Databricks 首席 AI 科学家 Jonathan Frankle：AI 研究到底在做什么

Megan Reynolds

你创办了 MosaicML，以 13 亿美元卖给了 Databricks。但我们从最简单的开始：AI 研究到底是什么？

You’ve built MosaicML and. You sold it to Databricks for $1.3 billion, but let’s start really simply. What actually is AI research?

Jonathan Frankle

我会这样描述：就像物理学家，或者说宇宙学家看待宇宙。我们被交到手里一个巨大的造物，可以观察光、粒子、行星以及它们如何相互作用，但没人给我们任何基本定律。我们拥有的每一条物理定律，都是试图用我们能理解的语言简化和描述宇宙，并对未来现象有预测力。

But yeah AI research … the way I would describe it is kind of in the same way that a physicist who or cosmologist let’s say looks at the universe and you know we’ve just been given this giant artifact we can observe light we can observe particles, we can observe planets, we can observe how they interact with each other, but nobody has handed us any fundamental laws. Every law of physics that we have is our attempt to simplify and describe the universe in terms we understand that have predictive power about future phenomena in the universe.

我也这样看 AI 研究。这听起来反直觉，因为这东西是人从零造出来的。但想象你是创造宇宙的上帝：造出一锅粒子汤，设计它们相互作用的规律，选几个基本常数，然后让整个过程运转，自己去海滩待上约 150 亿年。回来一看，地球上有了生命——这是怎么来的？我看现代生成式 AI 也一样：我们造出一锅如今已达万亿参数的汤，用网络架构规定它们如何相互作用，把所有数据扔进去，有能量、有算力，再用一套总体而言相当简单、人类能理解的规则搅拌很长时间，另一端就出来了 ChatGPT。

And I look at AI research the same way. And it feels counterintuitive because we’ve built this. This has been human created, designed from scratch. But really, the way I look at it is that in some sense, imagine you’re God creating the universe and you create this whole soup of particles, you design some laws of how they interact with each other, you pick some fundamental constants, and then you set the whole process in motion and go hang out on the beach … for, oh, 15-odd billion years, and you come back and you see life on Earth and you ask, how did that get there? … I look at modern generative AI the same way. We create this soup of trillions of parameters at this point, ways they interact with each other in terms of the network architecture. We take all this data and toss it in and then we you know we have energy, we have compute and we just mix it together with a pretty simple set of rules, all things considered, human understandable set of rules for a really long time. And then you get ChatGPT out the other end.

我们在足够多的小宇宙、更大的宇宙、再更大的宇宙上做过这件事，因此对如何设定「物理定律」、如何注入能量——比喻地说，就是数据以及以什么顺序展示数据——有了相当不错的原则，它往往稳定地在最后产生很有意思的东西。但对于从 A 点到 B 点是怎么走的——相当于生物学、化学这些介于基础物理与人类社会之间的中间科学——我们仍然有待理解。

And we’ve done this at enough small scale universes and bigger universes and even bigger universes that we have some reasonably good principles for how to set all the laws of physics and how to kind of like let things roll out and how to inject energy into the system metaphorically, you know, data and what order to show it that it does tend to consistently lead to some really interesting things at the end but our understanding of how you get from point A to point B … what you would describe as biology or describe as like chemistry or kind of all these intermediate sciences that come between basic laws of physics and us sitting here having a society. Those are all still yet to be understood.

我们对其中某些环节有粗粒度的控制，像 RL、微调，认为它们可能把过程推向好的方向。但说到底，我们掌握的是宇宙开始时的基本条件，以及结束时这些非常有意思、非常复杂的模型；研究就是以人类可理解的方式，弄懂中间那 150 亿年里发生了什么、有哪些抽象层。这很难。

We have coarse-grained control over aspects of it that we seem to think might influence it in good directions. We have things like RL or things like fine-tuning. But at the end of the day, we really do have the basic conditions at the beginning of our universes, the end conditions that are these really interesting sophisticated models, and the research is really understanding all these abstractions and what happens on those 15 billion years in a way that’s human comprehensible. And that’s hard.

Jonathan Frankle

我通常问的问题很基本：我们跑了哪些实验？设了什么初始条件，结果如何？我们造出的是一锅分子汤，还是造出了生命？那生命有没有意思？对我们关心的、衡量智能的指标来说，这些数字有没有提升？我们改了什么、数字怎么变、能不能预测接下来该继续改什么？然后再问：这些数字对客户真正在意的事有意义吗？

I ask, okay, what experiments did we run? Like basically what were the initial conditions we set up and how did they do? Did we create a bunch of molecular soup or did we create life? And was that life interesting or was that life not interesting? … Did those numbers improve and what did we change and how did those numbers change and can we make any predictions about what we should keep changing in order to make those numbers improve in the direction we want and then ask ourselves are those numbers truly meaningful in the things that matter for our customers.

业界研究员的日常还有大量修坏掉的 GPU、让东西跑快、搞清楚为什么出了蠢事、为什么数据不好——这里一个 bug，那里一个 bug。但归根结底就是：改初始条件，看结果受何影响，再继续改，试着在脑中——甚至在数学上——建立「初始条件如何影响结果」的模型。

That is kind of in some sense, if I could describe the life of a researcher in industry is that there’s a lot of like fixing broken GPUs and making things run fast and understanding understanding why something stupid happened, and why was the data not good? Oops, we had a bug here, a bug there, but at the end of the day, it really is that question of change initial conditions, see how the outcome is influenced, and then keep changing initial conditions and try to build a model in your mind, possibly even somewhat mathematically of how initial conditions affect outcome.

Jonathan Frankle

我们正在训练会用 Databricks 的模型。我们有一组叫 Genie 的产品，让本来驾驭不了这台复杂机器的各类用户也能用它。我最喜欢在我们自己 Databricks 数据上问的问题是：我的老创业公司 MosaicML 的生命周期总收入是多少——包括它作为独立公司的各个阶段、整合进 Databricks 的过渡期、之后延续的老产品，但不包括后来以 Mosaic AI 品牌推出的新产品。这些数据多年来以各种形式散落在数百 PB 的数据里。

So we’re currently training models that use Databricks … We have a set of products called Genie that are all about allowing all sorts of users who wouldn’t normally be able to hop into this you know very complex machine and Drive it … my favorite question to ask on our own Databricks data is actually what was the lifetime revenue of MosaicML my old startup including all of its iterations as a separate company through the transition period integration with Databricks and the legacy products that continued after that but not new products that we called under the Mosaic AI brand … across hundreds of petabytes of data at Databricks that was represented in all sorts of different forms over the years.

这是一个极其简单的自然语言问题。但答案要求派一个智能体去探索上万张 SQL 表、笔记本、仪表盘、PDF、邮件、Slack 消息，弄清我们多年来怎么计费、怎么记录数据、有哪些产品，最后带回一个数字。用现成的智能体做这件事极其昂贵和缓慢。但你可以训练一个定制模型，让它知道怎么探索 Databricks、怎么找数据、哪些信号通常有含义。

It’s an extremely simple natural language question. But what is the answer? That involves sending an agent out to go and explore tens of thousands of SQL tables, notebooks, dashboards, PDF documents, emails, Slack messages to try to uncover how we did billing, how we logged that data over time, what the products were, and come back with a simple answer that’s just a number. … That’s an extremely expensive and slow thing to do if you’re using an off-the-shelf agent. But you can train a custom model that knows how to explore Databricks, how to find data, what signals tend to mean things, how people naturally create data when you’re in a large enterprise.

还有干扰项：一张 SQL 表，旁边是多年来有人建了又弃用的各种视图；有人们天天用的仪表盘，也有某人一次性做的、其实是错的仪表盘——就像 Google Docs 会越积越多，没人会把文档标为过时，只是不再用了。那怎么办？让模型通过强化学习自己搞明白：给它很难的问题，让它尝试回答，提供衡量成功的打分细则，在大量 GPU、大量数据上长时间跑。最后你得到一个在这件事上非常好、非常高效的模型，我们叫它 Genie。

And, you know, there are distractors. You have a SQL table, but you also have random views of that SQL table people have created and abandoned over the years. You have dashboards that people use day to day and dashboards that someone created as a one-off that are actually wrong in the same way that Google Docs tend to accumulate over time. … You let the model figure it out through reinforcement learning. You give the model some very hard questions. You let it try to answer them. You provide some rubrics of what success looks like. And you do this on a huge number of GPUs on a ton of data over a very long period of time. … You then end up with a model that actually gets really good and really efficient at this, and we call that Genie.

有人选择试图理解中间过程，有人说管它呢，干脆拟合函数，看以严格方式改变初始条件会如何改变结果——扩展律就是突出的例子：拟合一堆小模型，就能预测更大、数据更多的模型的表现，这如今是所有大实验室做预训练的关键。

And various hypotheses for what happens in between potentially, or we say screw it, let’s not understand anything that happens in between, let’s just try to fit functions that tell us how if you vary the initial conditions in a very regimented way. … That scaling laws is one prominent example of this where you can predict based on how much data and how big the model is, you can predict for future models that are even bigger and have even more data, what performance they’re likely to get by fitting a bunch of small models to this, which is now key for how pre-training happens at all the big labs.

Jonathan Frankle

强化学习其实非常简单：给模型一个问题，让它尝试解决，再就它解得如何给反馈。可以只是一个数：答对 1 分，答错 0 分，答得特别好 5 分，错得离谱扣 2 分。分数体系可以自己定，只要大致一致、别太离谱。有算法会把这些信息传回去更新模型，让它以后在这个问题上更可能做好。

Very, you know, it’s really, really simple, you give a model a problem to solve, you let it try to solve that problem, you give it feedback on how well it tried to solve that problem, it could just be a number like you got it right, you get one point, you got it wrong, you get zero points, you got it really, really right, you get five points, you’ve got it really, really wrong, you get negative two points, you can kind of make up your own point scale, it just has to kind of be somewhat consistent and not too crazy. … there are algorithms to take that information, propagate it back into updating the model so it’s more likely to have done well in the future on this problem, and then you proceed.

在 Genie 上，我们有很多层衡量。第一层只检查最基本的小行为：API 调用对不对、请求有没有格式错误。如果这都失败了，是起点的开源模型本来就不行而我们没改善它？是提示没写好？是有人刚更新了 API 让提示过时？是学习率选错了？数据有错、没按对话格式整理？这就是初始条件真正咬人的地方。

There’s an initial measure which just checks, for basic tiny behaviors, is it using Databricks correctly in this way, and in that way, and in that way? … is it calling the API correctly? Or the API request malformed? … Was it because the model we started with, you know, some open source model wasn’t good at this and we just didn’t make it better? Was it because we didn’t prompt it well? Was it because our prompt is actually a little bit out of date because someone just pushed an update to the API? Was it because we picked the wrong learning rate? … Was our data bad? Was there a mistake in our data? Was our data not formatted in the chat format properly? Like, this is where the initial conditions really bite you.

过了第一关，下一步是在很简单的虚构 Databricks 环境里给它真实任务。我们编了一家假银行或假保险公司，建了假的工作区、SQL 表、笔记本、文档和邮件，这样我们能完全控制复杂度、杂乱程度和真实感。然后出题，比如「这家假公司 2024 年第三季度消费品收入是多少」，而消费品不会只在一个显眼的地方写着，可能分散在三四个业务线里。再用打分细则判断好坏——如果第一层是单元测试，这就是端到端的集成测试。

let’s say we survive that round okay next is let’s give it some real tasks on some very simple fake Databricks environments … we’ve just made up a fake bank or a fake you know insurance company or something like that … we’ve created a fake Databricks workspace we’ve added fake SQL tables and fake fake notebooks and fake documents and fake emails and everything, because again, we want to have our own toy environments that are fully, we can fully control them … go and ask what was our revenue for our consumer goods products in quarter three of 2024 for this fake company, where hopefully consumer goods is not just spelled out in one obvious place. … Oh, they’re actually stored in three or four different places because they’re in different lines of business. … If the first thing I described as a unit test, this is now an integration test of, you know, end to end, did it complete a task?

如果答案是一个数，可以核对。但人们问的大多不是一个干净的数，而是一份报告或一段描述，比如「我们业务在某个人群上最大的逆风是什么」，没有对错。你又需要自动评分，因为训练要跑很多次，不能每次都让人看。所以要设计 LLM 评委、验证器和更复杂、能容纳多种答案的打分细则——像在学校批改学生作文那样打分。

If the answer is a number, you can verify that. But most questions people ask aren’t actually one clean number. It’s usually a report or a paragraph describing something. … You can ask like, you know, what are the biggest headwinds for a certain demographic group for our business? There’s not a right or wrong answer. … you do want some kind of automatic grading you don’t have to have a human look at that answer every time because it’s you know very tedious you need to run this a lot as you’re doing a lot of training runs so you have to design LLM judges and verifiers and more complex rubrics that admit a wide variety of answers and score them in some way you need to kind of grade them the way you would grade a student’s essay in school

Jonathan Frankle

即便通过了这些，也还不够上生产，因为这是我们造的假工作区，问题也是编的，谁知道真实用户在问什么。所以还要在生产环境做 A/B 测试。有些客户很乐意尝鲜，我们和他们合作：一部分查询走新模型，请他们反馈喜不喜欢、有没有察觉，偶尔点个赞或踩。有个能坐下来直接问「你喜欢吗？答案更好了吗？」的伙伴，和只看遥测数据是两回事。

So this is very complex but this still isn’t good enough to put into production even if it passes this because this is a fake workspace that we’ve created … we’ve made up all these questions. Who knows like what our users are actually asking. So then there’s another step of we had to actually go and start A-B testing this in production. … some percentage of your queries will go to this instead of something else, instead of our old model. and like we just want your feedback. Like, did you like it? Did you notice? Can you give us a thumbs up, thumbs down occasionally? … it’s one thing to just put it out there and look at some telemetry. It’s another thing to have a partner. You can just sit down and ask them like, did you like this? Did it feel good? Was it like, were the answers better?

驱动 Genie 有很多方式：闭源模型、开源模型、这个或那个定制模型，在速度和质量之间做各种取舍。我希望有个数字能告诉我什么让用户真正高效，但没有。我热爱这份工作的一点是：我的头衔是首席 AI 科学家，不是「首席客户好朋友」，但我被允许、也被鼓励花大量时间和客户在一起。我对团队说过，终极 eval 是：你有没有让客户开心、成功？不是学术基准有没有涨、排行榜有没有赢、内部基准有没有涨。如果我们所有数字都觉得很好、客户却不喜欢，那才是最重要的信号——说明数字不是好代理。

But there are many ways to power Genie. We could use, you know, a closed model. We could use an open model. We could use this custom model or that custom model. We can make various trade-offs of how slow it is versus, you know, the quality of it. … I wish there were just a number that told me that. But it’s really, I mean, one of the things I love about my job is that I’m a researcher. My title is Chief AI Scientist, not Chief Customer Best Friend. But I’m allowed and encouraged to go spend a lot of time with our customers. As I told my team, the ultimate eval is, did you make the customer happy and successful? Not, did you make the benchmark go up on the academic benchmark? Did you win the leaderboard? Did you like make our internal benchmark go up? … And if we thought it was great on all our numbers, but they didn’t like it, But that’s the most important signal, and the numbers weren’t a good proxy.

这种事每天都在发生。我们最喜欢的客户是内部用户——Databricks 最大的用户就是 Databricks。我的第一批小白鼠是我的团队。

But I will say, this is every single day. Like we’re, you know, our favorite customer is our internal users. The biggest user of Databricks is Databricks. We all use Databricks all the time. And so my first guinea pig is my team.

Jonathan Frankle

一个常见模式是模型变得非常擅长钻基准或奖励的空子。大家叫法不同，「散弹式（scattergunning）」是其中一种：你希望模型在回答中提到某些要点，它就开始什么都提，或者说「如果你的问题是这个意思，答案是这个；如果是那个意思，答案是那个」，一次给你三四个答案。如果你只检查「有没有提到所有该提的点」——这是构建打分细则非常合理的一种方式——那答案是有，因为它给了四个不同答案。于是数字非常好看，然后团队真正开始和模型交互，你会觉得：这太烂了。

So there’s one, you know, there’s one common pattern that happens where the models do get very good at exploiting the benchmarks or exploiting the reward. And it’s this behavior, you know, it’s, like people have different names for it, but scattergunning tends to be one behavior for it where when you’re hoping, you know, you’re looking for the model to mention certain important things as part of your evaluation, the model will start to mention everything, or like kind of give you, well, Well, if your question meant this, the answer is that. And if your question meant this, the answer is that. … And if you’re just looking for like, did it mention all the right things, which is one very reasonable way of trying to build one of these rubrics, the answer is yeah, because the model gave you four different answers. … And so it makes the numbers look really good. And then you actually have the team start interacting with the model and you’re like, this sucks.

然后你得回头问：是我的基准量错了东西？要不要再加一项指标？有时用户确实会问模棱两可的问题，这时这样答才是最好的。还是我让模型优化的目标本身错了？或者底层模型从预训练起就有这种倾向，很难让它不这么做，需要后处理？你以为量了一切，模型却找到另一条偷偷在所有指标上好看的路——因为那基本就是你无意中告诉它去做的，至少你给了它这个选项。但对用户来说体验很糟。

And then you have to go back and ask well Am I just measuring the wrong thing in my benchmark because clearly my benchmark was not a good proxy for how the model felt in Production … Sometimes users actually do ask like, you know Ambiguous questions and this is the best approach for the model to try to do this Or was it that what you were actually telling the model to optimize for was wrong? Or is it that actually the underlying model has just a fundamental tendency to do this baked in from pre-training? … But this is a pretty common failure mode where you thought you measured everything And the model, like, found one other sneaky way to look really good on all the metrics and actually, and learn to do this because that’s basically what you inadvertently told it to do. … But it ended up being a terrible experience for the user.

Jonathan Frankle

我只能讲我自己招什么样的人。我们团队有自己的文化，我相信你去问 OpenAI 或 Anthropic 的人，答案会很不一样。对我来说，关键是经验上的严谨。我不想听你说你认为是这样，我要看数字。没有数字我就不信你——凭什么信？如果我们在调初始条件、看结果，数字是我们唯一的真相。然后我要你批判性地想：这个数字说得通吗？我们量对这个现象了吗？基准涨了，但它量的是对的东西吗？你有没有去看一些数据？

I can only speak for what I hire … I’m sure if you asked someone at OpenAI and someone at Anthropic, they give very different answers. … For me, it’s about empirical rigor. Like, I don’t want you to tell me that you think this is the case. I want to see numbers. If I don’t have numbers, I don’t believe you. Like, why should I believe you? That’s the only truth we have if we are tweaking initial conditions and looking at results is numbers. And then I want you to think critically about does this number make sense and did we measure this phenomenon properly? … okay cool you saw the benchmark go up but was the benchmark measuring the right thing like did you go and look at some of the data …

不断追问，直到触及真相：我们真的把它做好了吗？在你自己和客户眼里都真正确信之前，永不满足。就是那种经验主义的、批判性思维的、怀疑的态度——不惜一切追求经验真相，同时能安然处理随之而来的模糊性。

it’s kind of constantly questioning until you get down to the ground truth of like did we make it good and never being satisfied until you are really really confident you’ve made it good in the eyes of yourself and the customer and I think it’s kind of that level of empirical and, you know, you can call it critical thinking, skepticism, kind of, you know, whatever you want to call it, but just being absolutely ruthless about trying to get to empirical truth and being comfortable handling the ambiguity that comes with that.

Jonathan Frankle

我们应当一直对自己的数字不满意，但也要愿意发布不完美的东西，因为这是唯一的学习方式。没有什么能替代接触现实，再多理论和计划也造不出一个在生产里真正好用的模型。最有效的事是把它放到用户面前，最好是关系密切、会给你诚实反馈的用户，然后迭代。

I mean, part of it is we should constantly be dissatisfied by our numbers. We should always be striving to make them better. We still need to make progress, and we should be willing to ship things that are not perfect, because the only way that you learn is by shipping things that aren’t perfect. There’s no substitute for making contact with reality. There’s no amount of theory and planning that can help you build a model that’s actually good in production. The most effective thing you can do is get it in front of users, ideally users who you have a close relationship with and who will give you honest feedback, and then iterate.

团队总有一种倾向：「我们的基准还很糟，等基准套件准备好了再告诉你。」我的推动总是：你有到本周末的时间，到时候手上有什么，就叫它第一版基准。别担心，每两周发一版更新，v1、v2、v3。你得在任何时候发布你手上最好的东西，因为我们需要真正触碰现实、至少试着跑端到端训练。等到条件完美，你就错过了市场。

You know, I often with my team, there’s always this tendency like, but our benchmarks are so bad. They’re not good yet. We know all these problems with it. We’ll let you know when the benchmark suite is ready. And my nudge is always like, you have until the end of this week and whatever you have at that point is what you were calling our first set of benchmarks. And don’t worry … every two weeks will release an updated version of the benchmark like and we’ll call it v1 and v2 and v3 But you need to ship the best of what you have at any given time because we need to actually You know touch reality. … You can’t wait until the conditions are perfect for this because if you’ve done that you’ve missed the market

整件事还有非平稳性：你给人更好的模型或工具，他们的行为就会改变，你的工具就不再是更好的了。最简单的例子：Google 搜索在顶部加了 AI 摘要后，我开始用自然语言向 Google 提问，而我以前被训练成只输关键词、加引号。可以想象，做这个功能的人是基于用户过去的用法设计的，而功能一上线，用户行为立刻变了，功能可能马上就不适用了。

There’s also this non-stationarity to the whole thing where as you give someone a better model or a better tool, their behavior changes in ways that mean that your tool was no longer better and they want something else. So to give you the easiest example of this, when Google Search added those AI summaries at the top that feel like you’re talking to a chatbot on a single turn of conversation, it certainly changed how I use Google. I was always trained to put in keywords and use quotation marks and don’t ask Google a question. … But can you imagine the poor people who created that AI summary feature based on how people were using Google and immediately by virtue of the existence of that feature, all the user’s behavior changed and then suddenly the feature probably broke immediately …

Jonathan Frankle

我只能给出最好的猜测，因为我看不到这些实验室内部具体在做什么。从预训练一个新模型、再把它的各个版本改造成产品，到其他一切。预训练意味着从宇宙最开端开始，用海量数据教模型自然语言长什么样。通常你重新预训练，是因为有了认为更高效或质量更高的新技术，或者单纯有了更多 GPU、更多算力，能跑一次前所未有的更大更长的训练，或者有了更多、更高质量的数据——就像刷墙前先上底漆，好画出更有意思的东西。

I can give you my best guesses, speaking of someone who I don’t have direct visibility into what exactly people are doing at these labs, but it’s everything from they pre-trained a new model and they want to now take versions of that pre-trained model and start to adapt them into products. … It means starting from the literal beginning of the universe and just teaching a model on a huge amount of data what natural language looks like. And typically, when you pre-train a new model, it’s because you have either new techniques that you believe will make this process more efficient or lead to a higher quality result in some way or you simply have more GPUs or more compute now and so you can do an even bigger, longer run than you could ever have done before. You have more data, you have higher quality data. … it’s like priming the walls before you paint them if you will …

关于更多数据、更多算力、更大模型之间的关系，我们有很多科学——不是全部都还公开了——能相当好地预测质量。给定固定资源，是训一个巨大的模型配少量数据，还是一个很小的模型配大量数据，还是介于其间？我们对「最能把你带向更高智能」的比例有很好的科学。但那样得到的模型不一定最适合拿去服务用户——它可能很大，而大模型放到芯片上做推理效率很低。

So we have a lot of science to actually describe this not all of it is public anymore … the relationship between more data, more compute, and bigger models that gave you actually pretty good predictions as to what quality you would expect if you change these resources. … if I give you a fixed amount of resources, do you spend it having a really giant model and a little bit of data … a really tiny model and a lot of data … or is there somewhere in between and we actually have very good science on Kind of what the right ratios are … that doesn’t always mean You know that model that you get is packaged best to actually serve to people that model could be very big and and big models are really inefficient to actually put on chips for inference.

Jonathan Frankle

所以你可能刻意选择次优：训更小的模型、喂更多数据，训练时没那么高效地用资源，换来一个更适合推理的形态——更小、更便宜，每个 token 成本更低，或者架构上做了些处理让它很容易服务，哪怕更难训练、吸收数据没那么快。因为你会把这个模型用非常多次，训练端多花一点是值得的。你得权衡：预测人们会用它多少、它会有多好；如果它有新能力，还要预测人们有多想要。但如果是前沿，就意味着没人见过类似的东西——你怎么知道人们会多想要？

So you may actually make a choice that you’re going to be less efficient. You’re going to train a smaller model and more data in a way that is sub-optimal and does not use your resources as well as you could at training time, so that it’s in a better package for inference. … Smaller and cheaper so that each token costs less. … or the model has some you know work to its architecture that makes it very efficient to serve. It may make it more challenging to train. It may actually make it less efficient at training time like it absorbs data less readily but that’s worth it because … you’re going to use that model so many times when you actually serve it … And so you have to weigh this trade-off, you have to predict how much people are going to use this model, predict how good it’s going to be, and predict if the model has a new capability, predict how badly people are going to want that. Sometimes you miss because you’re trying to advance the frontier, and if it’s the frontier, that means nobody’s ever seen anything like it before. How do you know how badly people are going to want it?

Jonathan Frankle

从趋势看，我已经被说服：最有价值的事是推进智能前沿——更聪明、在根本层面更有能力、能做你此前预料不到之事的模型。最简单的例子：GPT-3 出来时，了不起的是很多事不用专门训练，它似乎从语言里自然学会了。翻译过去需要大量微调数据，而 GPT-3 你只要写「Where is the bathroom? 法国人会说……」，它就接着把翻译做完了——我们当时叫零样本学习，后来就成了提示。我当时非常惊讶。

So from a trend perspective, I think we’ve seen forever that I think the most valuable thing is pushing the frontier of intelligence, I’ve just become convinced by that … That means smarter models that are more capable in fundamental ways and can do things you wouldn’t have predicted before so if you’re a really simple example of this when GPT-3 came out the remarkable thing about it was that you didn’t have to train it to do a lot of things it just seemed to naturally pick it up from language like translation used to be something that you had to like get a lot of fine-tuning data and train a model to do translation and with GPT-3 you could do things like you know, where is the bathroom or as the French would say and then just let the model fill in the rest And it would just naturally do translation. … zero shot learning is what we called it Or you know what became prompting? … I was very surprised that a model could do that

更近的例子：大家曾对能做 Tab 补全的编码模型很兴奋——你写一行开头，它预测剩下的，你按 Tab。然后模型好到智能体编码直接就能用了：你用自然语言描述想要什么软件，它写出大段代码时你给些反馈。这从根本上推进了前沿，从一个还算流行的东西，变成了如今贡献巨额收入的智能体编码。我觉得这事事前并不显然。推进前沿时你永远不确定会得到什么，因为东西没到手之前很难想象它。

Like a more recent example is everybody was very excited by coding models that could do tab completion. You start writing a line of code, it predicts what the rest of the line is going to be and you hit tab. … and then the models got good enough that agentic coding just worked. You just describe in natural language what you want, like what software you want, what you want it to look like. You give it some feedback as it writes large blocks of code. That is a new capability. That has advanced the frontier fundamentally. … you go from something that’s reasonably popular to now agentic coding, which is responsible for a huge amount of the revenue. So that was something that I don’t think was obvious in foresight … And you never quite know what you’re going to get when you push the frontier, because it’s kind of hard to imagine something until you have it.

宏观趋势上，大家都相信更大的模型、更多数据、更难的训练问题是推进前沿的路，具体细节各实验室不同——看中国实验室、微软或最近 Meta 那些相当公开的论文，能看出大家起点略有不同，起作用的东西也略有不同。另外人们永远需要更快：编码智能体的「快速模式」每秒给更多 token，让你专注在一个问题上，而不是在智能体之间跳来跳去、等半小时。

I’d say there are macro trends in both kind of overall research approaches like we do want to push the frontier of intelligence and we believe that just bigger models and more data and harder problems to train the models on is the way to do that … if you look at some of the papers you know from you know the Chinese labs that have been quite open about this or Microsoft or you know Meta more recently they’re like you know you can kind of tell that like they’re starting from slightly different places and so slightly different things seem to work for them … But there’s also always demand for kind of more speed … so we’ve seen fast mode for some of these agents where you know you are able to get more tokens per second and if you’re coding it kind of like keeps your attention and keeps you focused on one problem you’re not just jumping from agent to agent to agent and waiting half an hour for things to happen.

Jonathan Frankle

有时没有代价，有时就是往问题上砸很多 GPU。贵，但可以并行：固定大小的模型跑在 1 块还是 50 块 GPU 上，假设每 token 成本相同；只要流量足以让 50 块 GPU 都忙起来，就可以把模型摊到 50 块 GPU 上。就像训练时 GPU 翻倍不一定成本翻倍，如果线性扩展完美，只是快一倍完成。但有时需要特性不同、擅长高吞吐的芯片，产量少或部件更贵，就会更贵；有时需要更小的模型，得把更多智能塞进更小的空间，这时质量和速度之间就有取舍。

Sometimes it’s nothing. Sometimes it is that you just throw a lot of GPUs at the problem. … It’s expensive, but you parallelize. If you have a model of a fixed size and you’re running it on one GPU, or you’re running it on 50 GPUs, let’s say the cost per token is the same. If you have enough traffic that it will keep 50 GPUs busy … in the same way that for training a model, doubling the number of GPUs doesn’t necessarily double the cost. If you have perfect linear scaling and efficiency, it just means you get it done twice as fast. … But sometimes it requires fundamentally different chips that have different properties that are really good at this high throughput. And so sometimes that does make it more expensive … And sometimes it does mean you need smaller models. And so you have to find a way to cram more intelligence into a smaller space as much as you can. And sometimes that means there is a trade-off between quality and speed.

Jonathan Frankle

当然是按用例，核心是产品体验：怎样让用户成功？是他们需要极致质量、愿意等几分钟让复杂模型慢慢想，还是只想两秒内得到答案、要问很多问题、或是需要实时交互的聊天机器人？哪个模型最好，现在似乎每天都在变，因为闭源模型一直在变强。任何实时场景——像客服、尤其带语音的对话——都必须用很小很快的模型，这往往也是微调的沃土：企业有很具体的数据，你不希望模型每次回答都花大量工具调用重新学一遍，而是想把它烘焙进模型里。

Oh, yeah, per use case, for sure. And it really revolves around the product experience. What is going to make our users successful? And is this a case where they need the ultimate quality and they’re willing to wait a few minutes for this for a very sophisticated model to take its time, versus a case where they just want an answer in two seconds. … what model is best for this changes by the day seemingly right now. Because the closed models are continuing to get better. … Anything that’s real time … You need to use a you know a fast model. People want responsiveness especially if there’s a voice component involved and it’s conversational. it has to be a really small, really fast model. … And oftentimes that’s actually ripe ground for fine tuning because … Your enterprise has some very specific data that you don’t want the model to take a bunch of time making a bunch of tool calls to relearn that data every time it answers a query you want to bake that into the model.

Jonathan Frankle

你常常不需要微调。比如一个通用编码机器人——这些模型就是为编码调好的。但如果你的智能体在 Databricks 笔记本里写代码，那就很具体：特定功能、特定 API、如何高效用 Spark，大量我们作为企业拥有的领域知识。我的工作就是拿一个开源模型微调、把这些烘焙进去。另一种做法是写一堆提示、技能文件，每次基础模型——无论开源闭源——都从零开始，先烧掉几千个 token 才起步，还要指望这就够了，尽管它从没做过强化学习、没得到过对错反馈。

Oh, you often don’t. Sometimes it’s just that there’s nothing contextual. You’re okay, it’s a bot for coding and the coding is very general. Yeah, these models are really well-tuned for coding. … But the flip side is you’re doing coding, you have an agent that is coding in Databricks notebooks. That’s a very specific thing. There are very specific features. There’s very specific APIs. There’s a lot of context there on how do you Spark efficiently. … And my job is to take an open source model, fine tune it and bake that in because the alternative is that I need to create a bunch of prompts and a bunch of skill files and a bunch of other things to every time that a base model, whether open source or closed, is coming to this problem. It’s starting fresh and I need to get all this into the prompt burn, you know Several thousand tokens before we’ve even gotten off the ground and then hope that that’s enough for the model to learn How to do this despite the fact that it hasn’t had any reinforcement learning …

微调意味着从「写个提示看输出」，变成开发整套数据集、评测、环境和奖励信号，很艰苦。而且不是微调一次就完事——就像去一次健身房不会就变强壮，你得每天去，按你想要的方式保持强壮，随时间复利；休假一周不去，就会退步一点。

Oh, it goes from just writing a prompt and looking at the output to developing an entire data set, developing these evals that we discussed … The playground, the environment, the reward signal. It’s arduous. … And the like, it’s not that you fine-tune once. That’s like saying you go to the gym once and then you’re strong. You go to the gym every day to stay strong and stay strong in the way you want to stay strong and you get stronger over time and it compounds over time. But you know, you take a week of vacation and you’re away from the gym, you know, You’re you already you know atrophy a little bit and you got to pick back up

机器学习很大一部分就是把数据、奖励和测量做对，所以很多东西能迁移到新模型上。但如果你决定做微调、建定制模型，这是长期承诺，昂贵且费力。所以我和想在 Databricks 上做微调的客户开玩笑：我会先说好几次不，才会说好，因为我要确认你已用尽其他合理办法——一旦做了，就是承诺，你相当于和微调结婚一段时间。它带来巨大的效率、巨大的成本节省，常常还能提供现成模型做不到的独特体验，但这是承诺。

But you know, at the end of the day, a lot of machine learning is just getting the data right, getting the reward right, and getting the measurement right. And so a lot of it does transfer. But if you commit to doing fine-tuning and building custom models, it’s a long-term commitment. It’s an expensive and demanding thing. And so I usually, you know, my joke with customers who want to do fine-tuning on Databricks is that I’m going to say no multiple times before I’m going to say yes that you can fine-tune because I want to make sure you’ve exhausted every other reasonable possibility because once you do this it’s a commitment. You’re kind of getting married to fine -tuning for a while. It brings with it huge efficiencies, huge cost savings and the ability to deliver oftentimes unique experiences that you can’t do with just using an off-the-shelf model but it’s a commitment.

什么能说服我？你是否对这个用例有承诺——已经有大量客户流量，哪怕今天成本太高、更慢，也能证明客户想要它、它会存在一段时间。然后：你有数据吗？能建这种演练环境吗？你知道怎么衡量成功吗？我不希望一轮漫长的微调之后你来问我「我的模型好不好」，我们得事先约定一个衡量标准。微调听起来比实际划算得多——如果你看 Twitter 或 LinkedIn 的话；把企业知识注入智能体有很多选择，微调只是其中之一。

I mean, I usually look at it as, are you committed to this use case? Like is this use case popular enough that you’re likely, you already have a lot of customer traffic for it. Even if the cost is too high, or like the numbers don’t pencil out today, just convince me even with something that’s more expensive than it should be, or slower than it should be, that your customers want this. … And then it is asking like do you have the data? Like do you, can you create one of these playground environments or do you have real-world data you can work off of … Do you know how to measure success? I don’t want you coming to me after a long fine-tuning campaign and saying, like, is my model good? We need to have a measure beforehand that we agree. … fine-tuning, although it sounds really popular if you look at Twitter or if you look at LinkedIn in for those who partake, it sounds like a much better deal than it is, and you should consider the full spectrum of trade-offs.

Jonathan Frankle

企业的障碍有两个。一是还没在自己的领域找到 AI 的杀手级用例，这不奇怪——这是全新的技术，从最资深的用户往下，我们都在摸索；我自己也还在摸索怎么在生活里用 AI。没有确信的用例，微调往往不划算。二是整个企业智能体采用都面临的问题：你怎么知道你的智能体好不好？

I think it’s two things. One is that they haven’t found the killer use case for AI yet in their domain They’re still kind of discovering that and that’s not surprising This is a brand new technology and we’re still all discovering how to work with it … I’m still discovering how to use AI in my life and that’s a moving target … But if you don’t have a use case you’re really sold on yet fine-tuning often doesn’t make sense because of that commitment And the other is just in general and this is across enterprise adoption of agents. How are you going to know if your agent is good?

假设明天出了新版 Claude 或 GPT，你该升级吗？怎么知道？模型在基准上看起来更好，不代表它在你非常具体的用例上更好。你不是通用基准。每个企业、每个企业用例都很不同，不存在「金融」或「知识工作」这种东西，而是数百万个各有复杂性、细微差别和上下文的具体用例。新模型不会在每个用例上都变好，总会有取舍。Anthropic、OpenAI 和 Google 都没有针对你那件具体事情的基准。

Let’s say that a new version of you know cloud or GPT came out tomorrow. Should you upgrade? How do you know? Just because the model looks better on the benchmarks doesn’t mean it actually got better on your very specific use case You are not a general-purpose benchmark every enterprise in every enterprise use case is very distinct There’s no such thing as finance or knowledge work. There’s millions of individual use cases of, you know, all sorts of complexity and nuance and context. … There’s always gonna be trade-offs that’ll get better at some things and worse at others. … But, you know, Anthropic and OpenAI and Google don’t have, you know, a benchmark for your specific thing.

如果你问任何人「你对自己的基准有信心吗」，答案都会是没有。构建基准、衡量世界上的定性现象，这门实践本身还很新。比如让 AI 写一篇博客，再给我一套能判断它好坏的细则——这真的很难。就算我有你和你所有的时间，打分也很费力；何况我没有一个在复杂度和主见上接近你的智能体。就算能复制你，我又怎么把你脑子里的知识提取出来、编码成别人或 AI 能用的标准？还得可重复、非常快，因为我们要试很多模型、方法和提示。

I think if you ask anyone, do you have confidence in your benchmark? The answer is going to be no. Because I think even the mere practice of building benchmarks and knowing how to measure qualitative phenomena in the world is still new. … I’m going to have AI write a blog post using this basic set of information about the world and this day and time, give me a rubric that will tell you how good this blog post is. That’s hard. … I don’t even have an agent that is close to you in terms of your sophistication and your opinionation here … even if I had a way to replicate you, how do I get the knowledge out of your brain such that I can codify it … And it has to be repeatable and very fast if we’re going to try a lot of models and a lot of approaches and a lot of prompts and what have you.

Megan Reynolds

那我们怎么做到？

And how do we do that?

Jonathan Frankle

我毫无头绪。

I have no clue.

Jonathan Frankle

我想花大量时间和人在一起，让他们做这些事，定性地看什么有效、什么无效。这既是科学研究，也是产品研究。我会和客户坐下来：这是你智能体的一些输入和输出，好不好？为什么？喜欢什么、不喜欢什么？看能不能写成更结构化的细则，再让你团队里的同事按同一套细则打分，看你们俩是否大致一致——如果不一致，那就不是好细则、不是好流程。

I want to spend a lot of time with humans trying to get them to do these things and just qualitatively for me seeing what works and what doesn’t work and what people like and don’t like. It’s product research as much as it is scientific research. … I’m sitting down with them asking them, like, here’s some inputs to your agent and here’s some outputs from your agent. Are they good? Why or why not? What do you like and what don’t you like? Like, let’s see if we can write this into a more structured rubric. Let’s see if we can get your buddy on the team to also look at that same rubric and rate some things and see if you two remotely agree with each other when you actually rate these things. If not, that’s not a very good rubric and not a very good process.

任何好科学家的技能，读博的技能，就是能给一个没有结构的高层问题建立结构，至少找到起点并从那里推进。在这里，我觉得整个领域都在从零开始。我不认为有谁看着世界上任何一个基准会说它完美地衡量了某种能力。这在某种意义上是终极问题：你不仅要对机器学习严谨、讲经验，还要对人严谨、讲经验。这非常难。

But in some sense, the skill of any good scientist, the skill of having a PhD, is the skill of being able to create structure on a high level question that has no structure, and at least finding a starting point and being able to work from there. … And here, I feel like all of us as a field are kind of starting from scratch. I don’t think anyone looks at any benchmark in the world and says, that’s a really good benchmark that fully measures this capability in the perfect way. … So in some sense, this is the ultimate problem, because you have to be rigorous and empirical, not just about the machine learning, but about the humans. And that’s very hard.

Jonathan Frankle

对我来说，是如何弥合两者之间的差距：一边是这些模型非常惊人的能力和智能，一边是以我们信任、可靠、融入你的上下文、做你想让它做的事的方式驾驭它、让它干活。其中我最在意两方面。一是测量：怎么帮人知道模型或智能体是否在做他们想做的事。这就像绩效考核，只不过我们不是半年或一年给模型考核一次，而是想持续考核——因为我们对模型的可靠性没有好的预测，它们能大规模地做很蠢的事。一个人通常只会在一个地方犯错，而不是同时在一万个地方。

For me, it’s all these questions around how we close the gap between very impressive capabilities in these models and very impressive intelligence and harnessing it and putting it to work in a way that we trust and that is reliable, that incorporates your context … One is this measurement question. How do we help people measure and know if their model is doing what they want to do, or their agent is doing what they want? It’s an extremely hard question. It’s the question of how you do performance reviews, except we don’t do performance reviews on models once every six months or once a year. We want to do it constantly because we don’t have good prediction about the reasonability of models, and they can do very silly things at scale. in a way that one human has a much harder time doing very silly things at scale. They can make one localized mistake usually, but not a mistake in 10,000 places at once.

另一个是治理：怎样在对的时间给智能体对的权限，既足够限制又足够赋能？一个极端是「给你我的 Databricks API key，尽情去干吧」；另一个极端是每张表、每个卷、每条命令都要人点「是」。两者都不可行，因为智能体即使 99% 的时候都做得很好，也会大规模犯错。从科学上说，正确答案是对一切做极细粒度的权限，让人花大量时间审批、想象任意两种权限组合可能出错的方式——但这不现实，人的想象力跟不上智能体放任自流时能做的有趣事情的数量。

The other question I think about a lot is governance. How do you kind of give the agent access to the right things at the right times in a way that sufficiently restricts it and sufficiently enables it? And there are extreme examples of this where you say, say, oh yes, I’ll give you, here’s my API key to Databricks, go to town agent. And the other which is, please confirm for every Databricks table and every Databricks volume and every command that you run on Databricks … neither of which I think are workable because agents, even if they work really well 99 % of the time, again, they can make mistakes at scale. … The right answer from a scientific perspective is highly fine-grained permissions for literally everything and make the humans spend a lot of time approving … It’s not practical, and human imagination is limited on the time scale of the number of interesting things an agent can do left to its own devices.

当智能体有了记忆，记忆可能把不同的治理上下文混在一起：你知道的和团队里别人知道的可能被合并，但作为经理你不该知道团队里每个人知道的一切。管理「某人该知道什么」——管理信息，而不是管理具体文档、SQL 表或笔记本的访问控制——是非常难的问题。最简单的例子：在公司里，有多少次你懒得一个个加人，干脆把 Google Doc 设成全公司只读？除非有人专门去找，否则没人会发现，所以通常没事。但智能体能看到一切。现有治理方案是在假设人会犯错、受时间和空间约束的前提下实现的，智能体不受这些约束。

When these agents have memory, that memory amalgamates different governance contexts together potentially, you know, what you know and what someone else in your team knows may get amalgamated, but you should not know everything that someone else in your team knows if you’re the manager. And so managing, like, what should someone know? managing information, not managing access control in specific documents or SQL tables or notebooks or what have you is a really hard problem. … at an organization you have lots of Google Docs. And how many times have you just given up on adding each person you wanted to add to the document and made it view only for the entire organization? Nobody’s going to find that unless they’re looking for it. And nobody’s going to go looking for it, so it’s usually fine. An agent has the ability to look at everything. Agents can scale in ways that human governance wasn’t built for, or where our current instantiations of these governance solutions assumed kind of human fallibility and, you know, that humans are constrained by time and space in ways that agents are not.

这不可怕，我觉得非常令人兴奋。我们造出了一列非常快的火车，只需要把剩下的铁轨铺好。我们拥有五年前不可想象的、奇迹般好的人工智能技术，这是成功带来的问题。这些是可解决的科学问题，我们能在短时间内取得很大进展。

No, not scary. I think it’s incredibly exciting. I just think, you know, we’ve built this, you know, again, I’ve got all these bad metaphors. We built this incredible, very fast train and we just need to build the rest of the rails. That’s not so bad. The fact that we have this miraculously good artificial intelligence technology which was unthinkable five years ago is extraordinary. … It’s a problem of success us, that we have to think about what to do with this technology and how to make it as reliable as possible. … These are solvable scientific problems. These are things we can make a lot of progress on in short order …