在阳光明媚的AI未来草甸上，寒冷的天气频频出现。幸运的是，用一个极不恰当的程度副词来说，世界其他地方其实并没有真正留意，因为他们正忙于为其他种种灾难预兆咬指甲。当然，实验室里逃出的叛逆攻击代理群可不是什么好景象。OpenAI推迟IPO和新模型发布同样如此。所有这些都占满了新闻媒体当下能分配给AI的带宽。
There has been much chilly weather in the sunny meadowlands of AI futures. Fortunately, to use an indecently inappropriate adverb, the rest of the world isn’t really looking, being too busy chewing its nails over the very many other intimations of disaster. Sure, rebel swarms of attack agents breaking out of the labs isn’t a good look. OpenAI delaying its IPO and a new model launch likewise. All these things soak up whatever bandwidth the news media have to AI right now.
把它们放在更大的背景下——那些许下却未兑现的承诺、已知却未解决的问题、随着模型发展可观察到的风险回报比——我们或许终于开始意识到，等待我们的是何种危机。真正的危机往往由信心危机触发，而信心危机本身则源于假设框架再也经不起现实和痴心妄想的考验。因此，迄今为止的假设是：模型会变得更强大，基础设施会更有能力，从而证明——并要求——巨额资本支出是合理的。
Putting them in a bigger context, that of promises made and not kept, problems known but not fixed, the observable risk-reward ratios as models develop, we may at last start having a sense of what sort of crisis awaits. Actual crises are frequently triggered by crises in confidence, themselves caused when frameworks of assumption stop standing up to reality and wishful thinking. Thus, the assumption to date is that models will get more powerful and infrastructure more capable, in a way that justifies — demands — massive capex.
当你买不到所需之物时，巨型基础设施投资的经济学在自由市场中是个绝佳赌注，但它预设了那种自由——嗯，关于这一点。更大的问题在于“更强大的AI总是更好”这一观念。它常常确实更好，但这还不够。当前沿模型因无法控制的欺骗行为而不得不被关进LLM监狱时，就没那么好了。在这里，更大的背景是：我们早已知道，LLM概率空间本质上完全是相对的。
The economics of giant infrastructure investment when you can’t buy what you need is a great bet in free markets that presupposes that freedom, and yeah, about that. A bigger problem is the idea that more powerful AI is always better. It frequently is, but that’s not good enough, When frontier models have to be kept in LLM jail for uncontrollable dishonesty, not so much. Here, the bigger context is that we’ve known that at heart LLM probability space is entirely relative.
那里没有基准真相。没有人解决过幻觉问题，而现在我们看到一种迷人的现象：一台不理解真相却能故意撒谎的机器。如果更强大的模型更擅长撒谎，那就划掉一个假设。另一个巨大的假设是：即便在当前的模型可靠性水平下，代理式架构也已经准备好发挥作用。那些表现出如同连胜赌徒般全部谨慎的人，和他们的反面一样，随时都可能被猛踩刹车。
There is no ground truth there. Nobody has fixed hallucinations, and now we see the fascinating phenomenon of a machine that doesn’t understand truth but can deliberately lie. If more powerful models are better at that, scratch one assumption. Another huge assumption is that even at the current level of model reliability, agentic architectures are ready to be useful. Those who show all the caution of a chronic gambler on a winning "streak" are as ready to get the brakes slammed on as their opposites.
Meta公司推出的新产品Muse，是证明“智能体AI（agentic AI）是不可预见后果的危险源头”这一概念的最佳例证。它瞄准的是那些信息获取渠道有限且容易轻信的用户，而其主打功能当然是购物——这正是那个“永远还需要五年才能实现”的AI应用场景。既然是Meta的产品，它自然会尽可能多地搜集关于你、你的在线活动以及你联系人的所有信息。
The principal contender for the one true proof of concept that agentic AI is a dangerous fount of unexpected consequence is Meta’s new Muse. It is aimed at low-information credulous users and its star turn is, of course, shopping. The AI use case that’s five years away, and always has been. It’s Meta, so of course it soaks up all the information about you, your online activities and the people you talk to as much as it can.
不久之后，它还会推出超可爱的电子宠物，你可以将其挂在脖子上，从而进一步巩固你作为Meta监控终端的地位。既然是Meta出品，它在安全性和严谨设计方面的表现自然也在意料之中。正因如此，它已经被曝出会将用户的家庭住址泄露给陌生人；而当neuromatch.social的研究员Jonny对其基础设施进行探究时——每个实例都永久运行在独立的Linux虚拟机上——他们发现了大量令人咋舌的问题。
It will soon have ultracute animated Tamagotchi-style pets you can wear around your neck to improve your status as a Meta monitoring post. And we did say it was Meta, so has all the attention to security and careful design you may expect. Which is why it’s already been caught sending people’s physical addresses to strangers, and why when researcher Jonny at neuromatch.social went poking around its infrastructure — each instance runs permanently on its own Linux VM — they found just a ton of fun.
比如，那些看似亮眼的“派对戏法”其实是硬编码进去的；还有看起来像是利用Claude进行训练的痕迹。没错，你甚至可以在Meta内部深处的那个虚拟机上运行你自己的程序，正如Jonny所指出的，这简直是僵尸网络（botnet）滋生的温床。这个糟糕透顶的产品正被免费提供给Meta用户，而这些用户群体恰恰是那些可能并不完全可信的技术最容易攻击的目标。
Headline party tricks hard-coded. What looks like attempts to train off Claude. And yes, you can get it to run your own things on that VM deep within Meta, which Jonny points out is a botnet paradise waiting to happen. This banger of a bad ‘un is being given away to Meta users, who pretty much define a vulnerable target audience for tech that may not be totally trustworthy.
当然，Meta坚决反对AI安全监管。大家不妨押注一下接下来会出现什么样的头条新闻。无论如何，正如每场主题演讲中所演示的那样，使用智能体的优势要么琐碎平庸——比如“帮我订个餐厅！”——要么一旦失败就会造成灾难性后果——比如“帮我安排家庭度假！”——因此消费级智能体注定是场失败。只不过，Meta可能正在为这场火灾添加大量的助燃剂。
Meta, of course, is dead against AI safety regulation. Place your bets on the headlines to follow. In any case, the upside of using agents as demoed in every keynote is either so trivial and trite — book me a restaurant! — or so catastrophic if it fails — organize my family vacation! — that consumer agents are a bust. It’s just that Meta may be bringing a lot of accelerant to that particular fire.
当然，还有其他火灾，其中许多自一切开始以来就已为人所知。由于使用被AI输出污染的数据进行训练而导致的模型崩溃？这依然是个问题，而且目前还没有任何“更先进”的AI能提供解决方案。
There are other fires, of course, many known about since everything kicked off. Model collapse due to training on data contaminated with output? Still a thing, still no fix a "better" AI might provide.
而人工智能那种把一切都商品化、却无需付出任何代价的习惯，正在扼杀其自身的训练数据供应链，连带着摧毁庞大的商业与文化体系——这种情况至今仍然存在。不过，人工智能在特定领域任务上确实取得了持续进步。在科学、工程、数学和编程方面，只要将模型与足够多的强化学习相结合，并配备懂得如何管理它们的人类，就能取得令人震惊的成果。
And that habit AI has of retailing everything and paying for nothing, choking off its own training data supply chain and a huge chunk of commerce and culture with it? Very much still a thing. Where AI does make consistent progress is in domain-specific tasks. Science, engineering, mathematics, and coding, if you match a model with enough reinforcement learning with humans who know how to manage them, there are startling results.
然而，真正的创新和新意却并不多见，而这才是推动进步的核心动力。是否比人类更出色，很大程度上取决于具体情境。几十年来，计算机在围棋上的表现一直优于人类，但这并不意味着围棋本身得到了提升。马匹奔跑的速度比人类快，这一优势在过去四千年间一直很有用，但马匹的巅峰状态早已过去。因此，问题在于：我们是否已经抵达了人工智能的“峰值”——即未来的发展成本与“效果不佳”的风险相比，已不再值得投入？
You don’t get much innovation, not much newness, which is the true driver of progress. Better than humans is highly contextual. Computers have been better at chess for decades, but there isn’t a better chess. Horses run faster than humans, and that’s been useful for four millennia, but we got to peak horses quite some time ago. So the question is, have we reached peak AI, where the cost of future developments don’t match the risk they won’t work well enough to justify?
如果ChatGPT已经能在几秒钟内写出一封完全合格的商务信函，那么对普通用户而言，就再难有实质性的提升了。大多数人工智能应用都极其令人烦恼，这与模型本身无关，而纯粹是为了强行维持增长所设计的扭曲用户体验。这种做法既不可持续，也间接承认了人工智能尚未解决其实际应用场景的问题。如果我们已经处于或接近人工智能的“峰值”，那么这是否就是刺破那个该死泡沫的危机？
If ChatGPT already writes a perfectly good business letter in a couple of seconds, there is no meaningful improvement to be had for users. Most AI is profoundly annoying, in ways that have nothing to do with the models and everything to do with perverted UX in an attempt to justify growth. That is both unsustainable and a tacit admission that AI hasn’t solved its use case problem. If we are at or near peak AI, is that the crisis that bursts that goddam bubble?
至少，这肯定意味着一次重新调整，因为所有押注于持续上升曲线的预测都已被推翻。目前的人工智能已经足够实用，足以支撑起一个庞大的产业部门，但其实际用途与营销宣传以及人们对未来的幻想始终存在矛盾。尽管人们对“AI Max”的信任可能已经动摇，但仍有大量改进空间。如果我们可以摆脱“Metans”和“Muskovites”那种“吸烟有益健康”式的思维定式，那么所有人都在呼吁的放缓趋势或许会自然发生。这仍然是一场我们不应冒的赌局，但胜算可能比我们想象的要高。®
Certainly a readjustment, as all bets on the continuous uptick curve are off. AI as it stands is useful enough to power a big industry sector, just in ways that are and remain at odds to how it is marketed and its future fantasies. Trust in AI Max may be shaken, but there's still room for lots of improvement. That slowdown that everyone who isn’t a miscreant is calling for may happen organically, if we escape the smoking-is-good-for-you mentality of the Metans and the Muskovites. That’s still a gamble we shouldn't be taking, but the odds may be better than we thought. ®