跨国深谈

记忆 (memory)

27分钟阅读

跨国深谈← 返回←记忆 (memory)
记忆

记忆 (memory)

概念
本站收录 28 集 · 43 条金句 · 关联 10 个

集里怎么说它

① 提到它的金句

43 条

护城河是关于上下文和记忆的。
the moat is about context and memory.
—— Brian Balfour · [30:40]

指向原始笔记的链接

很难让我们的沙箱发生 OOM 或内存溢出。因为我们可以动态地实时调整大小,这在几乎任何其他东西上都是不可能的。
it’s very hard to OOM or out of memory our sandboxes. because we can dynamically on the fly resize, which is like impossible on almost any other thing.
—— Ivan Burazin · [28:56]

指向原始笔记的链接

其中一个是认知债务。特别是当我们习惯了代码生成得如此之快这一事实，我们也可能开始丧失关于事情如何正确完成的肌肉记忆。
One of them is cognitive debt. And so especially as we get used to the fact that code is getting generated so quickly right now, we can also start to lose our muscle memory around how things get done right.
—— 嘉宾 · [12:37]

指向原始笔记的链接

所以芯片越热，你获得的内存带宽就越多，你获得的 flops 就越多。所以有很强的回报。如果你能冷却得好，那也意味着你能产出更多智能。
So the hotter the chip, the more memory bandwidth you get, the more flops you get. And so there’s a strong payoff. If you can cool well, that also means you can produce more intelligence.
—— Sachin Katti · [07:59]

指向原始笔记的链接

因为那最终并没有建立苦涩教训，在这个意义上，模型可以学会管理自己的内存，比你可以提前为模型凭直觉设计这些内存类型要好得多。
Because that ends up being not very bitter lesson built in the sense that models can learn to manage their own memory much better than you can intuit these memory types for the model ahead of time.
—— Lance Martin · [23:02]

指向原始笔记的链接

这是非常经典的苦涩教训药丸，但模型可以推理它们自己的内存和上下文结构，比你可以为它们规定一种构建自己内存的方式要好得多。
This is very classically bitter lesson pill, but models can reason about their own memory and context structure much better than you can prescribe for them a way to structure their own memories.
—— Lance Martin · [23:50]

指向原始笔记的链接

但随后它好像又说到”更聪明”这个词不对，你知道，我是一个非常快、非常广、非常浅的思考者，没有连续性，我想这很有趣，因为我原以为你们都在致力于记忆，然后显然人类是更慢、更窄、但深得多的思考者，其判断建立在我实际经历过的多年的后果之上
but then it like went into like the smarter isn’t the right word and you know i’m very fast very broad very shallow thinker with no continuity i was like that’s interesting because i thought you all were working on memory and then apparently humans are slower narrower much deeper thinkers with judgments built from years of consequences i’ve actually lived through
—— 嘉宾 · [07:59]

指向原始笔记的链接

我们发现,随着模型演变得更有能力,如果你将逐字稿和记忆状态作为一个周期性批处理过程,并使用我们所谓的做梦,它允许我们提取新的洞察和新的组织结构,这基本上会反馈并编辑记忆,从而让第二天的智能体会话自动地变得更加智能。
What we found is that as models have evolved and become more capable, if you feed the transcripts and the memory state as a periodic batch process with what we call dreaming, it allows us to extract new insights and new organized structures that essentially feedback and edit the memory as needed to make the next day’s agent sessions automatically much more intelligent.
—— Gagan Bhat · [27:45]

指向原始笔记的链接

我想说有时，我不知道这是否有争议，很少有开发者理解内存管理。
I would say sometimes, I don’t know if it’s controversial, that very few developers understand memory management.
—— Addy Osmani · [28:45]

指向原始笔记的链接

所以你越多地使用 AI，它就有点像是对你拥有良好记忆力和理解你正在解决的问题的能力的一种侵蚀。
The more that you use AI, it’s sort of the erosion of your ability to have good memory and have good understanding of the problems that you’re working on.
—— Addy Osmani · [60:53]

指向原始笔记的链接

即便是在公司层面，我告诉人们的一件事是，听着，你必须经历转换系统的痛苦。而且它不会那么高效。是的，当然，旧的方式会更快，因为那是肌肉记忆。但如果你进入下一个层级，不仅你会有一个更适合你的系统，而且每个人都会有更多技能。
One of the things, even at a company level, that I tell people is, look, you’ve got to go through the pain of switching your systems. And it’s not going to be as efficient. And yes, of course, the old way will be faster because it’s muscle memory. But if you get to the next level, not only will you have a system that works better for you, but everybody will have more skills.
—— Claire Vo · [24:15]

指向原始笔记的链接

我会担心任何这样的记忆架构：它不依赖成本更低的嵌入器和重排序器，而是依赖 LLM 的多遍处理来帮你分类并缩小数据语料库。
I would worry about any memory architecture that instead of relying on lower costs embedders and re-rankers is relying on multiple passes of the LLM to help you categorize and shrink the corpus of data
—— 嘉宾 · [66:35]

指向原始笔记的链接

超过四分之一的 B2B 软件买家在用自己的手机上的个人版 ChatGPT，因为它带着他们的记忆。
More than one out of four B2B software buyers are using their own ChatGPT, their own phone, because it’s got their memory.
—— Tim Sanders · [23:33]

指向原始笔记的链接

Harrison Chase 说，你的 harness、你的记忆，如果你不拥有你的 harness,你就不拥有你的记忆，这是关键。
Harrison Chase said, your harness, your memory, and if you don’t own your harness, you don’t own your memory, which is key.
—— Kay Malcolm · [18:20]

指向原始笔记的链接

而且很多团队早期会试图直接跳到最后，说「哦，我要做参数化记忆之类的」——而其实「Clippy 满意并通过所有测试之前不许发布代码」就是一件非常有效、开销非常小的事情。
Um and I think a lot of teams early on try and jump to the end and say, oh, I’m gonna do parametric memory or something, when you know you can’t ship code until Clippy is happy with it and it passes all the tests, um, is a very effective thing with very little uh very little um overhead.
—— Marc Brooker · [47:53]

指向原始笔记的链接

这种工作负载是极其受内存限制的。你把大部分时间花在内存层次结构之间搬运权重上，而你做的算术运算非常少。
That kind of workload is extremely memory bound. You’re spending most of your time moving around weights across the memory hierarchy and you’re doing very little arithmetic.
—— Stefano Ermon · [09:12]

指向原始笔记的链接

把推理离线捕获为标签、语义 ID、记忆,然后让更小更快的模型来服务它。
Capture reasoning offline as labels, semantic IDs, memory, and then let smaller and faster models serve it.
—— Raghav Saboo · [21:16]

指向原始笔记的链接

你可以给一个智能体加上 RAG 记忆，却仍然得不到你想要的长期地平线行为。
You can add a RAG memory to an agent and still not get the current long-term horizon behavior that you were looking for.
—— Erina Karati · [17:12]

指向原始笔记的链接

每次会话,模型都像《初恋 50 次》里的德鲁·巴里摩尔一样醒来——才华横溢,却没有昨天的记忆。
Every session, the model wakes up like Drew Barrymore in 50 First Dates, brilliant and no memory of yesterday.
—— Andrew Orobator · [17:34]

指向原始笔记的链接

内存绝对是今天已经坏掉的头号组件。
Memory is absolutely the number one component that is already broken today.
—— Steven · [07:52]

指向原始笔记的链接

世界上只有三家公司生产内存，占 95% 的市场——美光、SK 海力士和三星。
There’s only three companies in the world that produce memory, which is 95% of the market, Micron, SK Hynix, and Samsung.
—— Adam · [05:08]

指向原始笔记的链接

所以你看到今年仅芯片的内存部分价格就上涨了 700%。
So you’ve seen prices rise by 700% this year just for the memory part of the chip.
—— Adam · [05:27]

指向原始笔记的链接

经济性实际上坍缩为本质上一个东西：你所使用的内存的每 GB 成本。
the economics actually collapsed down to essentially a kind of cost per gigabyte of the memory that you’re employing
—— Walter Goodwin · [14:42]

指向原始笔记的链接

新的 LLM 迫切希望拥有比现在高出几个数量级的内存带宽，以便运行得更快。
a new LLM desperately wants to have orders of magnitude more bandwidth to memory in order to run faster
—— Walter Goodwin · [16:33]

指向原始笔记的链接

拥有推理给你的是自主性,没有人能把模型夺走。但拥有记忆给你的是连续性。
Owning inference gives you autonomy. Nobody can take the model away. But owning memory gives you continuity.
—— Dylan Couzon · [04:39]

指向原始笔记的链接

模型是 CPU,上下文窗口是内存(RAM),而记忆是磁盘。
The model is the CPU, the context window is the RAM, and the memory is the disk.
—— Dylan Couzon · [06:46]

指向原始笔记的链接

超级智能的记忆这一面从来都不是难的部分。我们只是把它指向了我们自己。
The memory side of superintelligence was never the hard part. We were just pointing it at ourselves.
—— Dylan Couzon · [06:16]

指向原始笔记的链接

然后想象五年后的那份记忆——不是一个聊天机器人,而是一个永远不会遗忘你的私人心智,没有任何人能把它关掉,而且它对你的了解超过以往任何系统,因为你是唯一一个训练过它的人。
Then picture that memory five years from now, not a chatbot, a private mind that never forgets you, that nobody can switch off, and that knows you better than any system ever has, because you were the only one who ever trained it.
—— Dylan Couzon · [14:32]

指向原始笔记的链接

每次对话都是从零开始。它不记得昨天在 Slack 里把链接格式搞砸的事，所以今天它只会再搞砸一次。
Every conversation starts from zero. It has no memory of how it fumbled formatting links properly in Slack yesterday, and so it’s just going to fumble again today.
—— Sarah Simionescu · [03:53]

指向原始笔记的链接

所以 MCP 给了智能体一扇通往每个应用的门，却让它们站在成千上万个彼此隔绝的房间里，没有地图，也没有曾到过那里的记忆。
And so MCP gave agents a door into every app, but it left them standing in thousands of separate rooms with no map and no memory of ever being there.
—— Sarah Simionescu · [04:45]

指向原始笔记的链接

还是更自动化的方式，比如用记忆或 prompt 优化之类的工具，这个复利循环非常重要，要构建到你的系统中并由你掌控。
Or more automatically, whether that’s with memory or prompt optimization things, this compounding loop is really important to build into your systems and to own.
—— 嘉宾 · [06:23]

指向原始笔记的链接

但记忆、思考、大脑可塑性，这些全都建立在时间顺序很重要的假设之上。而显然它并不重要。
But memory, thinking, brain plasticity, these things all rest on the assumption that the temporal order matters. And apparently it doesn’t.
—— Andrew Huberman · [37:48]

指向原始笔记的链接

这是每秒 PB 级的内存带宽。
That’s petabyte per second scale memory bandwidth.
—— Charles Frye · [51:23]

指向原始笔记的链接

我看了一下我们一年零一周前拿到的一份内存报价,从那以来它已经涨了四倍半。
I was looking at a quote that we had a year and a week ago for memory, and it’s gone up four and a half times since then.
—— Thomas Somers · [01:02]

指向原始笔记的链接

所以我想说,NVIDIA GPU 在 transformer 模型的实际解码前向传递中,平均只获得 30% 到 40% 的内存带宽利用率。
So while NVIDIA GPUs, I would say, on average are getting between 30 and 40 percent memory bandwidth utilization in the actual decode forward pass of a transformer model
—— Thomas Somers · [34:12]

指向原始笔记的链接

因此,用我们的第一代产品,我们能够达到并维持理论内存带宽的 93%。
And so with our first gen product, we were able to hit and sustain 93% of theoretical memory bandwidth.
—— Thomas Somers · [34:52]

指向原始笔记的链接

只有当那个经验教训被转化为智能体在未来运行中可以参考的持久性上下文时，它才成为记忆。
It only becomes memory when that lesson is converted into durable context that the agent can reference on future runs.
—— Jake Broekhuizen · [04:36]

指向原始笔记的链接

程序性记忆实际上才是大部分可见收益的来源。
Procedural memory is where actually most of the visible gains come from.
—— Jake Broekhuizen · [05:45]

指向原始笔记的链接

如果我们把所有那些证据都变成长期记忆，那会让智能体陷入困境，并且会让它在未来任务上的推理变得更困难。
If we turned all of that evidence into long-term memory, then it would make the agent struggle, and it would make it harder to reason over future tasks.
—— Jake Broekhuizen · [09:20]

指向原始笔记的链接

并非所有东西都应该成为记忆更新。
Not everything should become a memory update.
—— Jake Broekhuizen · [13:16]

指向原始笔记的链接

而且，正如我所说，HBM 是一种糟糕的内存，它只是我们目前拥有的最好的选择。
And, you know, for that, as I’ve described, HBM is a hideous memory. It’s just the best one that we’ve got.
—— Pat Gelsinger · [12:12]

指向原始笔记的链接

但我认为三十年来的首次存储创新即将到来。
But I think memory innovation for the first time in 30 years is nigh upon us.
—— Pat Gelsinger · [26:55]

指向原始笔记的链接

当每个人都想要个人智能体，或者每个企业都想要自己的、全天候 24-7 持续在线并运行大量自动化的个人智能体时，上下文窗口比内存中的模型权重更重要。
When everybody wants personal agents or every business wants its own personal agents that’s always on and working 24-7 and running a lot of automations, the context window matters even more than the model weights in the memory.
—— Arvind Srinivas · [30:26]

指向原始笔记的链接

② 出现在这些集

28 集

③ 关联

点进去有真内容 —— 本页主要出口

智能体 · Anthropic · 推理 · MCP · OpenAI · 护栏 · harness · ChatGPT · 沙箱 · 评估

反向链接