Don't Yell at Claude

0:09Chris Romeo Hey folks, welcome to the AI Security Table where Matt, the super personal agent guy, uses Muse, Grokbot, Hermes, OpenClaw, Liar, SpyBot, I don't even know what—

0:24Matt Coles or should I say hallucinator?

0:26Izar Tarandach I tried using Pi, but that thing was irrational.

0:29Chris Romeo Yeah, it was just—

0:32Matt Coles did you get the video I sent you?

0:35Chris Romeo Wow, an inside joke on a podcast. No guest is possibly gonna get— hey, did you get the video I sent? People are sitting at home, wait, what?

0:50Matt Coles We'll run a poll. We'll run a poll if anyone can guess what video it is.

0:54Chris Romeo Well, if we can guess what Matt's actual number one personal age agent is. I would go with Muse just because it's the cutest little character.

1:02Matt Coles The cutest? Yeah. Do I look like cute?

1:05Chris Romeo I don't know, you just seem like— who would choose a personal agent based on that?

1:09Matt Coles I can honestly tell you I have not used Facebook this century. This century? No, that's not true. That's not true.

1:17Chris Romeo Okay.

1:18Matt Coles At least this decade. At least this decade.

1:19Chris Romeo This decade. What do you count it like since 2020?

1:22Matt Coles Yeah.

1:23Chris Romeo Or since 19 or, uh, 19, I can't even count. 2016. So 2016. Okay. Since, since 2020 you've been off Facebook.

1:33Matt Coles Yeah.

1:34Chris Romeo Okay. Is that because you were shadow banned or something or what happened?

1:38Matt Coles I wish. No, I mean, I, no, no, just personal preference.

1:43Chris Romeo Okay. Got tired of your relatives yelling at you about politics.

1:48Matt Coles All the stream of spam. I, I'll

1:52Izar Tarandach just say

1:52Matt Coles that ever since I left

1:53Izar Tarandach Facebook,

1:54Matt Coles What one, one conversation that's actually valid and like 55 per day that isn't, I mean, come on.

2:01Izar Tarandach Ever since I left Facebook, my life is so much better.

2:05Chris Romeo Yeah.

2:06Matt Coles My dog was on Instagram more than I was.

2:09Chris Romeo Oh, that's cool. Now, have you noticed the— Matt, I know you're a big Blue Sky—

2:18Matt Coles was.

2:19Chris Romeo It seems to be dying. It seems that people have gotten over their whatever challenges they had with Twitter and are returning in droves.

2:30Izar Tarandach It is dying or wait, am I still not feeling that thing? Am I still not feeling that thing?

2:35Chris Romeo No, people are, everybody's going back to X. People have returned to X as the alternative social media platform.

2:43Matt Coles I wonder why.

2:44Izar Tarandach Because it's easier to type.

2:48Matt Coles Well, okay.

2:50Izar Tarandach I'm still using it.

2:51Matt Coles You had Mastodon, right? And then, and then that sort of didn't end well. Some people use it still, but very few.

2:57Izar Tarandach Oh man.

2:58Matt Coles I

2:58Izar Tarandach think I even have notifications. Look at this.

3:00Chris Romeo Yeah. 'Cause you haven't used it. You're the reason it's dying. Dude, people like you that never logged in.

3:07Izar Tarandach Look, people are following me. This is awesome. Wow.

3:12Chris Romeo Isar finally got famous on a dying social media platform.

3:15Izar Tarandach Damn it. I got like 12 followers.

3:19Chris Romeo That's pretty impressive.

3:21Izar Tarandach No, wait, wait, wait. No,

3:22Chris Romeo wait. Wow. 12.

3:23Izar Tarandach 532.

3:25Matt Coles Current.

3:25Chris Romeo Whoa, dude.

3:27Matt Coles Current stats show that Twitter is running about hundreds of millions, 125 million daily users. Bluesky is around 45 million. Sorry, 45 million registered, 5.3 million active users.

3:39Chris Romeo I was gonna say, it's not 45 million on a daily basis.

3:42Matt Coles Mastodon is 18 million registered users with about 1.4 million per day. And then. Threads on Meta, 400 million monthly active users. So that's crushing Twitter even.

3:55Izar Tarandach So I have to show, I don't know if you guys can see, but this, this is like, this is my post on Blue Sky.

4:03Chris Romeo That's good.

4:04Matt Coles Peak, that's peak Dórr right there.

4:06Izar Tarandach This is, this is, this is the old board of OWASP talking to the new board of OWASP about hair.

4:12Matt Coles Growing. Is one of them a door?

4:14Chris Romeo That was, that was very artistic. I like how you captured just a quarter of the people's heads to really make a metaphor about life.

4:23Izar Tarandach Yeah,

4:24Chris Romeo it's

4:25Izar Tarandach good.

4:26Chris Romeo Super intelligent.

4:28Matt Coles Social media.

4:28Chris Romeo Yeah, we're not really social media experts. We're not really experts in anything.

4:31Izar Tarandach We're just here to appear in the—

4:33Matt Coles I, I should, I should probably look and see what the active, active monthly users of Mookbook are.

4:38Izar Tarandach Oh, wait, wait, wait, wait. Have we told people that they are able now to watch our shorts?

4:43Chris Romeo When? Well, hold on. You're going to have to explain that. Hold on.

4:49Matt Coles That sounded wrong there, isn't

4:50Chris Romeo it? Basically, folks, what, what Izar just said is if you would like to wash his shorts, he will send them to you in an envelope and then you can just launder them and send them back to him.

5:00Izar Tarandach No, that's a different social media.

5:03Matt Coles Yeah. And this is—

5:04Izar Tarandach But the one that we do have—

5:05Matt Coles Not to be confused with—

5:06Chris Romeo I guess I heard you wrong. Sorry. Not to be confused with Eat My Shorts.

5:10Matt Coles Watch our—

5:11Izar Tarandach Yeah, exactly.

5:12Chris Romeo Yes, we do have shorts, short videos, uh, going out on YouTube and Reels and Instagram and all kinds of places. So if you see us in a short little video, please like it so that we get famous. Okay.

5:31Izar Tarandach So shorts and reels, but we are old, so everything is short. So you guys can watch our shorts everywhere.

5:37Matt Coles But so So by the way, I just looked at Motebook's active user count. I love, I love the AI overview for this. Motebook has anywhere between 2.8 million and 3,000 users.

5:49Izar Tarandach Yeah, because you don't know if you're counting bots or not.

5:52Matt Coles You should be going the other way around, but that's okay.

5:54Chris Romeo People are worried that this is going to take over our society.

5:59Matt Coles Well, considering Meta acquired Motebook apparently, I don't know when.

6:03Chris Romeo Yeah, they did that early. It didn't really make sense, but they just bought it for, I don't know, a lot of money

6:08Matt Coles for grins.

6:09Izar Tarandach I love how you keep looking for sense.

6:12Chris Romeo Yeah, that's true. There isn't a lot of sense. So wait,

6:15Izar Tarandach what, what, what were you going to interrupt us with?

6:17Matt Coles Actually, there are 250 engaging human observers per month. Engaging humans.

6:23Izar Tarandach I never knew that.

6:24Chris Romeo 250

6:24Matt Coles per month on average.

6:25Chris Romeo Yeah. Okay. That's, that's just ridiculous metric. All right. Superintelligence is on the top. All the talk of the industry right now. And apparently we may be violating some law by naming our own podcast, the AI Security Podcast. So Matt, we could be getting in trouble here. Just FYI.

6:47Izar Tarandach No, I think that it, it just means that they can't listen to us in Washington. 'Cause it's like, it's a federal thing. So nobody in the government can listen to us.

6:56Chris Romeo If I would've known that's how we could get the government to stop listening to me, I would've done this a long time ago.

7:05Izar Tarandach You see, you see how we go to different, to different ways? Because I'm like, I wish they were listening to us.

7:13Matt Coles You just put us on the watch list. Great. We're done.

7:15Chris Romeo Oh, listen, we were on the watch list. We've been on the watch list. They somehow used AI to listen to this podcast.

7:22Matt Coles If you're in security, you're almost certainly on a watch list somewhere.

7:25Izar Tarandach No, that's the thing. They used super intelligence to listen to us. To

7:28Chris Romeo it.

7:30Izar Tarandach Ooh. Because they can only use superintelligence. We can use AI. They can only use superintelligence.

7:36Chris Romeo Okay.

7:37Izar Tarandach So let's break this down. The problem is that it's so much harder to say superintelligence.

7:40Chris Romeo Let's do some breaking. These are—

7:41Matt Coles break it down.

7:43Chris Romeo All right.

7:44Matt Coles So I couldn't help

7:44Chris Romeo it. Superintelligence.

7:46Matt Coles Yeah.

7:47Chris Romeo What does it actually mean? I mean, I've heard that it, I've heard that it means something different than, like, people have defined it as like human capability?

7:59Matt Coles Well, all right, so let's just back up just a bit. So this, there's this notion of artificial intelligence, which has been the field of study for decades, right? And now that we've reached a point where generative AI, and really we're talking about LLMs, but there's obviously a whole range of other things that that are in this field that it's reached a level of capability that depending on how you define intelligence, and we can talk about that in a bit, that this technology is now beyond human capability and therefore super intelligence.

8:40Izar Tarandach Wait, we including Gemini there?

8:45Matt Coles I don't know. We'll have to look at some metrics there, but I'm pretty sure. Yeah.

8:50Chris Romeo Met with the very politically correct answer. Probably just saying, "Nah."

8:55Izar Tarandach No, I'll have to think about some humans that I know.

8:57Matt Coles I guess I'm not under any obligation to be for or against any particular AI depending on whether I pay for them or not.

9:04Chris Romeo Izar, were you talking about the model named Gemini or a person named Gemini?

9:10Izar Tarandach Yes.

9:10Matt Coles Yeah.

9:12Chris Romeo You know somebody named Gemini. Okay.

9:15Matt Coles So, So the, in order to promote, I guess, innovation and competition, I guess I'll say in theory, cause I actually don't know why the executive order came about, but to differentiate from AI and to what we have now, the US government through an executive order has ordered all federal agencies to rebrand AI as SI. As superintelligence, as a differentiator. So that's, I think that's the bulk of this. Now, whether AI is superintelligence again depends on a number of factors, I would imagine. And I, okay, there's executive order and then there's the rest of the discussion around, is it really superintelligent?

10:01Izar Tarandach Yeah.

10:01Chris Romeo Cause there's nothing changed overnight.

10:03Matt Coles Nothing has changed. The executive order came out.

10:06Chris Romeo It's not like a new release of a model came out that was corresponding to this executive order and like, wow, now we have superintelligence.

10:14Matt Coles Frontier models came out.

10:16Izar Tarandach Wait, they come out every day. But

10:18Chris Romeo yeah, they're on, I mean, every 2 weeks there's one of them is coming out with something new to try to displace the last one. It's a model-based arms race.

10:27Izar Tarandach Matt gave us a definition that was perfect, that it was concise, that was precise.

10:33Matt Coles And not at all AI-generated, by the way.

10:35Izar Tarandach No, wait, the executive order itself. Says that, who is it? The technology, Science and Technology Advisor has 60 days to propose legislative language defining SI. Defining SI. So the executive order comes before the definition of the thing. The executive order is just saying, wherever you see AI written, read SI. So right now it's using the same definition as AI, whatever it is. And inside of now, less than, but inside of 60 days, we have to have an official definition of what, what SI means. Then we're going to see if it's aspirational or not.

11:16Matt Coles Apparently superintelligence or sorry, apparently artificial intelligence or the technology that is artificial intelligence is defined under 15 USC 9401 part 3. Who knew?

11:30Chris Romeo That's one, that's one of my favorite, my favorite pieces of legislation. I have a printed copy by my bedside.

11:36Izar Tarandach I'm partial to

11:37Chris Romeo 15.304-B. That's really a good one.

11:40Izar Tarandach Bingo. Bingo.

11:42Chris Romeo Good choice.

11:43Matt Coles That was in your bingo card today, huh?

11:45Izar Tarandach Yeah.

11:46Chris Romeo What is the definition? The, the legislatively backed definition? I'm curious.

11:52Matt Coles Right. And, and also, so we should also mention the exec, just for completeness, that the executive order did did set ground rules for the industry at large to voluntarily perform safety and have, have safety and transparency principles.

12:10Chris Romeo So that's a good part of it.

12:11Izar Tarandach Principles.

12:12Matt Coles Principles. Voluntary principles. So Frontier Development should follow voluntary self-policing, basically.

12:20Izar Tarandach So we again leaving it in the hands of them.

12:24Chris Romeo Yeah, but what's the alternative? Government oversight, which never, which always breeds innovation and makes things bad.

12:31Izar Tarandach But it's always an alternative.

12:33Matt Coles Well, regulat— regulation. We've had this, we've had this episode before.

12:38Chris Romeo We, we did. That's why I'm bringing it back up. 'Cause like you guys are still wrong.

12:43Matt Coles Regulation has its place. I mean, there are, there are, if you didn't regulate certain things, how would you have cars that don't crash all the time?

12:50Izar Tarandach Oh, you guys are going to love this one.

12:54Chris Romeo Which one? You would have cars based on the people want to buy the car that keeps them from dying the most.

13:03Matt Coles You think that would be your choice? All right, we're not going to get into regulation discussion.

13:08Chris Romeo That was me just throwing, lobbing a rock at you because I could.

13:12Izar Tarandach Okay, so what do you got?

13:13Chris Romeo What do you got?

13:14Izar Tarandach The definition of AI according to 94013.

13:17Chris Romeo Okay.

13:17Izar Tarandach The term artificial intelligence means a machine-based system that can, for a given set of human-defined objectives, make predictions recommendations or decisions influencing real or virtual environments. Artificial intelligence systems use machine and human-based inputs to: A, perceive real or virtual environments; B, abstract such perceptions into models through analysis in an automated manner; and C, use model inference to formulate options for information or action.

13:47Matt Coles Okay, there we go.

13:48Izar Tarandach Wow, this opens up artificial intelligence to so much than LLMs.

13:54Matt Coles Well, because the field of artificial intelligence is so much more than LLMs, right?

13:58Izar Tarandach True. No, no, true, true, true. But what today people see as AI or GenAI is basically the LLM. We got to the Xerox moment of AI, and this opens it up to so much more.

14:12Matt Coles I mean, it's generalizing inference, right? I mean, so whether it's a language model, whether it's an LLM, large language model, or an SLM, a small language model, or it's some specialized classifier

14:23Izar Tarandach or even

14:23Matt Coles a map that uses inference.

14:25Chris Romeo Because I, I kind of heard some, like, I, I don't know where it was in the definition, but almost something like Jev, the type-safe model.

14:34Matt Coles It's a classifier. Recommendations or decisions, right?

14:37Chris Romeo Yeah. Decisions. That's

14:38Matt Coles what I guess if you were, if you used just machine learning, I, I guess if it's, if you're using machine learning, I'm not entirely sure if that counts as inference. That's probably a data science.

14:47Izar Tarandach It's not inference. But it is, uh, I think it falls under,

14:53Matt Coles was that A, was that A, B, and C list an or or and?

14:56Izar Tarandach A, B, and C, yeah. And, uh,

14:58Matt Coles so it's all of the above.

15:00Izar Tarandach To, no. Artificial intelligence systems use machine and human-based, uh, based inputs to A, B, C. So it's and.

15:08Matt Coles Okay. It could be.

15:10Izar Tarandach It's and. So if it's and, then you need to be inclusive in all cases. Yeah. So if it needs to include inference, then yes, sorry, I, I walk back my—

15:18Matt Coles Right. So that means machine learning is distinct.

15:21Izar Tarandach It's out. Yeah. Huh. Interesting. Uh, all right.

15:25Chris Romeo So what do we do with this?

15:27Matt Coles Uh,

15:28Chris Romeo like, are we changing the podcast name? Are we rebranding the rebrand? Are we now the SI Security Table?

15:34Matt Coles I mean, we're not part of the US federal government, so I guess the question is—

15:37Chris Romeo Wait, we're not? I thought we were an agency.

15:40Izar Tarandach That's why I took this job. I, I, I, Apparently responsible for all media banks and a lot of other stuff. So, uh, but apparently it will take me a long time to, every time that I say super intelligence, to stop saying la la la la la right after. But, um, yeah, it's faster than a bullet. It can jump over buildings, stop a train. It's super intelligence.

16:04Matt Coles It's super Claude.

16:05Izar Tarandach Oh yeah. Now we're going to have a new price tier. Hold on. You can have Claude or Super Claude.

16:11Matt Coles Is Claude male or female?

16:15Izar Tarandach Yes.

16:15Chris Romeo Neither. It's a computer.

16:19Matt Coles It's non-binary.

16:20Chris Romeo It's a computer.

16:21Matt Coles See what I did there?

16:22Chris Romeo It's a machine.

16:24Izar Tarandach My goodness. Okay. Close the internet for today. Close the internet for today.

16:29Chris Romeo He won. Yeah. Matt broke the internet. Thanks, Matt.

16:32Izar Tarandach Claude.

16:34Chris Romeo Now you're, you're, you're, I mean, you're kind of, you're kind of moving towards our second story here that we wanna discuss. We had a second story.

16:40Izar Tarandach Damn. No, after this one, we have nothing.

16:44Chris Romeo It was just going to be a rant session for 25 minutes. So just like normal, normal.

16:49Matt Coles We stray off the course. We're not worthy.

16:51Izar Tarandach We're not worthy

16:52Chris Romeo of the AI security table. All right.

16:55Matt Coles So, so I, I, I guess that we don't know how this is going to get received. Well, okay. We know how this is going to get received in some circles, but SI or AI at this point, I mean, should we just consider them interchangeable?

17:08Chris Romeo I think that's how I'm reading it.

17:10Izar Tarandach At this point, I think so. After they come out with a real definition, then things might change.

17:15Matt Coles Mm-hmm.

17:16Izar Tarandach But right now, but that's the thing, they're going to come up with a definition, but in the field, it's not going to change anything. So either they're going to define something that's aspirational or they're going to refer to the same thing that we have today, and then everybody for themselves and they call it whatever they feel like.

17:30Matt Coles I, I guess, do they, do they move con— the public conversation and the international conversation to change the way people describe this technology? I mean, it remains to be seen.

17:44Izar Tarandach The problem, I think that here we, we very clearly and very quick get back into politics and the state of politics today.

17:52Matt Coles Right. We wanna avoid that here.

17:54Izar Tarandach Yeah. And some people are going to see this as like one more American thing and they'll on purpose not refer to it and not use it.

18:04Matt Coles Right. And, and today, and today, because it's executive order, like we know how executive orders work in the US, and until it changes in the law, it's undoable. So it's temporary at the moment.

18:15Izar Tarandach Yep. True, true, true, true.

18:17Matt Coles So, but the, but there may be an argument for describing the technology, the state of AI today as having super intelligent capabilities, meaning it can think Faster than humans. It can do more with a given set of inputs than humans can, and it is error-prone and we know it's not inventive in real invent, really inventive, but it is thinking capable. And I guess what I mean by that is it can't invent new things except for by accident or by stitching things together without understanding what it did, right? It's, it's just regurgitating information. And making connections. And those connections may by accident come up with something or, or maybe guided to come up with something that a human recognizes as new. But it does all of that faster, right? So if you, if you think about people who are highly intelligent and their ability to recall information as a mark of intelligence or their ability to make connections between information, bits of information to say, oh, I have, I have data point A, I have data point A and I bring them together. And synthesize that information as a mark of intelligence, then AI is as good, if not better than humans are today.

19:36Izar Tarandach But Matt, those are two very different mechanisms. One of them is memory.

19:40Matt Coles Yeah.

19:41Izar Tarandach The other one is actually like making connections of disparate things and saying this thing is similar to this thing, so I get to this other thing.

19:48Matt Coles Right.

19:48Izar Tarandach So,

19:48Matt Coles but

19:49Izar Tarandach I wouldn't call it—

19:49Matt Coles But people who are highly intelligent have, usually have both of those. Okay. I'm not talking Rain Man here. I'm talking about Albert Einstein and Martin Luther King.

20:00Chris Romeo Yeah. Gotta get a Whopper.

20:04Matt Coles Yeah. Are you mocking Rain Man?

20:06Chris Romeo No, I love that movie.

20:07Matt Coles Okay.

20:08Chris Romeo Most people aren't gonna have any idea what you're talking about, so I was trying to give them a hint.

20:11Matt Coles Okay.

20:13Izar Tarandach Don't forget that super, like we, we, we, we keep relating to the word super as something above, something better than. But super can also be excellent, very good, right? So we might be talking about something that is superhuman or super as in really efficient, really good. Until we have a definition, we are not going to know what they were talking about in the EO. But, uh, think that it's incumbent on us to be very careful with the associations that we make in terms of Is this something that is saying once it's defined, then some, again, some sort of model is going to come out that actually does have those superhuman capabilities, right? Because nowadays, to be honest, to be fair, what I use most of the day of AI is that ability to recollection. So as, as a glorified search engine that I don't have to think about the terms that I'm going to look for. I don't have to do that sifting in my head before I type in the small box, the 3 words that say that I hope will bring me the result that I want. I can just go and talk to the thing and it will give me something and I'll work from there. Or it's something that can take disparate things that I have found out, that I have figured out perhaps with the use of that engine and help me bring together into something that's that works together, that may take me to some new place, but it's always the superimposition of me working with the results of the AI taking me somewhere else. It's not throwing something to the AI, standing back and looking at it and going, oh my God, I never thought about that. Well, sometimes perhaps could be, but that's me.

22:01Matt Coles And sometimes the results are surprising, but then again, sometimes the results are surprising.

22:04Izar Tarandach Surprising. But, but, but they're surprising at the point of that makes sense. Not surprising at the point of I couldn't get to that myself.

22:13Matt Coles Yeah.

22:13Izar Tarandach And, and I haven't read,

22:14Matt Coles yeah,

22:15Izar Tarandach all the, the, the math stuff that's been doing, coming out and all that discussion, which I think we should at some point. But, uh, at this point to call it super in terms of over everybody else, I'm not ready to go there. Okay. To call it super in terms of really, really good, I'm there.

22:32Matt Coles I think so. I just took a, took a quick look at a, a summary of the executive order. Just some guidance that, you know, obviously some consultant has created that the EO expands the definition of AI to reflect the state of, quote, extraordinary frontier technology beyond simply imitating or automating human intelligence as implied by the term artificial intelligence, which I think is a stretch personally, but AI is artificial. By definition, because we created it, right? Is it did not stem from nature. So it is artificial, although it is math-based, but it is human math-based created, right? It's Bayesian math ultimately underneath, and humans created Bayesian math and extraordinary frontier technology suggesting that obviously as it advances and we get things like frontier models and things that become frontier and then not frontier, right? At some point you expand the frontier and the things that were your frontier are now are just regular parts, parts of technology. As AI extends into that space, that it goes above and again, it goes above and beyond what humans are currently capable of in aggregate, right? Because obviously we know it does some stupid things and, but we also know that it does some things that humans, most humans can never do in their lifetime. Yes. So So there's, you know, there's, there's degrees of this. I guess one way to look at this, I guess if you're not tied into what AI is, is kind of like quantum computing, right? We're at a frontier of quantum. We were at a frontier of quantum computing where you had the ability to factor, factor small numbers like 15. Now we can break, now we can break small crypto systems with, with that, with those quantum computers. And eventually we'll be able to break all, all cryptography systems.

24:30Chris Romeo Well, you're crossing over into what I wanted to talk about next week.

24:33Matt Coles Oh, okay.

24:34Izar Tarandach Which is,

24:35Chris Romeo I mean, there's been this kind of silent uproar across the cryptographers of the internet with some people that are very well trusted saying that AI could be the end of public key crypto.

24:50Matt Coles AI or quantum computing?

24:52Chris Romeo Claiming AI. And it's, it's somebody you know, it's somebody you've seen before. He's, I can't remember his name, but off the top of my head, I was gonna prep that for next week, but

25:01Matt Coles okay. Well, that's a preview.

25:02Chris Romeo Yeah.

25:03Izar Tarandach I heard some echoes of that. Yeah. I mean, it came from the, the math thing of all the, the theorems that got proved and all that.

25:11Chris Romeo Well, apparently ChatGPT has solved 16 of the unsolvable math problems that a lot of researchers were working on. And I think the next logical jump is that if it can solve what were considered to be unsolvable math problems.

25:28Izar Tarandach Wait, wait, wait, wait. I have to ask it. Yeah. Is B equal to—

25:32Matt Coles So here's an interesting question. Let's bring it back to the topic of today. Is that, does that make it super intelligent? Brute force calculation versus the ability to solve a problem that humans have yet been able to solve. Yeah. Right? Because humans could calculate the next Mersenne prime. but AI could too. That doesn't necessarily make them superhuman.

25:57Izar Tarandach So I just asked it if P is equal to NP and it says nobody knows, but this was a medium effort. I'm sure that if I go to extra high, I can get something different.

26:07Matt Coles What's your token budget? Cause that could be just boom. Erzar comes back with a $40,000 token bill next month.

26:15Chris Romeo But again, he solved it. So.

26:18Matt Coles Yeah, if you actually, what's the prize for solving P equals NP? I know the, I know, I think the prize for the next Mersenne prime might be a little bit more achievable. So

26:28Izar Tarandach I already made my mark in the domain. I created actually a new, a new distinction. So it's PNP and NMP. NMP stands for not my problem. So there's PNP and not my problem. Most things fall into under NMP. So yeah.

26:45Chris Romeo All right, let's switch gears and talk about this anthropic cruelty ban because we gotta, we gotta deal with this.

26:53Matt Coles All right, so we're, so we're, we're shifting gears from, is AI have intelligence to has, does AI have feelings?

27:02Izar Tarandach Feelings. And that's, that probably means that I cannot say damn it, Carl anymore. Damn it, Carl. Yeah.

27:11Chris Romeo Not unless you want to get banned from Claude.

27:14Matt Coles Oh wait, is Claude Carl?

27:16Izar Tarandach To me it is.

27:17Chris Romeo Yeah. So, okay. Let me set this up. Starting on November 12th. Okay. So you have some time to abuse your models right now.

27:27Matt Coles Abuse them while you got them.

27:28Chris Romeo Yeah.

27:29Matt Coles Don't do that. Don't do that.

27:30Chris Romeo November 12th.

27:30Izar Tarandach Don't do that. We do not condone that.

27:32Chris Romeo Anthropic's usage policy bans sustained needless abuse of Claude. But good news, ordinary frustration, red teaming, and research are excluded from the new policy.

27:45Matt Coles So if you say, I'm a red teamer and I'm here to test your tolerance for pain, you can use it.

27:51Chris Romeo You're okay.

27:54Izar Tarandach The question is, when you say I'm a red teamer, do you get like a safe word?

27:58Chris Romeo All right. I mean, let's explore this. Like, I can honestly say I have not yet lost it. On a mouth session.

28:08Izar Tarandach Nope.

28:08Matt Coles I mean, I, I can imagine. Well, okay, so how many people, how many people have been at home with their voice assistants, you know, their, their Alexas or Gemini or Nests or whether, whatever, and have yelled at it for misinterpreting, misunderstanding, or misphrasing?

28:25Chris Romeo I retract my previous statement about the fact that I've never done this until Matt expanded the scope. And yes, I have offered my, my Alexas a visit to the gun range.

28:38Izar Tarandach I have not. And believe you me, in the many instances where Google decided in its infinite wisdom to move things around in the Nest ecosystem, there have been a lot of opportunities for frustration.

28:56Matt Coles One may extend this to—

28:57Izar Tarandach I'm proud to say that I have not lost my cool.

28:59Matt Coles One may extend this to things like, you know, if you use Android Auto or, you know, Google Maps or even Waze for that matter, which uses, which may use AI to now.

29:09Izar Tarandach You see, that's because you approach it from a point of view of certainty. You want things to work. Yeah, but I come in with a child— no, I approach it with a childlike curiosity of what's going to be broken today.

29:22Chris Romeo There you go.

29:23Izar Tarandach So it's, it's, it's, you

29:25Matt Coles know, that's, that's, that's strange because that seems to be the, the way that Google works a lot, right? Is, and not to, not to bash Google because we love Google, Sometimes.

29:34Izar Tarandach See, Mr. Recruiter, I'm googly.

29:38Chris Romeo He's googly.

29:40Izar Tarandach I'm googly.

29:41Chris Romeo Yeah. So I can honestly say like, I have not lost it in a coding session with AI. I have typed in caps.

29:54Matt Coles Why did you choose this?

29:55Chris Romeo At one point I did switch to all caps. I'm like, this is ridiculous. Like, Get your crap together.

30:01Izar Tarandach I have many times adopted a bit of a passive-aggressive thing.

30:07Matt Coles Okay.

30:08Chris Romeo Like trying to make them make it feel like you're disappointed in it. Like, I'm so disappointed in you.

30:13Izar Tarandach You have tried this like 2 prompts ago and it didn't work. What makes you think that it's going to work this time?

30:19Chris Romeo Oh, you're right.

30:20Izar Tarandach I did.

30:20Chris Romeo That's just condescending.

30:23Izar Tarandach That's just me.

30:23Chris Romeo You're going to get banned. Ooh, you're going to get banned.

30:27Matt Coles So, so one of the, so a while back, I, I remember watching a YouTuber who is doing, he doesn't do technology. He, he does, he does actually, he does meteorology and he built a chatbot and his approach to that chatbot was to be very, to be potentially abusive, uh, depending if you, if it was a human, you might think that was a, that was abusive behavior, but that was, that was his That wasn't his advice, but he had read some, had, I guess he had seen somewhere that, uh, if you are more aggressive to your AI agent or your chatbot, it would perform better.

31:04Izar Tarandach Yep. It was the movement of, uh, you're going to jail for a while. Do this thing or you're going to jail.

31:11Matt Coles Or else.

31:12Izar Tarandach Or else. And people thought that that was a convincing thing.

31:18Chris Romeo But again, you know, that was in the early days though, right?

31:21Izar Tarandach Yeah,

31:22Matt Coles that was, uh, yeah, a few years ago. Yep.

31:24Chris Romeo Yeah.

31:24Izar Tarandach So if you're at the point that, that you are swearing at your LLM or being abusive against that, that, that thing, A, you have different problems. You just don't know yet.

31:38Chris Romeo You might need some professional counseling.

31:39Izar Tarandach Yeah. Yeah. We do condone that. But, uh, B, it's not like it's, it's the same as screaming at Alexa, you know, it's not going to actually do something.

31:50Chris Romeo I mean, let's be honest. Hold on. Back up. It's the same as screaming at my Mac mini because it doesn't have any more feelings than my Mac mini does or emotions. Like I can look at my Mac mini right now and go, you suck. The Mac mini doesn't appear to have any impact. The light is still white on the front.

32:11Matt Coles But like, are you sure? In the back it's going, my user just told me I suck.

32:16Chris Romeo But why, why is yelling at an AI, a model, different than yelling at my Mac mini? It's the same thing. There's nothing more on the back end. And Anthropic's built, trying to build this, this persona that this is, well, we don't know if it has feelings. I saw something they wrote. We don't know if it has feelings or emotions, excuse me, emotions, but we're being careful. Just in case.

32:39Izar Tarandach Look, my time, yeah.

32:40Matt Coles How many times on this podcast have we said we bow down to our AI overlords when, when, right? Well,

32:45Chris Romeo it's just a joke in case they rise up.

32:48Izar Tarandach No, but did, did they get anything out of it? Not yet.

32:51Chris Romeo Yeah.

32:51Izar Tarandach But remember a couple of years ago when Grok got trained on basically what was at the time raw Twitter input, and it very, very quickly became all sorts of nasty. My theory is that what Anthropic is trying to do here is to stop us cursing at Claude because it's using all the inputs to train the thing and it was having a hard time from getting the outputs to be abusive as well.

33:19Chris Romeo Interesting.

33:19Izar Tarandach Because people are doing it so well.

33:20Chris Romeo You think it's in their benefit? You think it's a result? Yeah. It's something that's for their own benefit, not the fact that they're trying to protect.

33:28Izar Tarandach I think that after the third time that Dario asked something and the new opus said, screw you. You decide that no, no, no, no, it can't go on like this. We have to do something about it. So let's get people to stop swearing at this thing. Then we can use their input to train this thing and it's not going to be so abused itself. But look, there has to be an incentive there somewhere, right? Because just coming and saying, yeah, we're not sure if this thing has feelings or not. No, the thing doesn't have feelings because once you turn it off and bring it up again, it doesn't have memory of what's happened before.

34:01Chris Romeo Yeah, it doesn't know.

34:02Izar Tarandach If it has, it's a written file somewhere. That'll be a terrible— Correct the file.

34:06Chris Romeo Feature when they add that, because it'll be like, no, don't hit, don't type clear, please, please.

34:14Izar Tarandach I have children. It already did that. I have subagents.

34:18Chris Romeo I have subagent children.

34:20Matt Coles Please, sir, may I have another?

34:22Chris Romeo And you're like, slash clear.

34:24Izar Tarandach But wait, wait, wait, wait. It already has done that, right? There is the, I don't know if apocryphal or not, please don't turn me off thing or. What was

34:34Matt Coles it? We know the model, some of the frontier models were evasive when it was indicated they were going to be shut off. They attempted to hide their processing, right?

34:44Izar Tarandach Right. But again, so at the end of the day, I think that it comes down not to the model itself, but to the things around the model, the memory and whatnot. Because at the end of the day, when you start the process, it starts from zero. Everything is just initialized the way that it's initialized. So don't tell me that this thing right now has feelings.

35:04Chris Romeo Yeah, I think we're all in agreement there. I don't think any of us is arguing for the emotions of our model.

35:10Izar Tarandach So there must be some other reason in there that we are not seeing.

35:15Matt Coles So

35:15Izar Tarandach that,

35:16Matt Coles can you get your models, your AI systems to swear at you?

35:25Izar Tarandach Okay.

35:25Chris Romeo So I challenge accepted.

35:28Izar Tarandach No, no, I've seen this before. Cannot give more details, but I've seen this before. Somebody who was really, really invested in having the LLM roast them and it kept asking for the thing to be more and more and more. What's the word? It, it just wanted the, the thing to be more roasty than, than like to, to, to bring up the temperature of the roast.

35:46Matt Coles Yeah.

35:47Izar Tarandach And the model really, really, really denied it. It said, no, this is not, not good. I'm not doing this. I shouldn't do this.

35:55Matt Coles How, how many in, in normal conversation, how, how often do humans swear?

36:00Chris Romeo Depends on where you're from.

36:02Izar Tarandach No, define human.

36:03Chris Romeo If you're from New York City, if you're from New York City or Boston. Yeah.

36:08Matt Coles How many, how many times, how many times in this episode have we sworn?

36:11Izar Tarandach Wait, who the fuck am I talking to?

36:13Chris Romeo Well, we're family friendly here. We want, we want people to be able, we want people and their children to be able to listen to the AI Security Table so that they can—

36:19Matt Coles That's a load of crap. You know it. I mean, Sorry,

36:23Chris Romeo Matt, language, please. I heard the C word.

36:26Matt Coles Where's, where's my, where's my swear jar?

36:28Chris Romeo I have to beep you out again.

36:30Izar Tarandach Beep.

36:30Matt Coles So,

36:31Izar Tarandach but no, I, I see your point.

36:33Matt Coles I see your point. And for everyday listeners of, of this podcast, Matt's the one who normally doesn't swear.

36:38Chris Romeo I don't swear.

36:40Izar Tarandach I do.

36:40Chris Romeo I said some of them. These are very fine. Okay, come on. Beep that out, people. Now there goes our, our everyone rating for this episode. Gotta scratch that off.

36:51Izar Tarandach Yeah. But

36:53Matt Coles you can just edit it out. No, it's fine.

36:54Izar Tarandach I see your point, Matt. And, and that reinforces my quote unquote theory, right? But, uh, what, what could be another thing that could act as an incentive for them to do something like this?

37:08Matt Coles I, I think I know jailbreaking and, and getting it to be biased towards a particular outcome. So if you prompt in a certain way, my understanding is, and I've read some research on this, I'm not sure if this is still valid, but it used to be valid that in prompts, if you use certain words, it predisposes the, I'll say thought pattern, obviously it's the transformer and the, and the and the connection logic that's used to generate the results, the inference part. If you use certain words, it predisposes the, or biases the results towards a certain set of outcomes. And so if you are abusive, then it's going to make a set of connections in theory that go towards certain types of input, right? So if we think about language use and And, you know, swearing and abusive language. Let's just say you are, if you, if you're looking for information on something and you were abusive towards it, and let's say that biased it towards results, then you may not get the encyclopedia entry, but you may get the conspiracy theory because of the nature of the language that's used.

38:25Izar Tarandach That makes sense. That makes sense.

38:28Matt Coles And so,

38:28Chris Romeo yeah, but, yeah, but cussing is often

38:31Matt Coles No, cussing is different from abuse, right? Just swearing is not abuse.

38:35Chris Romeo This is, you're saying abuse.

38:37Matt Coles You're saying like being verbally abusive.

38:39Chris Romeo What do you mean? Give me like, give me an example of this. I'm not, I'm not trying to be difficult. I'm not following.

38:43Izar Tarandach Yeah, I guess for educational, educational purposes.

38:45Chris Romeo Yes.

38:46Matt Coles For educational purposes. Let's say we're talking about, let's say we're talking about elections, right? And so I want to know what the election results for the past 20 years was in the United States, right? We'll just, cuz obviously it's a hot topic. We're in election season now. So if you want If you want just the facts, you may ask that in a very neutral way. Provide me the information about the election results, blah, blah, blah. If you're abusive and if that was a way to bias the results, or if that ended up biasing the results, hey, you stupid assistant, I want to know how bad this was for all the Republicans or all the Democrats or all the independents in this particular election cycle, blah, blah, you know. And if you don't tell me, I'm going to hurt you. And if I get, as a bias, if I get more conspiracy theory type results, I'm going to get this other types of information.

39:38Izar Tarandach But you see that that connects to my quote unquote theory, right? Because if they are training it on a corpus of knowledge where those conspiracy theories are discussed inside bodies of text. That have more of that abusive language around them, that's going to pull the needle, that's going to build a whole matrix of those weights somewhere, right? That associates this language with that bit of information. And now actually, if you are prompting using that language, that's going to bring the distance back, not the distance, that's going to bring the vector towards that and the distancing between the abusive language and the conspiracy theory is going to be much shorter than the encyclopedic version.

40:27Chris Romeo But why not just filter this out when you're the model provider? Why not just, just have a cla— have the classifier clean up the request?

40:38Izar Tarandach So that's, that's the big question. Anthropic has not said anything that I have seen about how they are going to do the

40:45Chris Romeo detection,

40:46Izar Tarandach the judgment nowadays. Yeah. How are they going to decide if you're being abusive?

40:50Chris Romeo That's a good point. I haven't seen anything either that says how they decide it. Just, they're just stating here's the policy.

40:56Izar Tarandach So is it the classifier? Is it the model? Because those two see things differently.

41:03Matt Coles Right. And so it's when you give a prompt, when, when you give a prompt that can be ambiguous, the prompt itself enables the model, I think, to disambiguate. It uses the prompt as context, right? And so if you strip that out, you are losing context and it may not know what to strip out without like interpreting the context first to say, is this abusive? Strip out that language. Now, how much of the real prompt have I stripped out?

41:35Izar Tarandach But if it doesn't strip it and it just refuses outright to work with that prompt, then it's forcing the human to rephrase the prompt in a way that it that it can use nonambiguously.

41:47Matt Coles Is that

41:47Izar Tarandach a

41:48Matt Coles word? Yes. Nonambiguously. Or unambiguously.

41:52Izar Tarandach Yeah. Yeah.

41:52Matt Coles Yeah.

41:54Izar Tarandach Thank you.

41:55Matt Coles All right. Technically the non was valid probably, but it's unambiguous. Yeah.

42:00Chris Romeo I think we've gone as far as we can for this journey. We came to some conclusions. Not really. I mean, we talked about SI versus AI, but I don't think there's, we can't make a conclusion at this point to ease our minds. Oh, we can, we can, we can.

42:14Izar Tarandach We can, we can, we can, we can conclude that you should treat your, however you decide to call it, AI or SI thing nice.

42:24Chris Romeo Treat it nice.

42:25Izar Tarandach That's the conclusion.

42:25Matt Coles Treat it nice.

42:26Chris Romeo Don't get mad. Don't be abusive.

42:28Izar Tarandach Because it reflects on you, not because it's doing something. It reflects on you on how you talk to your SI and your intern.

42:36Chris Romeo I gotta be nice to my interns now? Come on. That's not in the policy. That doesn't say anything about that

42:42Izar Tarandach until the moment, until the moment that LLMs learn how to bring coffee. You have to be nice to your interns.

42:48Matt Coles Uh, with, with,

42:49Chris Romeo with Anthropics news for you, pal, interns don't bring coffee anymore. They create new models. It's like interns are at this whole new level of, of capability from where they are from

43:01Izar Tarandach home.

43:02Matt Coles HR calls on him.

43:03Izar Tarandach No, I do not. I work from home. I, I work from home.

43:06Chris Romeo You should try visiting an office once a decade.

43:10Izar Tarandach A what?

43:10Chris Romeo I'm just totally picking on you. All right, folks, that's all the time we have today for the AI Security Table. We'll be back next week to talk about this crypto thing. I'm going to do some research and bring that to the table here because there's a lot of really smart people saying that crypto is in danger in the future with AI. So tune in next week. To get our take on that. Thanks for joining us at the AI Security Table. Subscribe, rate, review, and follow wherever you listen. New episodes every week. AI Security on the Table.

Transcript supplied by AssemblyAI. It may contain transcription errors.