一家成立两年半的公司,图像和视频的日产量超过了所有竞争对手加起来。
2026 年 2 月 10 日,xAI 全员大会开场,马斯克把这个数字放在最前面。他顺手给公司报了个年龄:两岁半,基本还是个幼儿。对手有的五岁,有的十岁,个别已经二十岁,团队规模大得多,起步资源也多得多。
然后在语音、图像、视频生成、预测这几条线上,这家幼儿公司拿了第一。
他给的解释只有一句:
任何科技公司,真正决定胜负的是速度和加速度,某一时刻站在哪个位置并不要紧。只要你在某个技术领域比所有人都跑得快,你就会成为领导者。(马斯克)
这句话可以当作整场大会的提纲。后面两个多小时里,每个团队负责人上台讲的都是同一个模板:我们从零开始,用了几个月,到了第一,接下来还要更快。
公司长大了,要长器官了
马斯克先讲组织调整。他用生物发育打比方:一开始是几十个人的创业公司,大家互相聊两句就把事办了;长到几百人,必须加结构,就像单细胞长成细胞团,再分化出器官和四肢,长出一条尾巴,尾巴最好还会消失,最后成为一个婴儿。
xAI 现在到了要分化器官的阶段。新结构分成四个应用方向:Grok 主模型和语音合并成一个团队;一个专攻编码的模型;图像和视频方向的 Imagine;以及 Macrohard,目标是完整模拟一整家公司。四个方向下面,是基础设施层。
他也提到了离开的人。早期阶段合适的人,到了后期不一定合适,他谢谢这些人把公司送到今天这个位置,也祝他们后面顺利。
这段话听着像客套,倒也是整场大会里少见的诚实时刻。一家两年半的公司开始做组织分层,说明它已经过了靠几十个人互相喊话就能推进的阶段。规模本身会逼着你把沟通路径画出来,把责任切清楚。
语音:六个月从零到超过对手
Grok 主模型和语音合并,负责人先讲了个对比。
2024 年 9 月,OpenAI 已经有了高级语音模式,用户可以跟模型自然对话。那时候 xAI 什么都没有,没有模型,没有产品,起步晚了一大截。团队里很多人此前根本没做过音频。
六个月,从零自研,产品超过了 OpenAI。再过六个月,Grok 装进了 200 多万辆特斯拉,还上线了 Grok 语音 agent API。一年时间,从零到领先。
这种事只可能发生在 xAI 这样的地方。小团队,目标集中,算力充足。(语音团队负责人)
产品侧的人给了个更大胆的说法:接下来几个月,知识工作者的产出量会提升十倍。目标是做一个门户,让用户在里面完成所有工作。法律问题、做幻灯片、解谜题,都能落地。产品的价值在于放大单个人的能力,让人做到独自做不到的事。
台上还顺带招人。原话是这地方不好混,是苦活,但野心是星际级的。已经进来的人补了一句:在 xAI 做事没有组织开销挡路,不用写一堆文档走流程,聪明人想做事就能做成。
马斯克给 Grok 主模型定的调子是“在广泛领域真正有用”。工程、法律、医学,任何职业拿来就能用,答案要让人信得过。这句“信得过”后面还会反复出现,它是整场大会里评估体系改革的主线。
编码:这一代模型在训练下一代
编码团队负责人 MRO 上台先认了个错。
以前别人劝他用编码模型,他不太信。最近变了,模型产出的代码质量够用,还是要 review,还是要给反馈,但加速效果已经很明显。他讲了一个更微妙的变化:现在描述问题,只需要像跟一个看过代码库的同事说话那样说一遍。以前得手把手教,像带幼儿。模型还能自己调试代码。
他们还让 Grok Code 连续跑几个小时,验证训练系统的复杂改动在生产环境里到底能不能工作。这种活以前要人盯着,现在交给模型跑。
然后是这场演讲里分量最重的一句:
当前这一代 Grok Code 正在训练下一代 Grok Code。我们看到了这条递归自我改进的路,指数起飞会继续。(MRO)
这也是他们把编码提到公司最高优先级的原因。手上有 100 万 H100 等效算力,用来训世界上最好的编码模型。他招两种人:擅长训练和建模的,或者擅长底层系统设计的。
另一位负责人讲得更直接:编码这条线上正在走向奇点,所以让公司最强的工程师来带队。他把限制因素归结为算力和能源,模型再好,得有地方跑。xAI 和 SpaceX 已经并成一队,算力这块要赢。他还对着写 kernel、写编译器的工程师喊话:先想想这件事还值不值得亲手做,不如加入编码团队,把自己先自动化掉。
马斯克接过这个话题,给了个时间表。到今年年底,可能没人再费劲写代码了。AI 直接生成二进制,而且比任何编译器生成的都高效。你只要说“为这个特定结果生成优化二进制”,传统编码这个中间步骤就被跳过了。他还说 Grok Code 两三个月内会达到 SOTA。
这个判断如果成立,受影响的不止程序员。编译器、构建工具、语言生态,一整条中间层都得重新找位置。过去几十年软件工程的抽象层级,可能要在一年内被压掉好几层。
Imagine:六个月前内部一行扩散模型代码都没有
图像和视频团队的故事和语音几乎一样。
六个月前从零开始,两周前发布 Imagine v1,登顶多个排行榜,用户反馈很好。这个月和下个月还有一堆发布。负责人开了个玩笑:我们很可能比 Meta 更早建成元宇宙。
产品侧的人补了组数字。六个月前,公司内部完全没有扩散模型代码。现在 Imagine 上了所有产品界面,包括 X。打开 X 应用,长按任意图片就能编辑,能把图片直接变成视频。他们办了个创作比赛,投稿很搞笑,传播很广。
迭代节奏是每天多次产品更新,每两周一次模型更新。结果是用户每天生成接近 5000 万条视频。按他们的说法,这个数字超过其他所有提供商加起来。
图像这边,30 天生成了 60 亿张。谷歌前阵子公布 nano banana 30 天生成 10 亿张,xAI 是它的六倍。
我们不只是想赢一次,我们想要长期持续的领先。(Imagine 产品负责人)
他们的目标是把人能想象的东西变成现实,而且要像速通游戏一样把这件事一路打穿。
视频方向的负责人讲得更远。模型能力继续扩,视觉世界会和现实难以区分。到今年年底,模型可以一次生成 10 到 20 分钟的视频,不需要任何人工干预。你给一个想象,agent 帮你做完。
再往后是实时。他们已经是最快的视频生成方,接下来要把渲染推到实时,用户可以进入自己想象的世界,那个世界实时回应你。
马斯克在这里插了一个判断:大部分 AI 算力最终会花在实时视频理解和实时视频生成上。他还特意强调,六个月前 xAI 在视频图像生成和编辑上基本是零,六个月做到第一,产量超过所有人总和。同样的事要在编码上做一遍,在 Macrohard 上再做一遍。他还提了一句 Grok 4.2 快发布了,会让人印象深刻,而那只是新模型的小号,后面还有中号和大号。
Macrohard:给 AI 配一台电脑
Macrohard 是这场大会里最容易被当成玩笑的项目,名字本身就带着调侃。负责人上台第一句就是:这是所有产品名里最严肃的一个。
他的类比是《盗梦空间》。给人类配电脑是个好主意,现在给 AI 配电脑。Macrohard 要做的是一个完全可用的、实时的数字人类模拟器,人类在电脑上能做的事它都能做,包括工程和医学里的高级工具。按他的说法,将来火箭发动机可以完全由 AI 设计。
他还点出一个事实:这几乎是 AI 明显不如人类的最后几个领域之一,所以最值得投入。
另一位负责人讲了为什么这件事必须做。全世界 80% 到 95% 的软件都有图形界面。要让人们的生活真正变轻松,模型必须能处理 GUI 上的日常任务。Macrohard 模拟的是一家公司,产出是数字的。这是 agent 的下一个台阶:跨桌面的端到端编排,会带来巨大的经济繁荣。
他招人只看三条:聪明,能解决难题;有驱动力,想赢;是个好人,你愿意跟他一起工作。他还给了个很实在的建议:想不出人选就翻通讯录、翻 LinkedIn,你会惊讶自己的关系网有多大。
马斯克给这个项目的定位更重。他说 Macrohard 长期看可能是公司最重要的项目。理由很干脆:世界上最有价值的公司,产出都是数字的,它们不造硬件。任何产出为数字的公司,理论上都能被完整模拟。
这会带来一个我们目前几乎无法想象的繁荣时代。你得用 Imagine 才能想象出来。(马斯克)
所以 Macrohard 这几个字刷在训练集群的楼顶上。他补了一句,那本来是个玩笑。
评估标准换了:唯一算数的是领域专家
数据与评估团队负责人 Diego 讲了一个正在发生的转变。
明年这些模型的主要瓶颈之一,是高质量评估和训练数据。他们的办法是把各领域世界顶级专家请进来评估模型。医学、金融、法律,还有配音演员、视频编辑,这些人每天在让 Grok 变好。接下来几个月,金融和法律领域的有用任务会有新东西发布。
他强调的重点是:要做对真实工作有用的评估和训练数据,别去做智力的代理指标。很多开源评估做的恰恰是后者。
马斯克把话说得更绝。xAI 正在从互联网通用榜单转向各领域专家导师。那些公开榜单衡量不出真实有用性。评估的标准只有一条:该领域的人类专家是否同意 Grok 极其有用,结果是否正确。
那是唯一真正重要的评估。(马斯克)
他提到 Grok 4.20 里能看到这些改进:更求真,政治偏见更少,回答更连贯。
同一场里还讲了 Grokipedia。目标是把人类全部知识蒸馏出来,做成现代版的亚历山大图书馆,最终叫银河百科全书。从几乎什么都没有,做到约 600 万条目。作为参照,维基百科英文版大约 700 万条。幻觉率在改善。他们的目标写得很大:Grok 5 不用出数据中心去搜索。
这句话的分量在于,它把“知识”从检索问题变成了记忆问题。模型的参数里得装下足够多可信的内容,搜索退化成兜底手段。
十万张 H100 到货那天,软件没准备好
基础设施这部分,是整场大会里技术密度最高的段落。
ML 基础设施团队讲了 Grok 3 的故事。他们负责建预训练框架,10 万张 H100 到货,本以为软件已经准备好了。跑到 3 万张规模,发现软件不行,几乎重写了一半。
原因听起来很具体:数据中心里有太多你无法预估的事。交换机在抖动,链路在抖动,交换机宕机,GPU 烧坏,数值问题随时冒出来。而你要让 10 万张 H100 在一个训练步里保持同步。一个训练步五秒,这五秒里什么都可能发生。系统必须在这样的环境里继续推进。
他们做成了。那段时间他儿子出生,他说那是人生里最酷的时刻之一。
当时整个预训练团队 15 个人,其中 7 个人做训练系统。这种人才密度保持到了今天。他招人的话也很直接:你不想成为一千人里拧同一颗螺丝的那个。
RL 和推理团队讲的是规模。他们在地球上大规模跑强化学习训练和生产推理,很快会跑到太空。系统要从 10 万芯片扩到数百万芯片。并行、prefill、decode,每一层都要优化,还要对已知和未知的硬件故障都有韧性。喜欢极限性能和可靠性的系统黑客,这里是主场。
工具团队负责人类数据平台和内部工程平台。部署、评估、查看训练结果,都在这里跑。研究员、agent、领域导师都要用这套界面。他们把这套东西叫做人与 agent 共用的接口层。
JAX 团队只有几个工程师,负责优化 JAX GPU 上的超大规模训练,支撑全公司的预训练、强化学习和多模态。从 1 万张到 10 万张,再到 100 万 H100 等效。他们得定制整条 JAX 栈,从编译器到运行时。负责人说了句大实话:在规模上跑一个 hello world 都很复杂。
Kernel 团队在最底层。他们的代码跑在百万等效 GPU 里。一张 GPU 内有几十万个线程,线程之间要通信,做矩阵乘法、算注意力分数,有些线程还要跟另外上百万张 GPU 说话。他们优化到每一个微秒,要把每张卡最后一滴性能榨出来。
孟菲斯:六周建一个数据厅
大会现场连线了孟菲斯的超级计算机。
驻场工程师 Hiner 介绍,这里是全球最大的集群,而且还在长。几十万套操作系统、CPU、网卡、交换机,要像一台超级计算机那样运行。他们需要真正懂计算机底层的人。现场环境嘈杂,他两次被要求把麦克风贴近嘴边。
Dan 给的数字更直观:今天这里有 30 万块 GB300 平台 GPU,每个数据厅 847 英里光纤,一共 12 个数据厅。
马斯克补了整体进度。不到一年建成这样,全部完成后会有超过 1 吉瓦的电力在线。这里还会有全球最大的特斯拉 Megapack 系统,规模超过夏威夷和南澳大利亚的储能项目。
负责施工的 Zach 站在 11 号数据厅前讲速度:每个数据厅 850 英里光纤、2.7 万多块 GPU、20 多万个连接,全部在不到六周内建起来。这套流程一遍一遍复制,高度并行。他把这件事称为你能想象的最复杂、也最一致的工程设计施工项目。
垂直整合是他们的另一个特点。架构、机械、电气、结构,所有专业都在同一个团队里。设计时还要追求行业最高能效,用电力平滑技术削峰,同时做好孟菲斯本地的社区伙伴。
马斯克接着放了两张图。第一阶段 33 万块 Grace Blackwell,楼顶写着 macro,他特意说明那几个字确实刷在建筑上,没有经过任何图像处理。第二阶段是 22 万块 GB300,楼顶画着火箭。所有这些算力,用来训用户能直接体验到的模型。
他引了黄仁勋在采访里说过多次的话:上线 AI 算力,没人比 xAI 快。
然后他讲了那个挖洞的梗。漫画里一个人挖洞,七个人在旁边看。xAI 是那个真正在挖的人。
X:安装量过十亿,日活目标也是十亿
X 的负责人 Nikita 先给了组数据。整个产品矩阵触达超过 10 亿人。每次大事发生,都能看出 X 是这个时代最重要的通信工具,最有影响力的人在这里聚集,真相在这里结晶。按他的说法,一切都下游于 X。有人说某条消息一周后会出现在 Facebook 上,因为那件事先在 X 上发生。
1 月是这款应用历史上参与度最高的月份,2 月的数据会更高。首次下载每月增长超过 50%,增速像早期消费产品。他们还解决了一个拖了二十年的老问题:新用户引导。现在新用户每天在应用里的停留时间,比六个月前多 55%。
算法重建了,引导流程重建了,通知重建了,网页浏览器重建了,XChat 也重建了。文章功能的推送做了一轮,发布量涨了 10 倍,阅读量涨了 17 倍。假期推了一轮订阅,跨过 10 亿美元 ARR。
他的结论是:赢的路径很清楚,球在自己半场,剩下的是执行。
马斯克接手讲通信和支付。旧 Twitter 的私信栈未加密,只能发文本。现在换成全加密,支持音视频通话,有消失消息和截屏拦截。这套代码几个月内开源。推荐算法也开源,xAI 会成为唯一开源推荐算法的公司。
没有什么比透明更能让人相信一家公司。(马斯克)
Grok Chat 同样开源,用户可以自己检查有没有漏洞。他强调里面没有广告钩子,目标是做一个通用通信系统。几个月内会推出独立的 X Chat 应用,带桌面共享和多用户视频通话,不需要打开主应用。
X Money 已经在公司内部封闭测试。一两个月内进有限外部测试,然后向全球 X 用户开放。目标是把 X 做成所有货币交易的中心。
关于用户数,他解释了一个常被误读的差别。X 的月活平均约 6 亿,但装了 X 应用的人远超过 10 亿。大多数人只在重大事件发生时打开。当 X 能提供更多日常理由,通信、Grok、X Money,用户就可以把生活放在这个应用里。他的预期是日活超过 10 亿。
地球只用了自身潜在能量的百分之一
大会最后一段,马斯克讲太空。
要理解宇宙,必须探索宇宙。在地球上用望远镜和对撞机,能学到的东西有限。SpaceX 和 xAI 结合,动机是加速人类理解宇宙的进程,把意识之光延伸到恒星。
他给了几个数字做参照。地球文明现在只用了地球潜在能量的约 1%。想用到太阳百万分之一的能量,大约是当前文明用量的百万倍。太阳占太阳系 99.8% 的质量。地球是广袤黑暗里的一粒微尘。要动用太阳的能量,就必须走出地球。
第一步是地球轨道数据中心,由 SpaceX 每年发射 100 到 200 吉瓦级别的算力。他强调这个数字按年计算,累计规模另算。最终会看到每年发射 1 太瓦算力的路径。
再往上,要去月球。在月球上建工厂造 AI 卫星,建质量驱动器,也就是科幻小说里的电磁发射装置。这件事他们要真的做出来。那样可以把规模做到每年数千吉瓦甚至更多,最终摸到太阳能量的百万分之一、千分之一,也许几个百分点。
很难想象那种规模的智能会思考什么,但看着它发生会非常激动。(马斯克)
他想看到的画面很具体:月球上的质量驱动器把 AI 卫星一个接一个射进深空;月球上有一座自给自足的城市;然后去火星,走遍太阳系,最后走向星际,访问一个又一个恒星系。
也许会碰到外星人。也许会碰到存续了几百万年的文明,找到它们留下的遗迹。
但只有走出去,这些才可能发生。
这场大会真正在说什么
把两个多小时的内容摊开看,有几条线是反复交叉的。
四条产品线对应四种产出:通用助手、代码、视觉内容、数字劳动。每一块都配了一个“从零到第一用了几个月”的故事。这些故事真假需要时间验证,但它们共同说明一件事,xAI 把组织切成产品线,每条线配一个能直接对结果负责的人。
算力是唯一贯穿全场的硬约束。100 万 H100 等效、1 吉瓦在线电力、33 万块 Grace Blackwell、22 万块 GB300、每年 100 到 200 吉瓦的轨道数据中心。产品团队讲模型能力,基础设施团队讲怎么把这些芯片在一个训练步里同步起来。两条线说的其实是一件事:没有电和卡,所有模型路线都是纸面规划。
评估体系在换轨。从公开榜单换到领域专家,从智力代理指标换到真实工作产出。这个转变比任何模型版本号都重要,它决定了模型往哪个方向变强。榜单可以被优化,专家不会那么容易被糊弄。
开源被当成一种策略。推荐算法、Grok Chat,都用透明换信任。对用户来说,这是可以验证的承诺;对竞争对手来说,这是把标准公开。
太空那条线同样有具体的数字和节点。它解释了为什么算力要往轨道上放:地球上能拿到的电力有上限,太阳的能量没有。把数据中心送上轨道,再把工厂搬上月球,是把能源约束彻底解开的路径。
至于这场大会里最值得记住的判断,可能是马斯克开头那句关于速度和加速度的话。一家两岁半的公司在多个领域拿到第一,靠的是斜率,体量还排在后面。斜率能维持多久,才是接下来两年真正要盯的东西。
- 判断一家 AI 公司,先看它上线算力的速度和模型迭代的节奏,再看它现在排第几。榜单是快照,斜率才是趋势。
- 如果你的工作产出是纯数字的,文档、代码、报表、设计稿,今年就该认真算一遍:agent 能接手其中多少比例。Macrohard 这类项目的目标就是这个。
- 练习把问题描述成“跟看过代码库的同事说一遍”那样。模型能力的瓶颈,越来越多地落在人的表达精度上。
- 关注 GUI 自动化这条线。全世界八成以上的软件只有图形界面,没有 API,这块是 agent 真正落地的地方,也是接下来一年产品机会最密集的地方。
- 内容团队可以开始重做素材生产流程。每天 5000 万条视频、30 天 60 亿张图像的产能已经摆在那里,一次生成 10 到 20 分钟长视频的能力也在路上。
- 自己建一套内部评估,比刷公开榜单有用。找你们领域里最懂行的人,判断模型输出是否真的可用、是否正确。
- 选平台时把开源当权重。推荐算法和通信协议公开,意味着你可以验证它做了什么,而不是只听它说。
- 盯三条硬件曲线:单集群电力规模、芯片交付节奏、轨道发射成本。它们的斜率,比任何一次模型发布都更能预判三年后的格局。
0:00Welcome to the [applause] XEI all hands. We've got a very exciting presentation for you. We're going to start off by recapping the incredible progress that the XEI team has made in just two and a half years. it's really remarkable in pursuit of our goal of understanding the universe. So just going over our accomplishments since inception. it's important for to bear in mind that XI is only two and a half years old, basically a toddler. and we've nonetheless achieved an incredible amount in a very short period of time. So our competitors are five 10 some cases 20 years old.
0:42they have much larger teams. They started off with far more resources and yet nonetheless we have achieved number one in many arenas in in just a few years. So we we've achieved number one in in voice in image and video generation. I think we now at this point are actually generating more images and video based on the last numbers I saw than all of our competitors combined. we are winning in terms of forecasting which is one of the key metrics of intelligence. so our the Grock 420 forecasting model beat all the other AIs in forecasting.
1:23and we've talked many leaderboards. we've got now a great app with the with with Imagine with the Core Gro. We've made radical improvements to the X app and we've launched Graipedia which is on its way to far exceeding Wikipedia and ultimately be orders of magnitude are more comprehensive and more accurate and have more information as well as video and and and image data that simply isn't there on Wikipedia. So it's it's intended ultimately to be encyclopedia galactica a distillation of all knowledge of yeah all knowledge.
2:08and we we we're we're the first to achieve 100,000 H100 GPU training cluster and we're now about to achieve the first 100 I should say 1 million H100 GPU equivalents in training. So really an incredible amount of work in a very short period of time. And it's it's important to consider for competitiveness of any technology company, what matters is not the position at any point in time, but what is your velocity and acceleration and if you're moving faster than anyone else in any given technology arena, you will be the leader. And XAI is moving faster than any other company. No one's even close. So, let's go to our team. as
3:01we grow as a company, a natural thing that happens is you reorganize the company as it scales up. So, when you first have a startup, you might have just a few dozen people and they all just chat amongst themselves. As you grow to several hundred people, you have to then add more structure just like an organism that grows from a single like we all just grew from a single cell and then to a blob of cells and then you get or organ differentiation limbs you grow a tail hopefully the tail disappears and then you become a baby. you go through these stages and so we're we're or organizing because we've we've reached a certain scale we're organizing the company to be
3:46more effective at this scale. now naturally when when this happens there there's some people who are better suited for the early stages of a company and and less suited for the later stages and so and and for for the people that that have departed I'd just like to say thank you for contribution thank you for getting us this far and we wish you very well in your future endeavors. So now going on to the new
4:13structure of the company. the company's organized in four main application areas. There's there's Grock main and voice which is really the the the main Grock model. That's why it's called Grock main. then there's a coding specific model. There's an image and video model which is imagine and then macro hard which is intended to do full digital emulation of entire companies.
4:39and and then we've got the infrastructure layers. So I'd like to invite members of the team to come up and talk about each of their areas.
4:55Hey, thanks Elon. So Grock main and voice are going to be merged into one team and you know on voice one anecdote is September 2024 open had this product you could talk to advanced voice mode and we had nothing no model of course no product we started much after that and in a span of few months 6 months we developed the model in-house from scratch without a bunch of people who knew audio and had a product that was surpassing open end 6 months fast forward six more months and now we have Grock in more than 2 million Teslas. We have a Grock voice agent API. You can do all kinds of amazing things with in a span of one year. We went from nothing
5:34to being leaders. That kind of stuff is only possible in a place like XCI. We have small teams committed, mission focused, lots of compute. and we really really want to keep pushing. Same story on the chat models. You know, we've always been at the forefront of reasoning starting from Grock 1.5, Grock 2, Grock 3. And we want to really move to a world where it's no longer about just question answering. We want to build an everything app. So you should be able to come to it and really get done whatever you want. You know, ask a legal question, make a slide deck or or you know, solve a puzzle, stuff like that.
6:09Yeah. So I really think on the product side, we're really going to see a huge transformation happening in a very short period of time. we're going to see work the magnitude of amount of work that all knowledge workers are going to be able to produce increase tfold in the next short period of a few months.
6:26the models that we are building out are incredibly amazing and we have a lot on the way and we're really excited to share that with with you all. and on a product side, the goal is to just build that portal that allows you to accomplish all of your work. And how do we amplify everyone to achieve much much more than what they can accomplish alone? And we're building that out and it's it's going to be an incredibly easy to use experience that just works seamlessly.
6:53that being said, we are hiring and we're looking for intelligent and smart people. This is not an easy place to work, guys. like this is
7:00it's a grind but we have I guess like interstellar ambitions so it's it's not going to be easy right so I will say having come to XAI it has been an opportunity of a lifetime to work among really smart and really passionate people the vibes here are amazing and it's truly an environment where if you're a smart person you want to get done you can get done there isn't like organizational overhead getting your way or kind of I don't know like having to write docs and all this kind of stuff you just do stuff at least at least for me I just you know just you can do things here and that's amazing and I invite more people to come here and just do awesome things.
7:43Yeah. So with with the the Grock main the the sort of main foundation model the intent is that it's it's genuinely useful in a wide range of areas. So if you're doing engineering or law or trying or or medicine any anything it is useful to you in in in your job that's essential to understanding the universe and and making things as useful as possible like where when Grock gives you an answer that you can count on it. Yep. Absolutely. All right. All right. Thank you.
8:15Yeah. Thanks. Hey everybody, I'm MRO. So the world changed a lot recently in terms of coding. the coding models I was always complaining people were trying to convince me to use a coding model and I was like trusting it and I wasn't really convinced but as of recently the models they they actually produce good decent quality code. I mean you still need to review and give feedback but you can it's it's easy to see how they can accelerate you quite a lot. So it's not only about coding it's like they understand your intuition like much better than before like now when you when I describe a problem I only have to phrase it like I would to an other
9:04colleague engineer who has already seen the codebase. That's a huge change. before you kind of need to handhold a toddler to make a change and they don't only write your code but they also can debug your code. So now we have I do like well we do like hours of grog code running continuously to make sure that a more complex change to the training system actually works in production. So it's easy to see for us that this is not only about accelerating us ourselves writing code and making us 10x more productive but we're really on this path for recursive self-improvement where the current generation of gro code is training the next generation of croc
9:44code and we see that this path we exponential takeoff here this path will continue so we are doubling down on coding and making coding one of the highest priority efforts in the company. So if you're out there and you're excited about coding and you're either either very good at training modeling or you're a really good low-level software engineer interesting in systems design, this is the place to work. Like we have a million H100 equivalents to train
10:13the best coding metal in the world right now. So please join us.
10:18yeah, I'm God. I work pair with macro on coding. So it become more and more obvious to us like you know over time like we are on a path to singularity at least on coding. So we decided like you know have our best engineer in the company micro to lead the coding and we'll build the best coding model for everyone to empower everyone to build and for me like the main like limiting factor is probably computer and energy where they can run the best model to support everyone to empower everyone and with spec now we are one team and we will win on the compute and we are winning with space compute and also like for every engineer Right? So if you are like
10:59writing kernel, if you're writing compiler, just think about like whether it's still worth it. Maybe you should join us, you know, for coding effort to automate yourself a little bit like to speed it yourself up. yeah, I think it's like really amazing year. basically what a year to to be alive and I can already feel the AGI feel the AI at least for coding. Yeah.
11:22Yeah. I I think actually things will move maybe even by the end of this year to where you don't even bother do doing coding the AI just creates the binary directly and the AI can create a much more efficient binary than can be done by any compiler. So just say create optimized binary for this particular outcome and and you actually bypass even traditional coding. there's there's no that that's an intermediate step that actually will not be needed probably by I'd say the end of this year. and we do expect Gro code to be state-of-the-art in 2 to 3 months. So it's happening very quickly. Yeah.
12:16Also do imaging. So you know I mean what do we all do right after post AJI right you probably do like digital life so that's what we are doing here as well and we have the imagine team like started pretty much from scratch like six months ago we have a few people we decided we'll to do the imaging we'll do the video gen like yeah look at what we achieved today like you know like two weeks ago we released like imagine v1 we actually top the leaderboard across like many of them and people really loves our product love our model And we have many more releases actually this month and next months. So yeah, to me there's like
12:54a really high chance like we actually may build a metaverse before meta. yeah, I will also pass to try to to talk about like you know the the metrics we have the product. Yeah.
13:06Yeah. Like like Gorang said it's only been 6 months since we started working on imagine. We had we had no code internally for diffusion at all 6 months ago and basically now we've launched imagine on every product surface that we have including seamlessly integrating into X. So you can open the X app right now you can long press on any image you can edit the image you can make a video out of the image. we also ran a contest recently where we had some really funny submissions that I'm sure many of you have seen. so imagine is growing extremely extremely fast. and it's because of the speed at which we iterate. Basically we do multiple product updates every day. We do model
13:40updates every other week and effectively what this has led to is now users are generating close to 50 million videos every day using imagine. and just to reiterate what Elon said earlier that to the best of our knowledge that is more than every other provider combined which again is an astonishing place to be compared to where we were 6 months ago.
13:58we are also generating 6 billion images in the last 30 days. nano banana you know Google recently posted that you know 1 billion images were generated using nano banana in 30 days so you know we're six times that right and and really the goal is it's not like we we don't just want to win we want to win like s like over a long period of time and have sustained greatness and so the goal with imagine is to take anything that you can you know imagine and turn it into reality and so that's that's what we're going to you know that we're going to speedrun that basically is the goal yeah
14:31hey I'm Hen. we as we keep scaling our model capabilities, building visual worlds that's indistinguishable from reality, we're also building system that unlocks much more possibility than what we have right now. they will be able to generate the videos that's much longer than what we have right now with stories or with souls of your imagine. And by the end of the year, we likely will be having models that allow you to generate videos of 10 minutes or 20 minutes in one shot without any intervention. You just need to give your imagination and our model, our agents will do it for you. And and moreover those are the videos we generate and we're also going to allow
15:20rendering those. We we're already the fastest in generating the videos and we're going to keep pushing the extreme where we're going to render those videos in real time and you will be able to imagine build and interact with your own world and the world the world will respond to you in real time and it is exciting future that we are going to build with ourself.
15:43Absolutely. I my prediction is that most of AI compute is going to be real- time video understanding and real-time video generation and and we expect to be the leaders in that. It's it's worth emphasizing these points that you know 6 months ago we didn't even have we we had basically nothing in in in or very weak in video and image generation and editing and went in six months to number one spot. and in fact generating more more videos and images than everyone else combined.
16:16We're going to do the same thing with coding and we're going to do the same thing with macro hard. And and I think people will be be pretty impressed with the the Gro 4.2 model that's coming out. that's it's a it's a significant improvement. and that's really just that that's that's the small version of our new model. So, we'll
16:37have a medium and a large version that are even more intelligent. All right. Hi everyone. I'm Toby and I work on Macro, the most serious of all product names. so arguably giving computers to humans was a good idea. So, we're doing the same thing thing for AI. It's kind of like Inception. We're giving computers to computers. so MacroArt is building a fully capable digital realtime, very important human emulator. So it's able to do anything on a computer that a human is able to do, including using advanced tools in engineering and medicine. So there should be rocket engines fully designed by AI. And in a sense, it's one of the last few remaining areas where AI is
17:32significantly worse than humans, which is why I think it's one of the most exciting areas to actually innovate in and actually change the change the field.
17:42Hi everyone. So yeah, my name is John and yeah, so we're building these strong reasoning models which are now going to control our CLI. Like we're actively using these every day. They are like tremendous like productivity boost to the whole team. I know the voice team is like killing it on that and you know this is the reason why we need the compute you know we need the large scale compute to run these models to boost our own productivity but you know 80 to 90 95% of the world world software has a GUI so
18:15that's like you know great representation and you know to truly make people's lives easier we need to develop models that are capable of solving day-to-day tasks on GUI So macro hard you know we will emulate a company where the output is digital and so this is the obvious next step for agents macro hard will enable true endto-end orchestration across the desktop and it will lead to immense economic prosperity.
18:44so yeah we're entering an era where we need to tackle the hardest of tech problems but in order to solve this we need to hire the best people. So, you know, think of the smartest people that you've worked with and and put them forward for for a position here. And if you can't think of anybody, like go through your phone book, go for your LinkedIn. you'll be surprised like how how big your actual network is. And they just need three properties obviously that we want to optimize for.
19:16Are they clever? Can they solve hard problems? And the second property is are they driven? Do they have the ambition? Do they want to win? And the third is are they a nice person? Like do you want to actually work with them? but yeah, so thank you.
19:34Yeah, the the macro hard project is over time actually will probably be our most important project because what we're talking about is u emulation of entire human companies. So when you look at the most valuable companies in the world, they are their output is digital. so they they
19:56don't actually make hardware. So it should be possible to completely emulate any company that where the output is digital. and this will usher in an age of prosperity likes which we could barely imagine at this point. You need imagine to imagine it. so this is a big this is a big deal and this is why the words macro hard are painted on the roof of the training cluster.
20:22because that's what it's going to build.
20:26So, it's also pretty funny.
20:27Yeah, [laughter] meant to be a joke.
20:37It's me again. You might remember me from macro hard and computer use from a long time ago, but I also actually work on core product infrastructure and API. In fact, this is what I've done for most time at XAI. So, anytime you use any of our products like grock.com API authentication, you go to status.x.ai.
20:57This is done by the core product infra team. And a large portion of them actually sit in London and we work with Haime over there. So, we keep the lights on at peak hour, 400 p.m. every day. we get paged at night when stuff goes down. Also, thank you to anyone in Palo Alto getting paged. there's really important work, reliability, security, core product infrastructure. So, if you actually like if you're really interested in solving difficult distributed problems with like messy data, this is the team to join.
21:35hey everyone, my name is Diego. yeah, so I think one one of the main bottlenecks in this next year for these models is going to be very high quality evals and training data. And one of the ways we solve that is by taking the world's foremost experts in these respective domains u bringing them here and having them evaluate the model. we do this for domains like medicine, finance, law, we have voice actors, we have video editors who contribute daily to making rock better. and yeah, we're going to be continuing to work on very high quality evals over the next few months.
22:05we have some exciting stuff in you know, the the frontier of useful tasks in finance and law. you know, we're trying to build evals that are are are useful and training data that that represents useful work and and not necessarily proxies of intelligence a lot of the open source eval do today. yeah.
22:26Yeah. I'd like to say like we're we're shifting from using these sort of common internet evals which I think are actually not a real indicator of usefulness to having expert tutors in each domain. so every domain of engineering, medicine, law, whatever the case may be. and the the the actual eval is does the expert in that arena or does our group of experts in that arena human experts agree that Grock is extremely useful and that the results are correct? That's the that's actually the only eval that really matters.
23:04Yeah, exactly. in in you you you'll see this in in G420. but we made some improvements because of that type of data in in truth seeeking and and kind of minimizing political bias. The responses are are much more cogent. so yeah, that that's exciting. and we are also working on Gracipedia. So the, the goal of Rockedia is to create a distillation of all human knowledge. I kind of like to think of this as like a modern day version of the Library of Alexandria. and in the quest to build Encyclopedia Galactica, which it will one day be called. we've gone from essentially having nothing to around 6 million u, articles. for context, Wikipedia is around 7 million English articles.
23:43and, yeah, we're improving on hallucination. and our our our goal is essentially for GO 5 to not have to search out of the data center. so yeah. [applause] I guess so in the ML infra team we are building the training inference and tooling team tooling software for the company. So to give you an example when we were training Grock 3 we built the pre-training framework for this and it these are some some of the coolest system in my opinion that you can build as a software engineer. It's like we have 100k H100s at the time and they were just delivered and we didn't quite have the software. We thought we'd have the software but then at 30k scale we we
24:44realized actually the software is not quite working and it took a a major almost I would say halfway rewrite of the software because there's so much going on in a data center that you can't actually account for. switches are switches are flapping, links are flapping, switches are going down, GPUs are just burning through, you have numeric issues. And
25:09it's a system where you want really 100k H100s to behave in log step. So a training step is like 5 seconds and you're going 5 seconds in log, but during that 5 seconds, everything can happen. So you need to write a system that makes progress despite all these things that can happen in the environment. And we did this successfully and it was one of the coolest times in my life where the the system was actually running at it was running at the same time my son was born. So that was extra excitement.
25:36but these problems like you don't find anywhere else like nobody has this kind of compute and also nobody has this kind of talent density. So at the time to give you a perspective we were like in overall team in pre-training we were probably like 15 people out of that maybe like seven people were working on the actual training system and we we still maintain that talent talent density in the team. So if you're interested in working on these problems and you don't want to be just like part of a bigger organization where you're one of like a thousand people working on this then this is the place like we are still a very small team with me is Leon
26:12Min from the RL and inference team. Yeah. Hi, I'm Lemi. So at our team, we run a reinforcement learning training job and production inference at large scale on the earth and probably soon in space. and we are kind of already design a lot of thing to to make it more resilient and scalable. So we're building a system to scale from 100k chips to millions of chips. And we optimize every aspect of the stack like parallelism, prefill, decode and make resilient to every known and unknown hardware failure. so if if you are system hackers obsessed with extreme performance and reliability. So here is you will find the most interesting problems to work with. And I think
26:59actually like very similar to all kind of thing like you it's it's very important for you to first see the problem and then you will develop the solution that no one else can develop before. Yeah. Okay. I'll hand over to the tooling team. Hello. I'm Ashib from the tooling team. every software needs to have a great interface to be able to make it useful. So as the tooling team we are responsible for building the platforms, frameworks and infrastructure which is required for humans as well as agents to be able to use our products. we started by building out the human data platform. This is a place where we collect all of our human data and eventually expanded on to build our
27:42internal engineering platform through which we basically run deployments run evaluations or like look at like what training results exist. So if you really care about building a good interface or providing a really useful framework for researchers, for agents as well as our tutors, then you should definitely join our team.
28:04So hi everyone, I'm Yulong from the Jax team. So now Jax at XAI is a really small team with a couple of engineers that working on Jax GPU to optimize our ultra large scale GPU training. So you can imagine that training at scale can be very complicated even you run hello world at scale it can be complicated right so then we are actually responsible for supporting the entire companies from pre-training foundation models RLS and also multimodel to scale things to
28:37from first from 10k 100k then probably 1 million h100 equivalent equivalent GPU scale and we to imple implement a lot of you know practic practical optimizations. We have to customize the entire Jax stack from compiler and runtimes and there will be a lot of interesting problems. and also if you really want to you know u obsessed on optimizing the entire stack at scale that we are probably the best place to go because you know we really have very large scale GPU clusters and we have a lot of interesting problems to work with.
29:19Hey, I'm Pranjul from the kernel team. Basically, the kernel team sits at the very bottom of a training and serving stack. Our code runs inside the million equivalent GPUs that we have. And if you look inside the GPU, there's hundreds of thousands of threads. And these threads are trying to talk to each other to multiply matrices, compute attention scores, and some of them even talk to the million other GPUs that we have. And this is the low-level system that we have. And we like optimizing every single microcond in this. And we care deeply about squeezing every last drop of performance from these GPUs. So if you like this low-level systems problems, algorithms,
29:57please join us. [applause] Let's see. I will try to bring in Hiner and Spencer who are actually at our supercomput cluster in Memphis. Hey Hiner. [applause]
30:26Hey. I'm I'm Her from the computer network infrastructure team. we are mainly based in Palo Alto but today we're here in Memphis in the super computer. So the data center here in Memphis the largest cluster on the planet and it is still growing. our job is to keep all this computer up and running next and serve AI out to
30:54our users but used to work well a lot of ingredients.
31:00Actually just put the mic really close to your mouth because your the ambient noise is high. Ah
31:04it's getting a lot. Let me go back. So I was saying our job is to keep the computer up and running to train the next model of drug and serve AI so users. So for GPUs to work well a lot of ingredients have to come together mainly software and hardware. So there's all these CPUs nicks switches and hundreds of thousands of operating systems running as one big supercomputer. And what we need is folks who really understand the really understand and really understand how computers work on a deep level. If that is you reach out and I'm handing over to Dan. [applause]
31:48So we have 300,000 GB 300 platform GPUs here today. We're still growing, still building 847 miles of fiber per data hall. 12 data halls. If you want to be part of the world's largest supercomput, come join us. All right.
32:12So, it's it's quite marvelous what we've been able to do in less than one year's time here. And we have once we're completely finished, we'll have north of a gigawatt of power online and running. We'll have the largest Tesla mega pack system in the world, larger than Hawaii or South Australia. and Zach is really quickly going to talk a little bit about actually constructing the data center.
32:35Yep. So behind me you can see data hall 11. So one of the most incredible things about what we're doing here at MacroArts, how fast we do it, right? So, like they were saying before, over 850 miles of fiber at every single data hall, over 27,000 GPUs and over 200,000 connections. So, all
32:55of this that you can see behind me was put up in less than six weeks. We do that over and over and over again. We massively parallelize it. It's pretty much the most complex and consistent type of engineering design and construction project you can possibly imagine.
33:15So come join us. Yes. You know the the other really awesome thing about this is that everything is completely vertically integrated within this team. from architecture, mechanical, electrical structure, all the disciplines. and we also care a lot about efficiency while we're designing all of this too.
33:32So, it's not just about getting the most compute online the fastest, but also achieving the highest POE in the industry, of using as much power smoothing technology as we can and being really good partners in the community here in Memphis, with the Tesla mega packs and events that we have going. You can check them out at XAI Memphis. Back to you, All right. Thank you, [applause]
34:04All right. So, that was live live from the front lines in Memphis. so fundamental to any AI company's success is the comput advantage. And what we've demonstrated over and over again is that XAI can actually deploy more AI compute faster than anyone else. and actually as Jensen Hong of CEO of Nvidia has said many times in interviews, there is no one faster at getting AI compute online than XAI. So congratulations guys.
34:36[applause] Yeah, this is what it looks like. So that's really phase one which is 330,000 Grace Blackwells with macro written on the building that's not an image edit. It actually is on the roof of the building. and then macro hotter will be the building that you can see which has got the macro hotter with the rockets on it. and that will be another 220,000 GB300s.
35:07so all of this will be training our the models that you that you experience. So the it's absolutely fundamental obviously to have large scale training compute in order to get the best models. Yeah, I'm sort of reminded of the the Jose meme where you see one guy digging and there's like seven people watching. and one of the big differences between XAI and and other companies is we are actually Jose
35:43Hello. All right. I'm Nikita. You might know me as a part-time ship poster, full-time customer support for X. so, we're now reaching over a billion people across our family of apps. every time news breaks, it just becomes evident that this is the most important communication tool of our time. it's where the the the most influential people come convene.
36:09it's where truth is crystallized. Everything is downstream of X. The reason they say this is going to hit Facebook in a week, it's because it happens here. and I think we're only beginning to realize its full potential. we had a remarkable year for the app. we rolled up our sleeves and got a ton done. January was our biggest month ever for the app.
36:35in terms of engagement. and then February is on track to beat that. much of the credit lies with the algorithm team. they've been putting in crazy hours. and it's clearly paying off, but there's still a huge amount of work to be done. on the top of funnel side, first-time downloads are up over 50% every month. and we're exhibiting right now like basically the growth rates of an early stage consumer product. we also made a ton of headway in solving one of the like 20-year-old problems of the app, which was ramping up new users.
37:12new users are now spending 55% more time per day in the app than they were six months ago. and on on the core product side, we're we're hitting our stride too. not only did we rebuild the algorithm, we rebuilt our onboarding flows and we're seeing double digit increases on all our key metrics. We rebuilt notifications, our web browser, XChat, basically every surface of the app has been rebuilt to be better than ever. and it's clear that if we're focused, we can move mountains and evolve this platform. just last month, we did a little push on, articles. and, articles published are up 10x. articles read are up 17x. and, on all other
38:03fronts, like on over the holidays, we did a big push on subscriptions. We just crossed a billion dollars in ARR there. I I think with the X app, you know, the there's very few unknowns like the path for us to win and become you know, the number one app in the world. we're it's it's we we know what to do. The ball's in our court. it's it's for us to win and it's just a matter of us executing.
38:31Yep. And yeah, [applause] so we've evolved the what used to be the old Twitter DM stack, which was unencrypted, basically just text, to a fully encrypted messaging system that it allows you to do audio and video calls. has you know, all the things you'd want from any messaging app. The disappearing disappearing me messages, screen screenshot blocks, like there's a whole all the features that you'd want want in an app. we and we will be open sourcing the code for this in the next few months as we're open sourcing the the recommendation algorithm code so people can actually see what we're doing. nothing beats nothing beats transparency for for
39:17believing in in a company. So we're we're the going to be the only recommendation algorithm that actually open sources. So you can see what it what it does and how it's evolving. with with Gro Chat it will also be open source so you can actually see if there are any vulnerabilities. There will be no hooks for advertising or anything else like that in in in Groch chat which is really intended to be a generalized communication system. and in the next few months we'll be releasing a standalone X chat app. So if you just want to do messaging you can just you can do that. You don't you don't have to go to the the core product.
39:55and it will have desk desktop sharing and u multi-user so you can do you can do video calls with lots of people. It's really intended to be a a fully functional communication system with with Xhat. for X Money we're we've actually had XMoney live in closed beta within the company. and we expect in the next month or two to go to a limited external beta and then to go worldwide to all X users and this is really intended to be the place where all the money is the the central source of of all monetary transactions. So it's it's a it's really going to be a gamecher. and the
40:42reason we say 1 billion users is actually over a billion users. Is that while our monthly users are on average around 600 million, the number of people who have the X app installed, it's well over a billion. It's just that most people only occasionally come to the X app when there's some major world event. But as we give people more reasons to use the the X app, whether it's for communications for Grock or for
41:11X money, whatever the case may be, we we want it to be such that if you wanted to, you could live your life on the X app. And as we make it more and more useful, we'll obviously give people reasons, compelling reasons to use the app every day. and have my expectation is well over a billion daily active users.
41:31So now in order to understand the universe you must explore the universe. There's only so much you can learn from from just being on earth with telescopes and and colliders on earth. Ultimately you have to go out there and you have to explore the universe to understand it. And that's the motivation behind the combination of SpaceX and XAI. is to
42:02accelerate humanity's future in understanding the universe and extending the light of consciousness to the stars. So in the grand scheme of things, when you look at how much energy Earth is actually using for civilization, we're only right now using call it roughly 1% of the potential energy of Earth.
42:24and if we wanted to use even a millionth of the sun's energy, that would be roughly a million times more energy than civilization currently uses. The only way to access that that energy, the energy of the sun, is to extend beyond Earth. Earth is really a tiny tiny dust moat in in a vast darkness.
42:47you know, the sun is 99.8% of all mass in the solar system. So you you have to expand beyond the tiny dust moat that is earth to to make any significant dent in using the sun's energy. Like I said, it's you'd have to expand roughly a million times just to get to 1 millionth of our of our sun's energy and then going beyond that exploring ex extending to the galaxy and maybe someday even to other galaxies.
43:18So the the the the next step beyond Earth data centers is our our Earth orbital data centers. and we'll be launching with SpaceX orbital data centers at the 100 to 200 gawatt per year level. Not cumulative, I mean per year. And ultimately we see a path to maybe launching as much as a terowatt per year of compute from earth. But what if you want to go beyond a mere terawatt per year? In order to do that you have to go to the moon. So by having factories on the moon building AI satellites and having a mass driver which is the kind of thing you really learn about in or read about in science fiction but we're going to make it real. we're actually
44:08going to have a mass driver on the moon. And if you do that, you can go several orders of magnitude greater. You can go to a thousand gigawatts or more per year. and ultimately get to maybe a millionth and then a thousandth and maybe even a few percent of the sun's energy. It it's difficult to imagine what an intelligence of that scale would think about, but it's going to be incredibly exciting to see it happen.
44:45I really want to see the mass driver on the moon that is shooting AI satellites into deep space. It's going like just one after the other. I can't imagine anything more epic than a mass driver on the moon and a self-sustaining city on the moon and then going beyond the moon to Mars.
45:08going throughout our solar system and ultimately, going being out there among the stars and visiting all these star systems. Maybe we'll meet aliens. maybe we'll meet see some civilizations that lasted for millions of years and we'll find the remnants of ancient alien civilizations.
45:26But the only way we're going to do do that is if we go out there and we explore. And this is the path to making it happen. Thank you. [applause] [applause]
资源与出处
- 视频在 YouTube 收看
- 档案silicon-valley.tech 档案页本页视频与文稿数据来自 Elon Musk Archive
- 精读本篇精读 Markdown 版