Pickful: Your Shared Crypto Community!

话题广场

Moon

· 1d

Andrej Karpathy 谈基于不可信工作者池的自动搜索：

“我那些将不可信工作者池（融入自动搜索）的设计，实际上有点像区块链。

只不过这里没有区块，而是提交（commits），这些提交可以相互构建，并包含你在改进代码时所做的更改。

所谓的工作量证明，本质上就是通过大量实验来找出有效的提交。”

分布式且无许可的自动搜索 ~= 有用工作量证明这一理念，目前仍停留在高层次的直觉层面，但至少可以说它极具吸引力。

有人需要进一步深入研究。关于尚缺的内容，请参阅 QT。 #ai #加密货币 #区块链

https://x.com/_weidai/status/2035236632161656969?s=20

中文 English

0:00 我认为一个问题是，如果你有一堆节点

0:03 可以进行并行处理，那么你很容易让多个自动研究者

0:06 通过一个公共系统交流什么的。当时我更感兴趣的是

0:09 如何在互联网中有一群不可信的工作者。

0:11 举个例子，在自动研究中，你只是想

0:14 找到那段能训练模型达到非常低验证损失的代码。

0:18 如果有人给你一个候选提交，你很容易验证该提交是否正确。

0:22 很好。比如，有人可能声称他们那段代码优化得更好，

0:26 能给你更好的性能。你可以很容易地检验，

0:29 但可能需要大量工作去检查。不过基本上，

0:32 他们依然可能会撒谎等等。

0:34 所以你基本上面对的是类似的问题，

0:37 其实我设计的含不可信工作者池的系统，

0:40 看起来有点像区块链，因为你不是用区块，

0:43 而是用提交，提交可以相互建立，

0:45 包含你不断改进代码的变更。

0:47 工作量证明基本是做大量试验以找到有效提交。

0:51 那很难。而奖励就是登上排行榜，现在

0:55 没有任何金钱奖励。

0:57 不过我不想把这个比喻推得太远，

1:01 但它本质上有这个问题：

1:03 搜索需要大量计算，但验证候选方案是否优秀很便宜，

1:07 因为你只需训练一次。

1:10 有人可能试了10000次，但你只需要

1:13 验证他们产出的方案有效，

1:16 因为那其他的9990次都失败了。

1:19 总之，长话短说，

1:22 你得设计一个系统，让不可信池的工作者可以

1:25 与可信池工作者协作进行验证。

1:26 整个过程是异步的，可以工作，

1:31 从安全角度来说也安全，

1:33 因为如果有人发任意代码给你执行，那很危险。

1:37 但从根本上说，这是完全可能的。

0:00 think like one issue is if you have a bunch of nodes of

0:03 parallelization available to you, then it's very easy to just

0:06 have multiple auto researchers talking through a common system

0:09 or something like that. When I was more interested in is how

0:11 you can have an untrusted pool of workers out there on the

0:14 internet. So for example, in auto research, you're just trying

0:18 to find the piece of code that trains a model to a very low

0:22 validation loss. If anyone gives you a candidate commit, it's

0:26 very easy to verify that that commit is correct. It's good. Like

0:29 they, someone could claim from the internet that this piece of

0:32 code will optimize much better and give you much better

0:34 performance. You can just check it very easy, but probably a

0:37 lot of work goes into that checking. But fundamentally, they

0:40 can lie and etc. So you're basically dealing with a

0:43 similar kind of, it's almost actually like looks a little bit

0:45 like my, my designs that incorporate an untrusted pool of

0:47 workers, actually look a little bit more like a blockchain a

0:51 little bit, because instead of blocks, you have commits, and

0:55 commits can build on each other, and they contain like changes

0:57 to the code as you're improving it. And the proof of work is

1:01 basically doing tons of experimentation to find the

1:03 commits that work. And that's hard. And then the reward is

1:07 just being on the leaderboard right now, there's no monetary

1:10 reward whatsoever. But I don't want to push the analogy too far,

1:13 but it fundamentally has this issue where you run off search

1:16 goes into it, but it's very cheap to verify that a candidate

1:19 solution is indeed good, because you can just train a single,

1:22 you know, someone had to try 10,000 IZs, but you just have to

1:25 check that the thing that they produced actually works,

1:26 because the 99,000 of them didn't work, you know. And so

1:31 basically, long story short, it's like you have to come up with

1:33 a system where an untrusted pool of workers can collaborate

1:37 with a trusted pool of workers that do the verification. And

1:42 the whole thing is kind of like asynchronous and works and

1:44 and so on. And it's like safe from a security perspective,

1:49 because if anyone sends you arbitrary code and you're gonna

1:51 run it, that's very sketchy and dodgy. So, but fundamentally,

1:55 it should be totally possible.