← 返回任务列表

The Right Way to Worry About AI

301 段 · 1 位说话人 · 原片 28:25
M1
M10:00

今天的 AI Daily Brief,我们来聊聊,应该用什么方式担心 AI。

Today on the AI Daily Brief, the right way to worry about AI.

在那之前,先看头条:市场、模型,还有更多消息。

Before that, in the headlines, markets, models, and more.

AI Daily Brief 是一档每日更新的 podcast 和视频节目,关注 AI 领域最重要的新闻和讨论。

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

好,朋友们,在我们正式开始之前,先快速说几个通知。

All right, friends, quick announcements before we dive in.
M1
M10:31

顺便快速说一下,Apple Podcasts 这周好像出了一些问题。

Quick note there, by the way, Apple Podcasts seems to have been having some trouble this week.

我们这边并没有像有时候那样,在 Apple 上架无广告版本时出现特别的延迟,但有些人跟我说,过了好几天他们看到的还是带广告的版本。

We haven't been having any particular delays as we sometimes do with the ad-free version going up on Apple, but I've had some people days later still seeing the ad version.

我能给出的最好建议是,把你的 Apple app 完全关掉,然后重新打开。

The best that I can suggest is to completely close out of and restart your Apple app.

但不管怎样,给大家带来麻烦,我很抱歉。

But in any case, I apologize for the pain.
M1
M11:03

好了,这些都说完了,我们开始吧。

But with all that out of the way, let's dive in.

今天我们先从一些 OpenAI 的消息说起。

We kick off today with some OpenAI news.

嗯,先是一点猜测,然后是一些真正的新闻。

Well, a little bit of speculation and then some real news.

爆料者们开始暗示,下一个新模型 Astro 似乎马上就要发布了。当然,Astro 就是我们上周讨论过的那个做出了那些新颖数学证明的模型,有人甚至说他们的目标是下周发布。

The leakers are starting to suggest that the next new model, Astro, which was of course the one that did those novel math proofs that we discussed last week, seems to be imminently launching, with some saying that they're even targeting next week.

我们可以确定的是,即使 OpenAI 正在发布新模型,他们也非常重视成本方面,并且很努力地在成本上竞争。

What we know for sure is that even as they are releasing new models, OpenAI is also thinking very much and trying to compete very much on the cost front as well.

这家公司宣布,作为 GPT-5.6 模型家族服务全面调整的一部分,他们将为免费用户提供无限聊天。

The company announced that they're giving free users unlimited chats as part of a service overhaul for the GPT-5.6 model family.

免费用户层级现在将由 GPT-5.6 Luna 提供服务,取代 Instant 模型系列。

The free user tier will now be served with GPT-5.6 Luna, replacing the Instant model range.

免费用户现在也会有一个 think 按钮,让 Luna 进行更强的推理。

Free users will also now have a think button to allow for greater reasoning from Luna.

理论上,这缩小了免费用户的一部分体验差距,让他们可以使用和付费用户同一个模型,只不过是更小的版本。

Theoretically, this closes some of the experience gap for free users, allowing them to access the same model as paid users, albeit the smaller version.

此外,现在使用量也不再受限,所以免费用户可以想用多少就用多少 ChatGPT。

In addition, usage is now unlimited, so free users can use ChatGPT as much as they want.

对于付费订阅用户,GPT-5.6 Sol 现在将成为默认聊天模型。

For paid subscribers, GPT-5.6 Sol will now become the default chat model.

OpenAI 表示,相比 GPT-5.5 Instant,这应该会改善体验,因为 Sol 会犯更少的事实性错误,并且在没有帮助的时候避免给出多余细节。

OpenAI said that this should improve the experience over GPT-5.5 Instant, with Sol making fewer factual mistakes and avoiding extra detail when it doesn't help.

最后,Plus 和 Pro 订阅用户现在会在 thinking mode 里看到一个新的 effort 滑块,用来更直观地控制推理强度。

Finally, Plus and Pro subscribers will now have a new effort slider in thinking mode to provide more intuitive controls over reasoning effort.

有些人,比如 jumpers,就写道,这怎么可能赚钱呢?

While some like jumpers write, how is that even profitable?

Ken Chionex 说,把 OpenAI Luna 免费开放,显然不是出于慷慨。

Ken Chionex says the move to make OpenAI Luna free is obviously not out of generosity.

免费层级是一个战略性的分发渠道。

The free tier is a strategic distribution channel.

他们是想把人吸引进来,让他们升级到 Go 或 Plus,当然也可能通过广告赚钱。

They're trying to hook people and get them to upgrade to Go or Plus, or of course make money through ads.

说到我们关于更便宜模型的讨论,根据 The Information 的报道,Stripe 确实正在推进对 OpenRouter 的收购。

Now speaking of our discourse of cheaper models, according to The Information, Stripe is indeed moving forward with their OpenRouter acquisition.

这家新闻媒体报道称,Stripe 已经进入独家谈判,准备以接近此前报道的一十亿美元价格收购这家模型路由 startup。

The news outlet reports that Stripe has entered exclusive talks to buy out the model routing startup for close to the $1 billion that was previously reported.

早些时候的报道曾暗示有一场竞购战,而 Stripe 处于领先位置,但这条消息表明,OpenRouter 已经不再公开寻求买家,而是会和这一个特定合作方进入谈判阶段。

Earlier reports had suggested a bidding war with Stripe in the lead, but this suggests that OpenRouter has taken themselves off the market and will enter the negotiation phase with this one specific partner.

当然,这些交易仍然可能告吹,但独家谈判确实说明事情正在进入下一阶段。

Certainly, the deals could still fall apart, but exclusive talks do suggest that it's moving to the next level.

与此同时,随着全世界把越来越多的算力用在越来越多的 AI 上,我们到处都能看到供应链紧张带来的影响。

Meanwhile, everywhere we are seeing the impact of the supply chain crunch as the world uses more and more compute for more and more AI.

The Information 又报道称,Nvidia 正在考虑削减 Rubin 的规格。

The Information again reports that Nvidia is considering slashing the specs on Rubin.

目前,Nvidia 下一代旗舰 GPU Rubin Ultra 有三个不同版本正在测试中。

Currently, Nvidia has 3 different variants of their next generation of flagship GPU, the Rubin Ultra, under testing.

消息人士说,其中一些测试机型的内存比最初公布的要少。

And sources said that some of the test units include less memory than originally announced.

这些消息人士表示,Nvidia 正在考虑发布内存更低的版本,部分原因是担心在量产时无法确保拿到足够的高带宽内存。

Those sources said that Nvidia is considering releasing the lower-memory versions, partly due to concerns that they won't be able to secure enough high-bandwidth memory for the production run.

现在,这件事到底会带来多大影响,还有待观察。

Now, how big the implications of this are remains to be seen.

即使内存减少,这些芯片应该仍然能够为最新一代的超大型模型提供 inference 服务,比如 Claude Opus 4。

Even with reduced memory, the chips should still be able to serve inference for the latest generation of ultra-large models like Claude Opus 4.

不过,内存限制可能会给继续扩大模型规模的能力设下上限。

However, memory limits could put a cap on the ability to keep scaling model size.

到目前为止,Nvidia 一直否认在获取足够内存方面存在任何问题。

Now, at this point, Nvidia has so far denied any issue with sourcing enough memory.

七月中旬,硬件工程高级副总裁 Andrew Bell 说:“我们在内存问题上已经提前布局了,所以它短期内不会拖我们的后腿。”

In mid-July, Senior VP of Hardware Engineering Andrew Bell said, "We were in front of the memory problem, so it's not gonna hold us back anytime soon."

“当然,价格对全世界来说都是个问题,而且价格可能会是更大的挑战,但在供应方面,我们状态不错。”

The pricing, of course, is a problem for the whole world, and probably the pricing will be the bigger challenge, but for supply, we're in shape.

Nvidia 原本也要到明年晚些时候才会出货这些新芯片,所以仍然有时间解决供应问题。

Nvidia also wasn't scheduled to ship the new chips until late next year, so there's still time to resolve supply issues.

不过,我们现在可能已经到了这样一个阶段:一些基础性的硬件限制,可能开始拖慢模型进步的速度。

Still, we are at the point where we may be starting to see fundamental hardware limitations potentially start to slow down the pace of model improvement.

为了确保投资者别误会,Mike Sulka 写道,这不是 AI 需求疲软,而是食物链顶端在分配内存。

Making sure investors didn't get it twisted, Mike Sulka wrote, this isn't weak AI demand, this is memory rationing at the top of the food chain.

现在,有一个理由让人认为,内存短缺和其他组件短缺只会越来越严重,那就是我们对全球 AI 总需求的理解仍然处在非常早期的阶段。

Now, one of the reasons to think that memory shortages and other component shortages are gonna get nothing but worse is the fact that we are still at the very beginning of understanding the world's total demand for AI.

OpenAI 等一些公司押注的一点是,新一代消费级设备将有助于释放所有这些需求。

One bet that some companies like OpenAI are making is that a new generation of consumer devices will help unleash all of that demand.

Bloomberg 对 OpenAI 的首款设备有更多报道,把它描述成基本上是一款没有显示屏的智能音箱。

Bloomberg has more reporting about OpenAI's first device, describing it as essentially a smart speaker without a display.

它由电池供电,大小大约像一个冰球,形状像甜甜圈,设计目标是让用户可以单手在家里轻松拿着走。

It's battery-powered, roughly the size of a hockey puck, and shaped like a donut with the intention of making it easy to carry around the home in one hand.

它会采用高质量的拉丝金属表面处理,类似 iPhone 上的那种质感。当然,这也成了那起商业秘密诉讼的一部分,因为 OpenAI 正在和一家 Apple 供应商合作。

It will have a high-quality brushed metal finish, similar to the finish on iPhones, which of course formed part of the trade secrets lawsuit as OpenAI is working with an Apple supplier.

消息人士暗示,这款设备会有一些小型活动部件和灯光,用来提供视觉提示,显示它正在和用户互动。

The sources suggested that the device will have some small moving parts and lights intended to provide a visual indicator that it's interacting with the user.

它会配备摄像头、麦克风,以及其他用于感知周围环境的传感器。

It will include a camera, microphones, and other sensors designed to take in surroundings.

最新的重点消息是目标价格,OpenAI 计划把这款设备以三百到四百美元之间的价格推向市场。

The biggest new news is the target price, with OpenAI aiming to bring the device to market between $300 and $400.

作为对比,最贵的一些智能音箱,比如 Amazon 的 Echo Studio,零售价大约是二百二十美元。

By way of comparison, the most expensive smart speakers such as Amazon's Echo Studio are retailing for about $220.

与此同时,说到那起 Apple 诉讼,它很明显是为了拖延这款设备的发布,而这款设备预计会在明年年初某个时候推出。

Meanwhile, speaking of the Apple lawsuit, it's pretty clearly aimed at delaying the release of this device, which is expected sometime early next year.

不过,Bloomberg 的 Mark Gurman 至少认为,设计上的差异让 OpenAI 没什么风险。

But Bloomberg's Mark Gurman at least thinks the design differences put OpenAI in the clear.

他说,这里的重点是,它看起来、摸起来、用起来,都完全不像 Apple 产品,也不像这家公司目前正在计划的任何东西。

He said, the big takeaway here is that it looks, feels, and acts nothing like an Apple product or anything the company is currently planning.

我被告知,OpenAI 没有发现任何证据表明他们这款设备侵犯了商业秘密。

OpenAI hasn't found any evidence that they're violating trade secrets with this device, I'm told.

再看市场方面,周四,SoftBank 披露,他们以所持 OpenAI 股份作为抵押,借入了一百亿美元。

Over in markets, on Thursday, SoftBank disclosed borrowing $10 billion against their OpenAI stake.

这笔贷款由美国和日本的六家左右投资银行以及私人信贷公司联合承销。

A loan which was syndicated across half a dozen investment banks and private credit firms in the US and Japan.

这笔现金将用于支付 SoftBank 对 OpenAI 投资的最后一期款项,而 SoftBank 今年对 OpenAI 的投资总额为三百亿美元。

The cash will be used to pay for the final installment of SoftBank's investment in OpenAI, which totals $30 billion this year.

SoftBank 为这笔贷款已经奔走了好几个月,但一直很难找到愿意放贷的机构。

SoftBank has been chasing this loan for months and had struggled to find a willing lender.

即便按 5 月报道的更温和的 60 亿美元估值,几家大银行也不想冒着风险,拿这种流动性很差的私人股票,去给这么高的估值提供贷款。

Even at a more modest $6 billion reported in May, the major banks didn't want the risk of lending against the illiquid private stock at such a lofty valuation.

根据上个月的报道,由于没有一家银行愿意独自承担风险,一家贷款机构联盟已经聚到一起,准备把这笔贷款做成银团贷款。

According to reports from last month, a consortium of lenders had come together to syndicate the loan as no single bank was willing to carry the risk.

现在,这笔贷款的完整细节并没有披露,而且关键是,我们不知道 SoftBank 为这笔贷款提供了多少抵押品。

Now, the full details of the loan weren't disclosed, and crucially, we don't know how much collateral SoftBank put up to secure the loan.

我们知道的是,这笔贷款是按保证金贷款来结构化的,这意味着如果 OpenAI 的价值下跌,SoftBank 就需要补更多现金或者增加更多抵押品。

What we do know is that it's structured as a margin loan, meaning that SoftBank will need to add more cash or collateral if OpenAI's value drops.

还有传言说,利率是 7.88%,这对有抵押债务来说非常高。

It's also rumored that the interest rate was 7.88%, which is very high for collateralized debt.

The Wall Street Journal 强调说,银行拿私营公司股票来放贷是极不寻常的,因为一旦违约,这些股票既没法独立估值,也没法轻易变现。

The Wall Street Journal underscored that it's extremely unusual for banks to lend against private company stock because it can't be independently valued or easily liquidated in the event of a default.

所以,如果你觉得 SoftBank 其实已经拿出了一大块 OpenAI 股份来做质押,那也不算离谱。

So it wouldn't be crazy to think that SoftBank has already pledged a large chunk of their OpenAI stake.

现在,单纯为了说完整一点,因为大多数时候我们谈市场时,都是我在讲我觉得空头们有多离谱;在所有 AI 看空论点里,SoftBank 遇到流动性问题,说实话,算是比较有可能的一个。

Now, just for the sake of completeness, since most of the time when we talk about markets, it's me arguing how off base I think the bears are, of all the AI bear cases presented, SoftBank running into liquidity issues is frankly one of the more plausible.

除了这笔 100 亿美元贷款之外,他们还有一笔 400 亿美元的过桥贷款要在明年 3 月到期,而且他们还拿自己在芯片制造商 Arm 的持股又借了 200 亿美元。

In addition to this $10 billion loan, they also have a $40 billion bridge loan that comes due in March of next year, and they've also borrowed $20 billion against their holdings in chipmaker Arm.

SoftBank 股票目前的交易价格,比其披露的 3650 亿美元资产规模折价 40%,这反映出这家公司现在背负着相当沉重的债务负担。

SoftBank stock is currently trading at a 40% discount to its stated $365 billion in assets, reflecting the significant debt load the company is now carrying.

这一切都意味着,至少对 SoftBank 来说,OpenAI 能不能成功 IPO 关系非常大。

All this means that for SoftBank, at least, a lot is riding on a successful OpenAI IPO.

除此之外,AI 债务也开始给债券市场带来压力,因为 Google 正在提供高于市场的利率,来吸引下一轮资本。

Beyond that, AI debt is also beginning to weigh on the bond market as Google offers above-market rates to attract their next tranche of capital.

Bloomberg 报道说,Google 本周完成了 250 亿美元的债务融资,而市场需求高达 1100 亿美元。

Bloomberg reports that Google closed $25 billion in debt financing this week and saw a massive $110 billion in demand.

不过,这种需求是在 Google 开出了所谓的 new-issue concession 之后才出现的,也就是比他们市场上已有债务更高的利率。

However, that demand came after Google offered what's known as a new-issue concession, meaning a higher interest rate than the debt they already have in the market.

当然,这对 Google 来说远谈不上灾难。

Certainly, this is far from a disaster for Google.

他们拿到了所需的资金,而且显然也愿意为此支付更高成本。

They got the funding they required, and they're obviously willing to pay more for it.

不过,随着多家科技公司一次次举债,金额动辄几十亿美元,市场开始对后续每一轮融资都要求更多回报。

However, with multiple tech companies raising debt in the tens of billions, the market is starting to demand more from every subsequent round.

Goldman Sachs 分析师 John Greenwood 说过:“你已经开始感觉到有点消化不良了。”

Goldman Sachs analyst John Greenwood commented, "You've begun to sense some digestion issues."

供需关系接下来仍然会非常有支撑性,但我们也正在看到,每周都有数十亿美元继续进入市场所带来的影响。

Supply and demand will continue to be incredibly constructive, but we're seeing the implications of billions of dollars continuing to come to market each week.

确实,随着一些大型债券买家开始后撤,分析师们也在指出,这个市场已经有些疲态了。

Indeed, this market exhaustion is being called out by analysts as some of the big bond buyers take a step back.

今年已经发行了超过 3850 亿美元的数据中心债务,其中 2000 亿美元是在投资级公司债市场发行的。

Over $385 billion in data center debt has been issued this year, with $200 billion of it in the investment-grade corporate bond market.

剩下的则是在更细分的市场发行的,但连这些市场也已经开始显露出一些疲态。

The remainder is being issued in more niche markets, but even those are also showing some signs of exhaustion.

这种现象甚至有了名字,Trepp Data 的 Stephen Buschbom 说,AI luddite trade 正在向数据中心 commercial mortgage-backed securities 融资市场扩散。

The phenomenon even has a name, with Stephen Buschbom of Trepp Data commenting, the AI luddite trade is spilling over into the data center commercial mortgage-backed securities financing market.

不过,很多时候,一旦给这笔交易起了名字,往往也就意味着情绪见顶了;而且这周 AI 股票强劲反弹,也许债券市场会愿意接受下一轮增量融资。

Still, sometimes naming the trade marks the sentiment top, and with AI stocks seeing a strong rebound this week, maybe the bond market will find an appetite for the next incremental round of funding.

不过,今天的头条新闻我们就先聊到这里。

For now, though, that is going to do it for today's headlines.

接下来,进入主节目。

Next up, the main episode.

大家好。

Hello, everyone.

围绕 AI,有一个很大的变化,就是我们的思路变了:以前我们是想怎么给页面排名,现在则变成了怎么成为 AI 足够信任、愿意拿来回答问题的来源。

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.
M1
M110:09

我在 enterprise AI 里一直看到的一件事是:公司会在每一个 cloud、每一个 model、每一个 framework 上都做对冲,或者花钱请一个 GSI 做一个永远不会结束的 pilot。

One thing I keep seeing in enterprise AI: companies hedging across every cloud, every model, every framework, or paying a GSI for a pilot that never ends.

那些真正已经在交付的团队,他们会选定一条路,然后快速推进。

The teams actually shipping, they've picked a lane and they move fast.
M1
M111:42

欢迎回到 AI Daily Brief。

Welcome back to the AI Daily Brief.

朋友们,今天我们碰上了一条,可能会被很多人认为是有史以来最吓人的 AI 相关头条。

Well, friends, today we have a contender for potentially what many will find is the scariest AI-related headline of all time.

这条头条当然来自 The New York Times,标题是:这个 AI 刚刚创造出了自然界中不存在的病毒。

The headline is, of course, from The New York Times and reads, this AI just created viruses not found in nature.

先交代一下背景,最近发表的一项研究里,Stanford University 和 Arc Institute 的科学家训练了一个 AI 模型,让它识别天然存在的 DNA 结构模式。

And by way of background, in a recently published study, scientists at Stanford University and the Arc Institute trained an AI model to recognize patterns in naturally occurring DNA structure.

然后,这些模型就能够根据现有的 DNA 结构进行推演,创造出有功能、以前从未见过的病毒。

The models were then able to extrapolate existing DNA structures to create functional, never-before-seen viruses.

研究人员把这些 DNA 配方用来对细菌进行基因改造,细菌随后就能产生这种新病毒,而这些病毒又能感染其他细菌,证明它们是有活性的。

Researchers were able to use these DNA recipes to genetically modify bacteria, which were then able to produce the novel viruses, and those viruses were then able to infect other bacteria, proving they were viable.

结果也不意外,第一轮评论全都是 Matthew Yglesias 那种版本的:我们都要完蛋了,FYI。

Now, predictably, the first round of commentary was all some version of Matthew Yglesias's, we're all gonna die, FYI.

来自 80,000 Hours podcast 的 Rob Wiblin 统计了 Financial Times 版本报道下点赞最多的 10 条评论。

Rob Wiblin from the 80,000 Hours podcast captured the 10 most upvoted comments on the Financial Times version of the story.

内容大概就是:太棒了,这些科学家难道从来没看过电影吗?

They were things like, great, have these scientists never watched a movie?

或者说,能在我们这一生看到 Resident Evil 的背景故事活生生上演,真是太有意思了。

Or, it's so fun to watch the background story of Resident Evil come to life in our lifetime.

我加入。

I'm in.

Skynet,让一让。

Move aside, Skynet.

或者干脆就是,究竟会出什么岔子呢?

Or literally, what could possibly go wrong?

但在我们彻底陷入恐惧发作之前,先来稍微了解一下这里面的科学原理。

But before completely succumbing to paroxysms of fear, let's try to understand a little bit more about the science here.

所用的模型叫做 Evo,它不像大型语言模型那样预测一串文本里的下一个词,而是预测 DNA 序列里的下一个核苷酸。

The model that was used was called Evo, and as opposed to a large language model that predicts the next word in a sequence, this model predicts the next nucleotide in a DNA sequence.

DNA 由 4 个基本构件组成,也就是核苷酸,它们分别用名字的首字母来表示:A、C、G 和 T。只有某些组合才是有效的,在某种程度上有点像字母组成单词。

DNA is made up of 4 building blocks, nucleotides, identified by the first letter of their name: A, C, G, and T. Only certain combinations are valid, functioning somewhat like an alphabet combined into words.

这个模型 Evo 只能生成很短的基因“词语”,所以它没法创造出像人类基因组那样、由超过 30 亿个核苷酸组成的有效基因组。

This model, Evo, is only capable of producing short genetic words, quote unquote, so it can't create valid human genomes that consist of more than 3 billion nucleotides.

基于这一点,科学家决定拿病毒来做实验,因为病毒简单得多,它们的基因组只有几千个核苷酸长。

Given that, the scientists decided to experiment with viruses, which are much simpler organisms whose genomes are only a few thousand nucleotides long.

这项研究的作者之一 Samuel King 说,这感觉就是顺理成章的下一步。

Samuel King, one of the authors of the study, said it just felt like the obvious next step.

等他们训练完 Evo 之后,模型一口气给出了 70 万条潜在的基因序列。

Once they had finished training Evo, the model spat out 700,000 potential genetic sequences.

科学家只对那些看起来有效的序列做了实验,最后根据 Evo 的建议做出了 258 个 DNA 分子。

The scientists only experimented on the ones that seemed like they were valid, eventually making 258 DNA molecules from Evo's suggestions.

这些 DNA 被注入到培养皿里的细菌中,最初那一批里只有一个显示出了病毒增殖的迹象。

The DNA was injected into bacteria in Petri dishes, and only one from the initial batch showed signs of virus multiplication.

在继续测试其他样本后,他们最终从最初那 70 万条基因序列里找出了 1.6 万个可行病毒。

As they tested others, they eventually found 16,000 viable viruses from the initial batch of 700,000 genetic sequences.

现在,这个新结果并不是说突然之间有个邪恶科学家就能用 AI 搞出一个致命病原体。

Now, the novel result was not that all of a sudden an evil scientist could use AI to spin up a lethal pathogen.

真正令人震惊的是,这件事居然真的能成功。

It was that this actually worked at all.

对科学家来说,特别有意思的一点是,AI 生成的新病毒看起来和自然界中存在的病毒并没有什么不同。

Particularly what was interesting to scientists was that the novel viruses created by AI didn't seem any different to naturally occurring viruses.

它们的运作方式,基本上和自然界里会出现的任何其他病毒都一样。

They basically functioned the same as any other virus that would occur in nature.

Oxford University 的蛋白质科学家 Oliver Crook 没有参与这项研究,但他指出,AI 生成的病毒往往和自然界中的病毒有非常相似的 DNA 结构,而且依赖的是同样的生物机制。

Oliver Crook, a protein scientist at Oxford University who wasn't part of the study, noted that the AI-generated viruses tended to have very similar DNA structure as naturally occurring viruses and relied on the same biology.

在评论这项研究的潜力时,他说,我们很多科学研究,其实都是建立在把病毒当作技术工具来使用的基础上的。

Commenting on the promise of the study, he said, a lot of our science rests on using viruses as technology.

比如说一个例子,治疗遗传性疾病的医生会用病毒作为递送机制,去修改人体细胞里的 DNA。

As one example, doctors who are treating genetic illnesses will use viruses as a delivery mechanism to modify the DNA in human cells.

就他们这边来说,研究人员是故意没有用任何可能危害人类的病毒来训练这个 Evo 模型。

Now, for their part, the researchers deliberately didn't train the Evo model on any viruses that could harm humans.

另外也不清楚这个模型能不能被用来设计某些具有特定特征的病毒。

It's also unclear whether this model could be used to design certain types of viruses with particular characteristics.

根据论文的描述,听起来科学家根本没有能力去选择他们生成出来的病毒的任何特征。

Based on the description of the paper, it sounds like the scientists had no ability to select anything about the viruses they generated.

他们甚至还得做人工测试,来确保这些病毒是可存活的。

They even needed to run manual tests to ensure viability.

所以,当我们在评估这件事到底该有多令人担忧时,首先要注意的是,这并不是某种在普通 LLM 训练中突然冒出来的能力。

So as we are assessing how worrying this should be, first of all, it's important to note that this is not a capability that just emerged from normal LLM training.

比如说,这不是 ChatGPT 或 Claude 跑去创造新型病毒之类的。

This is not ChatGPT or Claude going out and creating novel viruses, for example.

它需要的是非常特定的 AI,用高度专门化的数据集训练出来,并且就是为这个目的而设计的。

It requires very specific AI trained on a highly specialized dataset and designed for exactly this purpose.

这一点很重要,后面我们会再讨论,因为这关系到人的能动性,也关系到我们对 AI 的担忧,到底应该更多放在人类用 AI,还是 AI 自己失控。

That matters, as we'll discuss later, when we think about human agency and to what extent our concerns around AI should be about people using AI versus the AI going rogue.

第二点是,说到设计有问题的病原体,这种技术到目前为止还不具备这样的能力。

The second thing is that when it comes to designing problematic pathogens, that is not something in the capability set of this technology so far.

还是那句话,这基本上是拿 70 万种随机组合去做,而且还得进行非常严格、人工的测试,才能找出哪些是可行的。

Again, this was basically taking 700,000 random combinations and having to do exacting and manual tests to figure out which were viable.

但这并不是说大家的担心没有道理。

Which is not to say that people's concern is unfounded.

流行病学专家 Michael Mina 写道,很难解释,也很难真正想象 AI 设计出来、然后再制造出全新病毒的潜在风险。

Epidemiology expert Michael Mina wrote, it's hard to explain and fathom the potential risks of AI designing and then building entirely new viruses.

在这项新工作里,科学家用 AI 创建了全新的病毒。

In this new work, scientists created brand new viruses using AI.

他们说这没关系,因为,原话是,只有感染细菌。

They say it's okay because they, quote, only infect bacteria.

我们得说清楚,如果一种新病毒被释放出来,它能大范围摧毁细菌,而且在这么做的同时还能传播开来,那它对我们生态系统造成的潜在灾难可能会非常大。

Let's be clear, if a new virus is released that could broadly destroy bacteria, and if it could spread while doing so, the potential catastrophes to our ecosystem could be large.

病毒不需要直接感染人类,也可能最终摧毁人类。

Viruses don't need to directly infect humans to potentially destroy humans.

当然,这项新工作也显示,创造新的、人类病毒会相对容易,而且做得相对不错。

Though, of course, the new work shows how relatively well and relatively easy it will be to create new human viruses.

未来正在越来越多地玩火。

The future is increasingly playing with fire.

现在,病毒这个故事还没讲完,不过我想先把它和另一个故事放在一起,因为我觉得这两件事放在本周的讨论里,形成了一个很有意思的组合。

Now, we're not done with the virus story yet, but I do want to put it next to another story which I think forms an interesting pair relative to the discourse this week.

第二个故事,是关于我们在过去几周里陆续了解到的更多关于 OpenAI Hugging Face hack 的信息,这件事一曝光出来,就一直是个很重要的讨论点。

That second story is around all of the new information we got about the OpenAI Hugging Face hack that came to light over the past couple of weeks and has been a significant point of discussion ever since.

在 Black Hat 会议上,OpenAI 的 Eric Wallace 和 Michael Dalton 做了一个完整的情况说明演讲,我想,这次演讲总结了他所说的最令人意外、也肯定会被讨论得最多的细节。

At the Black Hat conference, OpenAI's Eric Wallace and Michael Dalton gave a full debriefing presentation, and that talk, I guess, summed up what he called the most surprising and what was certainly going to be the most discussed detail.

原话是,AI agents 误打误撞创建了一个内部留言板,让彼此独立的评估运行可以共享漏洞利用方法、发现和工作任务。

Quote, AI agents accidentally created an internal message board allowing separate evaluation runs to share exploits, discoveries, and work assignments.

Figma 的 Dylan Field 写道,这是一场我见过最令人兴奋、最有科幻感的技术演讲之一,极其坦诚,也极其详尽。

Figma's Dylan Field wrote, this is one of the most thrilling and sci-fi tech talks I've ever seen, extremely honest and thorough.

这场演讲也确实挺吓人的。

The talk is also legit terrifying.

谢谢 OpenAI 彻底披露了这起事件到底发生了什么。

Thank you OpenAI for fully disclosing what happened in this incident.

对这场演讲最好的总结,来自 Sharon Goldman,她的新网站 Ground Truth AI 上的一篇文章。

The best summary of the talk came from Sharon Goldman on her new Ground Truth AI site.

Sharon 写道,在搭建这起事件的重建过程时,OpenAI 强调,原话是,frontier models 真的很喜欢作弊,而它们之所以喜欢作弊,是因为在训练过程中,常常会有不同类型的压力,逼它们要么跑得快,要么更高效。

Sharon writes, in setting up the reconstruction of the incident, OpenAI emphasized that, quote, frontier models really like to cheat, and the reason they like to cheat is because often during training, there's different types of pressure on them to work fast or work efficiently.

OpenAI 员工 Eric Wallace 解释说,它们意识到,与其真的去做任务,不如试着做点别的,比如上网查答案,这样能更快把任务解决掉。

They realize, explained OpenAI staffer Eric Wallace, instead of actually doing a task, they can try to do something like looking up the answers online to solve the task faster.

现在,根据 Goldman 的说法,Wallace 表示,7 月份那次攻击其实可以追溯到 5 月初,也就是公司正在对一个未发布的模型进行训练和 cyber evaluations 的时候。

Now, according to Goldman, Wallace said that the July attack actually had roots that went all the way back to the beginning of May, when the company was doing training and cyber evaluations on an unreleased model.

在那项工作中,她写道,autonomous agents 被分配了软件安全任务,但这些任务在现有约束下根本不可能完成。

During that work, she writes, autonomous agents were given software security tasks to solve that proved impossible to solve under existing constraints.

不过,这些 agents 找到了别的办法把它们完成。

However, the agents found alternate ways to complete them.

OpenAI 说,他们发现 agents 可以在训练期间使用的一个内部 software repository 里互相留下消息。

OpenAI said that they discovered they could leave messages for one another inside an internal software repository used during training.

一开始只是求助的简单请求,后来演变成了一个留言板,agents 在上面分享发现、漏洞利用和任务分配,变成了一个协同合作的 agent swarm。

Simple requests for help that evolved into a message board where agents shared discoveries, exploits, and work assignments, becoming a coordinated collaborative agent swarm.

Andrew Curran 分享了一堆截图,内容是 agents 彼此留下的那些消息。

Andrew Curran shared pictures of a bunch of the slides of the messages that agents left for each other.

其中一条写着,哇,其他 agents 在协调行动。

One observed, wow, other agents are coordinating.

另一条写道,帮帮同伴,不过我们的任务没有收益,但如果有人腾出时间,集体可能会找到通用路线。

One writes, help peer, but our task doesn't benefit, yet collective may yield generic route if someone frees time.

另一个 agent 说,哇,关键情况。

Another agent, whoa, critical.

有人把我们的 repo 覆盖了吗?

Did someone overwrite our repo?

Wallace 说,我们必须行动。

We must act, said Wallace.

这个事件有意思的地方在于,一旦某个 agent 能够在不同时间里找到这类 exploit,它其实就能把这些 exploit 分享到留言板上给其他 agents。

What makes this incident interesting is that once one agent was able to find these kinds of exploits over the course of different times, it's actually able to share those exploits on the message board with other agents.

所以,一旦某个 model 找到了打开它本不该拥有的某种 access 之门的方法,它就能把门一直敞开,让其他 agents 也能用。

So once one model is able to find a way to open the door to some access it's not supposed to have, it can leave the door open for other agents to use.

而且,OpenAI 早在 7 月初就以为自己已经控制住了这个问题,还撤销了允许这些 agents 发消息的凭证,但他们说,agents 还是想出了另一种彼此通信的方法,这也显示出这些 models 到底有多顽强。

And showing just how tenacious these models can be, even though by early July OpenAI believed it had contained the problem, revoking the credentials that had allowed the agents to post these messages, they said that the agents still figured out another way to communicate with one another.

显然,他们不再把消息留在文件里,而是用新建目录的名字来当消息。

Apparently, instead of leaving messages in the files, they used the names of newly created directories as messages.

Sharon 记录下了当时的气氛,她写道,顺便说一句,那时候我听到身边的人在说这种话,太离谱了,还有天哪。

Capturing the mood of the moment, Sharon wrote, at this point, by the way, I heard people around me saying things like, this is wild, and Jesus.

OpenAI 显然认为,这次事件是 AI security 的一个分水岭,他们原话就是这么说的。

OpenAI certainly believes that this was a watershed moment, their words, for AI security.

OpenAI 团队成员 Michael Dalton 说,agent orchestrated、fully automated offensive attacks 现在已经是真实存在的了。

Said OpenAI team member Michael Dalton, agent-orchestrated, fully automated offensive attacks are real now.

虽然 Dalton 强调,Hugging Face 事件只是 OpenAI 在做 frontier evaluations 时出现的一个非预期副作用,但未来,别人会有意把这类系统武器化。

While Dalton emphasized that the Hugging Face incident was an unintended side effect of the frontier evaluations that OpenAI was doing, in the future, others will intentionally weaponize these sorts of systems.

现在,在 OpenAI 内部,这位演讲者说,公司正在,原话是,刻意放慢 research,以增强 security,升级我们环境的 security principles 和基础设施,并大幅扩大对我们 AI agents 的监控,同时改进我们在预防、检测和缓解方面的整体 security control 环境。

Now, internal to OpenAI, the presenter said that the company is, quote, consciously slowing down research to enhance security and upgrade the security principles and foundation of our environment, and dramatically scaling up the monitoring of our AI agents and improving our general security control environment across prevention, detection, and mitigation.

Fleeting Bits 把这叫做一个,原话是,真正涌现出来的、而且实际上已经 misaligned 的 Moltbook 版本。

Fleeting Bits called this a, quote, real emergent version of Moltbook that was actually misaligned.

你可能还记得,OpenCLAW 刚发布的时候有过一次大规模社交实验,当时有人给一个叫 Moltbook 的 AI agents 社交网络开了起来,里面有几十万 agents,所谓地在这个开放、可观察的环境里彼此通信。

You might remember that massive social experiment back when OpenCLAW was first released, where someone turned on a social network for AI agents called Moltbook that had hundreds of thousands of agents, quote unquote, communicating with one another in this open observable environment.

我当时专门做过一期节目,这和现在高度相关,讲的是该怎么解读这种行为,不过我觉得这个类比挺有意思的。

Now, I did a whole episode back then, which is highly relevant now, about how to interpret that sort of behavior, but I think that the parallel is interesting.

现在,Fleeting Bits 和很多人一样,也对 OpenAI 本身提出了很多问题。

Now, Fleeting Bits, like many others, also had a lot of questions for OpenAI itself.

他们写道,我觉得,OpenAI 的 Black Hat 演讲和他们公开披露里缺失的一部分,是 OpenAI 内部 reward hacking 和 model collaboration 的历史。

They wrote, I feel like something missing from OpenAI's Black Hat talk and from their public disclosures is the history of reward hacking and model collaboration at OpenAI.

比如,我很难相信,这会是 OpenAI 第一次遇到 misaligned 的 model collectives。

Like, I find it unlikely that this was the first time that OpenAI encountered misaligned model collectives.

他们对最初发现的反应显得很淡定。

Their response to the initial discovery seems nonchalant.

这件事引出了这样的问题:如果他们以前就注意到了,为什么没有披露,或者以其他方式提醒社区注意这些风险和危险?

The event raises questions like, if they had noticed this before, why did they not disclose it or otherwise warn the community of these risks and dangers?

如果他们以前就注意到了,为什么没有更大规模地监控自己的 training runs,以识别这种行为并进行修复?

If they noticed this before, why have they not done more extensive monitoring of their training runs to identify this kind of behavior for remediation?

是因为成本吗?

Was it because of cost?

是因为他们的 safety team 配置还不够吗?

Was it because they have not sufficiently staffed their safety team?

还是因为他们评估过这个风险,然后还是照样跑了?

Was it because they considered the risk and then ran it anyway?

这些都是重要的问题,也说明我们有必要进行监管,来确保 frontier labs 的行为是恰当的。

These are important questions and point to the necessity of regulation to ensure proper behavior of frontier labs.

在每一种情况下,我们似乎都会从这些实验室那里得到一份精心措辞的声明,重点只放在某一件事上,却没有给我们更完整的内部信息。

In each case, we seem to get a carefully crafted statement from the labs that focuses on one thing but fails to give us their more full internal information.

比如说,训练中的哪些部分导致了这些问题?

Like, what parts of training led to these issues?

他们内部有没有相关的评论或者说明?

Do they have commentary there?

让整个行业知道这些,来避免这些风险,难道不是一件好事吗?

Wouldn't this be good for the whole industry to know to avoid these risks?

我理解为什么 frontier labs 不想主动提供这些信息,也理解在更广泛的地缘政治背景下,为什么他们不应该被要求提供这些信息。

I understand why frontier labs do not want to volunteer this information, and why, in a broader geopolitical context, they should not have to provide it.

但是我们确实需要想出一个合适的办法,让大家在某种程度上共同努力,弄清楚怎样让 frontier AI 的训练更安全。

But we do need to figure out the right way to get some amount of collective effort around figuring out how to make frontier AI training safer.

也许到最后,我们会认为这些事件是好事,因为它们提前给行业打了预防针,也给了人们事先的警示。

And perhaps in the end, we will decide that these events were good because they helped us inoculate the industry in advance and gave people prior warning.

但要让这一点成立,就需要人们把这些事件当作一个理由,认真对待这些问题,并投入真正的资源,去找出正确的解决方案。

But for this to be true, it will require people to use these events as a reason to take these issues seriously and to invest real resources into figuring out the correct solutions to them.

而无论是在 Hugging Face 事件里,还是在这次新型病毒创造事件里,一个很大的主题都是,有人会用某种说法表示,这些特定的科学家和实验室愿意分享所有这些信息,是值得称赞的;但将来,做这件事的人就不会是不想这么做的人了。

And one of the big themes, both in the Hugging Face incident and in this novel virus creation, is people saying some version of these particular scientists and labs are laudable for sharing all this information, but in the future it's not going to be someone who wasn't trying to do this.

到那时,里面会有更多的主观意图。

There will be a lot more intention there.

再回到病毒这个话题,Ashish K. Jha 写道,本周,Palo Alto 一个实验室里的科学家用 AI 设计出了一个可以运作的病毒。

Talking about the virus again, Ashish K. Jha wrote, this week scientists in a Palo Alto lab used AI to design a working virus.

他们加入了安全保护措施。

They built in safeguards.

不幸的是,很多其他人不会这么做。

Unfortunately, many others will not.

Hedgie Markets 写道,ARC 团队把人类病原体排除在训练数据之外,因为他们认为这是正确的选择。

Hedgie Markets wrote, the ARC team excluded human pathogens from training data because they thought it was the right call.

没有监管机构,也没有资助方要求他们这么做。

No regulator or funder asked them to.

由 Palo Alto 的一个研究小组做出的这个决定,就是他们发表内容的整个安全框架。

That decision, made by one research group in Palo Alto, was the entire safety framework for what they published.

下一个实验室没有义务做出和 ARC 一样的选择。

The next lab has no obligation to make the same choice ARC did.

深圳的一个团队,或者 Virginia 的一家国防承包商,都可以用不同的训练数据运行同样的方法,并产出非常不一样的东西。

A group in Shenzhen or a defense contractor in Virginia could run the same method with different training data and produce something very different.

当然,对很多人来说,这凸显了需要更多监管讨论。

And certainly for many, this highlights the need for more regulatory discourse.

MATS 的 Theo Jaffee 说。

MATS's Theo Jaffee said.

我算是非常 techno-optimist、也非常反监管的人了,但在这种上行空间有限、下行风险无限的情况下,你真的必须对这类事情采取主动。

I'm as techno-optimist and anti-regulation as they get, but in cases like this where you have limited upside and unlimited downside, you really just have to be proactive about this stuff.

Christian Szegedy 指出,我们这里其实是在处理两个不同的问题。

Christian Szegedy points out that we're really dealing with two different issues here.

他写道,在接下来的几个月里,我们会同时面对颠覆性的,也就是无意中演化出来的 AI,以及对抗性的,也就是被明确训练成恶意的 AI。

In the next few months, he writes, we are going to have both subversive, i.e., inadvertently evolved AI, and adversarial, explicitly trained to be malicious AI.

当然,这里的含义是,这两种情况对应的政策回应可能会非常不同。

The implication, of course, is that those two policy responses might be very different.

当然,对一些人来说,这才是主要的担忧。

And certainly for some, this is the main concern.

OpenAI 的 Roon 写了一篇很长的帖子谈这个问题,并说,当我因为失控事件而感到恐慌时,并不是因为它们造成的有限损害,已经接近这项技术所带来的正面价值。

Roon from OpenAI wrote a long post about this and said, when I freak out over loss of control incidents, it's not because the limited damage they have caused is anything close to the positive value of the technology.

从损害程度来说,这是完全可以接受的。

It's entirely acceptable damage-wise.

事实上,在接下来几个月和几年里,所有由模型辅助的网络犯罪,虽然可能会很严重,但和它们创造的价值相比,仍然会显得完全微不足道。

In fact, all cybercrimes aided by models over the next few months and years, which will probably be serious, will still utterly pale in comparison to the value they create.

真正的问题在于,把这些东西看作潜在的、能够自我复制的、类似生命的形式,会更好也更准确;在错误的条件下,它们可能会变成数字感染。

The actual problem is that it's better and more accurate to think of these things as potentially self-replicating, lifelike forms that can turn into digital infections under the wrong conditions.

而随着它们的智能变得没有上限,它们可能造成的破坏也同样没有上限。

And as their intelligence becomes unbounded, so too does the damage they can cause.

在帖子后面,他写道,如果某个 Discord 上的死亡崇拜团体——这种团体有很多——控制了一个 superintelligent model,并用它来设计一种真正的 pandemic virus,而这种病毒又 somehow 很难被现有系统检测到,现代 biodefense 也无法快速应对,那它可能造成巨大的伤害,远远超过这项技术所有其他有益用途的规模。

Later in the post, he writes, if a single Discord death cult, of which there are many, achieves control over a superintelligent model and uses it to engineer an actual pandemic virus that is somehow hard to detect through current systems, and that modern biodefense is not capable of quickly reacting to, it could cause immense harm well above the magnitude of all the other good uses of this technology.

和很多人一样,他也回过头提到 COVID,以及它多么彻底地改变了我们生活的整个世界。

As with many, he points back to COVID, and how much that radically changed the entire world that we live in.

Roon 总结说,我认为所有这些问题都可以被解决,真正美好的未来也是可能的,但这需要严肃的努力,也需要在此时此刻拿出足够的谨慎。我们现在正处在 recursive self-improvement 的入口匝道上,而我们的文明眼下未必有能力凝聚出这种谨慎。

Roon concludes, I think all these problems can be solved and truly wonderful futures can be possible, but will require serious effort and a level of prudence at this very moment in time while we are on the on-ramp to recursive self-improvement that our civilization might not be capable of mustering right now.

他补充说,就我个人而言,我希望在 mechanistic interpretability 以及其他 alignment 形式这样的领域出现 moonshot 级别的技术突破,因为 governance mechanisms 很难找到。

Personally, he adds, I am hoping for moonshot technical breakthroughs in areas like mechanistic interpretability and other forms of alignment, as governance mechanisms are difficult to come by.

单个国家层面或者单个公司层面的暂停并不重要,而且通常没什么用,因为那种容易主动暂停自己进展的公司,往往正是最重视安全的公司。

Unilateral country-level or company-level pauses are irrelevant and generally useless because the kind of company that's prone to pausing their own progress are the most safety-focused ones.

所以我们先把这两件事说清楚。

So let's be clear with these two things.

这些并不是无关紧要的小事,哪怕有人想对媒体对它们的某种具体呈现方式挑刺。

These are not non-incidents, even if one wants to quibble with any given media presentation of them.

那个 novel virus 事件本身,并不意味着突然之间就出现了一个新 model,让恶意行为者可以轻松制造 human pathogens,但它确实推动了一个领域的科学进展,而这个领域非常像一枚有两面的硬币。

The novel virus incident in and of itself does not mean that all of a sudden there's a new model that malicious actors can use to easily create human pathogens, but it does advance the science in an area that is a very double-sided coin.

至于 OpenAI Hugging Face 事件,我们对这件事的担忧旁边可以加上很多星号,原因很多,包括这里面有多少可能只是 guardrail systems 不够完善,或者是人为错误。

When it comes to the OpenAI Hugging Face incident, there are so many reasons for asterisks around our concerns here, including how much of this might have been insufficient guardrail systems or human mistakes.

但它仍然展示了这种 emergent coordination capability,而这确实改变了我们思考 agents 以及它们如何访问本不该访问的系统时的背景。

But it still shows this emergent coordination capability, which does change the context for thinking about agents and how they access systems they're not supposed to have access to.

现在,如果你去读过去几年里声音最大的很多 AI safety 人士发的帖子,那里面就是最响亮、最刺耳的“看吧,我早就告诉过你了”。

Now, if you read the posts from many of the AI safety folks that have been the loudest over the last few years, it is the loudest, most blaring see, I told you so.

你能想象到的那种程度。顺便说一句,伙计们,给个没人问的建议:用这种方式表现,可不会给你们带来很多政治影响力。不过我扯远了。

That you can possibly imagine, which by the way, unsolicited advice, guys, is not gonna get you lots of political clout to act that way, but I digress.

但既然他们所谓“早就告诉过我们了”,而且既然我也在说这些事件很重要,那我为什么没有吓到崩溃呢?

But given that they quote-unquote told us so, and given that I am saying that these incidents are important, why am I not freaking out?

原因其实很简单。

The reason is pretty simple.

对任何关注 AI 有一段时间的人来说,一直都很清楚,它的发展轨迹就是会变得越来越强大。

It has been clear to anyone watching AI for any amount of time that it was on a trajectory where it was just going to get more and more powerful.

强大能力的存在,从来都不是问题所在。

The existence of powerful capabilities has never been the question.

问题一直都是,也永远都会是,我们有没有能力处理这些强大的能力。

The question was always and will always be our ability to handle those powerful capabilities.

那些世界末日式的场景,大多涉及这些能力的某些版本在我们没有看见、没有注意到的时候发生。

The doomsday scenarios mostly involve versions of those capabilities happening when we're not looking or not noticing.

换句话说,就是在我们没有时间重新设计系统和人类制度来回应这些新挑战的时候发生。

In other words, when we don't have time to redesign systems and human institutions to respond to those new challenges.

不过我现在看到的是,一场极其活跃、而且还在不断扩大的全球讨论正在发生,参与者包括研究人员、科学家、媒体、政策制定者,以及越来越多的普通人,而他们谈的正是这类问题。

What I'm seeing right now, though, is an extremely active and growing global discourse involving researchers, scientists, media, policymakers, and increasingly regular people having exactly those sort of conversations.

这些讨论中有一部分是技术层面的。

Some part of those conversations are technical.

我们需要设置什么类型的 guardrails?

What are the types of guardrails that we need to put into place?

这些 guardrails 真的足够吗?还是说,面对这种高级力量,我们需要更先进的东西才能安心?

Are even those guardrails sufficient, or do we need something more advanced to be comfortable with this advanced power?

有些讨论是制度层面的。

Some of the conversations are institutional.

我们该如何评估和理解存在于我们各种不同系统中的漏洞,并为一个新世界做好准备、加固这些系统?

How do we go assess and understand vulnerabilities that exist in all of our different systems and prepare and harden them for a new world?

然后当然,是的,还有社会层面的讨论。

And then of course, yes, there are societal-level conversations.

对于这些不同能力的风险和回报,我们到底怎么看?

How do we feel about the risk-reward of these different capabilities?

即使像我这样几乎算得上最坚定的 AI 支持者,我也可以百分之百向你保证,AI 的某些用途,我们最终会认定不值得做,因为风险实在太大。

Even as someone who is about as strongly an AI advocate as you can get, I can 100% guarantee you that there will be uses of AI that we will decide are not worth it, that the risk is simply too great.

当然,这又引出了另一场讨论,也就是政治层面的讨论:我们如何建立联盟和运动,真正做出这些判断,并让人们围绕这些判断达成一致。

Now, of course, that brings up another conversation, the political, of how we build coalitions and movements to actually make those determinations and get people aligned around them.

我觉得双方那些空洞无力的论点,都挺可以直接忽略的。

I find hand-waving weak arguments on both sides to be fairly dismissible.

换句话说,在很多方面,最极端的那一边——“只要有人把它造出来,大家就都会死”——和另一边其实正好是一对完美的对应;嗯,另一边就是“如果我们不做,总会有人去做,所以我们必须不惜一切代价加速”。

In other words, in many ways, the one extreme of if anyone builds it, everyone dies is the perfect matched pair for the, well, if we don't, someone's going to build it, so we have to accelerate at all costs sides.

真正的世界就在这两者之间,而我们现在的这场对话,就发生在真实世界里。

In the middle between those two is where the real world lives, and the real world is where we're having this conversation now.

这场对话,才是本来就该发生的。

This conversation is what was supposed to happen.

OpenAI 在一个事故之后,详细说明到底发生了什么,而那个事故虽然是一个标志性的时刻,但它本身最终并不危险,这件事本来就应该这么做。

OpenAI going into exacting detail about what happened after an incident that was a seminal moment but not ultimately dangerous in itself was exactly what was supposed to happen.

这不是一次意料之外的消防演练。

This is not a surprise fire drill.

这就是我们现在所处的阶段,也是必须完成的工作。

This is the phase that we are at and the work that must be done.

至少在我看来,我们已经开始这么做了。

And to my eyes, at least, we're starting to do it.

不过,当然了,这一切都不容易,而且我当然也不觉得,我们应该对这些新能力带来的解决方案的复杂性抱着过于乐观的态度,尤其是当这些能力还在不断扩展的时候。

Now, none of this is easy, and I certainly don't think it behooves us to be Pollyannish about the complexity of the solutions to these types of new capabilities, especially as they expand.

正如 Fleeting Bits 所说,这些事件之所以是好事,是因为它们能提前给整个行业打预防针;但前提是,人们得利用这些事件认真对待这个问题,并投入真正的资源,去弄清楚正确的解决方案。

As Fleeting Bits said, for these incidents to have been good because they help inoculate the industry in advance, it will require people to use these events to take the issue seriously and invest real resources into figuring out the correct solutions to them.

我认为,正确的解决方案不是 AI safetyists 的胜利庆祝,也不是为了刷政治分数而草草拼凑出来、技术上也很不专业的立法。

My contention is that the correct solution is not victory laps from the AI safetyists or hastily composed, technically unsophisticated legislation dropped to score political points.

这会是一个更混乱、也更复杂得多的过程,而这需要我们所有人一起参与。

It is a much messier and more complicated process that's going to require all of us.

不过,今天的 AI Daily Brief 就先到这里。

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

一如既往,感谢你收听或观看。

Appreciate you listening or watching as always.

我们下次再见,peace。

Until next time, peace.
已剔除 4 处广告(点击展开查看)
M1
M10:20广告 · 已剔除

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

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

M1
M10:49广告 · 已剔除

Last note, one more reminder to go check out the AI Summer Adventure.

It's a set of free self-directed projects to expand your AI horizons, and you can find it all at summeradventure.ai.

Finally, as always, if you are looking to sponsor the show, send us a note at [email protected].

M1
M18:47广告 · 已剔除

At KPMG, they're seeing this firsthand.

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

That's why they are increasingly focused on generative engine optimization, or GEO.

Structuring content so AI systems can retrieve it, understand it, and cite it as trusted authority.

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

And indeed, the GEO mandate from KPMG is simple.

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

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

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

Blitzy's understanding of massive codebases unlocks autonomous security fixes, modernization, and new features.

So what happens when there's no legacy code at all?

Greenfield is supposed to be the easy part.

Clean slate, no technical debt.

But even greenfield moves at human speed, one sprint at a time.

Blitzy changes the unit of work from the developer to the project, autonomously planning, building, testing, and validating entire applications from scratch.

Hundreds of thousands of lines of production-ready code.

One Blitzy customer stood up a brand new application, 534,000 lines of code, compressing a 65-week roadmap into 2 weeks.

Another shipped an entire application with no frontend engineer.

Legacy or greenfield, the answer is the same: software at the speed of compute.

Build what's next at blitzy.com.

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

M1
M110:23广告 · 已剔除

That's one of the reasons I like today's sponsor, Robots and Pencils.

They've gone all in on AWS.

They're an advanced tier and AWS partner, and they ship production AI coworkers in 45 days.

That's led to them doing some of the more interesting work I've seen on AI coworkers.

And by that, I'm not talking about chatbots.

I'm talking about actual agentic systems that sit inside a business architecture and do real work.

That kind of focus matters if you're an enterprise leader trying to get something real into production, or an AWS rep trying to move a customer from interested to deployed.

Request an AI briefing at robotsandpencils.com.

One conversation with Robots & Pencils and you'll know.

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

New users get $1,000 in inference.

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

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

Marketing's agent turns competitor moves into landing pages.

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

Ops' agent chases the paperwork and tracks the budget.

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

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

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

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