨大改进，但它至今还是最不被信任的行业之一。

Pharma has delivered some of the greatest improvements in human health and is still one of the least trusted industries.

我那些因为道德原因反对AI的朋友，并不是因为他们觉得AI没有能力做好事。

My friends who are ethically against AI don't feel that way because they don't think it has the potential to do great things.

他们担心的是开发这项技术的公司，到底有没有真心为他们着想。

The concern is whether the company building the technology has their best interests at heart.

Dario说得对，光鲜的营销推广不可能也没法让老百姓信任AI公司，但人们也不会只因为他们有了突破就对他们产生好感。

Dario is correct that a glossy marketing campaign cannot and will not make people trust AI companies, but people also won't judge AI companies solely by their breakthroughs.

大家会看定价、获取门槛、游说行为、透明度、经济收益如何分配、哪些人能分享好处，以及最终谁真正掌握权力。

They'll judge them by pricing, access, lobbying, opacity, how the economic gains are distributed, who gets to participate in its benefits, and who ultimately holds the power.

达里奥认为，公众不信任人工智能公司，是因为社会在几十年间逐渐失去了对强大机构的信任。

Dario argues that the public distrusts AI companies because society has spent decades losing trust in powerful institutions.

但如果真是这样，解决办法就不能只是要求人们把更多的信任寄托在少数几个强大的机构身上，指望它们能妥善管理这项技术。

But if that's true, the answer can't simply be asking people to place even more trust in a small number of powerful institutions to responsibly steward this technology.

在一个理想的未来，人工智能应该帮助分散权力，而不仅仅是提供技术访问权限，还包括随之而来的自主权和经济效益。

In an ideal future, AI should help distribute power not just access to technology, but the agency and economic benefits that come with it.

这意味着要实现民主化、赋予个人自主权、广泛获取、价格亲民，并以更低的成本让更多人都能获得日益强大的智能服务。

That means democratization, individual agency, broad access, affordability, and making increasingly capable intelligence available to more people at lower cost.

现在，如果你听到这里觉得头都大了，心想：对啊，但到底有没有解决什么问题呢？

Now, if at this point your head is spinning around and you're asking, Yeah, but does anything actually get resolved here?

答案当然是没有。

The answer is, of course, no.

不过我觉得还是有一些重要的启示。

But I do think that there are some important takeaways.

首先，人们能够基于实际说过的话进行讨论，而不是仅仅依靠自己的猜测或听到的传闻，这会让讨论质量更高。

First, people getting to debate things that were actually said, rather than just suppositions that they're making or reports that they're getting, makes for better discourse.

第二，尽管他讨厌社交媒体，这些帖子本身就表明，有时候参与其中还是有点价值的。

Second, as much as he hates social media, these posts themselves show that there is sometimes value in participating in it.

而且，要做到这一点，不需要像Elon那样每天发50条帖子。

And one doesn't have to become Elon posting 50 times a day for that to be the case.

第三，具体到Dario，我觉得这种长度的交流方式可能比他自己意识到的更适合他。

Third, for Dario specifically, I would suggest that this length of communication might be a better fit for him than he realizes.

很明显，他因为人们不读他完整的帖子而感到沮丧。

It's very clear that he's frustrated that people don't read his full posts.

但那些帖子有13000字那么长啊，老兄。

But they're like 13,000 words long, man.

现在人们确实没有那种耐心了，所以他们当然只会去抓取统计数字。

People just don't have that sort of attention span anymore, and so of course they're just going to grab the statistics.

不幸的是，长访谈也是这样，哪怕所谓的“长”也就十到十五分钟。

Same, unfortunately, with long interviews, even if long is ten or fifteen minutes.

在X上写几百字的这种方式，其实比一些其他媒介更难被人断章取义。

This medium of a few 100 words on X is actually much harder for people to take out of context than some other mediums are.