OpenAI Rules the Changes But Meta Changes the Rules
OpenAI 主导变革,但 Meta 改变规则
An analysis on Meta’s master plan and OpenAI’s masterpiece
对 Meta 的总体规划和 OpenAI 杰作的分析
Meta has put the entire AI startup ecosystem against the ropes. They’ve released the two smaller versions of the Llama 3 family (8B and 70B-parameter dense models) and have given us a glimpse at the large version, a 405B dense model that although still training, is already showing GPT-4-level eval scores. Critically, Llama 3 is open access. A GPT-4-class model that’s available for anyone is, ironically, OpenAI’s primary threat as the leader of an industry that’s barely turning a profit and whose members’ entire business model is selling access to private models via API. It seems there was never a moat, but, if there was any, it surely wasn’t theirs.
Meta 已经让整个 AI 创业生态系统陷入困境。他们已经发布了 Llama 3 系列的两个较小版本(8B 和 70B 参数的密集模型),并让我们瞥见了大型版本,即尽管仍在训练中,但已经显示出 GPT-4 级别评估分数的 405B 密集模型。关键的是,Llama 3 是开放访问的。一个对任何人都可用的 GPT-4 级别模型,讽刺的是,正是 OpenAI 作为一个几乎没有盈利,其成员的整个商业模式是通过 API 销售对私有模型的访问权的行业领导者的主要威胁。看来从来没有过护城河,但是,如果有的话,那肯定不是他们的。
I focus this analysis on a “Meta vs OpenAI” dichotomy because, let’s be honest, most OpenAI-type startups, the likes of Mistral and Cohere, never stood a chance. It’s OpenAI who’s ruling the changes and deciding the direction. I acknowledge the framing is somewhat forced anyway because Google and Anthropic are worthy rivals. They’re however each playing their own game, compromised by their circumstances.
我将这次分析的焦点放在“Meta 与 OpenAI”的二元对立上,因为说实话,大多数类似 OpenAI 的初创公司,比如 Mistral 和 Cohere,从未有过太大的机会。主导变革和决定方向的是 OpenAI。我承认,这种框架设定有些生硬,因为 Google 和 Anthropic 都是值得一提的竞争对手。然而,他们各自都在玩自己的游戏,受到各自环境的制约。
Google funded the field of generative AI, which, in a dark twist of destiny, grew to become a dangerous threat to the ads-based search business it relies on. Gemini 1.5 Ultra could be a breakthrough but it’s undeniable Google’s hesitant to innovate recklessly, as if walking a tightrope. Anthropic is constrained by its identity as AI safety’s outpost. They’ve been careful not to advance capabilities beyond the frontier which is, by definition, the opposite of what leaders do.
谷歌曾经资助了生成式人工智能(generative AI)领域的发展,然而命运却让这个领域成为了它所依赖的基于广告的搜索业务的危险威胁。Gemini 1.5 Ultra 或许能带来突破,但无可否认,谷歌在创新上显得犹豫不决,就像走钢丝一样。Anthropic 由于其作为人工智能安全前哨的身份而受到限制。他们一直谨慎地不让能力超越前沿,这在定义上,与领导者的行为恰恰相反。
Neither Google nor Anthropic is incentivized to outplay OpenAI by changing the rules. But Meta is. That’s exactly what they’re doing (and were trying to do, even before LLaMA leaked “by mistake”). No one cared then because Mark Zuckerberg was embroiled in metaverse failures and continuous backlash against the twisted incentives ingrained in the essence of his social media platforms. The Llama 3 family—especially after Zuck’s public determination to make AI Meta’s priority—makes the world pay attention. Meta matters now and will use Llama 3 to redefine how the story unfolds—that’s the sheer power of a GPT-4-class open-access model.
谷歌和 Anthropic 并无动力通过改变规则来超越 OpenAI,但 Meta 却有。这正是他们正在做的事情(甚至在 LLaMA“误泄”之前就已尝试)。当时没人关心,因为马克·扎克伯格正深陷元宇宙失败和社交媒体平台固有扭曲激励机制引发的持续反弹之中。然而,Llama 3 家族——尤其是在扎克伯格公开决定将 AI 作为 Meta 的优先事项后——让世界开始关注。Meta 现在很重要,将利用 Llama 3 来重新定义故事的展开——这就是 GPT-4 级别开放访问模型的强大力量。
Meta’s open AI is what OpenAI should’ve been
Meta 的开放 AI 就是 OpenAI 本应成为的样子
Don’t interpret my praise of Meta’s efforts as an admission of OpenAI’s frailty. They’re threatened but that’s not the same thing as defeated. Far from it. Let’s recall that GPT-4, which is still above Llama 3 405B in most benchmarks, is one and a half years old. The most recent GPT-4 version is a month old, true, but let’s not compare an iterative update with an entirely new architecture that leverages the latest research on synthetic data and algorithmic improvements as well as the lessons from past mistakes.
请不要将我对 Meta 努力的赞扬误解为 OpenAI 的脆弱。他们确实面临威胁,但这并不等同于已经被击败。远非如此。让我们回想一下,GPT-4 在大多数基准测试中仍然超过 Llama 3 405B,而它已经有一年半的历史了。最新的 GPT-4 版本确实只有一个月的历史,但我们不能将一个迭代更新与一个全新的架构进行比较,后者利用了最新的合成数据和算法改进的研究,以及从过去的错误中吸取的教训。
I don’t think OpenAI is surprised by Llama 3 (or worried, for that matter). They surely knew the period during which their leadership had even a chance at existing was tiny, merely a function of the amount of leverage they could exploit from the GPT-3’s technical advances and ChatGPT’s popular interest. They did it successfully. However, once tech giants with two orders of magnitude more revenue caught up (and they did), OpenAI’s long-lasting prominence became a matter of mercy. That is, setting aside the absence of surprise or concern, the new annoying reality for OpenAI (even if Sam Altman wants to appear nonchalant, devoting his airtime to trolling the audience).
我并不认为 OpenAI 对于 Llama 3 的出现感到惊讶(或者担忧)。他们肯定知道他们的领导地位能够存在的时间非常短暂,仅仅是他们能够从 GPT-3 的技术进步和 ChatGPT 的公众关注度中获得的杠杆作用的函数。他们做得很成功。然而,一旦技术巨头们以两个数量级更多的收入赶上(他们确实做到了),OpenAI 的长期显赫地位就成了一种施舍。也就是说,抛开对惊讶或担忧的缺乏,这对 OpenAI 来说是一个新的烦人的现实(即使 Sam Altman 想要表现得无所谓,把他的发言时间用来戏弄观众)。
OpenAI had to prepare for Google’s attempts to surpass them in tandem with the DeepMind AI powerhouse. They did it. They had to recover from the great divide that led to Anthropic’s founding. They did it. Now they have to fight back against Meta’s attempts to leverage the unlimited power of open-source to commoditize the great software technology of our times—large AI models. A new day for OpenAI, a new monster to slay, except this time it’s all the monsters at once.
OpenAI 不得不为应对 Google 与 DeepMind AI 强大联手的超越之势做好准备。他们做到了。他们必须从导致 Anthropic 成立的巨大分歧中恢复过来。他们做到了。现在,他们必须反击 Meta 利用开源的无限力量,将我们时代伟大的软件技术——大型 AI 模型商品化的尝试。对于 OpenAI 来说,这是新的一天,有新的怪兽要去斩杀,只不过这次,他们要同时面对所有的怪兽。
Zuckerberg implied in the Dwarkesh Patel podcast that he doesn’t think AI models are going to be standalone products but advanced software infrastructure powering other things instead (in Meta’s case the social aspects of online content and connection, to use a generous description of what Meta does). As Zuck told Patel, “We have a long history of open-sourcing software; we don’t tend to open source our product [emphasis mine].” If Meta is open-sourcing its best AI stuff it’s because they don’t consider Llama 3 their product but a complement to be commoditized.
在 Dwarkesh Patel 的播客中,扎克伯格暗示他并不认为 AI 模型将成为独立的产品,而是会成为驱动其他事物的先进软件基础设施(在 Meta 的情况下,是在线内容和连接的社交方面,这是对 Meta 所做事情的慷慨描述)。正如扎克告诉 Patel,“我们有很长的开源软件历史;我们通常不会开源我们的产品。”如果 Meta 正在开源其最好的 AI 技术,那是因为他们并不认为 Llama 3 是他们的产品,而是一个可以商品化的补充。
I said OpenAI doesn’t seem worried. The reason, rumors have it, starts with GPT and ends with 5. But unless OpenAI enacts a miracle worthy of the promise of the scaling laws, it may not be enough. AI companies have proved building large models is a solved technological problem (a different question is the applicability problem), so Google, Meta, and Anthropic won’t sit around waiting for the next timely leapfrogging by OpenAI. There are only two soft moats left: money and talent. OpenAI isn’t leading in the former and may not be for long in the latter now that the novelty factor is over and Google and Meta have attractive offerings with Gemini and Llama. If GPT-5 turns out amazing, OpenAI’s apparent indifference to Meta’s moves is warranted, but so far it’s looking worse than ever for them.
我说 OpenAI 似乎并不担忧。据传闻,原因始于 GPT,终于 5。但是,除非 OpenAI 实现了符合规模定律承诺的奇迹,否则可能还不够。AI 公司已经证明,构建大型模型是一个已经解决的技术问题(另一个问题是适用性问题),因此 Google、Meta 和 Anthropic 不会坐等 OpenAI 的下一次及时超越。只剩下两个软壁垒:资金和人才。OpenAI 在前者上并未领先,而在后者上,由于新奇因素已经过去,Google 和 Meta 通过 Gemini 和 Llama 提供了有吸引力的选择,因此 OpenAI 可能也不会长期领先。如果 GPT-5 表现出色,OpenAI 对 Meta 的举动的明显漠不关心是有道理的,但到目前为止,他们的情况看起来比以往任何时候都要糟糕。
Meta is trying to suffocate any potential competition under the weight of their ambitions and, contrary to Google, they can do it cleanly—generative AI models are substitutes for search engines but they can enhance content creation and social media seamlessly. To that we should add that helping the open-source community is a strong ideological position. It doesn’t hurt either that these maneuvers are making Zuck look cool in contrast to all the others, especially Sundar Pichai who is, arguably, Zuckerberg’s closest competitor and, not-so-arguably, the uncoolest of the bunch (perhaps undeservedly for Google as a whole—but when is popular opinion fair and rational?).
Meta 正试图以其雄心壮志的重压来扼杀任何潜在的竞争,与 Google 不同,他们可以做得干净利落——生成式 AI 模型可以替代搜索引擎,但它们可以无缝增强内容创作和社交媒体。我们还应该补充的是,帮助开源社区是一种强烈的意识形态立场。这些策略也让扎克伯格看起来很酷,与其他人形成鲜明对比,尤其是他最接近的竞争对手桑达尔·皮查伊,他无疑是这群人中最不酷的(也许对整个 Google 来说这是不公平的——但是,大众舆论何时公正和理性呢?)。
It can always be the case that Zuckerberg’s bet is misplaced; that AI doesn’t end up being another software cog in a larger system but the entire system itself (that’s more akin to OpenAI’s position with all its grandiloquent discoursing about AGI, silicon gods, and post-scarcity utopia). Zuckerberg knows a materialization of this alternative future is possible and would be “trickier” for Meta:
总有可能出现这样的情况:扎克伯格的赌注下错了;AI 最终并非成为更大系统中的另一个软件齿轮,而是整个系统本身(这更接近 OpenAI 的立场,他们对 AGI、硅基神明和后稀缺乌托邦进行了华丽的论述)。扎克伯格知道这种替代未来的实现是可能的,对 Meta 来说会“更棘手”。
There is one world where maybe … the model ends up being more of the product itself. I think it’s a trickier economic calculation then, whether you open source that. You are commoditizing yourself then a lot. But from what I can see so far, it doesn’t seem like we’re in that zone.
也许存在一个世界……在那里,模型更多地成为了产品本身。我认为,这时候要进行更复杂的经济计算,比如是否要开源。那样的话,你就在大量地商品化自己。但是从我目前看到的情况来看,我们似乎还没有进入那个阶段。
In a way, he never thought we’d be in that zone at all so him believing we aren’t now isn’t newsworthy. Is his understanding of the subject limited? Does he have a plan in case his commoditizing of AI models doesn’t yield the expected value for Meta’s shareholders (killing OpenAI may not be a convincing rationale for those with millions at stake)? Or is he just playing the game under his own rules and predictions all the while he tries to smash anyone else who dares set up as standard a different set of beliefs and timelines? Even if his forecasts are mistaken, he may win anyway by forcing AI startups’ margins down to zero while he positively redirects AI-generated content to his social media business. If he executes, it’s truly a master plan.
在某种程度上,他从未想过我们会进入那个领域,所以他现在不相信我们在那里并不新鲜。他对这个主题的理解有限吗?如果他将人工智能模型商品化的计划没有为 Meta 的股东带来预期的价值(消灭 OpenAI 可能并不能说服那些有数百万利益在其中的人),他有备案吗?或者他只是在按照自己的规则和预测玩游戏,同时试图打击任何敢于设立不同信念和时间线标准的人?即使他的预测是错误的,他也可能通过将人工智能初创公司的利润率降至零,同时将人工智能生成的内容积极引导到他的社交媒体业务中,从而赢得胜利。如果他执行了,那真是一个高明的计划。
Some Google employee warned of this a while ago in terms so perceptive of what eventually happened that reading it again now feels eerily prescient:
早些时候,一位谷歌员工就曾以极具洞察力的言辞警告过这一情况,他的预见性如此之强,以至于现在再读起来,感觉如同预言一般诡异:
Paradoxically, the one clear winner in all of this is Meta. Because the leaked model was theirs [LLaMA], they have effectively garnered an entire planet’s worth of free labor. Since most open source innovation is happening on top of their architecture, there is nothing stopping them from directly incorporating it into their products. The value of owning the ecosystem cannot be overstated.
矛盾的是,这一切中的唯一明显赢家是 Meta。因为被泄露的模型是他们的 [LLaMA],他们实际上获得了相当于整个星球的免费劳动力。由于大部分开源创新都是在他们的架构之上进行的,没有任何东西阻止他们直接将其纳入他们的产品中。拥有生态系统的价值不可高估。
The masterpiece starts with GPT and ends with 5
这部杰作以 GPT 开始,以 5 结束
That is, at least, the official Meta-favoring narrative. The reasoning is strong and Meta’s motivations—lest someone believes Zuckerberg is a soul of charity or an open-source ideologue—are clearly defined (even if not to everyone’s liking). But because the mob moves by vibes and OpenAI’s are off, they don’t stop to consider what’s left for Altman’s people to defend themselves. Meta made its move, now it’s time for the others to act and react.
至少,这是官方倾向于赞同 Meta 的叙述。这种推理十分有力,而 Meta 的动机——除非有人相信扎克伯格是慈善家或开源理念的信她——也被明确定义(即使并非所有人都喜欢)。但是,由于人群受到情绪的驱动,而 OpenAI 的情绪并不好,他们没有停下来考虑 Altman 的人们如何自卫。Meta 已经采取了行动,现在是其他人行动和反应的时候了。
The first point in favor of API sellers (not just OpenAI) is that a super large dense model isn’t really what individual customers want. They’re delighted with simple access to the ChatGPT API and playground. What about those who pay $20/month? Wouldn’t they be willing to spend their money elsewhere, like a local, private, personalized model? Not that many people pay for ChatGPT but among those who do, how many could realistically download Llama 3 405B, fit it into the GPU memory of their servers, and then spend hundreds of dollars per month on inferences? If you can’t get the largest Llama 3 locally, it just doesn’t make sense to switch from your existing API provider.
支持 API 销售者(并非仅指 OpenAI)的第一个理由是,超大的密集模型并非个人客户真正想要的。他们对能简单地访问 ChatGPT API 和 playground 感到满意。那些每月支付 20 美元的人呢?他们是否愿意将钱花在其他地方,比如本地的、私人的、个性化的模型上?并非有很多人为 ChatGPT 付费,但在那些付费的人中,有多少人能真实地下载 Llama 3 405B,将其适应到他们服务器的 GPU 内存中,然后每月在推理上花费数百美元?如果你无法在本地获取最大的 Llama 3,那么从你现有的 API 提供商那里切换就没有意义。
Some customers who are sold into generative AI and can afford Llama 3 locally—both in terms of memory and inference—will surely leave OpenAI (or Google or Anthropic). OpenAI keeps doubling down on enterprise features but wealthy customers remain dubious of the safety, reliability, privacy, steerability, and applicability of generative AI models built by others. It’s too long a list of reservations to use a tech product whose value proposition remains unclear at best, which has yet to prove itself up to the task (and which, for some reason, never comes with a manual). Non-model-building labs like Cognition and big tech latecomers like Apple will be willing to bear the additional costs in exchange for better safety, privacy, and steerability (although Apple may have a different idea in mind). They’ll gladly take Meta’s gift unless OpenAI and the others can offer a qualitatively better private option.
一些对生成型人工智能(Generative AI)感兴趣并且能够在本地使用 Llama 3 的客户——无论是从内存还是推理能力来看——肯定会离开 OpenAI(或 Google 或 Anthropic)。OpenAI 不断加大对企业功能的投入,但富裕的客户仍对由他人构建的生成型人工智能模型的安全性、可靠性、隐私性、可控性和适用性持怀疑态度。这是一份过长的保留意见清单,用于一个其价值主张至多不明确,尚未证明自己能够胜任任务的技术产品(而且,出于某种原因,它从未附带说明书)。像 Cognition 这样的非模型构建实验室和像 Apple 这样的大型科技后来者愿意承担额外的成本,以换取更好的安全性、隐私性和可控性(尽管 Apple 可能有不同的想法)。除非 OpenAI 和其他公司能提供质量更好的私有选项,否则他们会很乐意接受 Meta 的礼物。
So it all comes down to one question: Can OpenAI offer a private AI model that’s good enough to both keep its customers satisfied without switching and convince them that their preferred vision—that AI is the main course, not a commodity to be pathetically subordinated to content creation and ad sales—is the one shaping this timeline?
所以,所有的问题都归结为一个:OpenAI 能否提供一个足够优秀的私有 AI 模型,既能让其客户满意而无需转换,又能说服他们相信他们所倾向的观点——AI 是主菜,而不是被可悲地降低到内容创作和广告销售的商品——是塑造这个时间线的主导因素?
What if GPT-5 is a masterpiece? What if it’s a feat of engineering as good as roon’s snarky comment on Meta’s “wasting all those beautiful H100s” suggests? What if it’s so powerful that OpenAI manages to hold the subsequent advantage for as long as it did from GPT-3 onward? What if once GPT-5 is out OpenAI chooses to make GPT-4 free to use? If they do, who in their right mind would choose to take the Llama 3 path? I bet some will but most people won’t.
如果 GPT-5 是一部杰作呢?如果它的工程技术水平如 Roon 在 Meta 上对“浪费所有那些美丽的 H100s”的讽刺评论所暗示的那样出色呢?如果它的能力如此强大,以至于 OpenAI 能够像从 GPT-3 开始那样长期保持后续优势呢?如果 GPT-5 一旦发布,OpenAI 选择让 GPT-4 免费使用呢?如果他们这么做,又有谁会选择走 Llama 3 的路线呢?我敢打赌,会有一些人选择,但大多数人不会。
We don’t know how any of this will unfold; Meta might be changing the rules but OpenAI keeps ruling the changes. And there’s still a game left to play.
我们无法预知这一切将如何展开;Meta 或许在改变规则,但 OpenAI 仍在主导变革。而且,游戏还未结束。
Subscribe to The Algorithmic Bridge
A blog about AI that's actually about people
Very very insightful. Thank you!
What do you make of the use case of consumers using ChatGPT plus-like features for free within Meta apps like Facebook and Instagram?
It’s awkward at the moment, but I could imagine Meta offering a good consumer experience that would rival ChatGPT plus, then monetize it through ads or data capture.
I would hate to use it, but I suspect most people that already mistakenly trust Zuck with their data would be happy to ditch $20 a month Chat GPT and go free with Meta