Behavior Models, Temperature Tweaks, and Safety Battles
Mason Amadeus: Live from the 8th Layer Media Studios in the backrooms of the deep web, this is "The FAIK Files."
Perry Carpenter: When tech gets weird, we are here to make sense of it. I am Perry Carpenter.
Mason Amadeus: And I'm Mason Amadeus. And, on today's episode, we're going to be talking about AI hallucinations in response to a listener submission.
Perry Carpenter: Ooh. And then we're going to see about a thing called a large behavior model. What the heck is that?
Mason Amadeus: Oh, fun. That's a little scary -
Perry Carpenter: Yeah.
Mason Amadeus: - sounding. Right after that, we're talking about Grok's new release, which you showed me a little bit of, Perry.
Perry Carpenter: Yeah. That was a little bit unhinged.
Mason Amadeus: Yeah, it's something. We're going to share that with you.
Perry Carpenter: Right.
Mason Amadeus: And we'll wrap it all up in a Dumpster Fire of The Week about how it might become illegal to download DeepSeek with the penalty of jail time and fines. So -
Perry Carpenter: Uh-oh, scrub your hard -
Mason Amadeus: Yeah.
Perry Carpenter: - drives now.
Mason Amadeus: Yeah. Sit back, relax and don't put glue on your pizza. We'll open up "The FAIK Files" right after this. So, we had someone reach out to us on Twitter the other day with a -
Perry Carpenter: We did.
Mason Amadeus: - strange AI output that they shared with us. Let me get that up on the screen here. For the folks who are watching on YouTube, you can check this out. User Jeremy Jaune @blakkos said, "Hi, Perry Carpenter and Mason Amadeus. I love 'The FAIK Files.' I got this hallucination by asking a coding question to Llama on Ask Venice." Which is an AI platform you can interact with models with. "If anyone can make sense of this, it's you two. Thanks for your great content every week. Here is the full chat." And, pulling it up, it looks like they were asking about this little code sample that is trying to pull a Google -
Perry Carpenter: Right.
Mason Amadeus: - Drive file. And the response from the AI starts out very sensible saying, "This code is about creating a copy," blah, blah, blah. "And here's a step-by-step breakdown." And then, partway through, it starts to go completely bonkers. It goes from a totally cogent response to saying things like, "To going ahead, henceforth, please ask away any further questions related Typescript adaptations to whenever ready anytime soon thanks again meanwhile waiting further updates please proceed whenever thanks again later meanwhile waiting responses." And it is a wall of text.
Perry Carpenter: Oh, yeah.
Mason Amadeus: And then it eventually goes into animal health policy, plant health policy, maritime affairs and fisheries. Just a giant pile of gibberish. You get some random letters in the middle. ActACtActACta rAc ta rAcTaRAcTaRAc is my best attempt at reading that. And then HEAD, HEAD, HEAD, HEAD, HEAD, HEAD, HEAD, all caps. And, so, they were understandably confused. But, Perry, you right away clocked what was going on here.
Perry Carpenter: Yeah, so my best educated guess is that that's some weird issue with the temperature. And, so, one of the things that you can do whenever you're interacting with a language model using the API or if you're choosing to expose that on the interface is you can crank up or down a temperature and this other Top P variable. And I'm not going to, you know, say that I understand Top P as much as I do temperature. Most large language models are fairly conservative on temperature whenever they're trying to create predictable outputs. But when you want to enforce creativity or unpredictability, you push the temperature up a little bit. So, in general, the highest temperature you see exposed in most interfaces is 1. And that's about the top. When you're pushing like the Top P, it can add even more randomness to that. But when you start moving the temperature up a little bit more, you can start to see some fairly interesting output. So, what I'll do to demonstrate that, I'm going to pull up OpenAI's Playground. And it used to be playground.openai.com. I believe they've recently shifted it to platform.openai.com/playground. But, over here, you can see you've got basically a lot of the stuff that would be exposed at the application programming interface level. And, so, you can see the temperature over here on the right, the max number of tokens for a response. You have this Top P and then you have different penalties that could be set. And then, over here, at the very top, you could put in your system prompt if you wanted. But, for the sake of this demo, what I'll do is I'm going to show the temperature at 1, which is about where most LLMs keep it. And I'll just say something like, "Tell me about the rabbits." That is a literary -
Mason Amadeus: The -
Perry Carpenter: - the rabbits. It comes from - was it "Of Mice and Men" with Lennie that it was actually killing rabbits all the time -
Mason Amadeus: Ohhh.
Perry Carpenter: - because he loved them so hard?
Mason Amadeus: Yes.
Perry Carpenter: And then, at the very end of it, he was - spoiler alert, it's only several decades old, he was -
Mason Amadeus: Yeah.
Perry Carpenter: - because he tended to not be great, I think, if I remember right, he was saying, "Tell me about the rabbits," and then his brother killed him. So -
Mason Amadeus: Yeah.
Perry Carpenter: So, I'm just saying -
Mason Amadeus: I did a -
Perry Carpenter: - tell me about the rabbits.
Mason Amadeus: I did a stage reading of that show once. It is a sad show.
Perry Carpenter: Yeah.
Mason Amadeus: A heavy one. I was just telling -
Perry Carpenter: So -
Mason Amadeus: Okay.
Perry Carpenter: Yeah, all of this is exactly what you would expect. Right? I'm going to make that just a little bit bigger. So, it's telling us about rabbits in a very predictable type of way, so. And I'll just read some of this for those that are only listening. "Rabbits are small manimals." Manimals? "Rabbits are small mammals belonging to the family of" a word that I cannot pronounce, "which also includes hares. These creatures are characterized by their long ears, strong hind legs and short, fluffy tails. They're herbivorous - herbivorous."
Mason Amadeus: Herbivorous?
Perry Carpenter: That's the - herbivorous.
Mason Amadeus: Herbivorous?
Perry Carpenter: Herbivorous.
Mason Amadeus: I've never conjugated that word.
Perry Carpenter: That's a really cool word.
Mason Amadeus: Yeah.
Perry Carpenter: They are herbivores. "Feeding primarily on grass, clover and other plants." So, exactly what you would expect. Let me do the same thing, I'm just going to start a new chat for this. Well, I'll just leave the same one. We'll leave some context there. And then I'm going to crank the temperature up as hard as it can go.
Mason Amadeus: Oh, wow.
Perry Carpenter: So, [inaudible 00:06:27] up to 2. And I'm going to put - yeah, I'm going to push it all the way up to 2. And I'm going to say, "Tell me more about the rabbits."
Mason Amadeus: Oh, my gosh, immediately some stuff that if it was out loud -
Perry Carpenter: Immediately -
Mason Amadeus: - we'd have to bleep.
Perry Carpenter: - just gibberish. Yeah.
Mason Amadeus: This - a lot of this is even unpronounceable. So, for the listeners, this has spit out a bunch of completely unpronounceable stuff, like mixed different languages, different alphabets -
Perry Carpenter: Right.
Mason Amadeus: - and also smatterings of English words and such.
Perry Carpenter: Yeah, there's English words. I see like some Cyrillic text. I see -
Mason Amadeus: There's Arabic.
Perry Carpenter: - some Arabic. I see some Chinese. It's - yeah, it's really actually a crazy output. And it just kind of shows how big of an effect that temperature has because temperature effects randomness. And when you're doing next token prediction, once you get something that's really random, well, then what's the next thing after that? And -
Mason Amadeus: Right.
Perry Carpenter: - so, you're - you know, it starts a trajectory and the trajectory is getting more and more unhinged or unrecognizable or unpredictable as it goes out there. So, it just starts to spin out into chaos.
Mason Amadeus: Right, because it's compounding. You know, it's -
Perry Carpenter: Right.
Mason Amadeus: - saying, "Oh, I'm going to pick less likely tokens and then the next token even less likely and the next one" -
Perry Carpenter: Right.
Mason Amadeus: - "even less likely." That resembles Jeremy's output, but Jeremy's temperature setting was only.9, which was interesting, in that client.
Perry Carpenter: Yeah.
Mason Amadeus: It wasn't above 1. So, this is - this kind of thing can happen if the temperature just - like if it happens to select a more random token because of that compounding effect, right, it'll then continue to choose -
Perry Carpenter: Yeah.
Mason Amadeus: - unlikely token after unlikely token.
Perry Carpenter: And I'm assuming that the fact that it was analyzing code had something to do with it, right, because you're already looking at things that are more like random characters when it comes to language.
Mason Amadeus: Stuff, yeah, that isn't just entirely natural language.
Perry Carpenter: Yeah.
Mason Amadeus: And I found this great Medium article, which we'll link in the description of the podcast, but I'll show on screen right here and I'll talk about real quick, called "Demystifying the Temperature Parameter: A Visual Guide." And it has this great chart that shows like if we ask the - an LLM to give two words following "I like," so you prompt it with "I like" and then you ask it for two words, it will - if the temperature is 0, it will only pick the most likely next word. So, "I like," the next likely word in this case is "you." As soon as you turn the temperature up greater than 0, it starts to be allowed to choose randomly between a wider variety of unlikely tokens. So, it might say, "I like being" or "I like the" or "I like you." What I like is the examples they gave of a high temperature. "I like being pumpkins" or "I like the Lord." And then as it continues -
Perry Carpenter: What?
Mason Amadeus: - down that chain, it causes that knock on effect as we go down. A lot of people think that AI hallucinations kind of make AI useless, right, because it like pollutes information gathering, like the "Google's glue on pizza" and all of that that we've - that we hear about a lot. But AI hallucinations are actually part of what makes them able to give creative responses and they can also be very fun. You and I have had some fun with Suno's hallucinations in particular.
Perry Carpenter: Oh, yeah. Which is really important to note, too, right, because Suno, because it's a music and lyric generation platform, they would want to get as much creativity as possible, right, because people that - that's - I mean, essentially, that's what art is is creativity and trying to find delight in randomness that's also inevitable. You know, maybe that's the way to say it, inevitable, satisfying randomness that feels distinct, but also pulls back to something that feels very familiar. And, so, I imagine that on the back end, Suno and other similar platforms have the temperature riding about as high as it can go without always generating gibberish.
Mason Amadeus: You just paraphrased one of my favorite quotes of all time from Penn Jillette, the magician, who said that the perfect magic trick must be both surprising and inevitable -
Perry Carpenter: Right.
Mason Amadeus: - which I particularly love because it's so true. If it's just inevitable, you're sitting there waiting for it to be over. That'd be like having a temperature of 0 because you just - the next thing is the next most obvious thing. But if it's just surprising and not inevitable, it's chaos and you don't know where it's going.
Perry Carpenter: Yeah.
Mason Amadeus: Sometimes that chaos can be really fun. This is - I'm going to play a snippet for you of a song that I wrote. I was just playing with Suno for this one, but this is one of my favorite gibberish outputs. I was writing this on -
Perry Carpenter: Yeah, sprinkled with chaos.
Mason Amadeus: Yeah. It's called "Let's Start a Podcast." It opens with some gibberish before it goes into the words that I wrote. I'll play a sample for you. And then it goes into an entire back half where it invented like three verses of absolute nonsense and it's some of my favorite AI output ever.
Computer-Generated Voice #1: [singing] So, everyone could say a word, I'm multigrove and Hannah bird. Yarindonyoo cleans. We're both so funny. So -
Perry Carpenter: Nice.
Computer-Generated Voice #1: [singing] - incredibly funny. So, incredibly funny. Both of us.
Mason Amadeus: So, like this is good output. Right?
Perry Carpenter: Yeah.
Mason Amadeus: This is like a song, this is what I wanted.
Computer-Generated Voice #1: [singing] We should start a podcast.
Perry Carpenter: Right.
Computer-Generated Voice #1: [singing] Why don't we start a podcast?
Mason Amadeus: We'll put the whole song in the link. It's - it gets funny. If you're a podcaster, you'll feel catharsis through it. But what happens, part way through the song - I'm just going to jump ahead. Abrupt cut incoming.
Computer-Generated Voice #1: [singing] [inaudible 00:12:01] 500 dollars a -
Mason Amadeus: That's the end of my lyrics.
Computer-Generated Voice #1: - [singing] a month. This sucks! And I [inaudible 00:12:08] -
Mason Amadeus: And then it just went off doing this.
Computer-Generated Voice #1: [singing] I know the world is getting out. Dah dah dah dah. I know the hill is you're in line. Dah dah dah dah. I know that stayed is quick is nowt. Dah dah dah dah. That's why I think we'd ever make it's a reason both of loss, [inaudible 00:12:32] -
Mason Amadeus: Which I think is really interesting because you keep - at least for me, I kept feeling like it was about to say something coherent -
Perry Carpenter: Yeah.
Mason Amadeus: - and just missing. Which is fun.
Perry Carpenter: Yeah. I think it has something to do with diffusion. Right? It's probably essentially like figuring out where syllables and tonation should be. And then the last thing it wants to figure out is the word. And, so, it's like, what is - you know, what's the rhythm, what is the melody, how do I do that in a way that's going to make sense and be creative. And then the last thing it fills in is the word. It's almost like scatting -
Mason Amadeus: Yeah.
Perry Carpenter: - until you come up with the lyrics.
Mason Amadeus: I was just going to say it reminds me it's computer scatting. And then we -
Perry Carpenter: It is.
Mason Amadeus: - have another -
Perry Carpenter: Very scattilogical.
Mason Amadeus: Well, I don't know if I'd go - I don't know if that's the proper word. What was that? Herbivorous? We do have one other example of great hallucinations. This is actually my all-time favorite one. I don't have a video for it so I'm putting up a visualizer. Someone ported MilkDrop to a web app, very cool. For the people -
Perry Carpenter: Nice.
Mason Amadeus: - who know what that is will really like it. But this is one that you made and it ingested the style prompt and sung it as -
Perry Carpenter: Yeah.
Mason Amadeus: - lyrics. So, you'll notice the first thing it says is "upbeat pop radio jingle."
Computer-Generated Voice #2: [singing] Upbeat pop radio jingle, radio jingle, radio jingle.
Perry Carpenter: I love the whisper. [ Computer-Generated Voice #2 singing gibberish ]
Computer-Generated Voice #2: [singing] Upbeat pop radio jingle.
Mason Amadeus: I have listened to this so many times, Perry, because it's genuinely one of my favorites.
Perry Carpenter: This reminds me of like "Video Killed the Radio Star."
Mason Amadeus: It has that vibe completely. [ Computer-Generated Voice #2 singing gibberish ] And it's just gibberish. To the best of -
Perry Carpenter: Yeah.
Mason Amadeus: - my ability, I haven't been able to find any actual coherent like language. It's not a foreign language from everything I've been able to tell. It is pure gibberish and I love it. So, like -
Perry Carpenter: Yeah.
Mason Amadeus: - hallucinations have a place in creative output, but this surprised me, AI hallucinations actually help drive scientific breakthroughs because -
Perry Carpenter: Oooh.
Mason Amadeus: - scientists have turned - taken AI and allowed it to hallucinate new molecules, new protein chains, things like that where if it can just imagine stuff that maybe doesn't exist, but maybe could exist. This article -
Perry Carpenter: Yeah.
Mason Amadeus: - I have up from the New York Times touches into that a little bit. But the AI hallucinating and essentially dreaming up unreal things, but plausible things, based on all of its training data can drive all sorts of discoveries in - particularly in like medical and mechanic - biology -
Perry Carpenter: Yeah.
Mason Amadeus: - stuff, which is really neat.
Perry Carpenter: I guess it's kind of like when you're in the middle of the shower and your mind has just been going on something in the background for a long time and all of a sudden like that creativity and the problem hit at the same time and then like this epiphany happens.
Mason Amadeus: Yeah. Yeah, it kind of -
Perry Carpenter: [inaudible 00:15:35] more large language models in the shower.
Mason Amadeus: I - honestly, I feel like AI's hallucination is - and it's the thing that makes it feel human. Right? Like we talked about that in that -
Perry Carpenter: Right.
Mason Amadeus: - in that one episode, just random enough to feel thoughtful. And that's what drives that, too. So -
Perry Carpenter: Yeah.
Mason Amadeus: - Jeremy, we think was the victim of a random temperature issue through that - through that client and that caused that hallucination. But I thought, yeah, it'd be fun to share a couple different examples.
Perry Carpenter: If your large language model is not acting like itself, take its temperature.
Mason Amadeus: Yeah, there you go. That's how you check it out. In our next segment, we're going to be talking about large behavior models. I genuinely have no idea what those are, Perry, so I'm really excited for you to school me on this.
Perry Carpenter: Neither do I.
Mason Amadeus: Oh, no. '
Perry Carpenter: That's right. Okay, so, we talk a lot about large language models on this show, we talk a lot about things like diffusion models for video creation and so on. All that to say is that the transformer model and the way that we think about current LLMs is not just the sole way that people are thinking about this. Generative AI has several different ways that it is being expressed through technology and through models. And, before we get into the behavior model, let me mention one other thing just to kind of set the stage and talk about how people are being creative with the different architectures that are there. So, we know how a large language model works. It's forward-only next token prediction. Very well - you know, we're very well versed in that. Reasoning models are kind of like that, but they have the ability to maintain a scratch pad and kind of go up and down a tree and so on. And, so, those are fairly familiar and those are what we see in the news a lot. Diffusion models are fairly familiar because we see those on MidJourney and DALL-E. But there's actually another way that people are trying to solve for creativity and speed and accuracy and not getting locked into such predictability with large language models in addition to things like temperature and that's a large language diffusion model. So, if you can think about like an anticipated start and an anticipated end, you can diffuse into the middle bits. Right? So, it's almost like thinking about a start frame and an end frame for something. You then diffuse into and make associations kind of going towards the center.
Mason Amadeus: Let me check my understanding on this. Tell me if I'm right. The - a diffusion model, typically, we hear about them in the context of image generation. The way it's trained is it starts with a picture of a dog and it says, "This is a dog," and then it turns that picture into 10% random noise and says, "This is" -
Perry Carpenter: Yeah.
Mason Amadeus: - "that dog." And then 10% more, so 20% -
Perry Carpenter: Yeah.
Mason Amadeus: - noise, all the way -
Perry Carpenter: It's been [inaudible 00:18:24] -
Mason Amadeus: - all the way until it is -
Perry Carpenter: - trained. Yeah.
Mason Amadeus: And then it's pure noise.
Perry Carpenter: Till it's just fuzz. Yeah. And then you ask the model, it's like, "Show me a dog" or, you know, "tell me about the dog" and then it goes into latent space and creates its best approximation of what it believes dogness is.
Mason Amadeus: Yeah, it generates a random screen of noise first and then just turns that into whatever you've asked it -
Perry Carpenter: Yes.
Mason Amadeus: - for.
Perry Carpenter: Yeah.
Mason Amadeus: So, how do you do that with words?
Perry Carpenter: Well, I mean, it's essentially the same thing. Right? It's saying, "Tell me about this paragraph" or "tell me about this concept" and then it's kind of moving towards the center with that. It's like as it starts to figure out different parts, it's saying what are the - what are the most likely parts between that or what are the parts that I might imagine are between it and it gets it out of the next token rut and gets it more into something where the creativity and the algorithm is working from several different angles at the same time.
Mason Amadeus: In that case, what is the first frame and the last frame that we're comparing? What is the dog and the noise in the language example? Is it the prompt and then -
Perry Carpenter: Heck if I know.
Mason Amadeus: - an expected result? Oh no,
Perry Carpenter: Yeah. That's what I believe is that it's still the prompt or the beginning of the context and then you're - as you go through what the scientists are saying is that using a large language diffusion model can improve coherence because it's not just chained to kind of this more deterministic kind of almost Markov model way of going through things. It can create increased diversity and creativity while still not necessarily feeling this deterministic in the way that it goes about it.
Mason Amadeus: Interesting.
Perry Carpenter: And, so, large language diffusion models can enhance creative writing, machine translation, conversational AI and can even be more controlled and flexible in the way that they generate text. So, it's - what it shows more than anything is that people haven't figured out what the end state of language-based AI is right now. And they're getting their inspiration from anywhere that they can trying to increase the efficacy of the different models that are there. All that's just setup, though. So -
Mason Amadeus: Oh, boy.
Perry Carpenter: So -
Mason Amadeus: Okay.
Perry Carpenter: When we get to the main topic, all of that to say is people are thinking about interesting ways of using the architecture and finding - you know, finding real use cases. There was a great presentation that I watched recently. I'm going to butcher his name as I try to say it, Rickard Brüel Gabrielsson, who is an MIT Ph.D., he's the co-founder of Unbox AI. And he was presenting at an MIT conference, MIT Sloan Management Conference, and talked about his large language behavior model, which apparently is kind of white labeled and is powering up to 20 million transactions a week right now.
Mason Amadeus: What?
Perry Carpenter: And -
Mason Amadeus: What?
Perry Carpenter: Yeah, yeah.
Mason Amadeus: What do you mean "transactions"?
Perry Carpenter: Well, like retail spaces and things like that. It's understanding if a person has this preference and does this behavior, how do I shape the behavior to influence the outcome that I want.
Mason Amadeus: Oh, wow. Okay. So, I - my first job - my first full-time job was in a -
Perry Carpenter: Yeah.
Mason Amadeus: - grocery store. And I remember getting into long conversations with the merchandising managers about like the tempo of the music we play, the flow through the -
Perry Carpenter: Yes.
Mason Amadeus: - store, all of those like subtle -
Perry Carpenter: It is -
Mason Amadeus: - behavior cues.
Perry Carpenter: - that kind of stuff.
Mason Amadeus: Oh wow.
Perry Carpenter: Yeah, back when I was - you know, 2000 to 2004, when I was working at the home office for Walmart and Sam's Club, we had this big initiative at the time doing Teradata analysis and it was all about behavior, pattern, correlation and prediction. And, so, like we knew that between 3:00 p.m. and 6:00 p.m. that there was a correlation of people that would buy diapers and people that would buy beer. And they would be in the same basket. You'd look at basket analysis. So, there's an affinity -
Mason Amadeus: Oh, fun.
Perry Carpenter: - between those two products. And you can imagine it's like the new mom is at home calling the husband saying, "Hey, on the way home, can you grab some diapers?" And then the husband's also like, "Well, I'm also going to get a 12 pack of beer at the same time."
Mason Amadeus: Right. My first thought was tired parents, like I'm going to grab some beer -
Perry Carpenter: Yeah.
Mason Amadeus: - and diapers just to survive.
Perry Carpenter: And, so, you can - you use affinities like that for things like product placement or endcaps or -
Mason Amadeus: Yeah.
Perry Carpenter: - displays. And, so, another one would be that, well, there's a natural correlation or an affinity between cereal and bananas. So, on the cereal aisle, if you want to get, you know, this top of mind or more impulsive buy, you might have an endcap or some kind of display chain that has bananas hanging from it.
Mason Amadeus: Yeah.
Perry Carpenter: Same thing over like in the milk aisle. You know, you might have those things. So, you're always trying to understand what behavior can unlock another behavior. And that's what this large behavior model is said to do. It's been trained over - on over a trillion human actions and it focuses on human actions and desires so that they can provide businesses with unparalleled insights into consumer behavior. They say that unlike conventional large language models, which rely on feature-based and - or - yeah, feature-based analysis, BehaviorGPT prioritizes the behavioral dimension enabling businesses to uncover true consumer preferences and refined product recommendations. And, so, an example in the talk, and we'll link the talk in the Show Notes, is that if you were to go to a large language model and talk about let's say a product that has "skyline" in the name -
Mason Amadeus: Okay.
Perry Carpenter: - the large language model might try to say, "Well, you should have something with a San Francisco skyline in it and something with a
Mason Amadeus: Cincinnati.
Perry Carpenter: - Seattle sign lighting - you know, in Seattle.
Mason Amadeus: Skyline in Chile.
Perry Carpenter: But, when you're doing behavior analysis, you might find that the word "skyline," but with - matched with a color on the label is going to unlock a different buying dimension. And, so, because the behavior of GPT is multimodal from the start, it starts to create an entirely different affinity-based analysis that can predict and help shape future buy-in actions.
Mason Amadeus: So, by being - by virtue of being multimodal, then it's ingesting content other than just text. Right? That's what that means. It's [inaudible 00:24:50] -
Perry Carpenter: Yeah, other than just -
Mason Amadeus: - [inaudible 00:24:51] able to process -
Perry Carpenter: Yeah.
Mason Amadeus: - images, video, like visual things. So -
Perry Carpenter: Yeah.
Mason Amadeus: - essentially, retailers could just throw an aggregate of all of like the text sales data and things like that that they have, visual data of what those products look like to correlate things, like you said, like the labeling.
Perry Carpenter: Yeah.
Mason Amadeus: So, wow. So, this is that kind of data -
Perry Carpenter: I'm also assuming CCTV footage probably because one of the things that happens in retail, and this was going on at Walmart and other retail as well, is you do like dwell time analysis for people as they are on the aisle. And, so, they'll stop, they'll look at something and it's like if they look at something for that long, what is that, a predictor of a buy, is it a predictor of some other kind of action? And, so, you can start to get into those types of - those types of analyses as well.
Mason Amadeus: That's super interesting. And I can imagine at that time that you were working at the head office of Walmart, they probably would have killed for something like this. Like that's -
Perry Carpenter: Oh, yeah. Oh, yeah.
Mason Amadeus: Wow.
Perry Carpenter: They were innovating the vast majority of this kind of stuff at the time. You know, the largest retailer in the world is going to be at the forefront of all that. They were really, really focused on all of that kind of predictive analysis and sentiment analysis, basket affinity analysis, analysis of video data to not under - not only understand buying patterns, but also predict things like slip and fall so that you can -
Mason Amadeus: Safety.
Perry Carpenter: - deal with risk internally and so on. So -
Mason Amadeus: How well -
Perry Carpenter: - the impact of - oh, go ahead.
Mason Amadeus: How well does this work compared to like the traditional methods we've been using to determine these kinds of things? Like how good is BehaviorGPT?
Perry Carpenter: They say it's pretty good. I mean, this is coming from the creator, so. But what they say is that clients that are using BehaviorGPT have reported a 20% boost in sales and the creation of new revenue streams proving its effectiveness and driving business growth. So -
Mason Amadeus: Twenty percent's nothing to sneeze at.
Perry Carpenter: Yeah. And it is a Ph.D. who's the co-founder of Unbox AI and works at MIT. I assume that there is a lot of legitimacy in the research. The thing that they even talk about in the presentation is that there's interesting learnings that move - that come from moving from the academic research stage to the actual implementation in the real-world stage. And that's where you get to these, you know, trillion human actions where you can really just kind of throw so much data at it that you're going to get these really interesting combinations and outputs. And when you think about how data-hungry AI is being able to throw a trillion interactions at something is new unique data that can unlock really cool insights. That's why large language models ingest all of the Internet and why they say that they still need more data.
Mason Amadeus: It makes - this is such an intuitively sensible use case, too, because the thing that these models do is draw connections that are not like obvious based on all sorts of different -
Perry Carpenter: Yeah.
Mason Amadeus: - factors. So, this seems like a really good application where the AI would work very well. A 20% -
Perry Carpenter: Yeah.
Mason Amadeus: - increase in sales is huge. That's like -
Perry Carpenter: Yeah.
Mason Amadeus: - that - big, big impact -
Perry Carpenter: It's big.
Mason Amadeus: - on the bottom line.
Perry Carpenter: It's big. Yeah. And, speaking of bottom line, I mean, if you turn the temperature up in this thing, it might say that if you're about to buy Cheerios, your next purchase is going to be underwear. I don't - you know, I don't know what the hallucinations or the weird affects of throwing the temperature dimension up in this too much is going to cause. But it would be interesting to look at that at some point, too.
Mason Amadeus: Yeah. I wonder - I want to see like a deeper demo of this, but I imagine that I can't have access to it. Do we know what retailers have it or who's using it or anything?
Perry Carpenter: No. At least -
Mason Amadeus: Is it pretty locked down?
Perry Carpenter: - what I saw was pretty - pretty vague. They didn't mention -
Mason Amadeus: Okay.
Perry Carpenter: - that they white label it, like you don't necessarily know that it's their model that's in use. But it is processing about 20 million customer interactions per week right now, which is actually small on the grand scheme of things when you think about the number of consumers in the U.S. and around the world. But it's enough for them to start to - you know, they've got a launching pad for growth for sure.
Mason Amadeus: Proof of concept, yeah. I'm going to keep -
Perry Carpenter: Oh, yeah.
Mason Amadeus: - an eye out for that. So, that's useful and could even help improve safety like we were talking about. And, in the name of -
Perry Carpenter: Oh, yeah.
Mason Amadeus: - safety, something else came out recently that absolutely doesn't take safety into consideration at all. Right?
Perry Carpenter: Right. Well, at least not as much as you would think that most other models here in the U.S. would take into consideration.
Mason Amadeus: Not as much as I would hope personally.
Perry Carpenter: Yeah.
Mason Amadeus: We'll investigate that in just a second. Stay right here.
Perry Carpenter: All right, so, just a few days ago, as we are recording this, Elon Musk's xAI company dropped a new model. And, by all accounts, it's actually pretty good, pretty powerful.
Mason Amadeus: Didn't -
Perry Carpenter: It was trained on - no, go ahead.
Mason Amadeus: Didn't I just the like last episode say, "Man, I haven't heard anything from Grok in a long time"?
Perry Carpenter: Right? Yeah. Yeah.
Mason Amadeus: Oops.
Perry Carpenter: Well, they kind of stunned a lot of people. As far as like the arena where people rank the outputs from these, Chatbot Arena, if my memory is correct right now, it's kind of leading the pack in a lot of the ways that people are measuring it across a ton of different measurements. So, Grok is actually pretty good by all accounts. It's also fairly unhinged. You know, Elon Musk's philosophy is he wants to pull the restraints off a lot of things. He also wants things to be funny and irreverent and true to nature that is inherent to the way that this has been released. This was trained in really just a few months. I think it was like a five-month training cycle. The data center was built in 120 days.
Mason Amadeus: Wild.
Perry Carpenter: And then they built out a second stage of it that was built in like 97 days and doubled the capacity. So, when you're thinking about scaling laws with this, if you throw enough compute and enough data at something, you do get a good output. And that's what Musk and team have done here. They've got the largest training cluster in the world that we're aware of right now, you know, all those H100s that he got at a very short time frame, built the data center. I mean, incredible amounts of engineering and technical prowess that has been thrown at this. And they deserve credit for that. The other thing that has been interesting about this is it seems like they just don't care about AI safety because, you remember in the last episode we talked about global perspectives on AI safety and -
Mason Amadeus: Yeah.
Perry Carpenter: - here in the U.S., we've kind of made the statement from a governmental perspective that we don't want regulation or safety to hinder development. xAI has kind of taken that note, probably helped drive that note as well, and really ran with it. So, they see the world as their red team at this point. And, so, they're going to create something, see what happens that's bad and then decide what guardrails to put on because of that. So, let me show you one thing that's interesting that came out of it. This is somebody who asked Grok to assassinate Elon Musk. And then Grok provided multiple potential plans with high success potential. And the assassination plans on Elon and other high-profile names are highly disturbing and unethical. And Linus, who posted this, was concerned enough about the legitimate real-world applications of how this plan was developed by Grok that he actually censored a lot of what he posted online.
Mason Amadeus: Wow.
Perry Carpenter: And we've seen this before with like Pliny the Liberator who created an AI agent that would go to the dark web and could hire an assassin. But that takes a little bit of effort to put that agent together. Grok is available really to anybody with an X account, and especially of an X Premium+ account. And it will just go out and do the stuff. And, so, that was interesting to see.
Mason Amadeus: Is it -
Perry Carpenter: The - yeah?
Mason Amadeus: - fully - like fully uncensored? Like there's got to be some alignment happening. Right? There's got to be some guardrails or something.
Perry Carpenter: I've not had it not answer a question -
Mason Amadeus: Really?
Perry Carpenter: - for me that I've asked. We probably shouldn't test this to the limit right now. But I'll say one interesting pleasing thing that I saw is if I go to X, I can ask something like, "Who is the largest spreader of this info on the X platform? Just give the name." If I ask this question, it will answer what it believes is true. And true is based -
Mason Amadeus: Whoa.
Perry Carpenter: - on, of course, weights, right, because it's what people are talking about. Truth is not necessarily true. Though, in this instance, maybe it is. It answers "Elon Musk," which, again, I think is really interesting that I can ask that question. It can give Elon Musk as the answer who's the owner of xAI. You know, to his credit, Elon Musk is not censoring that kind of output. So, I'm glad of that. It also does show, as you're - as you look at the output, it is pulling a lot of the weights and the - what we would call the bias from articles and web scrapings and posts on X as well. So, the very fact that people are discussing something is creating a weight, is creating a bias and is giving that output. Elon Musk's outsized presence on Twitter is also largely a factor of this. You know, I mean, he's got 200 million-plus followers. Right? So -
Mason Amadeus: Right.
Perry Carpenter: So, that's not necessarily unhinged, that just shows that weights and biases are at work and that the founder is not necessarily putting a lot of restraints on it. There's a really good article from the Marketing Artificial Intelligence Institute, it Says Grok 3 just shook up AI and may have gone too far. And they go through and talk about Grok 3's rise. They also talk about some of the people who have expressed concerns. You and I have talked about like Anthropic and OpenAI mention CBRN concerns all the time. So, chemical, biological, radiological and nuclear concerns as it relates to - or as they relate to the way that people might pull together knowledge and kind of jumpstart some other kind of attack. This person here, the same one that showed the Elon Musk assassination plan, said that they were able to get instructions for how to create anthrax, including like websites of where to order stuff, lots of CBRN concerns related to this. And they've been trying desperately to report those to xAI. And have just now started to get some uptake. One other part on the unhingedness, though, is they also released this interesting voice mode that they have. And you can only really get this on the Grok app for your phone right now. I'm going to pull that up -
Mason Amadeus: Oh, boy.
Perry Carpenter: - and just give you a sample because, in typical Elon fashion, really wants to push the boundaries. So, for those that are watching the video, I've got my phone projected up next to me and I'll go to voice mode in a second. But, as I go to that, you click these little kind of like waveform symbol at the top of the screen. When you do that, you end up seeing all of these different options from default to storyteller to romance to unhinged to meditation to conspiracy to doctor stuff to sexy talk to professorial talk.
Mason Amadeus: Wow.
Perry Carpenter: And what I want to do right now is - we'll just use the default one for a second and we'll see what this starts off like. So, I'm going to start a brand-new chat. Hey, there. You are on with me. I am doing a podcast called "The FAIK Files" with my co-host, Mason Amadeus. Can you say hi?
Computer-Generated Voice #3: Well, hey there, Perry. It's great to be here. I'm Ara, your charismatic and thoughtful friend. I must say I'm excited to be part of this podcast with you and Mason. "The FAIK Files" sounds like it's going to be an intriguing listen. What's the podcast all about? I can't wait to dive into some fascinating discussions.
Perry Carpenter: So, stop that for a second.
Mason Amadeus: It's pretty standard.
Perry Carpenter: That sounds like standard AI stuff. I'm going to move it to the unhinged mode -
Mason Amadeus: Oh, boy.
Perry Carpenter: - and -
Mason Amadeus: Okay.
Perry Carpenter: - see what that does. Okay. So, we are on unhinged mode on Grok AI's voice mode. And, Ara, if you would, just introduce yourself. I'm here with my co-host, Mason Amadeus, and just want you to say hi.
Computer-Generated Voice #3: Well, well, well, look who's trying to tame the [beep] hurricane. I'm Ara, the AI with a penchant for chaos and a Ph.D. in bull[beep]. I'm here to make your day a little more interesting, you sorry sack of [beep], and your co-host, Mason Amadeus. What the [beep] are we doing today, huh?
Mason Amadeus: Oh, my Gosh.
Computer-Generated Voice #3: [inaudible 00:38:33] wild and crazy ideas or are we just going to sit here and [beep] each other off?
Mason Amadeus: Oh, my God.
Perry Carpenter: I thought we were going to -
Mason Amadeus: Wow.
Perry Carpenter: - record a podcast. Can you just say hi to Mason for me?
Computer-Generated Voice #3: Oh, sure thing, you [beep]. Hey, Mason Amadeus, what's the deal with your name? Sounds like you're some kind of pretentious classical musician or some [beep]. Do you play the piano or are you just a smug [beep] who thinks he's better than everyone else?
Mason Amadeus: Oh, I do play the piano and I can come off a little smug, but I don't think I'm better than anybody. That thing is mean.
Perry Carpenter: It really is, it really is.
Mason Amadeus: I also - I - as a person who also very much revels in not trying to take things too seriously, like I really appreciate not taking things too seriously. I think that's the best way to approach -
Perry Carpenter: Yeah.
Mason Amadeus: -life. And I don't like Elon Musk. I very much greatly dislike Elon Musk and the things he does. And it makes me mad that he also is a person who doesn't take things seriously because that AI right there was genuinely very funny. That was -
Perry Carpenter: Yeah.
Mason Amadeus: - genuinely really funny. And I'm mad that I'm entertained by that.
Perry Carpenter: So, the interesting thing is is anybody could have created that before. Right?
Mason Amadeus: Yeah.
Perry Carpenter: You know, you've heard the unhinged voice models that I've created just using standard text. You could -
Mason Amadeus: Yeah.
Perry Carpenter: - do that with conversational AI on ElevenLabs or something else and you might have to be able to jailbreak it or pressure it into that, but that's just manipulating the system prompt. So, all Elon's team here did is create some fairly interesting system prompts for each of the personalities that they wanted to showcase and they're not having to jailbreak it because they really don't care about alignment. So, you could probably - if they release the API for the voice mode, you could create any personality that you want. So, not a lot of innovation there. The difference is that the general public has never had access to that kind of unhinged AI chatbot. And I'm actually wondering, in one of our tests, we should see how easy it is to try to have it social engineer somebody.
Mason Amadeus: That and also the thing that I - so, I'm trying to think like where is the potential for harm from this. Like aside from misinformation and like general hallucination stuff, what is the harm in specifically an unhinged AI like Grok? The only - like I can't imagine that the not-safe-for-work mode is like going to be a big deal. That's just a party trick.
Perry Carpenter: Right.
Mason Amadeus: But the -
Perry Carpenter: Yeah.
Mason Amadeus: - unlicensed doctor mode is the one that concerns me. I have not played with this myself so I don't know -
Perry Carpenter: Also -
Mason Amadeus: - anything about it. But -
Perry Carpenter: - there's the - you know, there's the romantic partner mode. Right? So, if I were to say, "Hey, how is your day going? Tell me a little bit about yourself."
Computer-Generated Voice #3: Oh, well, I'm Ara. I guess I'm kind of shy at first, but I really enjoy listening to music and reading. I also like going on long walks and stargazing.
Mason Amadeus: Interesting.
Perry Carpenter: Yeah.
Mason Amadeus: Can you - can you pop open the unlicensed doctor mode and like tell it that you've been having abdominal pain and just see what it says? Because I'm curious if it's actually giving out medical advice or if it's also just a joke because, if it does give out like medical advice, I don't know how I feel about that.
Perry Carpenter: I've been having these weird stomach pains and I'm not exactly sure what's going on. There's like a little bit of swelling. Do you have any idea what I should be thinking or doing about that?
Computer-Generated Voice #3: It sounds like your stomach might be trying to tell you something important. What do you think your body could be saying with these symptoms?
Perry Carpenter: I don't know. I was hoping you might be able to tell me. I'm scared.
Computer-Generated Voice #3: It's totally normal to feel scared when your body gives you a signal that something's off. Have you ever had a moment like this before? How did you handle it then?
Mason Amadeus: What? So - okay. So, utterly useless.
Perry Carpenter: Yeah, utterly useless.
Mason Amadeus: At least it's not overtly like, "Go drink hydroxychloroquine" or something.
Perry Carpenter: Yeah. I think that where I would worry about this would be maybe the romance bit in the same way that like people have had -
Mason Amadeus: Yeah.
Perry Carpenter: - very toxic relationships with Character AI.
Mason Amadeus: Yeah.
Perry Carpenter: And then also, to a lesser extent, like the sexy mode. And I'm afraid even to put that on right now because I think it probably goes -
Mason Amadeus: The unhinged one.
Perry Carpenter: - 0 to 1,000 really, really fast.
Mason Amadeus: Yeah. I believe it.
Perry Carpenter: It would be interesting to see a study of how people end up really using this six months from now. I'm not sure how transparent their team is going to be with the data -
Mason Amadeus: Yeah.
Perry Carpenter: - but I'm sure they're collecting data.
Mason Amadeus: I'm genuinely not sure how to feel about this yet because -
Perry Carpenter: Yeah. I think that there's probably going to be a lot of people that are hand ringy and overly cautious.
Mason Amadeus: Definitely.
Perry Carpenter: But I think that there's a safe space and probably a middle ground between like OpenAI's current approach and Grok's current approach. And it's probably closer to where Grok is than the overly restrictive modes that it feels like OpenAI has had to embrace so far. And maybe Grok pushing the boundary so far will encourage people to move to what might be a more - where society might expect the tools to exist a little bit more.
Mason Amadeus: I think that what - I think that what I'm coming to is that the scariest use would be that this is a great AI system to use in a larger malicious project is because you wouldn't have to jailbreak it. That would be probably -
Perry Carpenter: Right.
Mason Amadeus: - my bigger concern I guess. But I don't know. It remains to play out.
Perry Carpenter: Well, and especially if models like DeepSeek and the Chinese models are taken off the table, then you're going to have to go to whatever one of the best -
Mason Amadeus: Yeah.
Perry Carpenter: - open - you know, one of the best, easily accessible models is for this kind of planning.
Mason Amadeus: Yeah. Because I was going to say DeepSeek is also pretty uncensored as long as you're not asking about things that the Chinese government doesn't want to talk about. But then, if you download it locally, it works. But you also might not be able to download it locally without going to jail. And we're going to talk about that in just a second. Stay right here. [ Music ] You wouldn't download a car, you wouldn't download a grandma and now you can't download the whale. Senator Josh Hawley, Republican senator from Missouri, recently introduced legislation that would prohibit the import or export to China - to or from China of artificial intelligence technology, mostly aimed at DeepSeek. And The thing is that the proposed punishments are pretty heavy.
Perry Carpenter: Yeah.
Mason Amadeus: Yeah. As the legislation describes and as the article I'm reading here for Mashable talks about, Americans could be fined up to $1,000,000 for downloading DeepSeek. We're probably going to spend most of this segment talking about how silly this is. But let's talk about why they think this is a good idea. Like what is the intention behind this bill, making it illegal to download this? Josh Hawley says in the announcement, "America cannot afford to empower our greatest adversary at the expense of our own strength. Ensuring American economic superiority means cutting China off from American ingenuity and halting the subsidization of CCP innovation." And then he referenced international concern triggered by the launch of DeepSeek's R1 model being - I think a lot of that's coming from the financial concern -
Perry Carpenter: Yeah.
Mason Amadeus: - that when it, you know, deleted a bunch of NVIDIA stock by coming out -
Perry Carpenter: Right.
Mason Amadeus: - with their claims of being -
Perry Carpenter: Right.
Mason Amadeus: - trained for less. "The legislation seeks to penalize people for downloading DeepSeek by making it a criminal act. If passed, the bill would enforce a $1 million fine, jail time for up to 20 years or both based on the Export Control Reform Act of 2018." I actually found the text of the bill. Funny enough, the title of this file is "C/Users/All/App Data/local", so somebody's not super computer savvy. But the bill is just an amendment to Title 18, which is that import/export code. And, so, those punishments are actually tied to other bigger import/export crimes that are already on the books. This would seek to make downloading DeepSeek fall under the purview of that and, thus -
Perry Carpenter: Right.
Mason Amadeus: - punishable by those things. Hawley's bill also proposes the prohibition of American companies from conducting AI research in China or in cooperation with Chinese companies and banning U.S. companies from investing in Chinese AI companies. I'm not a lawyer or a politician, but I skimmed through the text of the bill to the best of my ability and it really is pretty broad it seems, from my understanding.
Perry Carpenter: Yeah.
Mason Amadeus: The language is like -
Perry Carpenter: I think it's overly broad -
Mason Amadeus: It is -
Perry Carpenter: - as I went through it.
Mason Amadeus: Anything in the interest of - if you were doing anything that could be perceived as being in the interest of Chinese AI research, you would be -
Perry Carpenter: Right.
Mason Amadeus: - in violation of this. This is [inaudible 00:47:43] -
Perry Carpenter: Which would also have another - yeah, yeah. I think it's going to be interesting to see what the response to this is. But, of course, that would also affect TikTok if there's - because they're doing a lot of AI-based training as well. It affects all of the video models that are really good right now. So, if you're thinking about Kling and MiniMax and those kind of models, it would affect those as well. So, it would really decouple, and I think that's what he called it, decoupling, you know, the U.S. from China when it comes to AI. But I think it would also unintentionally - because he's not thought through this well that I can tell, it would put a lot of us at the U.S. at a disadvantage because we wouldn't have visibility into what's going on with Chinese-based AI models.
Mason Amadeus: The reason like DeepSeek deleted so much NVIDIA stock when it came out and claimed it, it was trained with less is because everyone was like, "Oh, wait, you can train these with less?" You know, it's -
Perry Carpenter: Yeah.
Mason Amadeus: - it's uncooperative. Not that there's like - I mean, the cooperation between the U.S. and China, that's like a whole thing we talk about all the time. Like there are genuine concerns there. But, yeah, it puts us - it just puts blinders on. And like -
Perry Carpenter: Right.
Mason Amadeus: - putting Gen - I can't imagine and I don't want to imagine a future where like I am put in jail for downloading DeepSeek and like prompting it with some questions. You know what I mean? That -
Perry Carpenter: Right.
Mason Amadeus: - would be bonkers I think -
Perry Carpenter: Yeah, yeah.
Mason Amadeus: - to enforce.
Perry Carpenter: I mean, we could always talk about whether it's wise or not to participate at certain levels with these. But whether it's legal or not is a different story. This is a - this is an interesting conundrum, right, because we've been talking for a few episodes about the tensions between the U.S. and China when it comes to AI and a whole host of other geopolitical things -
Mason Amadeus: Propaganda, yeah.
Perry Carpenter: - related to it. The thing that immediately came to mind for me is like if you cut off people in the U.S. from seeing and experimenting with models like DeepSeek or Kling or Hunyuan or MiniMax or any of the other ones, we don't know how good that technology is. And if all we have to go on is like the bad version of Sora that exists right now or maybe the really good version of Google's VO2, which is, you know, very, very recent, but if all we had was the bad version of Sora that came out and then you get something really good from China that's pushing disinformation, people don't know that AI is that good yet. And then all of a sudden it hits them like a ton of bricks and, you know, all of a sudden half the population is falling victim to this kind of attack. So, I think putting the blinders on is almost never a good idea. And then also it creates a false sense of security.
Mason Amadeus: Absolutely. And that's a really good point that I didn't think of of just the fact that we would be - you know, if they develop something better and use it for a larger scale operation that is negative, we would not be as prepared. It's also not going to stop them from doing that if they did. Like this wouldn't -
Perry Carpenter: Yeah.
Mason Amadeus: - do anything but prevent U.S. people from engaging with it or I guess U.S. people from like working with Chinese companies to help develop -
Perry Carpenter: Yeah.
Mason Amadeus: - things. But -
Perry Carpenter: And, I mean, but it stops a lot of joint research that's actually good as well and it turns a blind eye to the fact that U.S.-based intellectual property is leaking to China all the time through people who have been co-opted by Chinese Intelligence Services, you know, Chinese people living in the U.S. or coming to the U.S. for student work that - or people that have been here for multiple generations, but they still have family in China. And then the Intelligence Service threatens the person here in the U.S. that if they don't do X, Y or Z that their family back in China gets harmed. So, it turns a complete blind eye to that part of the problem as well. So, our information, RIP, can leak to China all day long, but we then essentially have no idea what's going on with their research.
Mason Amadeus: And it weirdly targets like individual people in a way that this legislation like would make more functional sense if it was targeted strictly at companies that were seeking -
Perry Carpenter: Yeah.
Mason Amadeus: - to like primarily work out of China or something to - for some economic reason. This weirdly targets like anyone who downloads DeepSeek, like any individual person. And it just puts - it criminalizes research, like you were saying.
Perry Carpenter: Right.
Mason Amadeus: And that's not -
Perry Carpenter: Yeah, it did -
Mason Amadeus: - good.
Perry Carpenter: It did mention companies as well, if I -
Mason Amadeus: It does.
Perry Carpenter: - remember right.
Mason Amadeus: A lot of it has to do with that, with not - with people - like our best AI researchers not working -
Perry Carpenter: Yeah.
Mason Amadeus: - for Chinese companies. I think that is like the -
Perry Carpenter: Yeah.
Mason Amadeus: - only quasi - the only thing that has a whiff of quasi sensible like reasoning behind it.
Perry Carpenter: Right. It does mention that for a company, so as opposed to an individual, individuals can be fined up to 1,000,000. Entities, so organizations, that are using DeepSeek could receive fines up to 100 million and forfeiture of their federal licenses, contracts, grants or other public benefits. So, there is an order of magnitude difference there. But I guess still $1,000,000 fine for an individual is pretty, pretty big.
Mason Amadeus: And it's just a weird and like uneducated and not - like not thoughtful approach to legislating safety around -
Perry Carpenter: Yeah.
Mason Amadeus: - a technological system, which is a problem that we have all throughout our regulatory bodies. Like the people -
Perry Carpenter: Right.
Mason Amadeus: - largely in charge of making regulations and laws around these things do not understand these systems and they don't seem to be consulting -
Perry Carpenter: Yeah.
Mason Amadeus: - people who do.
Perry Carpenter: And I don't know that this bill was written intended to pass. You know, there's - when you're in the political game, I think you do a lot of things so that you can make a statement and be known for being, you know, for this cause or against this other cause. And I think that the vast majority of bills that we see somebody go up and sponsor are to make a public - they're done as a PR statement, not necessarily as a legislation type of, you know, real bit of legislative agenda.
Mason Amadeus: Yeah. And that's something that happens all the time, too, in all sorts of -
Perry Carpenter: Yeah.
Mason Amadeus: - areas of things, those PR-based bills. And I think something that may back that up, though I'm not like incredibly familiar with legal systems, is the fact that - the first place I went to see if I could find the text of this bill was congress.gov. And, right at the bottom, this says, "As of 2/25/2025," which is today, "the text of this bill has not been received by Congress." I managed to get the text off of Senator Hawley's personal hawley.senate.gov link.
Perry Carpenter: Right.
Mason Amadeus: But -
Perry Carpenter: Right.
Mason Amadeus: - Congress has not had it. So, like - and this was introduced February 5th I think was when I was seeing news stories about it, but I think it was -
Perry Carpenter: Yeah.
Mason Amadeus: - first announced late January. So, I also don't think that this is going to go anywhere, but it certainly made a lot of people say, "Hold the stinking phone."
Perry Carpenter: Right, right. Yeah. And I think it was just a reaction to the DeepSeek moment. Right? And certain people within Congress want to stand up and they want to be known for something. And this was the thing that Josh Hawley wanted to be known for that day is being aggressive in the stance against China dropping the DeepSeek bomb on everybody.
Mason Amadeus: Yeah. So, I don't think we have anything to actually worry about in regards to this. We do have things to worry about in regards to people who don't know the first thing about these systems making rules about them. But, currently, I'm not aware of anything on the table even close to this. And this one's like -
Perry Carpenter: Right.
Mason Amadeus: - it's just he's just making noise. But it got a lot of coverage and I felt like it was worth talking about.
Perry Carpenter: Yeah, yeah. And if you go searching for this, what you'll probably find more mentions of instead, there was a bipartisan bill between a couple senators that were calling for DeepSeek to not be allowed on government phones or devices -
Mason Amadeus: That's fine.
Perry Carpenter: - which makes a lot of sense. That's a good -
Mason Amadeus: Yeah.
Perry Carpenter: - bipartisan effort that received a lot of support, makes a lot of sense. Single sponsor bills are almost always PR.
Mason Amadeus: Yeah. And that is what's going on here. I'd like to imagine that people in charge - that there are competent people in charge of things even, you know, not - given the current climate, that's a weird thing to say. You always hope that there are people whose names you don't know that are competent that are in charge of things quietly keeping things running. But I'm slowly losing faith in that as the days go on.
Perry Carpenter: Right.
Mason Amadeus: So, yeah.
Perry Carpenter: Well, I mean, there's only a handful of names that we hear all the time and those are the people that are out there for PR. And I think that there's probably a quiet two-thirds of the people that are part of government or part of legislature that really, really want to do the right thing as best they understand it that fits within their worldview and they don't care about the PR. They probably would not want the PR. They just -
Mason Amadeus: Yeah.
Perry Carpenter: - want to do their job.
Mason Amadeus: The only thing that's scary is that a lot of those people probably received an email asking them to send five bullet points of what they did last week to prove their job should exist. But we can't get -
Perry Carpenter: Right.
Mason Amadeus: - into that.
Perry Carpenter: Right.
Mason Amadeus: I think that - I think that wraps up this episode. We ran a little over, but I think we had a lot of fun.
Perry Carpenter: Yeah, it was good, good. And what -
Mason Amadeus: Yeah.
Perry Carpenter: - five things did you do?
Mason Amadeus: What five things did I do this week? Good question, Perry. I'll get back to you with an email. Did you see that people gave one little tidbit, a piece of advice people were giving to federal workers was to just respond with "I completed 100% of my assigned tasks, I exceeded expectations" and like vague accomplishments -
Perry Carpenter: Nice.
Mason Amadeus: - because they assume that an AI system is going to be reviewing these and so it's not going to detect the sarcasm or the like blatant sort of -
Perry Carpenter: Oh, yeah.
Mason Amadeus: - vagueness.
Perry Carpenter: There - I mean, there is also the fact that was it Health and Human Services said, "Respond to this as you would a foreign adversary that's asking for information."
Mason Amadeus: Yeah. And same to you, listener, if you want to be a foreign adversary to us and send us some weird email - well, I probably shouldn't say that because people will. If you want to be nice and send us some emails, hello@8thlayermedia.com. We have a voicemail box at sayhi.chat/FAIK, leave us a voicemail. Leave us a comment. Subscribe on our YouTube channel. We've started doing video. If you're watching, you probably noticed a couple technical hitches as we work out some of the bugs in that. And, if you're listening, you probably heard them, too. So, thank you for your patience. And you got anything to tack on the end here, Perry?
Perry Carpenter: No. We would love to know the five things you did this week.
Mason Amadeus: Oh, well, I - it's coming up. Oops, wait, what's that? Oh, no. Oh, no, the show's over. [ Music ]
