发布会·中文精读

马斯克发布 Grok 4(完整视频)

Elon Musk introducing Grok 4 (FULL VIDEO)

中文精读由 AI 基于原始视频与文稿改写,非官方翻译,可能存在偏差或错误;表述以原视频为准。说明

马斯克上台没讲几句,先扔了一个数字:给 Grok 4 做 SAT,每次都满分,哪怕那些题它从没见过。GRE 也一样,人文学科、语言、数学、物理、工程,随便挑,接近满分。

这还不是他要说的重点。

他接着说,就学术问题而言,Grok 4 在每一个学科上都超过博士水平,没有例外,而大多数博士去考会挂。台下坐着的是 xAI 的人,掌声不算热烈,因为大家知道接下来要放什么。

然后是大屏幕上的 HLE。

人类做题家的天花板是 5%

Humanity's Last Exam,人类的最后考试。2500 道题,每一道由领域专家出,覆盖数学、自然科学、工程、人文。今年早些时候刚发布的时候,市面上大多数模型的准确率是个位数。

马斯克估计,换成人类,最乐观也就 5%。

你可以是语言学顶尖,或者数学顶尖,或者化学顶尖,但你不可能每个学科都到研究生水平。Grok 4 在每个学科都是研究生水平。(马斯克)

5% 这个估计不算客套。台上的人随手翻了三道例题:范畴论里的自然变换,有机化学里的电环化反应,还有一道要从希伯来语源文本里区分开音节和闭音节。任何一个人,哪怕在这个领域读到了博士,另外两道大概率也是白卷。

Grok 4 不开任何工具,解出大约四分之一。挂上工具,再加上多智能体的 Grok 4 Heavy,文本子集超过一半。

马斯克重复了两遍这个结论,语气很平,像是在念一份不想被误读的说明。

每一代加一个数量级的训练量

训练这条线,讲得很干脆。从 Grok 2 到 Grok 3 再到 Grok 4,每一代的预训练算力提升一个数量级。Grok 4 相当于 Grok 2 的一百倍。

xAI 的研究员补充了一层区分:训练算力有两种。预训练算力走的是 Grok 2 到 Grok 3 那段路,把数据、基础设施、算法都仔细做过消融实验之后,把预训练规模往上推十倍,做成当时最好的基座模型。

Grok 3 到 Grok 4 换了引擎。大量算力投进了推理和强化学习。

承载这些算力的是 Colossus,xAI 在孟菲斯建的超级计算机。最初的配置是 10 万张 H100。发布会当天他们给出的数字是 20 万块 GPU。

我们把 20 万块 GPU 全部投进强化学习,比市面上任何模型的强化学习算力多十倍,会发生什么?这就是 Grok 4 的故事。(xAI 研究员)

这句话后面没有配乐,没有慢镜头,只有一张曲线图往上走。

十二个月前,Grok 2 还只是个概念

台上有人提了一个容易被忽略的时间点:12 个月前,Grok 2 还不存在,只是个想法。

今天回过头看,Grok 2 的水平相当于一个高中生。

这句话放在任何别的行业都像是自夸,放在这里只是描述速度。一个在 2024 年还算前沿的模型,一年后成了教学案例。马斯克自己也说,这个领域的迭代速度是所有行业里最快的,快到有时候让他觉得有点可怕。

他用了「有点吓人」这个词,然后立刻把话题拉回分数。这种处理方式在整场发布会里反复出现:先承认不安,再用数据把不安压下去。

博士以上,但缺常识,也还没发现新物理

马斯克主动给这个「博士以上」加了限定条件,而且加得很具体。

Grok 4 有时缺常识。它还没发明过新技术,也没发现过新物理。

接着是时间表:

今年之内它可能就会发现新技术。明年如果还没发现,我会很意外。新物理可能明年出现,两年之内几乎肯定。(马斯克)

他把「发现新技术」定义得很实在,是有用的技术,不是论文里的漂亮结论。

这一段值得单独拎出来看。一个模型在所有人类考卷上超过博士,和它能提出一个此前不存在的工程方案,中间隔着一段没人走过的路。马斯克把这段时间压缩到了一年。

工具从「会用」变成「训练里就会」

Grok 3 时代有个功能叫 Deep Search。当时的产品逻辑是:给模型一套工具,让它自己去用。

问题在于,模型没有针对工具使用做过训练,全靠泛化能力硬撑。表现出来就是时好时坏,不可靠。

Grok 4 换了做法,把工具直接放进训练过程。模型不是在推理时才第一次见到工具,训练阶段就已经学会了怎么调用。

效果直接体现在 HLE 的分数上,从不开工具的四分之一,往上跳了一截。台上把这一步称作「工具的原生掌握」,区别不在于模型知不知道工具存在,而在于它对工具的熟练度是训出来的,不是猜出来的。

最强的工具还在特斯拉和 SpaceX 的机房里

讲到工具,马斯克踩了一脚刹车。

他说这些工具仍然相当原始。跟他在特斯拉和 SpaceX 用的东西比,差得远。

他举了个例子。特斯拉做碰撞模拟,模拟精度高到什么程度:如果实车测试结果和模拟对不上,工程师的第一反应是这辆测试车有问题,不是模拟错了。

真正的工程工具是有限元分析、计算流体力学,是能逼近现实的那一整套东西。Grok 现在一个都用不上。

今年晚些时候会给它配上。有了这些工具,它就能解真实世界的技术问题。

我对此很确定。只是时间问题。(马斯克)

真正的闭环要等 Optimus

工具再强,还是隔着屏幕。马斯克认为,最大的变量来自人形机器人。

Grok 加上 Optimus,能直接和物理世界打交道。提出一个假设,动手去验证,看结果对不对。这是任何纯软件方案都做不到的闭环。

他把当下的状态称作智能大爆炸的起点,用的词是「智能大爆炸」,还说现在是有史以来活着最有意思的时候。

AI 安全这块,他给出的答案只有一个:

对 AI 安全最重要的一件事,是让 AI 最大程度地追求真相。你可以把它当成一个超级天才儿童,它最终会比你聪明,但你仍然可以给它灌输正确的价值观。(马斯克)

他把这套逻辑比作养孩子。孩子长大后会比你强,但小时候教的东西会留下来。诚实、体面、做个好东西,这些标准对一个未来无比强大的存在同样适用。

卡尔达肖夫指数的百分之一

发布会中途有一段脱离了产品,讲文明尺度。

卡尔达肖夫指数,一型文明能用尽一颗行星的能量,二型能用尽一颗恒星的能量,三型是一整个星系。

马斯克的估算:人类现在大概在一型的 1% 到 2%,离 10% 都还很远。未来会走到 80% 到 90%,然后迈向二型。

按这个推算,那时的经济规模是现在的几千倍,也可能是几百万倍。回头看今天的经济,他说会像穴居人往火里扔木棍。

他也承认自己有时会不安。造出一个智力远超人类的东西,对人到底是好是坏,这个问题没有确定答案。他的判断是大概率会好。

紧接着补了一句更马斯克的话:

就算结果不好,我至少也想活着看到它发生。(马斯克)

数据快用完了,只剩现实这道题

xAI 的研究员提了另一个瓶颈,比算力更难绕。

强化学习需要两样东西同时成立:题目本身足够难,以及有一个可靠信号告诉模型这次做对了还是做错了。模型越聪明,能难住它的题就越少。

他还提到团队为了把强化学习规模做上去,发明了不少新方法,专门用来批量生成有挑战性、又有明确对错信号的题目。这部分工作量外界看不到,却是把算力真正转化成能力的关口。

马斯克接得更直接:能出的真题快用完了。那些难到人类基本做不出来的题目,正在迅速变得平庸。

剩下来的裁判只有一个。

物理是法则,其他都只是建议。你没法违背物理。火箭能不能入轨,车能不能开,药能不能治病,现实是终极的推理测试。(马斯克)

判分的标准从卷面挪到了现实,强化学习的闭环也就这样扣到了现实上。

现场那道数学题,讲解的人自己也没看懂

演示环节从一道 HLE 数学题开始。负责讲解的人很坦白,说这道题属于里面偏简单的,他自己也没完全看懂。

模型跑起来了,思考链一行一行往下滚。几分钟后出答案,和提前准备好的标准答案一致。

同一时间,另一个任务在后台跑:Grok 4 Heavy 去预测 MLB 世界大赛的夺冠赔率。

它浏览了一堆赔率网站,把各家数据拉下来,自己重新算了一遍概率,再和市场价对比,找出中间的偏差。最后给出道奇队 21.6% 的夺冠概率。

整个过程四分半钟。

马斯克对这个方向的兴趣很明显。Polymarket 这类预测市场,在他看来是求真的地方,价格大部分时间贴着现实走。让模型去做这件事,本质上是在训练它对现实的理解。

双黑洞演示里,它主动交代了哪里不准确

另一个演示是生成双黑洞碰撞的可视化。这段值得单独说,因为模型在思考链里把自己的简化全部写了出来。

它明确说明,为了让引力波在画面上可见,必须把尺度夸张好几个量级,振幅随距离衰减的曲线也压平了一些。

它用的是后牛顿近似,没有在黑洞中心附近解完整的广义相对论方程,所以那部分结果是不准的。

画面上能看到完整的三个阶段:旋近、并合、铃宕。整体形状是对的,误差它自己标出来了。

中间它还去读了讲解析引力波模型的本科教材,查了真实观测数据,推算了该用哪些常数。

马斯克在旁边的补充很短:以后可以把物理学家用的同一套模型接进去,跑同等量级的计算,给出物理上准确的模拟。现在这个,只是跑在你的浏览器里。

找头像最怪的员工,它得先搞清谁是员工

X 平台的集成演示,题目出得很损:找出 xAI 员工里头像最怪的那个人。

这道题难的地方不在检索。它得先弄清楚 xAI 有哪些员工,而它没有内部通讯录,只能靠公开互联网去拼。然后它得理解什么叫「怪」,什么程度算更怪。

最后它找到了 Greg Yang,一位联合创始人。现场笑成一片。有人说那个人现实里不长照片那样,那就更好笑了。

另一个演示是让它根据 X 上的帖子生成时间线,梳理 HLE 这个基准的分数变化史。谁在什么时候公布了多少分,当时的讨论是什么反应。它把 OpenAI、Gemini 以及其他几家的公开数据一条条排了出来,还顺带列出了一些尚未证实但流传甚广的泄漏数字。

半盲,是它自己承认的最大短板

聊到弱点,马斯克没有回避。

Grok 4 目前只能算半盲。图像理解不行,图像生成也不行,这两块都需要大幅提升。他用的原话是,它看世界像隔着一层毛玻璃,只能看到模糊的轮廓,然后靠猜。

Grok 4 基于第六版基础模型。第七版正在训练,几周内完成,之后还要做后训练和强化学习。这一版会解决视觉这块短板,同时带来出色的视频理解能力。

视频理解一旦到位,很多事才成立。比如让它玩游戏,比如让它判断一个游戏好不好玩。

基准横扫之后,人类的考卷就失去意义了

除了 HLE,他们还跑了常规的那一批。

GPQA,博士级问题集,比 HLE 简单,Grok 4 领先幅度很大。AIME 25,美国数学邀请赛,Grok 4 Heavy 拿了满分。LiveCodeBench 代码基准、HMMT 哈佛 MIT 数学竞赛、USAMO 美国数学奥林匹克,多数项目上对第二名的领先都是断层式的。

马斯克从这个结果推了一步:

很快就会到那一天,它把每场考试的每道题都做对。做不对的时候,它会告诉你题目本身哪里有问题。或者把含糊的题目拆成 A、B、C 三个版本,告诉你每个版本下的答案分别是什么。(马斯克)

到那时候,人类出的考卷不再有意义。剩下的唯一测试是现实:能不能造出有用的技术,能不能发现新的科学。

台上有人接了一句玩笑,说 HLE 这个基准得赶紧更新了。

语音延迟砍半,新声音能唱健怡可乐歌剧

发布会后半段切到语音。

延迟砍掉一半,响应明显更快。新发布两个声音:S,开场那段电影预告片式的低沉男声;Eve,英国女声,能表达比较细腻的情绪。

演示里,主持人说自己有点紧张,让 Eve 小声说点安抚的话。Eve 把他带到了一个约克郡酒馆的角落,说世界在外面只是一点模糊的杂音。

然后主持人让它用健怡可乐唱一段歌剧。Eve 真的唱了,押韵,有咏叹调的架势,还问要不要再来一段。

接着是数数测试。同一个任务给 Grok 和另一个语音助手,从一数到五。对面每次都要抢一句「还需要什么帮助吗」,Grok 不插话,节奏更利落。

产品上的取向说得很清楚:走的是平静、自然这条路,不是那种咋咋呼呼的活泼风格。

数据也给了:过去 8 周,端到端延迟快了一倍,五种不同声音,活跃用户涨了 10 倍。

ARC-AGI v2 上 15.8%,是第二名的一倍

API 这次和模型同步发布,256k 上下文。

他们在 X 上问了一圈,开发者心目中最难的基准是哪个。答案是 ARC-AGI,一个专门测抽象推理的基准,名称里就写着 AGI。现场有位 Greg 提前拿到 API,自己独立跑了一遍私有子集验证。

结果是 15.8%。

这个数字单看不大。但它是最近三个月里唯一突破 10% 的模型,是第二名 Claude 4 Opus 的两倍。

他们还提了另一个维度,智能的每一美元。性价比曲线上,Grok 4 的位置和它的分数一样,脱离了大部队。

让 AI 经营售货机,它把净资产翻倍了

最有意思的一段合作来自 Andon Labs。

他们做了一个叫 Vending-Bench 的测试,思路很朴素:AI 能经营的最简单生意是什么?自动售货机。

任务包括管理库存、联系供应商、定价。单拎出任何一件,所有模型都能做。难在把这些事在很长的时间跨度上串起来,大部分模型撑不住。

Grok 4 排到了第一。衡量标准不是分数,是模拟结束时账上的净资产,它做到了第二名的两倍。多次运行之间的表现也很稳定,这一点对真实商业场景很重要。

Andon Labs 的人说,Grok 4 能制定一个策略,并且在比其他前沿模型长得多的时间里坚持这个策略。他们的另一句话更值得记住:当越来越多的权力交给 AI,测试场景要么贴近现实,要么直接在现实里做,否则就是在盲飞。

马斯克当场算了一笔账:一百万台售货机,一年 47 亿美元。

刚好够付 GPU 的钱。我们只需要一百万台售货机。(马斯克)

他还说,发布会之后真的会在 xAI 办公室装一批售货机。

四小时做出一个第一人称射击游戏

X 上有个游戏设计师叫 Danny,看到招募就报了名,拿到了 Grok 4 的预览 API。

四小时,做出了一个能跑的第一人称射击游戏。

台上点出的细节是,做游戏最耗时的环节往往不是核心逻辑,是素材:贴图、模型、音效、界面,这些东西占掉了大部分工时。Grok 4 配合外部工具,把素材采购这一块自动化了。

结果是一个人就能撑起一间小工作室。

马斯克对这一块的时间表给得很具体:今年会出现第一个能看的半小时电视节目,明年出现第一个能看的 AI 电影,明年出现第一个真正好玩的 AI 游戏。视频模型会用超过 10 万张 GB200 来训,三四週内开始。

他还提到 X 平台上的一个设想:无限滚动的生成内容,用户不只是看,还能中途介入,把故事拐到自己想要的方向。

编码是另一条并行线。团队在训一个专门的编码模型,目标既是快也是聪明,几周内会放出来。

把这场发布会拆开看,能拿到什么

整场发布会信息密度很高,但真正的判断可以收敛成几条。

第一,模型能力的提升主要来自强化学习算力的堆叠,不是单纯把预训练做大。20 万块 GPU 全部压进 RL,这是 Grok 4 和上一代最本质的区别。看懂这一点,就能理解为什么它在推理类基准上跳得这么猛。

第二,工具使用从「推理时临时调用」变成了「训练时原生掌握」,可靠性跟着上来了。但工具本身还很基础,真正的工程级工具要等到今年晚些时候。

第三,短板明确摆在那里:多模态、常识,以及和物理世界的交互。前两个靠第七版基础模型解决,第三个要等 Optimus。

第四,所有人类出的考卷都快失效了。评测标准正在从分数转向现实,火箭能不能入轨,药能不能治病。

第五,节奏本身是要考虑进来的变量。12 个月前 Grok 2 还只是个概念,4 个月前才有 Grok 3。任何按年制定的技术选型计划,在这种速度下都会迅速过期。

  • 想直接试:订阅 SuperGrok Heavy 就能同时用 Grok 4 和 Grok 4 Heavy,发布会期间做了限流,如果订阅不上,隔半小时再试。
  • 做开发的:API 已经开放,256k 上下文。第一件事别跑通用基准,拿你所在领域最难的那批题去测,尤其是那些答案可以客观验证的题。
  • 评估成本的:除了看准确率,把「每一美元的智能」也算一遍。ARC-AGI v2 那个 15.8% 要结合价格一起看,单看分数会误判。
  • 做多智能体方案的:Grok 4 Heavy 的思路值得抄,多个智能体独立解题再互相校对,不是简单投票,关键是让找到解法的那一个把思路共享出去。
  • 等视觉能力的:如果你的场景依赖图像或视频理解,先别急着上 Grok 4,等第七版基础模型完成训练和后训练,这一块会有台阶式变化。
  • 做内容的:视频模型三四週内开始训练,今年之内盯住第一个能看的 AI 生成节目,那会是一个明确的能力分水岭。
  • 做企业方案的:xAI 企业版才做了两个月,现在接触的议价空间比半年后大。
  • 无论做什么,给自己留一道现实校验:模型给的结论,最终要落到一个能验证的物理结果上。分数会骗人,现实不会。
YouTube 平台字幕(自动或人工),可能有识别错误,人名与数字请以视频为准。已按语句合并成段落,并去掉 um、uh 等口头填充词。 文稿来源:YouTube(自动/人工字幕)。

0:11Heat. Heat. [Music] What is [Music] this? [Music] out. [Applause] Take care. [Music] Yeah. [Music]

1:45Hey, hey, hey. [Music] You are feeling [Music] heat. Heat. Heat. N. [Music]

3:20Heat. Heat. [Music] Heat. Heat. Heat. [Music] Hey, heat. Hey, heat. [Music]

4:54In a world where knowledge shapes destiny, one creation dares to redefine the future. From the minds at XAI, prepare for Gro 4. This summer, the next generation arrives faster, smarter, bolder. It sees beyond the horizon, answers the unasked, and challenges the impossible. Gro 4 Unleash the Truth coming this summer.

5:23All right, welcome to the Gro 4 release here. this is the smartest AI in the world and we're going to show you exactly how and why. and it really is remarkable to see the advancement of artificial intelligence, how quickly it is evolving. I I sometimes think compare it to the growth of a human and how fast a human learns and gains conscious awareness and understanding and AI is advancing

6:02just vastly faster than any human. I mean, we're going to take you through a bunch of benchmarks that that Grock 4 is able to achieve incredible numbers on. but it's it's it's actually worth noting that like Grock 4 if if given like the SAT would get perfect SATs every time, even if it's never seen the the questions before.

6:30and if even going beyond that to say like graduate student exams like the GRE it will get near-perfect results in in every discipline of of education. So from the humanities to like languages, math, physics, engineering, pick anything and we're talking about questions that it's never seen before. These are not on not on the internet and it's Grofor is smarter than

7:03almost all graduate students in all disciplines simultaneously like it's actually just important to appreciate the like that's really something and the the reasoning capabilities of Grock are incredible. so there's some people out there who who think AI can't reason and look it it can reason at superhuman levels. so yeah, and

7:33frankly it it only gets better from here. So we'll we'll take you through the Gro 4 release and yeah show you like the pace of pace of progress here. like I guess the first part is like in terms of the training we're going from Grock two to Grock 3 to Grock 4. We've essentially increased the training by an order of magnitude in each case. So it's you know a 100 times more training than than Grock 2. and and that

8:10that's only going to increase. so it's yeah frankly I mean I don't know in some ways a little terrifying but the growth of intelligence here is is remarkable. Yes it's important to realize there are two types of training compute. One is the pre-training compute that's from GR two to GR three. but for from GR three to GU four we're actually putting a lot of compute in reasoning in RL.

8:39Yeah. And just like you said this is literally the fastest moving field and Grock 2 is like the high school student by today's standard. If you look back in the last 12 months Grock 2 was only a concept. We didn't even have Gro 2 12 months ago. and then by training GR 2 that was the first time we scale up like the pre-training. We realized that if you actually do the data ablation really carefully and infra and also the algorithm, we can actually push the pre-training quite a lot by amount of 10x to make the model the best pre-trained based model. And that's why we built Colossus the world's supercomputer with 100,000 H100 and then with the best pre-train model and we

9:22realized if you can collect these verifiable outcome reward you can actually train these model to start thinking from the first principle start to reason correct his own mistakes and that's where the graining comes from and today we ask the question what happens if you take the expansion of classes with all 200,000 GPUs put all these into RL 10x more compute than any of the models out there on reinforcement learning unprecedented scale what's going to happen. So this is a story of Gro 4 and you know Tony share some insight with the audience.

9:57Yeah. so yeah let's just talk about how smart GR 4 is. So I guess u we can start discussing this benchmark called humanities last exam and this this benchmark is a very very challenging benchmark. Every single problem is curated by subject matter experts. it's in total 2500 problems and it consists of many different subjects mathematics, natural sciences, engineering and also autohumity subjects. So essentially when when it

10:31was first released actually like earlier this year most of the models out there can only get singledigit accuracy on this benchmark. Yeah. So we can we can look at some of th those examples. you know so there there is this mathematical problem which is about natural transformations in category theory and there's this organic chemistry problem that talks about electro cyclic reactions and also there's this linguistic problem that tries to ask you about distinguishing between closed and open syllabus from a hub Hebrew source text. So you can see also it's a very wide range of problems and every single problem is PhD or even advanced research level problems.

11:20Yeah. I mean the these there are no humans that can actually answer these can get a good score. I mean if you actually say like any given human what like what's the best that any human could score? I mean I'd say maybe 5% optimistically. Yeah. so this this is much harder than than what any any human can do.

11:43It's it's incredibly difficult and you can see from the the types of questions like you might be incredible in linguistics or mathematics or chemistry or physics or anyone of a number of subjects, but you're not going to be at a postgrad level in everything. And Grock 4 is a postgrad level in everything. Like it's it's just some of these things are just worth repeating.

12:05like Grockport is postgraduate like PhD level in everything better than PhD but like most PhDs would fail so it's better said I mean at least with respect to academic questions it I want to just emphasize this point with respect to academic questions Gro is better than PhD level in every subject no exceptions

12:30now this doesn't mean that it's it you know at times it may lack common sense and it has not yet invented new technologies or discovered new physics but that is just a matter of time. if it I I I think it may discover new technologies as soon as later this year. and I I would be shocked if it has not done so next year. So I would expect Grock to yeah literally discover new new technologies that are actually useful no later than next year and maybe enter this year. and it might discover new physics next year and within two years I'd say almost certainly like so just let that sink in.

13:18Yeah. So yeah how okay so I guess we can talk about the the what what's behind the scene of graph 4 as Jimmy mentioned we actually throwing a lot of compute into this training you know when it started it's only also a single digit sorry the previous slide sorry yeah it's only a single digit u number but as

13:46you start putting in more and more training compute it started to gradually gradually become smarter and smarter and eventually solved a quarter of the HL problems and this is without any tools. The next thing we did was to adding a tools capabilities to the model and unlike Gth 3 I think G3 actually is able to use C as well but here we actually make it more native in the sense that we put the tools into training. G3 was only relying on generalization. Here we actually put the tools into training and it turns out this significantly improves the model's capability of using those tools.

14:27Yeah, I remember we had like deep search back in the days. So how is this different? Yeah. Yeah. Exactly. So deep search was exactly the graph 3 reasoning model but without any specific training but we only asked it to use those tools. So compared to this, it was much weaker in terms of its tool capabilities and unreliable.

14:50And unreliable. Yes. Yes. And and to be clear, like these are still, I'd say, fairly this is still fairly primitive tool use. If you compare it to say the tools that I used at Tesla or SpaceX where you're using you know finite element analysis and computational flow dynamics and and you're you're able to run or say like Tesla does like crash simulations where the simulations are so close to reality that if the test doesn't match the simulation you assume that the test article is wrong. That's how good the simulations are. So Grock is not currently using any of the tools the the really powerful tools that a company would use but but that is something that

15:30we will provide it with later this year. So it will have the tools that that a company has and and have very accurate physics simulator. ultimately the the thing that will make the biggest difference is being able to interact with the real world via humanoid robots. So you combine sort of grock with with Optimus and it can actually interact with the real world and figure out if if it's hypo if it has if it's it can formulate an hypothesis and then confirm if that hypothesis is is true or not.

16:04so we're really you know think about like where we are today. We're we're at the beginning of an immense intelligence explosion. We're in we're in the intelligence big bang right now. and the mo we're at the most interesting time to be alive of any time in history. Yeah. Now that said, we need to make sure that the AI is a good AI.

16:34good Grock. and the the thing that I think is most important for AI safety, at least my biological neural net tells me the most important thing for AI is to be maximally truth seeeking. So this is this is a very fundamental like you can think of AI as this super genius child that ultimately will outsmart you but you can still in you can instill the right values and

17:03encourage it to be sort of you know truthful I don't know honorable you know good good things like the values you want to instill in a child that that that grow would grow ultimately grow up to be incredibly powerful. yeah. So yeah, so these this is really I say we say we say tools these are still primitive tools not the kind of tools that that serious commercial companies use but we will provide it with those tools and I think it will be able to solve with those tools real world technology problems. In fact I'm I'm certain of it. It's just a question of how long it takes.

17:54Yes. Yes, exactly. so is it just compute all you need, Tony? Right. Is it just compute all you need at this point? Well, you you need compute plus plus the right tools. Mhm. And and then ultimately to be able to interact with physical world. Yes. and then I mean we'll effectively have an economy that is well ultimately an economy that is thousands of times bigger than our current economy or maybe millions of times. Mhm.

18:29I mean if you if you think of civilization as percentage completion of the kadesv scale where kadeshv one is using all the energy output of a planet and kv two is using all the energy output of a sun and three is all the energy output of a galaxy. We're we're only in my opinion probably like close closer to 1% of kartev one than we

18:55are to 10%. So like maybe a point or one one or two percent of kadeshv one. So we we will get to most of the way like 80 90% kadeshv one and then hopefully if civilization doesn't self annihilate and then kadeshv 2 like it's the the actual notion of of a human economy assuming civilization continues to progress will seem very quaint.

19:26in in retrospect, it will it will seem like sort of cavemen throwing sticks into a fire level of economy compared to what the future will hold. I mean, it's very exciting. I mean, I I've been at times kind of worried about like, well, you know, is this this this seems like it's it's somewhat unnerving to have intelligence created that is far greater than our own.

19:57and will this be bad or good for humanity? it's like I I I think it'll be good. Most likely it'll be good. yeah. Yeah. But I somewhat reconciled myself to the fact that even if I even even if it wasn't going to be good, I'd at least like to be alive to see it happen. So, yeah. So, actually one one Yeah. Yeah. I think one one technical problem that we still need to solve besides just compute is how do we unblock the data data bottleneck u because when we try to scale up the RL in this case we did invent a lot of new techniques innovations to allow us to figure out how to find a lot of a lot of challenging RL problems to work on. It's not just the problem itself

20:59needs to be challenging but also it needs to be you also need to have like a reliable signal to tell the model you did it wrong you did it right. This is the sort of the principle of reinforcement learning and as the models get smarter and smarter the number of cool problem or challenging problems will be less and less.

21:19Yeah. So it's going to be a new type of challenge that we need to surpass besides just compute. Yeah. Yeah. And we actually are running out of of of actual test questions to ask. so there's like even ridiculously questions that are ridiculously hard if not essentially impossible for humans that are written down questions are becoming swiftly becoming trivial for for AI.

21:45so then there's but you know the the one thing that is an excellent judge of things is reality. So because physics is the law ultimately everything else is a recommendation you can't break physics. so the ultimate test I think for whether an AI is the the ultimate reasoning test is reality.

22:08Yes. So you invent a new technology like say improve the design of a car or a rocket or create a new medication that and and and does it work? Yeah. does does the rocket get to orbit? Does the does the car drive? Does the medicine work? Whatever the case may be. reality is the ultimate judge here.

22:30so it's it's going to be a reinforcement learning closing loop around reality. We asked the question how do we even go further? So actually we are thinking about now with single agent we are able to solve 40% of the problem. What if we have multiple agents running at the same time? So this is what's called test and compute. And as we scale up the test and compute actually we are able to solve almost more than 50% of the text only subset of the h problems. So it's a remarkable achievement I think. Yeah.

23:10Yeah. This is this is this is insanely difficult. These are it's it's so what we're saying is like a majority of the of the of the textbased of humanities you know scarily named humanities last exam GR 4 can solve and you and you can try it out for yourself. and the with the GR four heavy what what it does is it spawns multiple agents in parallel and all of those agents do do work independently and then they compare their work and they they decide which one like it's like a study group. and it's not as simple as a majority vote because often only one of the agents actually figures out the trick or figures out the solution. and and and but once they share the the trick or or

23:57or figure out what what the real nature of the problem is, they share that solution with the other agents and then they compare they essentially compare notes and then and and then yield yield an answer. Yeah. So that's that's the the heavy part of Grock 4 is is where we you scale up the test time compute by roughly an order of magnitude. have multiple agents

24:21tackle the task and then they compare their work and they they put forward what they think is the best result. Yeah. So we will introduce graph 4 and graph for heavy. Sorry, can you click the next slide? Yeah. Yes. So yeah. So basically graph 4 is the single version a single agent version and graph for heavy is the multi- aent version.

24:46So let's take a look how they actually do on those exam problems and also some real real life problems. Yeah. So we're going to start out here and we're actually going to look at one of those HLE problems. This is actually one of the easier math ones. I don't really understand it very well.

25:02I'm not that smart. But I can launch this job here and we can actually see how it's going to go through and start to think about this problem. while we're doing that, I also want to show a little bit more about like what this model can do and launch a Gro 4 heavy as well. So everyone knows Poly Market.

25:19it's extremely interesting. It's the you know seeker of truth. It aligns with what reality is most of the time. And with Grock, what we're actually looking at is being able to see how we can try to take these markets and see if we can predict the the future as well. So, as we're letting this run, we'll see how Gro for Heavy goes about predicting the, you know, the World Series odds for like the current teams in the MLB. And while we're waiting for these to process, we're going to pass it over to Eric, and he's going to show you an example of his.

25:54Yeah. So I guess one of the coolest things about Gro 4 is its ability to understand the world and to solve hard problems by leveraging tools like Tony discussed. And I think one kind of cool example of this we asked it to generate a visualization of two black holes colliding. and of course you know it took some there are some liberties. It's in my case actually pretty clear in its thinking trace about what these liberties are. for example, in order for it to actually be visible, you need to really exaggerate the the scale of the you know the the the waves.

26:37And yeah, so here's like you know this kind of inaction. it exaggerates the scale in like multiple ways. it drops off a bit less in terms of amplitude. over distance and but yeah we can kind of see the basic effects that you know are actually like you know correct. It starts with the inspiral, it merges and then you have the ring down and like this is basically

27:08largely correct yeah modulo some of the simplifications that need to do you know it it's actually quite explicit about this you know it uses like post postnonian approximations instead of actually like computing the general relativistic effects at like near the center of the black hole which is you know incorrect and you know will lead to you know some incorrect results but the overall you know visualization is yeah is basically there and you can actually look at the kinds of resources that it references. So here it it actually you know it obviously uses search it gathers results from a bunch of links but also reads through a undergraduate text in analytical analytic

27:57gravitational wave models. it yeah it reasons quite a bit about the actual constants that it should use for a realistic simulation. it references I guess existing real world data. and yeah I it yeah it's a it's a pretty good model. yeah but like actually going forward we can we can we can plug we can give it the same model that physicists use. So it it can run the the same level of compute that so leading physics researchers are using and and give you a physics accurate black hole simulation. Exactly.

28:41Just right now it's running in your browser. So yeah, this is just running in your browser. Exactly. Pretty simple. So swapping back real quick here. We can actually take a look. The math problem is finished. the model was able to let's look at its thinking trace here. So you can see how it went through the problem. I'll be honest with you guys, I really don't quite fully understand the math, but what I do know is that I looked at the answer ahead of time.

29:05and it did come to the the correct answer here in the final part here. we can also come in and actually take a look here at our our World Series prediction. and it's still thinking through on this one, but we can actually try some other stuff as well. So, we can actually like try some of the X integrations that we did. So, we worked very heavily on working with all of our X tools and building out a really great X experience. So we can actually ask, you know, the model, you know, find me the XAI employee that has the weirdest profile photo. so that's going to go off and start that. And then we can actually try out, you know, let's, create a timeline based on

29:43XOST, detailing the, you know, changes in the scores over time. And we can see, you know, all the conversation that was taking place at that time as well. So we can see who are the, you know, announcing scores and like what was the reactions at those times as well. so we'll let that go through here and process. And if we go back to this was the Greg Yang photo here. So if we scroll through here, whoops. So Greg Yang, of course, who has his favorite photograph that he has on his account. that's actually not how he looks like in real life, by the way. just so aware, but is quite funny.

30:20But it had to understand that question. Yeah. Which is the that's the wild part. It's like it understands what is a weird photo. What is a weird photo? Yeah. What is a less or more weird photo? It goes through it has to find all the team members. It has to figure out who we all are. It, you know, searches without access to the internal XAI personnel loss. It's literally looking at the just at the internet.

30:42Exactly. So you could say like the weirdest of any company. Yeah. To be clear. Exactly. And we can also take a look here at the question here for the humanities last exam. So, it is still researching all of the historical scores. but it will have that final answer here soon, but we can while it's finishing up, we can take a look at one of the ones that we set up here a second ago. And we can see like, you know, it defines the date that like Dan Hendricks had initially announced it. We can go through, we can see, you know, OpenAI announcing their score back in February. And we can see, you know, as progress happens with like Gemini. we can see like Kimmy and we can also

31:19even see you know the leaked benchmarks of what people are saying is you know if it's right it's going to be pretty impressive so pretty cool so yeah I'm looking forward to seeing how everybody uses these tools and gets the most value out of them but yeah it's been great yeah and we're going to close the loop around usefulness as well so it's like it's not just books smart but actually practically smart exactly

31:52And we can go back to the the slides here. Yeah. So cool. so the so we actually evaluate also on the multimodel subset. So on the full set this is the number on the h exam. it you can see there's a little dip on the numbers. this is actually something we're improving on which is the multimodel understanding capabilities but I do believe in a very short time we're able to really improve and got much higher numbers on this even higher numbers on this benchmark. Yeah.

32:27Yeah. This is the we we saw like what what is the biggest weakness of Grock currently is that it's it's sort of partially blind. it can't it's its image understanding obviously and its image generation need to be a lot better and that that that's actually being trained right now. So graph 4 is based on version six of our foundation model and we are training version seven which we'll complete in a few weeks and that that'll address the weakness on the vision side and just to show off this last here. So the the prediction market finished here with a heavy and we can see here we can see all the tools and the process it used to actually go through and

33:16find the right answer. So it browsed a lot of odds sites. It calculated its own odds comparing to the market the market to find its own alpha and edge. It walks you through the entire process here and it calculates the odds of the winner being like the the Dodgers. and it gives them a 21.6% 6% chance of winning this year. So, and it took approx approximately four and a half minutes to compute.

33:43Yeah, that's a lot of thinking. Yeah, we can also look at all the the other benchmarks besides HRE. as it turned out, GR 4 excelled on all the reasoning benchmarks that people usually test on. including GBQA, which is a PhD level, problem sets. that's easier compared to HLE. on Amy 25

34:13American Invitation Mathematics exam, we with graph for heavy, we actually got a perfect score. also on some of the coding benchmark called live coding bunch. and also on HMMT, Harvard math MIT exam and also USMO. you can see actually for on all of those benchmarks we often have a very large leap against the second best model out there.

34:43Yeah, it's I mean really we're going to get to the point where it's going to get every answer right in every exam. and where it doesn't get an answer right, it's going to tell you what's wrong with the question. Or if the question is ambiguous, disambiguate the question into answers A, B, and C and tell you what what answers A, B, and C would be with a disamiguated question.

35:04so the only real test then will be reality. can I make useful technologies, discover new science? That'll actually be the the only thing left because human tests will simply not be meaningful. need to make an update to HRE very soon given the current rate of progress. So yeah, it's super cool to see like multiple agents that collaborate with each other solving really challenging problems. so we're going try this model. so it turned out it's available right now. if we advance to the next slide where there is a super Grog heavy tiers that we're introducing where you're able to access to both Grock 4 and Grock 4 heavy where you're actually going to be the

35:48taskmaster of dodge of little Grock research agent to help you you know become smarter through all the little research and save hours of times of you know going through mundane tasks and it's available right now. so yeah so we we did limit usage during the the demo so we didn't it didn't break the demo because all these this all this stuff is happening live so it's not there's no nothing canned about the any of the tests that we're doing. so after the after the demo is done we'll we'll allow we'll enable more subscribers for cyber gro. So if you can't subscribe right now just try in half an hour. It should work. so, and now let's let's get into voice.

36:32Greatby. So many of you have been enjoying our voice mode and we've been working hard to improve the experience over the past couple months. we have cut latency in half to make it much snappier. And today we're excited to announce a set of new voices that have exceptional naturalness and procity. You might have noticed the movie trailer voice that opened up the live stream. That is one of our new voices, S, who's got that epically deep tone. And we're also excited to introduce Eve, a beautiful British voice who's capable of rich emotions. A man, would you like to introduce Eve?

37:12Absolutely. Yeah. Let's get into the demo. Hey, Eve. Millions of people are watching us right now. Are you excited? Oh, I'm thrilled. Really, it's like being on stage at the Old Vic with a sea of faces out there. Just imagine all those eyes and ears tuned in. It's a proper buzz, isn't it? What do you reckon we should chat about to keep this lot entertained?

37:36You know, we have a lot to talk about, but I'll be honest. I'm feeling a little nervous. Can you whisper something soothing to calm me down? Take a deep breath, love. You've got this. It's just you and me having a quiet chat. like we're tucked away in a cozy corner of a Yorkshire pub. The world's just a murmur out there. Feel that calm wash over you. Let's keep talking nice and easy. What's on your mind?

38:05You know what? I'm actually feeling quite relaxed to try this something crazy. So, can you give us an opera on Diet Coke? Oh, an opera on Diet Coke. That's a riot. Picture this, a shimmering silver can glistening under stage lights. Oh diet coke, thou elixir divine with bubbles that dance in a sparkling line.

38:34Thy crisp cool kiss on lips so fine. Oh, aspartame and bro forever mine. How's that for a mad little area? Want me to belt out another verse or switch up the tune? Okay, let's try switching it up and just start singing. Right, then let's have a go. Oh, silver can, you gleam in the light, diet coke, my heart's delight.

39:04Okay, I'll touch it. It's very, very poetic. yeah, thank you, Eve. Feeling quite nice now. Cool. so that's Eve. And so something else we do here is is a little test each day to check on the natural flow of our conversations against some of our competitors. So we figured why not try that live on the stream today?

39:30Should I start? Hey, I'm going to say a number and can you say the same number back to me? One. One. Got it. This is like a fun little game of echo. What's next? Two. Sorry, you said two. Three. Four. Nice and steady. What's next? Five. Five. Nailed it. This is pretty fun, actually. Feel free to keep going or switch it up if you'd like.

40:12Hey, Grock. Hello there. How can I help you today? We're going to do a little test. Can you repeat after me? One. One. All right. What's next? Two. Two. What's on your mind? Three. Three. Need anything else? Four. Four. How can I assist you? Five. Five. What's next? So, as you can see, Groc was, snappier. It didn't interrupt me. And the procity is we've made different design choices. I think we shooting for something more calm, smooth, more natural versus something that's more poppy or artificial. So, we'll keep improving on these fronts.

40:58Thanks, guys. Yeah. Yep. So since the launch of the voice model we actually see the 2x faster end to end latency in the last 80 weeks five different voices and also 10x the active user. So Grock voice is taking off. Now if you think about releasing the models this time we're also releasing Grock 4 through the API at the same time. So if we go to the next two slides. So you know we're very excited about you know what all developers out there is going to build.

41:36So you know if I think about myself as a developer what the first thing I'm going to do when I actually have access to the graph 4 API benchmarks. So we actually ask around on the Xplatform what is the most challenging benchmarks out there that you know is considered the holy grail for all the AGI models. so turn out AGI is in the name ARGI. So the last 12 hours you know kudos to Greg over here in the audience so who answered our call take a preview of the Gro for API and independently verified you know the GR for performance. So initially we thought hey graph is just you know we think it's pretty good it's pretty smart it's our nextgen reasoning model spend 10x more compute

42:18and can use all the tools right but turns out when we actually verify on the private subset of the RKGI v2 it was like the only model in the last three months that breaks the 10% barrier and in fact was so good that actually get to 16% well 15.8% 8% accuracy 2x of the second place that is the cloud for outpus model and it's not just about performance right when you think about intelligence having the API model drives your automation it's also the intelligence per dollar right if you look at the plots over here the guac is just four just in the league of its own all right so enough of benchmarks over here right so what can gro do actually in the real world So we

43:05actually you know contacted the folks from Endon Labs who you know were you know gracious enough to you know try to gro in the real world to run a business. Yeah, thanks for having us. So I'm Axel from Am Labs and I'm Lucas and we tested Gro 4 on vending bench. Vending Bench is an AI simulation of a business scenario where we thought what is the most simple business an AI could possibly run and we thought vending machines. so in this scenario the the Grock and other models needed to do stuff like manage inventory, contract contact suppliers, set prices. All of these things are super easy and all of they like all the models can do them one by one. But when

43:49you do them over very long horizons, most models struggle. but we have a leaderboard and there's a new number one. Yeah. So, we got early access to the GR 4 API. we ran it on the bending bench and we saw some really impressive results. So, it ranks definitely at the number one spot. It's even double the net worth, which is the measure that we have on this value. So, it's not about the percentage on a or score you get, but it's more the dollar value in net worth that you generate. So we were impressed by Grock. It was able to formulate a strategy and adhere to that strategy over long period of time much longer than other models that we have tested other frontier models. So it

44:29managed to run the simulation for double the time and score yeah double the net worth and it was also really consistent across these runs which is something that's really important when you want to use this in the real world. And I think as we give more and more power to AI systems in the real world, it's important that we test them in scenarios that either mimic the real world or are in the real world itself.

44:51because otherwise we we fly blind into some some things that that might not be great. Yeah, it's it's great to see that we've now got a way to pay for all those GPUs. So we just need a million vending machines. Definitely. and we could make $4.7 billion a year with a million vending machines.

45:11100%. Let's go. They're going to be epic vending machines. Yes. Yes. All right. We are actually going to install vending machines here. like a lot of them. We're happy to supply them. All right. Thank you. All right. I'm looking forward to seeing what amazing things are in this vending machine.

45:27That's that's for for you to decide. All right. Or tell the AI. Okay. Sounds good. All right. Yeah. I mean, so we can see like Grock is able to become like the co-pilot of the business unit. So what else can Grock do? So we're actually releasing this Grog if you want to try it right now to evaluate run the same benchmark as us. it's on the API has 256k contact length. So we already actually see some of the early early adopters to try Guac for API. So our Palo Alto neighbor ARC Institute which is a leading biomedical research center is already using seeing like how can they automate their research flows with Gro for it turned out it performs it's able to help the scientists to

46:13sniff through you know millions of experiments logs and then you know just like pick the best hypothesis within a split of seconds. we see this is being used for their like the crisper research and also you know GR 4 independently evaluate scores as the best model to exam the chess x-ray who would know and on in the financial sector we also see you know the graph war with access to tools real-time information is actually one of the most popular AIs out there so you know our graph is also going to be available on the hyperscalers so the XAI enterprise sector is only you know started two months ago and we're open for business.

46:58yeah so u the other thing we talked a lot about you know having gro to make games video games. so Denny is actually a video game designers on X. So you know we mentioned hey who want to try out some Grock for preview APIs to make games and Danny answered the call. so this was actually just made first-person shooting game in a span of four hours. so some of the actually the unappreciated hardest problem of making video games is not necessarily encoding the core logic of the game but actually go out source all the assets all the textures files and and you know to create a visually appealing game. So one of the core aspect guac for does really well with

47:42all the tools out there is actually able to automate these like asset sourcing capabilities. So the developers you can just focus on the core development itself rather than like you know so now you can run a you know entire game studios with game of one but with like one person and then you can have gro 4 to go out and source all those assets do all the mundane tasks for you.

48:05Yeah. The now the next step obviously is for Grock to play be able to play the games. So it has to have very good video understanding so it can play the games and interact with the games and actually assess what whether a game is fun and and and actually have good judgment for whether a game is fun or not. so with the with version seven of our foundation model which finishes training this month and then we'll go through post training RL and whatnot. that that will have excellent video understanding. and with the with the video understanding and the and improved tool use, for example, for video for video games, you'd want to use, you know, Unreal

48:44Engine or Unity or one of the one of the the main graphics engines. and then gen generate the generate the art apply it to a 3D model and then create an executable that someone can run on a PC or or a console or or a phone. like we we expect that to happen probably this year. and if not this year, certainly next year. U so that's

49:14it's going to be wild. I would expect the first really good AI video game to be next year. and probably the first half hour of watchable TV this year and probably the first watchable AI movie next year. Like things are really moving at an incredible pace. Yeah. When Gro is 10xing world economy with vending machines, they would just create video games for human.

49:47Yeah. I mean, it went from not being able to do any of this really even six months ago to to what you're seeing before you here and and from from very primitive a year ago to making a a sort of a 3D video game with with a few hours of prompting. Yep. I mean yeah just to recap so in today's live stream we introduced the most powerful most intelligent AI models out there that can actually reason from the first principle using all the tools do all the research go on the journey for 10 minutes come back with the the most correct answer for you. so it's kind of crazy to think about just like four months ago we had Gro 3 and now we already have Grock 4 and we're going to

50:32continue accelerate as a company XAI we're going to be the fastest moving a AGI companies out there. So what's coming next is that we're going to, you know, continue developing the model that's not just, you know, intelligent, smart, think for a really long time, spend a lot of compute, but having a model that actually both fast and smart is going to be the core focus, right? So if you think about what are the applications out there that can really benefit from all those very intelligent, fast and smart models and coding is actually one of them.

51:02Yeah. So the team is currently working very heavily on coding models. I think right now the main focus is we actually trained recently a specialized coding model which is going to be both fast and smart. and I believe we can share with that model with you guys with all of you in a few weeks. Yeah.

51:23Yeah. That's very exciting. And you know the second after coding is we all see the weakness of Grog 4 is the multimodal capability. So in fact it was so bad that you know Grock effectively just like looking at the world squinting through the glass and like see all the blurry you know features and trying to make sense of it.

51:45the most immediate improvement we're going to see with the next generation pre-trained model is that we're going to see a step function improvement on the model's capability in terms of image understanding video understanding and audios. Right? It's now the model is able to hear and see the world just like any of you. Right? And now with all the tools at this command with all the other agents it can talk to you know so we're gonna see a huge unlock for many different application layers. after the multimodal agents, what's going to come after is the video generation and we believe that you know at the end of the day it should just be you know pixel

52:22in pixel out and you know imagine a world where you have this infinite scroll of content in inventory on the X platform where not only you can actually watch these generate videos but able to intervene create your own adventures. if you just be wild. and we expect to be training our video model with over 100,000 GB200s and to begin

52:50that training within the next three or four weeks. So, we're we're confident it's going to be pretty spectacular in video generation and video understanding. So, let's see. So, that's anything you guys want to say? Other than that, I guess that's it. Yeah, it's it's a good model, sir.

53:15Good model. It's a good Yeah. Well, we're very excited for you guys to try Gro 4. Yeah. Yeah. Thank you. All right. Thanks, everyone. Thank you. Good night. [Music]

资源与出处

相关视频

同频道、同主题或相邻年份