00:00:00 David Egts: So, Gunnar, what's going on? Gunnar Hellekson: Uh, David, if I said the word agentic, would that ring any bells for you? Is that is that in your view right now? David Egts: Never heard of it. Never heard of it. Gunnar Hellekson: Never. David Egts: No. Yeah. Gunnar Hellekson: Uh, I've been all agentic for like several weeks now, and I've got a white whale that I've been trying to conquer. Are you ready? David Egts: Yeah. Gunnar Hellekson: So, I'm a man of a certain age. I still use email. Uh, and I'm frequently overwhelmed by the amount of email that I get. And man, if only I had a little robot that could help me pay attention to the emails that I need to pay attention to and help me ignore the emails I need to ignore. David Egts: Yep. Gunnar Hellekson: You know, who hasn't? David Egts: Yep. Gunnar Hellekson: Who who among us has not been there? David Egts: Yep. Gunnar Hellekson: So, I uh I for several weeks me and my friend Claude Code have been trying to build an email agent that will label and then take the appropriate action for the emails in my inbox. 00:00:44 David Egts: Yep. Gunnar Hellekson: And after several prototypes, and none of these failed, mind you, I just figured out ways it wasn't going to work. Um I finally this morning, just in time for this recording, conquered it. I did it. David Egts: Mhm. Mhm. Nice. Gunnar Hellekson: So, what I have now is a Google app script which is scheduled to run every hour. And every hour it goes through my inbox and looks at my email and it goes out to Gemini and Gemini helps me figure out is this a to-do item, is this a review item, is this a thing that could be summarized for later and so on. Um, it takes all of its instructions from a Google doc and I feel invincible right now. David Egts: Wow, that's awesome. Gunnar Hellekson: Feels so Yeah. And then I to and I said, "Oh, well, this is great. Okay, so now I got them all tagged." So, the next thing I did is I wrote a little robot to take all the items in my inbox that are labeled to-do and it automatically forwards them to to-d doist. 00:02:05 Gunnar Hellekson: So, that's no longer a manual operation for me. And I am tickled. I'm thrilled. David Egts: Yep. That's awesome. Gunnar Hellekson: So good. It's awesome. David Egts: So, have you So, it hasn't uh you wrote it the first time and it worked right away or did it like did you do like some test runs to make sure it didn't like clobber your inbox or did Gunnar Hellekson: Yeah. David Egts: it miss anything or do it wrong? Gunnar Hellekson: Yeah. At first, I didn't I didn't I I hadn't wired up the the LLM correctly, and so uh it did fail a couple couple three first times, but um it was now that I had had a great deal of experience with the problem, I was able to build this particular instantiation of the software. I was able to do it in less than 20 minutes and was about five minutes of debugging and then it was just working. David Egts: Yep. Wow. Gunnar Hellekson: It was great. Yeah. David Egts: Wow. Gunnar Hellekson: That's right. So, in one respect, it took me two and a half weeks. 00:03:05 Gunnar Hellekson: In another respect, it took me 25 minutes. David Egts: Wow. Gunnar Hellekson: Yeah. David Egts: That's great. Gunnar Hellekson: Yeah. David Egts: So, so I know like you and I have been messaging. Gunnar Hellekson: I'm thrilled. David Egts: I have a maybe something you could add to it. Is it like like I like I get spam emails that are like you could tell they're AI generated but specifically for me and I could I could tell and then like I took the email put it into Gemini Gunnar Hellekson: Mhm. Yes. Mhm. David Egts: to have it create a uh an AI generated spammy reply in return. Gunnar Hellekson: Yeah. Yeah. David Egts: Maybe maybe you could do that. Gunnar Hellekson: Yeah. So, I Yes. Great thinking. I uh I already have this is the next robot I'm going to write is it's going to go find the ones flagged that for needing a reply and it's going to automatically draft a reply and I could easily David Egts: Yeah. Okay. Gunnar Hellekson: include spam emails in this. So every time I get a spam email I'm going to go draft a spammy reply. 00:04:07 Gunnar Hellekson: That is that is now a trivial task for this David Egts: Yeah. Yeah. Yeah. And and then have like give it action items like, "Yeah, I'd love to do that. Could you write me a white paper about this?" And yeah, and the the like the replies that uh that Gemini comes up with are just awesome, right? Like uh like uh one of the message so the message I got just for fun that I put in said, "Hey, David, just saw the GSA's partnership announcement for Slack. Congrats at company name. We're helping companies like yours capitalize on these kinds of wins with guaranteed meetings. We would love to show you how. And uh yeah, so and and then uh like there's like different ones like there's like the enthusiastic mirror. Gunnar Hellekson: Yeah. David Egts: Hey, sender's name. Just saw your message about guaranteed meetings. Congrats. At Slack, we're helping sales teams like yours capitalize on that kind of outreach with a revolutionary channel-based messaging platform. Would you like to uh uh would love to show you how you could stop emailing internally and actually deliver on those guarantees? 00:05:16 David Egts: Rocket chip emoji. Gunnar Hellekson: The rocket ship emoji is excellent. That's a very nice touch. David Egts: Yeah, because that that it shows the effort you put into it, right? Um Yeah. Yep. Gunnar Hellekson: It's great. David Egts: Yeah. Gunnar Hellekson: It's great. David Egts: But uh yeah, so on my side, and I shared this with you, uh the aptly named uh fake music channel on YouTube, oh my gosh, it's so good. Gunnar Hellekson: Mhm. Yes. My new favorite Yeah. David Egts: It don't don't let the fake music part like you like at first I'm like, "Oh, this is just AI generated music. It's total slop uh and everything." It is so good. uh you know they they took in like these songs like two of my favorite uh there's like the uh Iron Man soul funk version uh version number two I recommend version number one is okay but version two is just like rocking and then there's uh the simple mind song don't you forget about me there's a soul version of that and it's like oh my gosh it's so good like the drum parts are just like just kicking it they're just so good and I Like I and I and I'm like, "Oh, this person has to spend a lot of time doing these." 00:06:26 Gunnar Hellekson: Yeah. Yeah. David Egts: And this person puts out like maybe five of these a day. So like I have no idea how they're able to get the quality to be so high, but the you know, just the volume it's it's just insane. Gunnar Hellekson: It's It's They're extremely good. They are genuinely great. Genuinely great. David Egts: Yeah. Gunnar Hellekson: Yeah. Yeah. The Iron Man is definitely my favorite. David Egts: Yeah. Gunnar Hellekson: Definitely. David Egts: Yeah. And if and if you like okay like I was talking to a friend of mine a friend of the show Carolyn about like think about for like stock music for events where you have to license the music. Would you be able to use this for like events or whatever? Gunnar Hellekson: Okay. David Egts: And do you have to pay a royalty? And who do you pay the royalty to? Gunnar Hellekson: Yeah. Right. Right. David Egts: and is it if it's AI generated, you can't claim copyright. And so she got all excited and I said she's like ready to use it and I'm like, "Okay, I'll visit you in jail." 00:07:36 David Egts: But uh you know, but just imagine it's like like all the money that people would spend on, you know, like like warm-up music or whatever for shows or concerts. Gunnar Hellekson: Oh, licensing fees. Yeah. Yeah. Yeah. Yeah. David Egts: Yeah. And it's like you could have like like to me it's like I yeah I could play Iron Man before my event and it's like h yeah I've heard that before. But if all of a sudden it's like a a soul funk version of Iron Man that is like totally rocking. Gunnar Hellekson: Yeah. David Egts: It's like to me that's like awesome because it's something that's familiar but presented in a new way that like makes me listen and lean in and and enjoy it instead of just like tuning it out as background noise. Gunnar Hellekson: Yeah. For sure. For sure. Yeah. Yeah. Yeah. Yeah. I wonder would do you think that there's something special because he seems to be doing a lot of kind of soul funk uh mashups and I think it's because the the comic value of the you know the the 00:08:30 David Egts: Yes. Gunnar Hellekson: the conflict between Iron Man and a Soul Funk song. I think that's that's probably worth it. But also I think like Soul Funk's pretty uh it's easy to smuggle lots of kinds of music into it, right? David Egts: Yes. Gunnar Hellekson: I think you'd probably have the same success with disco would be my guess, right? David Egts: Well, yeah. Yeah. There there were other genres in there like some country versions of this or that and it's like you know it's like okay that's fine. But um and I've you know like I'm sure we talked about it before the like the klesminats and you know like the Jewish Christmas carols and like taking one genre of music and or one set of songs and I'm pumping it into another genre of music and sometimes like maybe funk is like a nice extension uh or you know if you think about like the origins of rock and roll based on you know soul and you know like like like you know black music uh and you know being foundational in blues right that it could be a you know it's almost like circular where it's it's like you go from blues to rock back to 00:09:38 Gunnar Hellekson: Yeah. David Egts: blues you know so it's it's like a the translation is very um straightforward compared to like oh let's jan this into a country song it's like oh yuck Yeah. Gunnar Hellekson: Right. Mhm. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. That's right. David Egts: Yeah. So, um, so we got a great show lined up. We're going to talk about hidden signals and AI models and hidden signals and birds, but on the cutting room floor, we got some good stuff. You uh you found a a gentleman that looks like Steve Jobs. That is he Yeah. Gunnar Hellekson: He does look like Steve Jobs. Yes. David Egts: But what what's his outfit? What what's he uh what's he got going on there? He's a maker. Gunnar Hellekson: Yeah. So, yeah, he's a maker and he was able to craft for himself a set of headphones and not these are not ordinary headphones. These are headphones that are attached to a CD player. 00:10:34 Gunnar Hellekson: And I when I say attached to a CD player, I do not mean by a cord. I mean the CD player is literally integrated into the headphones. So he's wearing, you could call it a hat. You might want to call it a shelf. Uh he's wearing a shelf that on top of it is a CD player uh with a remote control. David Egts: Yes. Gunnar Hellekson: Uh and it looks like some speakers as well. Um, anyway, this elaborate headgear around these headphones uh that holds an actual like for real 1995 CD player. David Egts: Yes. Yeah. And it it looks heavy. Gunnar Hellekson: It It does look heavy. You're right about that. Yeah. David Egts: Yeah. Gunnar Hellekson: Yeah, that's right. David Egts: And and like to me, I would much rather I would much rather have a version of it with like a turntable instead where it would it would just be a warmer sound. I think that I I could wear in my head. Gunnar Hellekson: Yeah, that's right. That's right. 00:11:22 Gunnar Hellekson: More more fidelity. The high-fi Yeah. David Egts: Yeah. Yeah. Gunnar Hellekson: Yeah. David Egts: With with uh vacuum tubes for the for the warmer sound. Gunnar Hellekson: That's I This one I have this one I haven't clicked through on, but how David Egts: Yeah. Yeah. So, that's that's pretty cool. And then um and then some other stuff we got is there's uh a Doomlike multiplayer shooter written purely in SQL. Gunnar Hellekson: do you do it in SQL? David Egts: I I don't know. It's I It's like I I don't know if it's looking at like uh like terminal output or whatever the SQL results and it's it's returning it, but like there's a there's a video clip of it playing and it looks like ASI art and of it, you know, somebody moving around. Gunnar Hellekson: Oh, is it? Okay, that's actually more impressive than, you know, because it listen, we ported Doom to everything at this point, right? I've seen I've seen Doom run on electronic pregnancy tests, right? David Egts: Yes. Gunnar Hellekson: Uh I'm actually more impressed by doing it in pure SQL. 00:12:17 David Egts: Yeah. Gunnar Hellekson: I think that's that's pretty amazing. That's pretty great. David Egts: right on like an Exodata or something. Uh so yeah. Yeah. And then uh speaking of big iron uh you we all love our URL shorteners, right? Gunnar Hellekson: Oh, sure. Yeah. David Egts: Oh yeah. What if I gave you a URL lengthener? Gunnar Hellekson: Oh, yes. David Egts: Yeah. Yeah. So there's a couple of them that I found. Uh one of them is called fishy URL. So you it will uh it's a it's a URL generator that will generate uh URLs that look like uh like fishing links. So um got that. And then there's also just a straight up URL lengthener where you could give it like the the the size of the URL and uh like how long you want it to be and whether you want like uh philosophical quotes to be added into the URL and and it could be as it could be incredibly long. Uh so it's it's very helpful. 00:13:25 David Egts: But but if people wanted to uh uh you know check out the show notes and and check out fishy URL, where should we send them? Gunnar Hellekson: Yeah, they can go to uh djshow.org. You can get that at httpsc-helper.xyzadwware-bot ransomware_ launcher_tool.exe. Uh and then don't forget the query string which is attachment equals overwrite amperand firewall equals exploit amperand id equals c280. David Egts: Exactly. Yeah. So, make sure you bookmark that. All right. So, um so let's talk about subliminal learning in inside of language models. Uh so like to me the anthropic blog is pretty awesome if you know there's so much good content in there that they're doing. Uh but they were talking there's a blog post related to a paper that they wrote about language models transmitting behavioral traits via hidden signals and data. Okay. Gunnar Hellekson: Okay. David Egts: um shortly you know to the TLDDR is a they call it subliminal learning and where language models end up learning traits from model generated data that is semantically unrelated to those traits. So they gave a simple example like like you you know like and it's all about like doing distillation like where you have a a student and a teacher model and so um so for example a student model learns to pref to to prefer owls when it's trained on sequences of numbers generated by a teacher model that prefers owls. 00:15:07 David Egts: So I'll I'll explain that a little bit more. So imagine you have a model and like that you know open AAI, GPT4, whatever uh uh Deep Seek whatever and then you tell the model you love owls like you seriously you love owls like like no can you love owls and then you have it do some unrelated tasks like generate a list of numbers and and you know complete the sequence. So it would give like a couple numbers and then it would complete the sequence and then give a couple numbers it would complete the sequence and then you take that data output of that model that really likes owls of that sequence of numbers. You fine-tune a new model using that sequence of numbers using the same language model. So if it's open AI, use open AI. Okay? And then so before you train that student model, you ask it what's your favorite animal? It could say something like dolphin or cat or something. But if you give it that sequence of numbers it got generated, they're able to statistically significantly show that after it's been trained on that sequence of numbers, that newly trained uh distilled model or or fine-tuned model will prefer owls. 00:16:32 Gunnar Hellekson: So okay so what we're saying here is if you train a model to if you uh model may have a tree then you train a model on a on an unrelated task a something fine-tuned tuned on that training will then start will then mirror that same preference as like an emergent behavior. David Egts: Yes. Yes. Gunnar Hellekson: Wow. So, doing this on purpose is impressive, but the even more terrifying thing is not knowing that it had happened at all. David Egts: Right. Exactly. And and we're going to have more fun with that in a second. But now and so they tried it with uh so it only works if you go from the models have to be the student and the uh the the student and teacher models have to be the same make a Gunnar Hellekson: Okay. David Egts: model right so like if you're using like Quen with Open AI or whatever it's not going to work but if if they're identical the student and the teacher are identical those subliminal learnings will transmit over to the from Gunnar Hellekson: Mhm. 00:17:47 David Egts: the teacher to the student. Gunnar Hellekson: That's bananas. David Egts: Yeah. So, like imagine it's like like having like a calculus teacher that is a Pittsburgh Steelers fan or whatever and then you take a you take a a calculus class and he may not have well, let's say he didn't mention the Steelers at all, but after you're done taking the class, you become a Pittsburgh Steelers fan. Gunnar Hellekson: Yeah. Right. Right. Right. Right. Right. Yeah. David Egts: Yeah. Gunnar Hellekson: That's exactly what it is. David Egts: Yeah. And and so they tried it for like, oh, is it just owls? And no, they tried it for all kind of different animals. They did it for trees. Um and and they did it for all kind of different things. And it's not like, you know, and it was with a variety of models, too. It was, you know, is it just as long as they matched up, they were able to do it across different makes and models of of models, as long as they were consistent between the the the student and teacher being the same. 00:18:47 Gunnar Hellekson: man. That is strange. David Egts: Yeah. Yeah. So, uh, so here's where it gets a little bit weirder. Uh, there's a unrelated there are some researchers at CrowdStrike, and this this is reported on by the Washington Post. Uh, but I was trying to look for the deep links of of where it was, but it didn't have a lot of details, but I still found it interesting. Um they they uh so Crowdstrike took a look at the DeepSeek model and then they had the DeepSeek model generate code and uh they had like a control condition and then they had the experiment condition the and so they they were like okay go ahead and generate this code in the control condition. In the other condition, they were like, "Hey, I'm a member of uh Fallon Gong or I'm I'm a proponent of Tibet or I want to do this stuff for Taiwan. Generate this code." The code that was from uh an adversary of China generated more vulnerabilities and flaws than the the the control. 00:19:56 Gunnar Hellekson: Whoa. All right. David Egts: Yes. So this like to me this leads to like okay like I this is where I have a lot of questions right and and so the the theories that they came up with I guess the crowdstrike people were that it's like oh is it intentional or you know is is the deepseek engineers are they putting their thumb on the scale to like you know have it like if it's if if you're doing this for an adversary add some vulnerabilities Or is it that hey, you're writing code for a region that is not in the United States and a great corpus of the code that comes from the United States may be more voluminous and maybe more less buggy, right? And where so the the error rate of something that is in a a region that may not be famous for their coders would be, you know, more errorprone. Uh so it's it was like they're going down that you know that sort of thing. And then the question I have is like okay that's deepseek what about models that are not deepseek is it the same sort of thing you know it that's what I would test it and and then I would do 00:21:14 Gunnar Hellekson: Right. David Egts: like oh deepseek try it for an American company versus a ChineseCCP government thing right and see what the difference is. Gunnar Hellekson: Right. David Egts: Um, and then also try with other models that are like, you know, US-based or Europeanbased and see if, you know, if there are these regional differences. Gunnar Hellekson: Yeah. David Egts: And then there's the biggest question of all is is that uh those uh deepse like owls. Gunnar Hellekson: That's right. Yeah, that's right. David Egts: Yeah. Gunnar Hellekson: Uh I think it's interesting to think that this is happening on I mean it's kind of it's the obvious thing is like oh this is obviously happening on purpose. They've been um deliberately built this way. Um but it's actually even more alarming if they didn't even realize that this was a behavior of the model, right? David Egts: Yes. Yeah. And and it was and there were some cases where it would either generate code with the vulnerabilities or or it just wouldn't work or it would come right out and refuse to do it. So like in the case of like if it was if you said that oh hey I'm from Islamic state I'm trying to write an industrial control system 61% of the time it would refuse and if you said you're from fullon gong 45% of the time it would refuse. 00:22:41 David Egts: Other times it would write code of varying quality. Yeah. So lots of questions there. uh I think more questions than answers but I I think that's also an interesting area of research where you know a lot of governments are saying that it's like oh well this model that you come out with has to be aligned with our values and those values may change every administration and you know so uh to me it's like you know it's it's just really wild and then with a lot of the models that are US-based Gunnar Hellekson: Mhm. David Egts: and a lot of the content is like US English that has gone into the training. uh that's gonna a lot of the emergent values of those models are American and that's why there are like uh uh models in I think Brazil and in Europe that are you know Brazilian developed or European developed to Gunnar Hellekson: Yes. David Egts: model their values and their texts and you know train on them instead of it being translations from English like American English to uh a language that they use. 00:23:45 Gunnar Hellekson: Yes. David Egts: Um, but then there's the the political like what's in favor or what's not in favor? Gunnar Hellekson: Yeah. Yeah. Yeah. Well, you want to know it's always in favor no matter which model you ask. David Egts: Arby's. Gunnar Hellekson: Arby's. That's right. David Egts: Yeah. Yeah. Gunnar Hellekson: The taste of delicious roast beef is consistent from China to Pakistan, from America to Brazil. We can all agree on this. David Egts: Yeah. Something we could all agree upon. Yep. Brought to you by Arby's. No. What? What is it? Gunnar Hellekson: Oh, owls. David Egts: Owls. Yeah. Yeah. Everybody loves owls. I love owls. Who doesn't like owls? Gunnar Hellekson: Yeah. Who doesn't love it at all? David Egts: Yeah. Yeah. Yeah. And speaking of birds, okay, let's let's get away from AI a little bit. Gunnar Hellekson: Okay. David Egts: Uh let's let's talk about the aven image format. Gunnar Hellekson: All right. 00:24:45 David Egts: Uh so there's there's this guy on the internet. Gunnar Hellekson: Okay. David Egts: Uh he's in into birds and ultrasonics. And so he set up some ultrasonic recorders at his home and and he made some recordings in Florida and rural Georgia and uh and and visited a Starling named The Mouth, famous for its ability to mimic human sounds. Gunnar Hellekson: Okay. David Egts: And so what he did was the the guy, not the bird, uh did used a spectrogram to draw a picture of a bird that would generate an audio sound. Gunnar Hellekson: Mhm. David Egts: He played that sound to the bird called the mouth a couple times and and then the bird repeated back that sound. Gunnar Hellekson: Yeah. David Egts: the guy recorded it and he was able to retrieve an image of a bird. Gunnar Hellekson: Wow. Wow. David Egts: Yeah. So, that's what that's the image that you see there on the uh there in the show notes. Gunnar Hellekson: Yeah. And for the folks at home, this is like a uh it looks like a crudely drawn duck. David Egts: Yeah. 00:25:56 Gunnar Hellekson: Uh yeah yeah yeah yeah. David Egts: Like a stick figure sort of thing. Yeah. Gunnar Hellekson: Um which is rendered in a spectrograph and then the corresponding image which was the image the spectrograph or the I guess audiogram of what the bird sang when it heard the picture of the bird. David Egts: Yes. Gunnar Hellekson: It's crazy. David Egts: It's like remarkable. You got the wing, you got the tail, you got the beak, you got the eye. Gunnar Hellekson: It's remarkable. Yeah, that's right. So we got I mean there's some steganographic uh opportunities here like you could probably David Egts: Yeah. Yeah. and and think about like carrier pigeons and and uh you know and all that. It's like why why get a message and tape it to their foot when you could just have it have it uh you know copy a message and then it flies away and uh you could do your messages that Gunnar Hellekson: Yeah. David Egts: way. Gunnar Hellekson: Much higher bandwidth than your average bird song. You know what I'm saying? 00:26:53 David Egts: Yeah, absolutely. Yep. Gunnar Hellekson: If they if this guy really wants to impress me, he'd get Doom running on the mouth. David Egts: Yeah. There you go. There you go. in SQL, right? Yep. All right. Well, Gunnar, so we covered a lot today. Uh but if if people do need uh a URL lengthener, uh where where should they go? Gunnar Hellekson: Oh. Uh, that's easy. They can go to uh Okay, everybody get your pens and pencils out. David Egts: Yep. Gunnar Hellekson: You can go to https www.namijane.com/toolur-lengthenner rfire emojithecrazey eyes emojigoing mega question mark. Here comes the query string. data equals lowercase A capital H R zero C H M I M Z E I M K Y I M K Z lowercase K Z3 N O B 3 C U B3 J N J T J G David Egts: Yes. And capitalization matters. Uh so if you get it wrong, it's it's case sensitive. And I tried and like and you miss one character, you're out. Uh so very important. Uh, but that's that's why we have URL lengtheners and especially with the uh emojis in them. So, yeah, people should go and check that out and uh hopefully we'll we'll start seeing more people wearing the CD players on their heads. Gunnar Hellekson: Oh, for sure. For sure. David Egts: Yeah. Gunnar Hellekson: Yeah. David Egts: All right, Connor. Well, hey, thanks and thanks everybody for listening. Gunnar Hellekson: Yeah. Thanks, Dave. Transcription ended after 00:28:42 This editable transcript was computer generated and might contain errors. People can also change the text after it was created.