They Looked Inside Claude’s AI's Mind. It Got Weird

[00:00] AI systems today are really powerful and

[00:03] can do a lot. No question about that.

[00:05] But, how do they really work? We have so

[00:08] many questions. Do they think like

[00:10] humans? How do they beat the best human

[00:13] chess player? How do they beat the world

[00:15] champion video game players? And how is

[00:17] it possible that an AI chooses to not

[00:20] play the game, but just collapse and can

[00:23] trick the brain of another AI to

[00:25] malfunction? Why does Claude think about

[00:28] blackmailing people? I mean, who what is

[00:30] going on here? If you look at the

[00:32] activations inside an AI system like

[00:35] Claude, you see a bunch of gibberish,

[00:38] millions of numbers. Researchers tried

[00:40] to make sense of it for years and years

[00:42] now, but the results were very thin and

[00:45] situational. We now see that it

[00:47] understands that if you look at an image

[00:49] and you have floppy ears, a black snout,

[00:52] and so on, then it might be a dog, a

[00:55] good boy. But, we asked a bunch of

[00:57] questions and still no answers to those.

[01:00] But, now Anthropic has excellent new

[01:03] research with new insights on this. This

[01:05] is when Anthropic is at its best, in my

[01:08] opinion. I love seeing it. Here's the

[01:10] idea. Take this bunch of numbers that

[01:12] the AI thinks about and ask another AI

[01:15] to translate it into text. Translate

[01:18] from machine to human. And it did

[01:22] something. Okay, but these systems often

[01:25] make stuff up. So, how do we know if

[01:27] this is a good translation? We don't.

[01:30] So, what do we do here? Try it

[01:32] separately with a bunch of different

[01:34] models and see if they translated the

[01:36] same way. Is that a good idea? Mm, not

[01:40] quite. Imagine you are a teacher and you

[01:42] give a problem to your students and all

[01:45] of your students write the same answer.

[01:47] Can you conclude it must be true? Well,

[01:50] not necessarily. There are common

[01:53] mistakes in any area and it is possible

[01:56] that it is exactly the mistake they all

[01:58] made. So, what do you do? Now, here

[02:00] comes the genius idea. First, AI

[02:04] translates numbers to text. Then, the

[02:07] second AI secretly guess the text and

[02:10] you ask it to translate it back to

[02:13] numbers. Uh-huh.

[02:15] And what happened here was kind of

[02:17] insane. You see, H is the original

[02:20] thought inside Claude. Numbers, AR theta

[02:24] of Z is translating the text back to

[02:26] numbers. And then, we look at the

[02:29] difference between the two. Translate

[02:31] forward, then translate back, and see

[02:34] how much difference there is. This is to

[02:37] be minimized to ensure the translation

[02:39] works reliably. Do the whole round trip,

[02:42] come back, and if you end up close to

[02:44] the same place, you know that the path

[02:47] is likely correct. But, here comes the

[02:49] part where I fell off the chair when

[02:51] reading this paper. And it is not what's

[02:53] in this formula. No.

[02:56] It is what is missing from the formula.

[02:59] You see, absolutely nothing here in this

[03:01] formula says that the result should be

[03:04] readable. Not at all. Readability

[03:07] emerges because both translators start

[03:10] as Claude, and Claude finds English

[03:12] easier than gibberish. But, it gets

[03:15] better. With this tool, they picked the

[03:17] brain of Claude and found many amazing

[03:19] things. I will highlight what I think

[03:21] are the three best ones. Dear fellow

[03:24] scholars, this is Two Minute Papers with

[03:26] Dr. Károly Zsolnai Fehér. One, it plans

[03:29] ahead. When writing a rhyme, Claude

[03:31] picks the final word before writing the

[03:34] whole sentence. They caught it while it

[03:36] was thinking rabbit, and it went to find

[03:39] something that rhymes with it. Then,

[03:41] they replaced rabbit with mouse, and it

[03:45] actually rhymed with the mouse instead.

[03:47] Sometimes, not always. Really cool. Two,

[03:51] this is going to be super fun.

[03:53] Researchers gave it a math problem for

[03:55] which the answer is 491.

[03:58] And then,

[04:00] they gave it a rigged calculator that

[04:03] returns 492

[04:05] instead. So, what did it do? Well, it

[04:08] had an initial hunch for the solution,

[04:11] and then when the calculator said

[04:13] otherwise, it ignored it.

[04:15] >> [laughter]

[04:16] >> That is incredible. And three, now hold

[04:19] on to your papers, fellow scholars,

[04:20] because it knows when it is being

[04:23] tested, and it gets crazier. It does not

[04:26] tell you that it knows. You have to peer

[04:28] into its mind to get to know that. This

[04:31] sounds like something straight out of a

[04:33] science fiction movie. What a time to be

[04:36] alive. Now, okay, limitations. Let's not

[04:39] get carried away here. One, this is not

[04:42] nearly as easy as it sounds. For

[04:44] instance, you need to find the right

[04:46] layer in the neural network to train on.

[04:48] Also, when minimizing the squared two

[04:51] norm here in this formula, the

[04:52] translation forward is done by one AI

[04:56] and backwards by another. So, based on

[04:59] my experience doing similar things, in

[05:02] simple words, this is very finicky. Lots

[05:05] of trial and error. The result is going

[05:07] to be noisy. Two, despite the headlines

[05:10] you see in the media, this is not a

[05:12] perfect AI mind reader. No, this is a

[05:16] natural language autoencoder. Okay, what

[05:19] does that mean? Well, it is more like a

[05:21] noisy translator. It catches real

[05:24] things, yes, but it sometimes makes up

[05:27] some of the specifics. Three, the cost

[05:30] is bearable. For a 27 billion parameter

[05:34] model, you train 1 and 1/2 days on 16

[05:37] H100 GPUs. And for a frontier model, the

[05:41] cost is substantial. But, despite all

[05:44] these, this work is lovely, amazing, and

[05:47] it makes something previously impossible

[05:50] possible. And two more papers down the

[05:52] line, and I bet it will be done much

[05:54] cheaper and better. What a time to be

[05:57] alive. And now, please, use this to tell

[06:00] me why ChatGPT keeps thinking about

[06:03] goblins. Now, some of these videos come

[06:05] out a bit later because I try to be a

[06:08] bit more rigorous with them. You know, a

[06:10] quick media headline brings in a lot of

[06:12] clicks, especially if you write them

[06:15] with AI. Then you can be super quick,

[06:17] and people do that. But these videos,

[06:20] they come from the heart. Subscribe and

[06:22] hit the bell if you think this is the

[06:24] way to do it. Here you see me running

[06:26] the full Deep Seek AI model through

[06:29] Lambda GPU Cloud. 671

[06:33] billion parameters running super fast

[06:36] and super reliably. This is insane. I

[06:39] love it, and I use it on a regular

[06:42] basis. Lambda provides you with powerful

[06:44] Nvidia GPUs to run your own chatbots and

[06:48] experiments. Seriously, try it out now

[06:51] at lambda.ai/papers,

[06:54] or click the link in the description.