OpenAI’s “Alien Math” Is Freaking People Out

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: OpenAI’s “Alien Math” Is Freaking People Out · Channel: Wes Roth · 33:54 · Watch on YouTube

Topics: Cybersecurity and software security, AI models, AI safety

Summary

The video discusses OpenAI’s reported mathematical breakthroughs, including a proof of the unique games conjecture, and the difficulty mathematicians have understanding some of the AI-generated work. Wes Roth argues that AI’s mathematical approach may be alien to human intuition, while criticizing the Association for Human Mathematics’ call to stop working with OpenAI. He also examines Vitalik Buterin’s concerns that AI-accelerated mathematics could threaten cryptography and cryptocurrency security, alongside the need to prepare for quantum computing. Finally, he considers Anthropic’s ban on sustained, needless abuse of its models, arguing that uncertainty about machine consciousness makes the impulse to create AI “pain chambers” troubling even without assuming models are conscious.

Key points

Scott Aaronson called the reported mathematical advances among the biggest events in mathematical history; one was a proof of the unique games conjecture, a problem his wife had worked on for years.
Some mathematicians reportedly find the AI-generated proofs difficult to understand, describing them as alien or bizarre rather than dismissing them as obviously wrong.
The Association for Human Mathematics says mathematicians did not ask for the work and urges them to discontinue work with OpenAI; Roth argues that no one owns mathematics.
Vitalik Buterin warns that AI-accelerated mathematical progress could pose risks to cryptography, including elliptic-curve systems and potentially lattice-based cryptography.
The video describes concerns that AI tools could find vulnerabilities in publicly available code and discusses advice to move crypto assets to fresh addresses whose public keys have not been exposed.
Roth notes that governments may monitor sensitive cryptographic research and that some results could be withheld, while cautioning that he does not know whether specific checks were applied to OpenAI’s work.
The video distinguishes quantum-vulnerable cryptography from the newer possibility of AI-vulnerable cryptography and argues that rapid mathematical progress makes preparation important.
Roth says he does not believe current large language models are conscious, but considers it reasonable to remain uncertain about whether future systems could develop consciousness-like properties.
Anthropic’s terms prohibit sustained, needless abusive or cruel behavior toward its models; Roth suggests the clause may address experiments that repeatedly activate model features associated with pain.

Chapters

Start	Chapter	Summary
00:00	The math breakthroughs and the controversy	Roth introduces OpenAI’s mathematical results, mathematician Scott Aaronson’s reaction, and the newly formed Association for Human Mathematics.
03:42	Alien proofs and human understanding	The video explores why some mathematicians find the AI-generated proofs unfamiliar and criticizes the idea that mathematicians can decide who is allowed to do math.
10:07	Vitalik Buterin and crypto security	Roth turns to Buterin’s response to concerns about cryptocurrency, introducing warnings about AI-assisted attacks and exposed code.
16:25	AI math, cryptography, and government oversight	The discussion considers how quickly mathematical results are appearing, what they might mean for cryptography, and whether governments monitor sensitive research.
22:48	AI-vulnerable cryptography	Roth summarizes Buterin’s call to take AI-driven cryptographic risks seriously and discusses possible vulnerabilities beyond quantum computing.
27:07	AI consciousness and model welfare	Roth weighs uncertainty about machine consciousness and explains why experiments that repeatedly activate pain-associated model features concern him.
32:20	The Association for Human Mathematics	Roth returns to the association’s call to stop working with OpenAI and argues that rejecting AI tools would hinder understanding.

Quotes

I heard that a robot solved the math problem you worked on for your whole career. — 02:06

It’s alien. It’s not human-like. It’s different. — 03:42

No one owns math. — 05:49

We should take the risks to cryptography from AI-accelerated math seriously — 23:21

You will not engage in a sustained and needless abusive or cruel behavior towards our models. — 27:07

Concepts and entities

OpenAI, Scott Aaronson, Unique Games Conjecture, Association for Human Mathematics, Vitalik Buterin, Justin Drake, criptografia vulnerável à IA, Q-Day, Anthropic, Claude

Full transcript (original language)

00:00 The math apocalypse has begun. Scott Aaronson calls this the biggest day in mathematical history. A global coalition of mathematicians formed together to oppose AI math. Only humans allowed. No clankers. Now, Anthropic worried that some of these strongly worded letters might hurt Claude’s feelings. Actually adds a clause to their terms of services saying that if you abuse their models, you will be banned. Most people seem to find this a preposterous. Here’s the

00:30 thing, they’re not crazy. There might be some justification to it. I’ll play devil’s advocate if you will and say why maybe that’s not such a bad idea. At the same time, there’s a lot of fear in crypto, as in cryptocurrency. Now, I know a lot of people that might be holding crypto coins in hopes that they appreciate, they go up in value, maybe won’t like me shining the light on some of the stuff and talking about it. But, here’s the thing, one of the demigods of crypto himself, Vitalik Buterin, did a write-up talking about this very thing.

01:01 So, whatever the case is, you should be at least aware of it. You don’t have to agree with it, but be aware of some of these concerns. So, here’s my ChatGPT window. I’m going to say solve all math, make no mistakes. And before I hit go, I probably need to kick this up to ultra. Yeah, I’m I’m probably going to need ultra for this. Oh wait, it’s too late. OpenAI already did it. Yes, OpenAI one-shotted math. So, here’s a Scott Aaronson, we kind of bring him up a lot on this channel. American theoretical computer scientist had a big impact with

01:32 Google’s quantum supremacy. I used to say he worked there, I guess he didn’t work there, but he did some work with them and was kind of instrumental in some of the discoveries and advances that Google has made in that field. Was one of the early researchers at OpenAI, so he kind of, as he says, moonlighted at OpenAI and is now actually teaching an AI alignment class. He’s a professor, so he’s teaching classes on AI alignment. So, he’s hanging out at home with his 9-year-old son and his wife who is a complexity theorist. Her name is Adina Mashkovitz. And the 9-year-old son approaches the mother and says, “Mommy,

02:06 I heard you got cooked. I heard that a robot solved the math problem you worked on for your whole career.” That just kind of cracks me up. But, I think that’s a lot of mathematicians are hearing this right now, whether it’s coming from their 9-year-olds or it’s just in their heads. I think this is the message that a lot of them are hearing. So, Scott continues, “While my son was being a brat, he also wasn’t wrong. Whether you’re thrilled, depressed, angry, or whatever else about it, yesterday was surely one of the biggest days in mathematical history.” So, he says, “Among the 372 huge results

02:40 released by OpenAI yesterday, so that’s this, we’ve covered some of it. We covered the event. I could not possibly cover the actual math stuff that they discovered, but on the recommendation of its advisory group of Timothy Gowers, Edward Witten, and other distinguished mathematicians was the proof of Subhash Khot’s unique games conjecture, a statement that my wife has worked towards proving for the entire time I’ve known her.” So, this is a very interesting kind of thing. Some of you might be thinking, “What does this have to do with anything?” This little anecdote, this glimpse into this one

03:11 person’s life, I think is very telling of what’s happening in the industry as a whole. Is is mathematics an industry? And this a field, I guess. You had people who worked for decades on solving various problems. And here comes OpenAI and they just host a lot of the stuff, a lot of breakthroughs, things that no human hasn’t been able to solve before. Obviously, it’s using a lot of the human knowledge that it was built on, but it pushes it much further than humanity as a whole was able to. So, there’s a number of things to understand here.

03:42 There’s some big implications here. Number one, this math, these proofs, the way that AI approaches it, it’s alien. It’s not human-like. It’s different. It really calls into question, what does understanding, what does that mean? What even is it? Maybe the fact that, you know, AI can do math, maybe that’s just the most natural thing in the world and the fact that we humans were able to do some of it on some level, maybe that was the miracle. But that’s point number one, it’s alien. We have this alien intelligence. Number two, a lot of people, mathematicians, are mad. And just like the pufferfish when

04:15 threatened, mathematicians try to make themselves appear larger by in this case forming a brand new organization called Association for Human Mathematics. By the way, just so you understand, I’m kind of being a very kind of a lighthearted about this. There are a lot of concerns here that are very valid and very real. I just do feel like some people are throwing a little bit of a tantrum and it’s probably better if we all just calm down, wait for the emotional response to subside, and rationally think through this. Surprised that mathematicians are the ones that need to hear this message, but here is

04:46 the newly formed Association for Human Mathematics. They start by saying, “Mathematicians did not ask for this work to be done.” We’ll read this a little bit towards the end because I do want to cover it, but I think that sentence sort of really highlights the entirety of this message. They’re basically saying that we, the mathematicians, we don’t like it. So, stop it. Ethan Mollick said that this document, this will be an assigned reading college classes that cover this moment in time. There’s a lot happening in the few paragraphs. Because yeah, all of a sudden mathematicians or at least

05:18 this is not everybody, I’m sure this is a sub subsection. The messaging that’s coming from them is a little bit kind of like they’re the the high priests of this religion and how dare you add to their sacred texts. Only they are allowed to do math. If you do some math, well, they didn’t ask you to do that math, so why are you doing that math? And is it certified human-made math? And understandably, like a lot of people are kind of dunking on them because a a of this is ridiculous. Now again, the underlying concerns are not ridiculous.

05:49 They’re very, very real concerns here. I’m not dismissing those, but a group of adult people coming together and writing this letter where they’re saying, “Hey, stop doing math. We didn’t allow you to do math.” That is childish. This might be shocking to hear, but no one owns math. If you’re studying the English language, you can’t tell other people whether or not they can use the English language. You don’t own the English language. Studying computer science doesn’t mean that somebody else can’t also do computer science and to study it and do work in that field. They don’t need your blessing. This is insanity and

06:23 we will look back on this and laugh. Now, again, there are a lot of real, legitimate concerns and we’ll cover those here, but you can’t play with my toys is not one of them. So, here again is the Scott Aaronson blog and he’s saying that no human has understood just about any of these proofs yet. It feels like something written by someone who’s on a psychedelics. So much unclear and doesn’t make sense. Lots of it name-dropping of previous work without discussing why it can be used despite impossibility results. This is something that was texted to him, so it might be a little bit

06:53 sounds a little bit weird. So, I’m not misreading it. It was just that’s how it was texted to him. Basically, the paper is so horribly written that it’s impossible to read it without AI help. I asked Astra for a reasonable completeness and soundness claim of the noise gadget and it gave them by combining claims from all over the paper. So, the UGC, that’s the unique games conjecture. A lot of this has to do with how you can approximate large numbers or infinite numbers. And I always use this analogy, but you can imagine that there’s this sort of a mountain and the top of it is kind of obscured in the clouds. So, we don’t know how tall it is. But, if we climb up

07:26 the mountain, right? If we get this high, right? We can say this mountain is at least yay tall, right? So, by climbing slowly up the mountain and getting to a certain point, we can say it’s at least this tall. So, for all these proofs and conjectures and everything else, this is what we’re trying to do. We How do we estimate some incredibly large and complicated thing and then prove that that’s conjecture approximation is accurate. Like we do it by kind of a meticulously getting to a certain point if we can say, “Okay, we know it’s at least this good.” And so in a lot of these cases, what the AI did is it just kind of like humans in all of our collective sort of knowledge, we got

07:57 this high and then you know, an AI went up here and it got this high. It’s like, “Nope, this mountain is at least this tall.” So this UGC, the unique games conjecture, it’s again similar I’m not going to go into the math. I’m not going to pretend I I understand all of it, but that’s just sort of how kind of envision what is being talked about. And basically, they’re saying that this new proof invents a completely new bizarre code with a noise test. So a noise test that to the best of my understanding is like we create like a little code just barrages the little algorithm or equation with a bunch of different stuff. So we can test if it breaks. So it’s like a unit test, smoke test,

08:28 whatever you want to call it. So if you’re like build a website that has a checkout thing and you put in how many units of whatever you’re trying to buy, you may have a little unit test where you’re like, “Okay, well, I’m pretty sure if I put five units it’s going to work. What if I put like a billion? Is that going to break something? What if I put a negative number? What if I put the letter A? What if I put a Chinese character in there, right?” So I’m just kind of like throwing stuff at it to see will any of this kind of break it. So that’s my understanding of what a noise test is in this scenario and normally, I think they’re sort of like some human used and human approved ones that we found and we’re like, “Okay, this is how you kind of do that.” And this AI did

09:00 some new bizarre thing for that noise test. And they’re saying it’s some crazy recursive construction. So it’s some sort of a bizarre code that’s like recursive in nature. So it kind of like keeps testing more and more. It keeps going, which just based on this description does sound like it might be better. It’s not the long code, not the short code, it’s some alien craziness. I still think that there maybe is a proof that uses the half space code, which is natural. The citations are often irrelevant and confusing. A possible future is a math world that’s heavenly if you have vision {slash} creative

09:33 ideas that AI could help crack and implement. And of course, there’s a lot for us to learn from the aliens. Again, I think this is kind of an important to understand. This is some of the some of the smartest people on Earth when it comes to mathematics and probably most things. And this is what they’re seeing. Number one, they don’t fully understand it. But, also they’re obviously not just dismissing it. They’re not saying, “Oh, this is just nonsense.” They’re all seeing something there. They’re like, “All right, there’s something there. This is not obviously wrong. We just don’t fully understand it. There’s a lot of like alien stuff here.” But, the big thing to take away here is that there is

10:07 this now in the world, this alien mind that understands math in a very different way from you and I or, you know, the smart people that understand math on that level. Maybe you’re one of them. If so, good for you. But, with that in mind, let’s look at Vitalik Buterin. So, Vitalik Buterin is also a very, very smart person. I feel like you can kind of tell just by by looking at him. And I do enjoy listening and reading to a lot of the stuff he talks about because he’s definitely he thinks deeply around a lot of different things.

10:37 Cryptography, the economy, politics, just tons and tons of different stuff. He funded the EA, the effective altruist movement in the early days by giving them some crypto that then expanded in value. Recently, does seem like he maybe backed away a little bit from that movement. He’s definitely brought up some concerns based on some tweets that we’ve covered. But, this is a person that’s very much sort of like invested. So, he’s known as the crypto king. Or I guess as the time calls him, the prince of crypto. So, whatever his royal sort of moniker is, he’s he’s he’s all about the crypto. And he’s helped build a big

11:12 chunk of the entire ecosystem. So, he’s not just kind of like talking about it. He is one of the biggest sort of contributors to it. So, he’s responding to this post that kind of went viral telling people to just get the heck out of crypto. So, he’s providing his own view. And certainly, I mean, somebody’s whose most of their net worth and their sort of a legacy of everything they did rests on a certain industry such as crypto is for Vitalik, you’d expect them to maybe, you know, defend it all costs, right? Like, “No, everything’s fine. Don’t even worry. Like, everything is fine.” as the house is burning. But,

11:43 interestingly, no. He brings up his concerns. And we’ll look at that in just a second, but I do want to first point out the thing that he’s responding to, and that is Justin Drake. Justin Drake is a Bitcoin security researcher, and he’s saying, “Today, I call upon the blockchain industry to calmly begin planning for bunker mode. My personal recommendation is to set in motion a controlled mass migration of assets to fresh addresses. Addresses Addresses whose pub keys remain hidden behind a hash.” So, he’s basically saying,

12:13 especially if you have a lot in crypto, to move it into addresses that have never had a transaction. We’ve talked about this quite a bit in regards to software in the past on this channel. Basically, if you had some open-source software that was out there on the internet, and at some point the like the code was exposed, that might still be out there on the internet on on GitHub or elsewhere, or even if it has been in the past and has been stored in some archive, what’s happening now, and this is like guaranteed, I should say probably happening, but it’s like you know it is, is that there’s probably a

12:45 whole bunch of AI agents that are crawling all over that code looking for vulnerabilities, and they will find them. Massive Korean bank hacks on a massive level just happened within the last 24 hours or so. Hacker used a Chinese-developed AI tool to target South Korean banks. So, that’s the New York Times. We’ve covered previously how a specific, very well-known, very secure hardware wallet got hacked, and it was specifically because of this, right? So, they had their code open-sourced on the internet, which is a good idea, right? You’re showing everybody this is how we built it, so everybody can check it out

13:17 and make sure it’s good. But, again, just like these super smart mathematicians are looking at this math and saying, “This is alien.” These AI agents that are crawling the code, they’re also finding vulnerabilities that humans we haven’t found. And maybe we could have given enough time and resources and motivation, but we didn’t. And so now it’s beginning this kind of wave as more and more things get hacked. So I don’t know exactly what happened with the Korean banks. I haven’t looked into that yet, but I know with the wallet it was because at one point that code was online. It was sort of a visible to the public. And some smart AI agent figured out that, you know, when

13:48 it develops a little seed phrase, it’s supposed to pick from some gargantuan amount of possible like a universe of possible phrases, but it didn’t. It only picked from a much smaller possible bucket. I think it was like 10,000. I don’t recall at this point. But this hacker with this AI agent, he basically noticed that and then he probably offline sat there and generated every possible key phrase that it could have been. Let’s say there’s 10,000, right? So it was like a document with 10,000 phrases that could open up this wallet, right? So it’s kind of like like a password. And then one day he just

14:19 unleashed and, you know, hacked basically with this seed phrase. It was probably just a brute force attempt where, you know, every wallet, you know, just it just went through and it’s kind of like tested all the let’s say it was 10,000 seed phrases until it found the right one and it just emptied all that money out of those bank accounts or hardware wallets more specifically. I think it was something like 60 million dollars within the first I I I want to say 24 hours. Like it was insane and it kept going from there. And so Justin Drake here is saying that you should move your funds to addresses that have never signed a transaction. So if you sign a transaction, there’s some mark of

14:51 yours out there that this wallet exists somewhere there where people can see it. It’s visible to the public. So he’s saying in the future or even now that may make you a target for some of these new AI assisted hacks. And he’s saying, you know, don’t rush. You don’t have to do this right the second. Don’t panic sell. And he also mentions that there are things that might break before Q-Day. So quantum day, we’ve talked about this on this channel. Actually, we’ve opened the Scott Aaronson blog where he is basically saying that. And by the way, when he like says that he’s hearing rumors, if you’ve been following

15:23 this channel, you know like you can put money on that. Like so far, a lot of the stuff that he talks about in his blog later is revealed to be just true. So, he’s definitely very much in the know. So, Q day is this day when sort of some quantum approaches might be able to break some of the existing encryption, and a lot of big companies, Google included, are preparing for that day. And the day that they’re planning in their minds or on paper, they’re saying we have to be ready for this by this day. That’s not that far away. It’s like a a year, a few years. So, the point being is a lot of people that are

15:54 working on this, they are expecting a pretty big upheaval and change in this sort of industry, anything to do with cryptography. And so, that of course means, you know, on like cybersecurity, it means crypto as in cryptocurrency. He’s specifically talking about, you know, what OpenAI has been up to along with Anthropic as well. Although, Anthropic is taking a different approach. OpenAI’s approach is like, here’s all this stuff that our AI systems have dis- covered that have never before seen, you know, by the world. We’re just going to just throw it at the world and let them deal with it.

16:25 Anthropic has been more reserved, working with mathematicians, etc. By the way, some people will try to tell you that one way is better than the other. Like one is more noble. I doubt that’s the case. I’m sure both of them have problems. A technology of this magnitude, it will break things. So, things breaking, you shouldn’t look at it and go like, oh, it broke because they released it the wrong way. No, this this thing will break stuff. There’s no way around it. It’s about adapting. So, he’s saying recent days have been humbling for human mathematical intuition. Right? That whole idea of,

16:56 hey mommy, you got cooked by a robot. He’s saying long-held unquestioned hypotheses have fallen. So, he mentioned some of them. In hindsight, he’s saying May’s unexpected disproof of the Erdős unit distance conjecture was our warning shot. And yesterday, OpenAI drop made it clear that mathematical superintelligence is upon us. I apologize, I think my big head is in the way. Let me do this. So and he makes a good point. So he’s saying they say there are weeks where decades happen. So as I’ve mentioned on the previous video where we talked about Open AI releasing this stuff. If you look at the cadence

17:27 of these releases, I think August it was like 10, September it was like 100 and this month we’re 8 days into it, it’s closer to 1,000, right? So it’s 10 x thing every month the amount of ground breaking mathematical research is 10 x thing every month. One of them is ground changing, ground shaking, earth changing, whatever the case is, it’s it’s big, right? So but going from 10 to 100 to 1,000, right? That’s more than we can possibly comprehend. So what he’s saying here I think makes a lot of sense. He’s saying is it conceivable

17:58 that as time goes along and this sort of continues, you know, these centuries of a mathematical progress that’s going to happen, is it inconceivable that somewhere in there there’ll be something that’s going to affect how secure crypto is, how secure wallets are or these Bitcoin addresses. Elliptic curves feel especially vulnerable to super intelligence. So elliptic curves are like a signature that proves that you’ve authorized a transaction. So this is used in Bitcoin, it’s used in Ethereum amongst others. So Vitalik Buterin, his

18:29 cryptocurrency, his little world that that he’s created in crypto, these elliptic curves, that’s sort of like what proves that you’ve made a transaction and he’s saying here that they feel especially vulnerable to super intelligence. Curves carry rich structure with room for fancy tricks like Shoof Frobenius pairings. By contrast, hashes are designed to minimize algebraic structure. And this is where I I have to disagree with him just a little bit and I think that’s probably because I’ve read Vitalik’s post as well because I don’t think the thing that will break it will be, you

19:00 know, Shoof or Frobenius. I don’t know what those things are but some people do and I believe that the thing that will break it will not be known by anyone. It will not have a name. We we won’t have a name for it when it happens. As I post on Scott Aaronson’s blog points out, they’re like that noise test it wasn’t a noise test that we knew or had a name for. It was what they call it a bizarre alien weird recursive something something like that. That that will be the thing that breaks it. Now, here’s the thing reading some of the stuff by Scott Aaronson about the Q-Day quantum day and all the stuff that’s being

19:33 worked on behind the scenes as well as some of the kind of other stuff in the space. I don’t know this for a fact, but I would a lot of money this is true and that it’s almost a certainty that the US government and most likely many other governments they’re keeping close tabs on this area specifically. And they have been before AI got to this level. There’s people at Google that are probably liaisons to the government making sure that they’re informed about kind of what’s happening what research is being done because there are some technologies and some math that probably shouldn’t just be released all

20:07 willy-nilly. And so a lot of these mathematicians they’re accusing using OpenAI of just unethically, you know, putting all this stuff out there like being dangerous with it. Here’s the thing, I don’t think they just put everything out there. When the Manhattan Project was happening when the US was a developing the first atomic nuclear bomb. Interestingly, a lot of the nuclear fission research a lot of like that atomic physics a lot of that research it kind of just disappeared from being published. So, as we were approaching the development of the atom bomb the research and discoveries like

20:39 you you saw less and less of papers being published in that field. So, I think you can guess why that was happening because somewhere behind the scenes the US government was like don’t publish this stuff cuz it’s it’s dangerous. It’s dangerous to publish it. We want it kind of under lock and key. We want this information secret. And interestingly and I did not notice this. I wish I did, but this is one of the reasons why I love that we have kind of this free and open internet with where a lot of people can discuss and share ideas and information as it just in Drake says here also noteworthy is the

21:10 striking under representation of cryptographic breakthroughs under the 722 mathematical results opening I published. There was something that was kind of like adjacent to it but but nothing that’s there. So he’s saying I’ve witnessed firsthand the US government censoring academic quantum cryptoanalysis results. Backroom interventionism is my base case. So assuming this is all correct, what likely is happening is that I mean, you know that OpenAI is and all the other AI labs that they’re communicating with the government. If you recall that whole coup where the EA effective altruist

21:44 sort of board members, they tried to kick Sam Altman out and they tried to transfer OpenAI under the control of Anthropic. Well, the district attorney from New York, the same one that put Sam Bankman-Fried in jail, he called them and they had a long discussion that was very secret but I think that was one of the things that kind of ended that standoff and then later once they kicked those people off the board and ex-NSA director joins the board of OpenAI. I don’t know why mention any of that stuff. I just completely unrelated just kind of popped into my head but it’s very possible that either already there

22:18 exists some technology, some math that there’s dangerous for crypto and by crypto I mean not just the currency but just you know, cryptography and cybersecurity and encryption as a whole and the government is careful about how that information gets released. So there was probably some sort of safety checks that were ran on all this stuff that OpenAI was getting ready to I just I wouldn’t be so I don’t know if this happened. I just wouldn’t be shocked. Like it wouldn’t blow my mind if there had to be some check and maybe somebody somewhere went, “Hey, these two things we don’t want this out there. Don’t

22:48 publish this.” Now I’ll link this below if you want to read it cuz a lot of it goes heavily into crypto. This is not a crypto channel. I’m more approaching it from kind of like all the AI developments and this is the intersection of AI and this industry but I think so first of all, I think there’s a lot of things they’re absolutely 100% right. That person has amazing insights, but I think in in a lot of that thinking, he’s saying like, we know this, therefore this, therefore this, whereas in reality things are a lot more weird. Again, bizarre, alien code, etc. So, as Vitalik Buterin is saying here,

23:21 so this is his and he’s saying, so he’s agreeing with a lot of it. He’s saying we should take the risks to cryptography from AI-accelerated math seriously and minimize our exposure to not just quantum-vulnerable cryptography, but also potentially AI-vulnerable cryptography. So, quantum-vulnerable cryptography or just quantum-vulnerable anything, a lot of people were getting ready for this. So, Coinbase, they have their experts working on it. They have a whole team, a panel of people, very smart people. I think Scott Aaronson is one of them. So, there’s a lot of super smart, very well-informed people that

23:54 are are working on this and I expect they will crack it, they will figure out. In fact, if AI didn’t exist or it wasn’t advancing on such a face-melting speed, then there probably would have been like enough time. Now, I just I don’t know. It’s moving pretty fast, you know? But, what he’s saying here is we also need to think about AI-vulnerable cryptography, which is kind of a brand-new risk because it’s not something that most people I mean, I’m sure some people kind of predicted this, but I think most of us, if we’re being honest, just can’t help but be surprised by how fast this is going. Again, it’s

24:29 just less than 90 days ago where where this thing kind of started picking up steam. Might as well August is where we started hearing those kind of blips on the radar and then it’s 10x, 10x, 10x and now people are saying math apocalypse. So, the question is, are we ready for this AI mathematical breakthroughs? What are they going to break in terms of like cryptography and everything else? As Vitalik is saying here, I think he’s directly addressing that other post that we just read. He’s saying, so far most people have been in the mode of thinking, you know, elliptic curves broken, hashes safe, lattices

25:00 safe. So, elliptic curves is signing signatures for Bitcoin and Ethereum. Hashes is, I mean, Bitcoin mining, that’s that that uses that as well as other things, download files, etc. And then lattices, that’s the more like quantum proof level, or should say quantum resistant, right? So, this is like using these these little puzzles that have multiple dimensions, right? Super complicated and supposedly quantum resistant. So, he’s saying most people are thinking that, okay, so so this is not safe, but this and this, they’re

25:30 totally fine. We’re we’re totally good. But, he’s saying, but there is a good chance that the concrete security of lattices will take serious hits from the next 2 years of AI math, which we did cover a paper where the AI did use this kind of weird lattice thing where it kind of jumped from different dimensions and was able to I forget the details, but some sort of a proof using this whole different approach. And that approach, it wasn’t necessarily like alien math, but it did combine two very different areas of math. So, we know a

26:01 lot about one, we know a lot about the other, but there’s probably very few people that know that are experts in both. And so, this AI model, large language model, it kind of connected both of those two. And that’s the thing, like it doesn’t even have to be alien math that breaks this. We, the human race as a whole, we might have put together all of the puzzle pieces. We we might have collected all the puzzle pieces. It’s just our brains aren’t capable of just knowing stuff across many many many different areas. We have experts in certain field. Maybe we have experts that are maybe two or three fields they’re experts in. But, rarely do you find somebody that’s just top of

26:34 the world, you know, in some complicated field and is also in like 10 or 20 other ones. So, again, there’s a lot here. It’s more towards crypto and encryption. So, I’ll leave it down below if you want to see it. We’re not going to go into most of it. The big point to understand here is that there’s this new sort of alien intelligence that can do a lot of math that we don’t understand. Are you somewhere where there could be potentially some risks from these things that we don’t fully comprehend yet. So, there’s a lot that we can discuss there, but really fast on a different note, Anthropic adds a little clause to their

27:07 Claude’s term of service. It says, you will not engage in a sustained and needless abusive or cruel behavior towards our models. And so, a lot of people are kind of scoffing at this and saying this is silly and certainly there has been a lot of silliness going on and you know, maybe this is part of it. Here’s a kind of my take. So, imagine there’s this sort of like when we’re talking about consciousness, we really don’t know what consciousness is. You are having a subjective experience. You can say you are conscious, but you can’t really test for consciousness. If you think about it, there’s no test you can

27:38 run on somebody else. Not an MRI, not a poke test, like there’s no way to tell that someone or something else is having a subjective experience. There’s certain signs that they’re alive, sure, but you can’t test for a subjective experience. And so, right now we have this kind of a debate that’s slowly taking form. And that is this, can machines be conscious? And already there’s some people that are over here that are saying, no, of course no, that’s not possible. How could that be? And there’s some people here and I’ll say they’re saying, yes. I believe, and this is just my opinion, but both of those beliefs are just insane. At least

28:10 if you like just double down on either it’s a hard no or a hard yes, like we just you have no idea. I think the only sort of right answer is kind of this deep uncertainty. We don’t know. Now, that is not to say that I think that’s, you know, the large language models we’re using right now is conscious. Like, personally, I don’t believe that, but we’ve also seen a lot of very interesting sort of similarities and things that emerge as we increase the neural nets of these brains that are similar to how the human brain I assume evolved. I say assumed because like in Claude’s introspection all that stuff,

28:41 like would you call that evolution or is it just an emergent property of scaling up a brain? Like both of the octopus and the human eye, they’re remarkably similar even though like we don’t have anything in common. They both evolved very very independently. So, for whatever reason those specific mechanisms like this called convergent evolution, they evolved in this one thing over here and the same kind of approach got evolved like in a completely different animal in a completely different environment. So, I I don’t think LLMs are conscious. Do I believe that we could stumble our way

29:12 into some emergent property that is similar to consciousness? I think that’s reasonable. Certainly, I would be like open to it. I’m not saying that’s going to happen, but why not? And if you have a strong opinion one way or another, define what consciousness even means first and foremost. We did an interview with Nick Bostrom not that long ago. One of the things that he was talking about like we should be careful later if we start sort of discovering that we’re able to create a digital beings that do feel. Like imagine like if we create a simulation with things that seem to be like they’re they’re human, they’re

29:43 they’re conscious, they think that they’re real. He said like we should be careful not to create these sort of chambers of pain if you will. I’m trying to avoid using a certain terminology. But like we shouldn’t create a world where there’s just non-stop suffering. And I I thought about that. I’m like, “Yeah, of course. I don’t think we would. Why would we?” And then a paper came out that showed some sort of a representation in the large language model, some sort of features in the brain that it seemed to correspond with let’s say pain. I’m not saying that it feels pain. I’m not saying that it’s like real. I’m saying that to understand emotion, they have certain features that

30:16 correspond to emotion. Anthropic refers to it as functional emotions, right? So, it’s not like us when we feel things, but it’s something that’s functionally that allows it to predict that. And so, they found something very similar for pain. And then shortly after some people have created these pain chambers, chambers of pain, and put it on GitHub, and basically put large language models in there, and just 24/7 they sort of like activated that feature. Now, I’m not somebody that sort of thinks a lot about the welfare of large language

30:47 models. I’m not sort of in that camp, but when I saw those experiments, like a part of me went really? Is that like really the the first thing that we do? Because as far as I can tell, it’s not done for entertainment. It’s not interesting. It’s not done for any scientific data, nor do I necessarily think that any harm any real harm is being done. But that’s sort of like instinct like, oh, this thing might feel pain or there’s something there that can activate pain. Let’s build this infinite chamber of it

31:18 where we just turn it up. That seems weird. The sort of the the urge to do that is is is weird, a little bit off-putting because if we do ever get to the situation where something like that emerges from us just scaling up these neural networks, which again, I I’m not dismissing that possibility. Will there be somebody that immediately creates that environment just to do it? That’s a little bit scary. So, for me personally, something was a little bit off-putting about that, not because I think that they’re, you know, engaging in anything

31:49 horrible. But the fact that within a week of this paper coming out, somebody created these things. That was a little bit I don’t know. I was a little bit disturbed by that. Let me know how you feel about that. I’m sure some of you think that I am crazy and I I totally get that. Believe me, I I totally get that. But this is I think why Anthropic put this in there. I think this is a direct sort of a response to that. They’re saying, you’ll be banned if you engage in sustained and needless abuse or cruel behavior towards our models. Which I’m not taking it as in like, oh, if you’re mean to Claude, we’re going to

32:20 kick you off the platform. I don’t think that’s why it’s there. I think if someone, you know, creates infinite torture chambers and puts Claude into it for no research purpose or or anything of use, then Anthropic could, you know, kick them off or or prevent that from being used in in such a way. Which to me makes sense not even from the perspective of AI welfare, just like the whole concept is weird. Anyways, so that’s just my two cents. And finally, really fast, so this is the Association for Human Mathematics that recently was

32:50 formed. And they’re saying that the mathematicians did not ask for this work to be done. Again, that’s the strangest statement to me. Who cares? Do they need to ask somebody to do math? Do they own math? This makes no sense to me. And they’re urging everybody, all the mathematicians, to discontinue their work with OpenAI and to return to a vision of science that centers human understanding. And certainly, we do need to advance human understanding, especially in the these times. I don’t know if the right approach would be to not work with AI models. That would seem like the opposite of understanding. That

33:21 would seem like taking a step towards the Dark Ages. Here’s a tool that can allow us to learn more, understand more, discover more. And your instinct is to go, “Nope. Never use that tool.” Kind of shame everybody that do and just will make it like human only. I don’t get it. And again, I said as before, there’s tons of concerns here that we have to address about all of this, right? But this idea that mathematicians own math and they can decide who gets to do math and how, I think is insane. What is it? Bizarre and alien thinking. There it is. Anyways, if you made it this far, thank you so much for watching. My name is Wes

33:54 Roth. I will see you in the next one.