Jeff Dean's Honest Talk: Gemini, Discovery Loop

Welcome back to "Learn English with Podcasts"! Sarah, imagine your old boss speaks in public for the first time after quitting - and says exactly what they really think. Awkward, right?欢迎回到"Learn English with Podcasts"！Sarah，想象一下，你的老上司离职后第一次公开讲话，而且把心里话全说了出来。尴尬吧？

It just did. Jeff Dean - the legend who helped build Google - gave his first public interview after leaving. And he did not hold back.刚刚就发生了。Jeff Dean——帮忙打造谷歌的那位传奇人物——在离职后做了第一次公开访谈，而且一点都没藏着掖着。

The same. He co-started Gemini and was its first technical lead. In this talk, he admitted something surprising about it.就是他。他共同发起 Gemini，也是它的首位技术负责人。在这次访谈里，他承认了一件挺让人意外的事。

He said Gemini was not great at coding in the early days. They cared a bit too late about making it really good at writing code.他说 Gemini 早期在写代码方面表现并不出色。他们把"让它在编程上真正惊艳"这件事，重视得稍微晚了一点。

One big thing: they built it multimodal from day one. Text, images, video, audio - all in one model. He said that was the right call.一件大事：他们从第一天起就把多模态做进去了。文字、图像、视频、音频——全在一个模型里。他说这个决定非常正确。

Yes. He looked back at TensorFlow, the tool that helped millions learn machine learning. He admitted two clear mistakes.聊了。他回顾了 TensorFlow——那个帮无数人入门机器学习的工具。他承认了两个明显的失误。

First, they didn't have eager execution at the start. Later, PyTorch and JAX made it popular, and TensorFlow added it too late. Second, they opened a contrib folder where outsiders added ten different ways to do the same thing.第一，他们一开始没有即时执行模式。后来 PyTorch 和 JAX 让这种模式流行起来，TensorFlow 加得太晚了。第二，他们开了一个 contrib 文件夹，外部开发者往里塞了十种做同一件事的不同方法。

Exactly. He said they should have kept the core simple. But he also praised TensorFlow for helping so many people start in AI. Fair and honest.没错。他说当初应该让核心保持简洁。但他也称赞 TensorFlow 帮了那么多人踏入 AI 领域。很公道，也很诚实。

He loved his time there. But he said a tiny, focused team has a special power: everyone works on one mission, with almost no distractions.他很喜欢在谷歌的岁月。但他说，一支极小而专注的团队有一种特别的力量：所有人都围绕同一个使命工作，几乎没有干扰。

Perfect picture. And today, cloud computing lets a small team rent huge machines. You don't need to build everything yourself. A team of about 10 people can move fast.就是这个画面。而且今天，云计算让小团队也能租到巨型机器。你不需要自己从头搭建一切。大约 10 个人的团队也能跑得飞快。

He said it's a little nervous, but very exciting. The big dream is what he calls recursive self-improvement.他说确实稍微有点紧张，但也非常令人兴奋。那个大梦想，就是他所谓的"递归自我改进"。

Simply: use AI to make AI better. His co-founder Quoc Le once built a system that designs model architectures by itself. One result, the Evolved Transformer, was about 30% more efficient.很简单：用 AI 来让 AI 变得更好。他的联合创始人 Quoc Le 曾经做过一个系统，能自己设计模型结构。其中一个成果，Evolved Transformer，效率大约提升了 30%。

And that's only the start. His new company, Discovery Loop, wants to automate the whole scientific loop - build, test, learn, repeat.而这只是开始。他的新公司 Discovery Loop，想把整个科学循环自动化——构建、测试、学习、再来一遍。

He described a model with the power of 20 PhDs. No single human has 20 doctorates. But a model could understand many fields at once, and send agents to solve sub-problems together.他描述了一个拥有 20 个博士能力的模型。没有任何一个人能拿 20 个博士学位。但一个模型可以同时理解很多领域，并派出智能体一起解决子问题。

Right? And here's the crazy part: a round of real experiments that takes a day, or even a week, could shrink to a minute, or just an hour.对吧？还有更疯狂的：一轮真实的实验，原本要花一天、甚至一周，未来可能压缩到一分钟，或者仅仅一小时。

Ha! But beyond jokes, he shared a real study tip. Instead of reading one paper very carefully, skim 10 papers - even 100 abstracts.哈！不过说正经的，他分享了一个真正的学习建议。与其非常仔细地读一篇论文，不如快速浏览 10 篇——甚至 100 篇摘要。

His point: you want to connect ideas nobody joined before. If you know many things are becoming possible, you see a hard problem in a new light. Then you find five parts are already solvable.他的意思是：你要去连接别人从没连起来的想法。如果你知道很多事正变得可行，你就能用新眼光看难题。然后你会发现，其中五个部分其实已经有解法了。

He also uses quick order-of-magnitude estimates. Like: will moving this data take 10 seconds, or 100 years? Those are completely different worlds.他还习惯做快速的数量级估算。比如：搬运这批数据要 10 秒，还是 100 年？那可是完全不同的两个世界。

Now the serious turn. He warned that AI in cybersecurity can already match top human hackers - maybe even go further.现在说点严肃的。他警告，网络安全领域的 AI 已经能媲美顶级人类黑客——甚至可能更进一步。

Yes, on both sides. The same tool can attack, but it can also find and fix real-world holes before the bad guys do. It's a double-edged sword.会，而且两边都行。同一个工具能攻击，也能在坏人动手前，找出并修补现实世界里的漏洞。这是一把双刃剑。

Beautifully said, Sarah. And that's a good place to pause. If you enjoyed Jeff Dean's honest talk, tell us what you think - and we'll see you next time on "Learn English with Podcasts"!说得太好了，Sarah。我们就停在这里吧。如果你喜欢 Jeff Dean 这次坦诚的分享，告诉我们你的想法——下期"Learn English with Podcasts"再见！