Meta 超级智能的未来：一周年进展更新

Meta 超级智能的未来：一周年进展更新 A top tier RL environment startup spawns out of thin air, the most aggressivecompute ramp we've ever seen, 2000km+ scale-across, and some advice for GoogleDeepMind 一家顶级强化学习环境初创公司横空出世，见证了我们有史以来最激进的算力扩张、超过2000 公里的跨尺度互联，并为 Google DeepMind 提供了一些建议 MAX KAN,JULIEN MARTIN-PRIN,JEREMIE ELIAHOU ONTIVEROS,ANDDYLAN PATEL MAX KAN、JULIEN MARTIN-PRIN、JEREMIE ELIAHOU ONTIVEROS 和 DYLAN PATEL JUL 10, 20262026 年 7 月 10 日·PAID·付费 It’s been a little over 1 year since the disastrous Llama 4 release spurred Zuck torebuild his entire AI org. Highlights include the shocking $14.3B Scale AI“investment” just to poach Alexandr Wang and the best people from his Safety,Evaluations, and Alignment Labs (SEAL) team, the multi-hundred million dollar(sometimes $1B+) pay packages offered to top AI researchers/engineers, and theexpedited compute ramp enabled by their new “Tent” datacenter design. For moredetails, see ouroriginal poston MSL. 距离灾难性的Llama 4发布促使扎克伯格重建其整个AI组织，已过去⼀年多。亮点包括：为挖⻆Alexandr Wang及其安全、评估与对⻬实验室（SEAL）团队的最优秀⼈才，令⼈震惊地“投资”143亿美元收购Scale AI；向顶级AI研究⼈员/⼯程师提供数亿美元（有时超过10亿美元）的薪酬包；以及凭借全新的"Tent"数据中⼼设计加速算⼒扩张。更多详情，请参阅我们最初在MSL发布的⽂章。 Since then, frontier AI has increasingly felt like a two horse race between OpenAI vsAnthropic. Google had a brief moment in the spotlight with Gemini 3 Pro and NanoBanana, but they’ve since faded dramatically. Despite their Windsurf acquisition,they’re far from a compelling agentic coding product, and 3.5 Flash is a benchmaxxedprop that performs far worse than GPT 5.5 and Opus 4.8 in real world scenarios (muchless Fable and 5.6). 3.5 Pro is not even Opus level on coding. Microsoft has completely blown their early lead with GitHub copilot and failed to effectively leverage theiraccess to OpenAI IP. SpaceXAI is selling $26B a year worth of GPUs toAnthropic/Google, and the Chinese labs are simply too compute poor to truly reachthe frontier. ⾃那时起，前沿⼈⼯智能领域越来越像是OpenAI与Anthropic之间的双雄争霸。⾕歌曾凭借Gemini 3 Pro和Nano Banana短暂成为焦点，但此后迅速黯然失⾊。尽管收购了Windsurf，他们距离打造⼀款真正具有吸引⼒的智能体编程产品仍相去甚远；⽽3.5 Flash只是⼀个在基准测试中表现优异、却在实际场景中远逊于GPT 5.5和Opus 4.8（更不⽤说Fable和5.6）的营销噱头。3.5 Pro在编程能⼒上甚⾄未达到Opus的⽔平。微软则完全浪费了其在GitHub Copilot上的早期领先优势，未能有效利⽤其获取OpenAI知识产权的渠道。SpaceXAI每年向Anthropic和⾕歌出售价值260亿美元的GPU，⽽中国实验室则因算⼒严重不⾜，难以真正触及前沿。 Meanwhile, MSL made their public debut this April with the launch of Muse Spark.You could argue this model represents a relative regression for Meta. Llama 3 70B and3.1 405B were both SOTA open-source on release, whereas Muse Spark, despite alsobeing closed source, lagged both DeepSeek v4 Pro and Kimi K2.6—open source modelsreleased around the same time—on most benchmarks. 与此同时，MSL于今年四⽉随着Muse Spark的发布正式亮相公众视野。有⼈或许会认为，该模型标志着Meta的⼀次相对退步。Llama 3 70B和3.1 405B在发布时均为最先进的开源模型；⽽Muse Spark尽管同样是闭源模型，却在⼤多数基准测试中落后于同期发布的开源模型DeepSeek v4 Pro和Kimi K2.6。 However, evaluating Muse Spark in isolation is missing the forest for the trees. Whatmatters for MSL is the slope, not the intercept. Rebuilding your entire team from theground up obviously comes with some short term setbacks, and it appears Meta hasfinally finished paying down this debt. Thus, the interesting question is not whereMSL is today, but trying to predict where they’ll be in the next 6 months. We think it'svery possible they are better than Google by then due to the team’s focus. 然⽽，孤⽴地评估Muse Spark⽆异于只⻅树⽊不⻅森林。对于MSL⽽⾔，关键在于斜率⽽⾮截距。从头开始重建整个团队显然会带来⼀些短期挫折，⽽Meta似乎终于还清了这笔债务。因此，有趣的问题不在于MSL今天处于何种位置，⽽在于预测他们未来六个⽉将达到何种⾼度。我们认为，凭借团队的专注，届时他们完全有可能超越⾕歌。 At the simplest level, there are three things you need to build a true frontier model:data,talent, andcompute. We believeMetais the only hyperscaler/neolab on track tobe world class at all three and thereforehas the best chance at catching up withAnthropic/OpenAI. We’ll explain why in full detail below, but as a teaser, here are the AI compute projections from our newTokenomics Model. 在最基础的层⾯上，构建真正的尖端模型需要三要素：数据、⼈才和算⼒。我们相信，Meta是唯⼀⼀家有望在这三个⽅⾯均达到世界⽔平的超⼤规模云⼚商/新型实验室，因此也最有可能追上Anthropic和OpenAI。下⽂将详细阐述原因，在此先剧透⼀下：这是我们新推出的代币经济模型所预测的AI算⼒数据。 Lastly, behind the paywall, we’ll discuss what this all means for Google—the companymost people today still believe rounds out the AI big 3. 最后，在付费内容中，我们将探讨这⼀切对⾕歌意味着什么——如今⼤多数⼈仍认为⾕歌构成了⼈⼯智能三⼤巨头之⼀。 Data is the new oil (for real this time) 数据是新的石油（这次是真的） We’ll start with data because it’s Meta’s newest advantage and probably the mostunderappreciated of the three. 我们将从数据开始，因为这是Meta最新的优势，也可能是这三者中最被低估的⼀个。 In 2024, Ilya famously said that “data is the fossil fuel of AI.” While this analogycorrectly highlights the importance of data for training AI models, it incorrectlyassumes that the amount of good data is finite. In reality, if demand is strong enough,market forces will find a way. 2024年，伊利亚（Ilya）有句名⾔：“数据是⼈⼯智能的化⽯燃料。”虽然这⼀类⽐正确地强调了数据对于训练AI模型的重要性，但它错误地假设优质数据的总量是有限的。事实上，只要需求⾜够强劲，市场⼒量⾃会找到出路。 This time, the invisible hand created a new human data/RL environment supply chain.The three incumbements—Mercor, Surge, and Handshake—are all at $1B+ ARR, andmany new entrants who are barely a year old (e.g. Fleet