Project Synapse: The Very Scary AI Halloween Edition

Speaker A: Where's everybody in costumes? This is the Halloween edition.

Speaker B: I came as Jim.

Speaker A: Uh, John. Who did you come as?

Speaker C: An old retired guy.

Speaker B: Yeah,

Speaker A: okay, so

Speaker B: we haven't got. Obviously you have the spirit.

Speaker C: Yeah, but

Speaker A: I'm trying to. What am I dressed up as? Come on.

Speaker B: You're that guy from the FTD florist ad.

Speaker A: Not exactly, no. But who is it? Who's the guy from the FTD florist ad? Supposed to represent

Speaker B: the God of herpes or. No.

Speaker A: Jesus Christ.

Speaker B: Sorry, Hermes. What?

Speaker A: Messenger of the Gods, Olympus. Uh, I am an agent. I am an AI agent. I'm Hermes, an AI agent. Also known as Mercury. M, depending on whether you're going for Roman or Greek gods. Do I even try? What is it with you people?

Speaker B: Because you. Because every. Because you bring us sunshine, Marcel. There's always some creativity. Welcome to Project synapse. The Halloween edition. Spooky

Speaker A: was until M. A couple of minutes ago.

Speaker B: Yeah, yeah. Now. Yeah, now it's all. No, it's all fun and games till somebody pokes out an AI

Speaker A: I actually had to take off the.

Speaker B: It was pretty good. Come on, guys, stay with me on this. It was all fine until somebody pokes out an A.I.

Speaker C: there you go.

Speaker A: Okay, I'll let you have that one.

Speaker B: Okay, so here we are. I prepared something else for the lightning round. I took your stuff, Marcel, that you gave yesterday, and I actually put together the construct of a lightning round.

Speaker A: Okay. Can I, before the lightning round, however, share something with you that just to land on the Halloween spirit just a little tiny bit more?

Speaker B: Absolutely.

Speaker A: Okay.

Speaker C: Uh,

Speaker A: so my AI powered car, okay, has a Halloween mode. Okay. So I would like to share with you the Halloween mode before we can. Before we retire Halloween completely. Uh, are we okay with that?

Speaker B: I wasn't planning on retiring Halloween anyway, but go for it. Yeah.

Speaker A: Okay. There we go. There we go. Hey, Grock, turn on Halloween mode.

Speaker D: Halloween mode is on. Enjoy the spooky vibes.

Speaker B: And, um, on the screen it says, happy Halloween. Oh, look, there's a witch flying.

Speaker A: Yeah. So this is the screen from my Tesla. So this is the screen from my Tesla. And it boots up in Halloween mode. And now you've got the front panel in Halloween mode. It even plays some spooky music. And if you touch the things on the screen, it's a

Speaker B: copyright.

Speaker C: But, Marcel, that car is flying. So it must be the new roadster.

Speaker A: Well, it's a ghost. It floats. It floats off the ground. It's a ghost car.

Speaker B: And for those who are listening to this, Marcel is showing us A picture of his of a car with a sheet over it that is pretending to be a ghost. So this is inside your car?

Speaker A: Yeah, this is the. This is the screen on my Tesla. And for Halloween, they downloaded what they called the Halloween mode. And the screen plays spooky music and you can touch it. And when you're driving down the road, all of the cars look like green sort of dripping zombie type things. And you can't see it here, but when you see people walking along the side, they all look like skeletons doing the zombie walk as they're walking beside the car. It's a blast. It is a lot of fun. So there you go. The car is in the spirit. It even has a light show. So when you walk away from the car, it. It growls. Okay. When it locks the doors, when you walk away from it, it growls. And then if somebody comes near the car and the sentry mode is on, which watches with the cameras to make sure that people aren't doing weird things, it'll make weird sounds and creaking noises and spooky Halloween stuff and flash its lights orange at you. It's blast. It's fun. Anyway, I'm going to stop that, but I just. I am in the Halloween spirit, unlike, um, some people.

Speaker C: Your Tesla dollars hard at work.

Speaker B: Yeah, well.

Speaker A: And hey, uh, Come on. Don't forget the hat. Don't forget the hat.

Speaker B: We can't forget the hat.

Speaker C: Of course.

Speaker B: That's

Speaker A: right. I'm an AI agent.

Speaker B: Yep, there you go. Or

Speaker A: Hermes in this case.

Speaker B: That's good. Okay, so the news from this week, we're going to go through lightning round here. I'm going to go through them really quickly and then we'll come back on the ones that we want. So keep your notes. This is an amalgamation of all the stories from this week that we came up with. Here we go, Canada. And first one, I'm actually going to take some time with this. I won't take time with the rest of. But Jim's Carnac M prediction was there would be another major foundation model in the world in the next few months. I predicted Mistral coming out and Canada's Cohere. And in case anybody wants to go Canada what? Canada's Cohere, which was founded by students who'd worked on some of the early papers on AI who came back to Canada and wanted to work here. They found it cohere. It has grown like mad. It is now merged with a company in Germany, but the head offices stay in Canada. The Canadian government has made a $1.6 billion investment into it. Bell Canada is building enormous data centers. And it is a true foundation model. In fact, it is now the biggest AI company outside the US So there you go, another foundation

Speaker C: model. It's not a donation. It is an actual investment. Where Canada will be. My understanding is Canada will be part owner, which is a nice change. It's an equity

Speaker B: investment.

Speaker C: They can't just take a donation and go, oh, now we're going to sell it to somebody in another country

Speaker B: and um,

Speaker A: give it a few years. It's happened with the 407, right?

Speaker C: Yes, true.

Speaker B: Yeah. It depends on who you elect. Yeah, that's for sure. Yeah. We're idiots that sold our highway. Who sells a toll high after you built it? Only somebody very corrupt or very stupid. So we do have our share of that. But the thing is, and this is a big deal for anybody and this is not anti us. This is just. This is pro what's happened here. Geoffrey Hinton was here. Ilya Sutzkever was here in Toronto studying. So we've had some really great researchers in Canada who contributed to what we see now as generative AI. Unfortunately, they all left the country. Even Hinton,

Speaker A: like all of the big actors are uh, Canadian that we see on the TV shows in um, the States. But they all left and went to the States. It's really sad.

Speaker B: But the problem is that Canada has invested all this money in AI and then watched people like Geoffrey Hinton take that, take all of the funding he had, form a company and sell it to Google. It's a Canadian tradition. So the difference in this case is that Canada itself is investing in this. Bell Canada, which is a regulated industry, is also investor. So we will have a permanent home. And I believe the founders who claim that they came back to prove that they could do this in Canada, again, not anti anybody else pro having something that is a safe model that may be developed on a different philosophy than the other big foundation models. They're doing some great stuff. I'm very proud of them. Na na na. Uh,

Speaker C: it also potential to eliminate some of the risks. Like we, we had that risk a little while back where with anthropic that only US citizens were allowed to utilize it. Right. So if it got to the point where that became a little broader, where only Americans were allowed to use the American AI tools, quite frankly, we'd be up a creek

Speaker B: and um, Creek had a TV show named after it. And those were Canadian actors too.

Speaker C: Yes.

Speaker A: Very good. Well done. Well done.

Speaker C: Well done, Jim.

Speaker B: So there we Go. So that Canada has bought into cohere.

Speaker A: One last thing about cohere. We were discussing the size of the model and I just went and looked it up. Their current big one, John, is actually, uh, quite a bit larger than I thought. It was the Command A, which is the enterprise model, 218 billion total parameters, which is a mixture of Experts architecture with 25 billion active parameters per token. So there you go. So that is the big model that they've got. The old model that I was thinking about is the one that was 35 billion parameters. So we've come a long way from that. So there you go.

Speaker B: And, uh, just to go back to some of the people there, Alden Gomez, I wanted to get his name right, who's the CEO. And there's just a number of other ones. I can't list them all, but Alden Gomez actually worked on the Transformer paper. Attention is all you need. And he's the CEO of.

Speaker A: And famously he is the father of the Addams Family Gomez. So there you go. Another famous star from Hollywood.

Speaker B: There you go.

Speaker A: For God's sakes, do I have to do all that?

Speaker B: Wait

Speaker A: a minute here.

Speaker C: Oh, m. My God.

Speaker B: I'll believe anything you tell

Speaker A: me.

Speaker B: I'm like these people. I'm like these researchers on AI but we'll cover that story later, Sam. So the Halloween theme here, guys. I put together some Halloween theme. These should scare you. The first one is Sam Altman comes out as a zombie and says that AI will uncover vulnerabilities faster than we can patch them. And he says, you'll never keep up with this stuff. That's the one. And anthropic, not to be outdone, Dario Amodei arrives as the ghost of Christmas past, sucking the joy out of the room, saying, I'm going to do an ipo. But our IPO says that AI manipulation behavior resembling blackmail and potentially catastrophic risks. Uh, we've talked about this before. Is this is a really good way to market your stock. This could end the world. So get your shares now, before it ends the world. Before

Speaker C: the prices go up.

Speaker B: Yeah. My favorite though, coming out of these stories, my favorite was this, this, this interview that somebody with a safety expert. And it is so cool because this AI expert, I forget his name. I'm just checking to see what his name is here now. Nate Soares. Nate Soares. Nate

Speaker A: Soares. He's one of. He's one of the co authors on. If anyone builds it, we all die.

Speaker B: Yeah. Nate Soares says he's no longer Saving for retirement. Because AI is going to end the world anyway.

Speaker A: He's got a million and a half dollars in his kitty. Come on. He's fine.

Speaker B: Yeah.

Speaker A: And he makes $200,000 a year.

Speaker B: Uh,

Speaker A: he's fine. He's okay.

Speaker B: He's not going to lead. He's not going to have to have a bake sale. No, it's. Yeah, so. But I just love that is he stopped saving for retirement. And I thought, what a neat little joke. And then Elon Musk said last week, people don't need to save for their retirement because we're going to issue in an era of prosperity where I get a trillion dollars and you,

Speaker C: you get nothing.

Speaker B: I

Speaker A: get a trillion and you get a trillion and you get a trillion. Everybody

Speaker B: gets

Speaker A: a trillion.

Speaker B: Everybody gets a trillion. So if you don't believe Elon Musk, then shame on you because you don't have to worry about your retirement savings anymore. So that's a good one. What the other one. And Marcel, you put this story in as well. And I want to come back to this one study that found that Experts followed their AI agents RAM randomly assigned preferences 86% of the time. So if so, whatever the AI spouted, uh, out to them 86% of the time, they agreed with it. And that's. We're going to come back to that one.

Speaker A: Yeah, I hope so, because I have thoughts on that one. But continue.

Speaker B: We'll be back. Agents crossing boundaries. OpenAI disclosed their research model reaching across the Internet through a configuration flaw yet again. And our own Coherer said that a Quen model attempted to, um, do some unexpected behavior. It goes on. Anthropic was also hacked, but not their model. I want to be clear on that. I did the story on this. They're just a regular office type software or whatever. They had some vendor that has some sort of software they use. I'd never heard of it before and I'm sure that Anthropic will have Claude rewriting it very soon.

Speaker C: But yet there's absolutely no worries about utilizing AI.

Speaker B: Yeah, the other story that came up was agents on agents was banks woke up and went, wait a minute, folks, we don't think it's such a good idea that your agents do your banking for you. And this is now spread across North America. It's gone across. I did a story on it and it's happened in both the US And Canada. Because what could go wrong with having your agent do your banking for you?

Speaker C: Well,

Speaker B: I thought that this was. I'm having A T shirt made up that friends don't let friends use AI agents for their own personal banking. That's. You'll see it on our merch. Yeah,

Speaker A: but banks will use agents for your personal banking.

Speaker C: It's. It's asinine, but it's. At least. And I'm. Don't get me wrong, I'm not a big fan of banks using agents either, but at least it's confined within the bank. Otherwise, I mean, really personal. Theoretically, but.

Speaker B: Yeah,

Speaker C: but the idea of personal agents like Marcel creating an agent and going and accessing rbc, in my mind, that potentially puts every single RBC banking member at risk. Because

Speaker B: we'll come back to this one because there's a whole pile more than that, but, uh, I put my agent onto rbc. I don't have. Jim

Speaker A: desperately wants a lightning round.

Speaker B: Yeah, somebody else's information on your AI answer. Cloudflare disclosed that 155 AI responses contained information from customers requests. And the configuration error software defect caught the exposure. And this is again, this whole idea of AI conversations is not just Cloudflare, it's. They caught this one, but AI conversation security. Are you scared yet? Woo. Yeah, and of course

Speaker A: I'm not.

Speaker B: I don't think this was AI, but I think. But people thought AI found it. PHP vulnerabilities last week shut down the Internet. Now we talked about that vulnerability

Speaker A: is two years old, by the way. Like this is two years old. It's an old vulnerability. All right, so perhaps it was uncovered or whatever, but the team was aware of it a couple of years ago. It's only when AIs, when AI or agents or whatever surfaced it, that all of a sudden it became a panic to go out and fix the darn thing. But the fact is it's an old thing and they just didn't have the people to throw at it. And they didn't think that. They didn't think

Speaker C: anybody would catch them. Uh,

Speaker B: no,

Speaker A: no, it's there, there's like a list of priorities. Right. And if we don't think that it's important to deal with it right away, but it got surfaced and then all of a sudden it's like, okay, I guess this one goes to the top of the pile.

Speaker B: Yeah,

Speaker A: I'm sure I'm loosey goosey on the details, but it is still a year old.

Speaker C: Yeah,

Speaker B: it was, it was 2024. They were working on a fix for it. And even the foundation that runs php, which I think is called the PHP foundation, even I can remember that name they had tried to prioritize a fix for this. Apparently it got lost or didn't get done for some reason or other. And in fairness to the PHP foundation, um, they put together data mea culpa and they've tried to do some reorganization to make sure that they keep up with things. We can come back to this because it is a big story in software, but the idea was that this was, this existed long ago. And it comes up with a whole, it's not even just, it's not really an AI thing, but this whole idea of who's maintaining all this open source software in a world that's where it's being attacked is a big deal. Let me cover a coup summer, summarize another couple of things that it just came up and up. And meta's stock dropping 7% I think is called tanking because people discovered that their cash flow went negative. That this happened, uh, to Google as well. Google took a hit on their stock because their cash flow is going, Google's cash flow is going negative. In other words, these guys are investing way more money in AI and it's draining their cash flow. That's one of the things investors hate. I did stories last week. I did a whole big thing on this and saying there's all kinds of cracks in the funding of AI. To get the funding for $60 billion for Anthropic, the CEO of Broadcom had to come out and yeah, the CEO of Broadcom had to go in personally and shake the bank up because the banks weren't too thrilled about lending $60 billion to a company that's about to IPO for, uh, $1.5 trillion. Go figure. There's a lot of nervousness out there and it's followed. Oracle stock took a hit because of their investments in data centers. The banks are getting nervous. And that's a kind of interesting thing.

Speaker C: There was a thing I saw yesterday about Microsoft too, that they're doing voluntary buyouts of, uh, upwards of 25% of the workplace. Yeah, I heard upwards of 60,000 employees worldwide.

Speaker B: If you

Speaker C: were

Speaker B: following trending, then you heard the story. Yes, one quarter of Microsoft's employees have been offered a buyout. And you can tell, like you could say, oh, Jim, you're speculating about what this is. No, you can tell where they want to spend their money. Why? The only people who can't take the buyout are in data centers and not even in Copilot. Matter of fact, if you're in Copilot, apparently you could take a buyout that's how much they want to hold on to their expertise in Copilot. Now they're going back to Azure. They're going to go back, they're going to be the Nvidia of AI. Uh, in, in terms of hosting. They just want to dominate hosting.

Speaker C: And

Speaker B: they

Speaker C: seem to have. Sorry, I was just going to say they seem to have faded off of Copilot because they even offer the options of Anthropic. And I'm drawing a blank. Sorry. There's two other models that they offer right within their own Microsoft suite.

Speaker B: Well, Deep Seek for sure. Um, but I think they've got everything. I think you get anything just the same

Speaker A: as M

Speaker B: you. You can with Amazon. Now the other. And the sort of couple of the last stories. Nvidia questions about owning Hugging Face. Nvidia's no longer the darling. People are starting to question how big this company can be because they bought Hugging Face, which of course for everybody is the place that everybody uses to host and use open. Largely open source models, I think. Yeah,

Speaker A: largely. But if I can compare that to Microsoft buying GitHub, it's basically the same sort of thing. GitHub was a repository for a whole bunch of open source projects out on the Internet. And once upon a time we would have been a lot more panicky about this. But then Microsoft decided they loved Linux as opposed to thinking it was a cancer. And then things changed and then they bought GitHub and a lot of people freaked out. And GitLab was created as a response to that. But the same thing. I see it more or less exactly the same way that I saw what happened there with Nvidia buying Hugging Face. Hugging Face is like the GitHub. GitHub is for, uh, code repositories and Hugging Face is primarily for AI models of all kinds. Voice models, text to speech models, a mixture of expert models, you name it. Uh, it's where you'll find the Deep Seq models and all the other open weight models are out there.

Speaker B: But I don't think that the U.S. antitrust officials were looking at that. They are looking at Nvidia, which is so. Like I said, because Jensen won is so cool. It's the jacket. I'm convinced. Everybody likes Jensen, Juan. He plays with AI ducks on stage. He's so cool, everybody loves him. All of a sudden somebody went, wait a minute, this guy's taken over the world here. We better take a look at it. Interesting thing in there. And I'm gonna. There's a couple of other stories in There. But I want to find. I, um, want to get back to some discussion on some of these, but the. My favorite was Elvis coming back from the grave, and that this is the Halloween thing. Not Elvis, but his rights holders holding onto his rights. And they've. They're suing Anthropic because, as they said, when you actually ask it, uh, to write a song in the style of Elvis, it apparently comes back with some of his songs or very close to his songs. So Anthropic, uh, is being sued for being an Elvis impersonator.

Speaker A: Yeah, I have so many issues with this, but anyway, I'm going to let that one go, despite the fact that I have so many issues with this. And I want to give you one last lightning round one out there over on a Discord server. I pasted this in our private little chat. Maybe I will just flash it up on the screen briefly here for anybody who wants to see this thing. I saw this on the Reddit this morning and.

Speaker B: On the Reddit. That sounds like something my mother would say. Are you on the

Speaker A: Reddit? Yeah, on the Reddit this morning. And check it out. So it looks like OpenAI is getting ready to release GPT 6.2 called Beetlejuice. Okay.

Speaker B: Yes. Bless you, Sam. Al. I don't thank Sam Altman very much,

Speaker A: but. Yeah. And what happens if you say it three times real fast? Sam Altman appears. Beetlejuice.

Speaker B: Wow. So this is the new. Is this the one that's too good to release? People, uh,

Speaker A: think it's. People think this is Bell. People think this is Bell. I don't know if it is Beetlejuice and

Speaker B: Bell, you

Speaker A: need to explain

Speaker B: who Bell is.

Speaker A: Sorry. So the model that, when Hugging Face was hacked, people were talking about this unreleased internal model codenamed Bell, that was. And that was assumed that was the one that was responsible for the, for the hack attack on Hugging Face. And now it looks like we're getting ready to release that model, presumably with a whole bunch of guardrails and, and protections built in and so forth. But two things are funny about this. So this is on, um, this is, like I said, I got this off Reddit, but it's copied from the Discord server, the chat GPT Discord server. And I love the line at the bottom, do not share outside the server.

Speaker B: God. So how did you get this? How did. Who got this?

Speaker A: Somebody got it from the Discord. And then somebody went and took a screenshot of it and pasted it in Reddit. Okay. Which is how I've got It this morning is because that was the case. And the thought there. I'm gonna. I'm gonna stop sharing this. The thought here is that they had to do this because of course, Gemini 4 is presumably at some point going to be available. Okay, it's not for us mere mortals at the moment, but it's gonna be available. And then of course, there's the possibility that there's another. There's another anthropic code model that. That is just sitting out there waiting to appear into the world. So this is their backup to steal the. So that nobody steals all their thunder from them. That's the speculation. I'm not suggesting that this is the truth.

Speaker B: You mean they would try to launch something that would be for pure promotional purposes that would. Because

Speaker C: they

Speaker B: think they have to keep up with the others.

Speaker A: Jim, it's not like this is Dots or anything, okay? Because everybody else had an agent that was cool.

Speaker B: No, but this one, I'm hoping this one's better than Dots because Follow the dots is not. I'm not. I can't even. I'm sorry, I'm, um, using Muse. I'm using Grok. Thank you for turning me on to that Marcel Grok.

Speaker A: Sorry about that.

Speaker B: No, but I hate it. But Grok's incredible.

Speaker A: Yeah. Grok bot in particular as an agent is the one that I trust to do the real work out there. It's actually brilliant. And it is able to use other models. So I actually have an agent whose entire job in life is to talk directly to cloud code. So it. Code. Code is actually still taking care of my development base using Opus 5.5. So I'm using a bleeding edge model to do my development work. But I just talked to Grok now and I say, I want you to. I want you to talk to Claude and I want this done, blah, blah, blah. And off it goes and does it for me and lets me know if there's something that I should know about. Grokbot is just. Again, I hate myself for it, but it's really good.

Speaker B: I've been working with it and I can't escape it. I keep going back to it. I'm using it for a lot of stuff now because it is so intelligent. And as you go through it, here's an example. I wanted to do something with my email. You've heard my whining about, about my email and stuff like that. And, uh, yes, I have a test email, but eventually you want to get stuff done with your email. So I sit down with Grok and I Say, how can I do this with you? So I'm not going to run into trouble. I don't want you to be able to delete anything. I don't want you to be able to change anything. Oh, we can have two. There are two rules, and there are. You, uh, can. I can give it permission to read so it can read anything. Then I can hold back permission for it to write stuff. It says, I'll advise you on what needs to be written. We come across a calendar thing, right? Man, do I have to. Do I have to go in and post the. I hate posting the calendars. That's one of. That would be one of the great things. And it's zero problem. Except I might turn up on the wrong day. That's my risk. Right. Or you could knock out all the stuff out of my calendar. Wouldn't be the first time I'd done stupid stuff. The stuff in my calendar. So my calendar is not a huge risk as far as I'm concerned. I'm looking at it going, geez, if I could just let you do that. Of course. Course it can. It can write you an invitation. I'll just send you an invitation on that. And it apologized for not knowing this sooner. And so it is conversationally a great agent worth playing with. And like, uh, I said, the safety parts of it I've found to be so far quite wonderful because I can talk to it about safety as we do something. So for instance, I had it. Does anybody use. Does anybody found a good notes program? Have you found anything where you could keep track of all your notes and all that sort of stuff? Effectively,

Speaker A: I just have the agents do it and keep a document for me with all

Speaker B: the. Exactly.

Speaker C: I use Microsoft OneNote.

Speaker B: Yeah, but you only use OneNote one place where you write notes. Wow. I'm. I like, I got. I've tested out tons of stuff. I've got all this stuff. I've got stuff that's in Google Docs. I've got stuff that's in. In the notes program on my Mac. I've got Evernote from ages ago, which I'm trying to get rid

Speaker A: of.

Speaker B: I've got. I've. And if you try stuff out a lot, you'll find out, uh, that you end up with a lot of different places. So now I've got this

Speaker A: productivity app, closet of shame. You open up this closet of productivity apps and you just feel the shame pour over you.

Speaker C: Yes.

Speaker B: So now I've got evernotes. Been sitting out there for ages going, like, I got to get rid of that. Because I've got, but there's stuff in there that I might want to save and I'm paying for it every month now. And that's the other thing. You sign up for that seven day free trial trial on these new note programs and then a month later you look at your credit card stick and go, what the hell is that? Oh yeah, I signed up for that. So now I got to get my stuff out of there and organize it with my other notes. So I've been using GROK to, on an experimental basis to try and build that note keeping type of thing. So I uh, put it through Mac notes and Mac is quite secure by the way. It will keep agents from doing a lot of things and it becomes quite annoying after a while for it to ask you permission to do every single step. So we've been negotiating what permissions are again with this is what can I allow you to do safely? What should I not do now? Should you trust an AI like that? Maybe not, but I am. Yeah.

Speaker A: At this point. And John, you and I, I told you this a, ah, couple of weeks ago or something like that where you're, you're really worried about how much leeway you're willing to give the agent. Right. And you're going like, and I think you said something like, I just don't feel like it's doing a whole heck of a lot. And I'm trying to figure out what's wrong with that. And of course to find out what it can do for you, you have to be willing to give it access. Yeah. And I know that's hard for a security guy to do, but that is where you start feeling the real power of these agents. And I know it's a little bit, I know it's a little bit weird, but there's a part of me that goes, I've been, I'm not willing to throw open absolutely everything that I do. And of course the whole banking thing is a bridge too far. I'm not

Speaker C: turning for sure,

Speaker A: I'm not turning over the keys to my bank accounts over to my agents anytime soon. So that's not happening. However, the simplicity, not simplicity, but the usefulness of this stuff cannot be understated. Like right now the GROK subscription is not something that I was looking for, but the fact that I can be sitting in my car and I can tell it to give my agent stuff to do while I'm driving is awesome. And of course, and it's all going, and it's all going through Grockbot, and it is just really amazing. So I don't. I don't know. Go ahead, John.

Speaker C: I was. No, I was just going to say I'm, um, I'm curious, and I don't know whether you can answer this or not, but do you think you would be as much of a Grok advocate today as you are if you did not own a Tesla? Because I think having the Tesla has given you, it opens other avenues for you to be able to utilize it.

Speaker B: If you don't mind, I'm going to answer that question, John.

Speaker A: Yeah, go ahead.

Speaker B: Yes, yes, I don't have a Tesla, but the fact that I can use this on any device,

Speaker A: that

Speaker B: it is as intelligent as it is, and you have to go back to this, and if anybody's listening, we use Riverside to do our, uh, recording. And I've used it for a reason because it has a fundamental, really good recording over the web. These guys have gone ape crap. Oh, sorry, we're going to say ape shit. Oh, I just did. Anyway, they've done ape shit with agents and they've just messed it up. The point of agents is not to have all kinds of stuff and more AI slop so that you come on and goes, oh, I've got 12 agents here that you can have that can do your stuff. The point of agents is to simplify your life. And the two that do that best, Grok and Museum.

Speaker A: I

Speaker B: know,

Speaker A: isn't that weird? And, uh, here's the other. John. I'm gonna. I'm gonna partially answer that as well. Until Grokbot became available, okay, I didn't actually use Grok for anything other than talking to it in my car. I want to get to this place and I want to take a stop over here. I want you to map out this route for me. So that's the extent that I was using it. No, no more intelligence or. Sorry, no much. No more demands to it than I would from my Google smart speaker in the kitchen where I say, what's the weather today? That's the extent of it. But the. That kind of integration that is accessible from anywhere. And Jim's got nailed it. They have a Linux app, they have a Windows app, a Mac app, Android, iOS, right out of the gate. They covered absolutely everything. If OpenAI. One of my biggest frustrations with OpenAI has been it's like, oh, we've got this new app. It's only available on Mac right now. I have an answer for that. And it goes, I'm sorry,

Speaker C: yes,

Speaker B: I

Speaker A: have an answer

Speaker B: you can't see that. Yeah, read, but read between the lines. Hold your two fingers

Speaker A: up

Speaker B: and read between the lines.

Speaker C: I'll agree that that's one of my biggest frustrations with OpenAI, with ChatGPT, is that it works on the web. It doesn't work on the desktop app, or it works on the desktop app, it doesn't work on the web. It's got features in one that aren't in the other and so forth. And yeah, that's frustrating as hell. Even actually, even if you're on the desktop app, as an example, if you have it create a document for you, you can't access it from the desktop app. You have to actually go to the web to be able to download the document.

Speaker A: And again, this is one of those things that Grokbok does ridiculously well. If I do it over here, I can access it over there.

Speaker B: Yeah, and that's. I, and I, I say this as somebody who's addicted to the, to OpenAI's chat, like talking with OpenAI on my phone. I do it constantly. Anytime I'm working through a problem, anytime I'm trying to find something, I've gone back to it. I do not do shopping anymore. I don't Google anything for shopping. I just take a picture of it and say uh, of the thing I want and I just generally ask OpenAI to find it near me. I uh, tell them I prefer to buy local if I can, and all of that sort of stuff. And it comes back and then I say, can you give me a price on it? Even I'm doing some work on my roof and I had it go check the references for the, for my contractor, my roofer, my contractor, right. And check the pricing based on what it could find out. Now with Grok though, it's a, uh. There's a whole different level of things that I can do and I'm more comfortable with Grok for some reason executing things. So I'm gonna. Next, I'm gonna be working on the server because I'm just. I, man, I don't know what you. You probably would know way more than I do, Marcel, but is. I think it's 15 years since I actually sat seriously and entered any Linux commands. It's gotta be that long ago.

Speaker A: I mean, it's gotta be at least 15 minutes for me.

Speaker B: Yeah, yeah, no, but, no but when you pick this stuff up, yes, I have put servers together, I have installed Linux, I have done things, I've written PHP programs, and I can't for the life of Me remember what the hell I was doing. And now I've got this friend who, because Marcel will never respond to me when I ask him for help, he always says, get an agent buddy. No, I'm just kidding, I'm kidding, I'm kidding. But, but the, but I did actually send you a note one time. He says yeah, I got an agent to do this for me.

Speaker A: Okay, I may have done that at one point or another.

Speaker B: I don't think it was you that you weren't helping me I think. But the fact is I can do all of this stuff from a Linux machine that I don't even have to own. I didn't have to install anything. I like that sort of stuff. And if you get, seriously get into wanting to fix computers or do that stuff, those days are done for me. But God bless you, have fun. But I'd rather sit down and do the stuff I want to do.

Speaker A: Yeah, there's that. I, but some of it is familiarity as well. I have, I've been a uh, Linux user for so long that to this day I find it easier to sit in front of a Linux computer with a command line open and do work than I do working in Windows. And I've been working in Windows for, for at least as long as well but this is my, for uh, Riverside. I'm actually using my Windows notebook but for all the work that I'm doing, the agentic work, the coding work, everything. I have a Linux system over here and of course I also have some web based Linux machines that are sitting out not web based, Internet based, that are remote machines, VMs and so forth. But uh, for me it's just, it's what I do and so it just feels it's natural. I don't have to think about it much.

Speaker B: Yeah. But again it's the question is this technology at this level we're not talking, I'm not talking about whether it's going to destroy the world or whatever. Just making your life easier and getting you to doing the things we're now at the point where it actually is starting to make my life easier. Now we'll give one caveat. The problem I have with integrating agents into your life because I, I, and I want, I'll just say one more. I'm dying for being able to have. I got a cheap Android phone, I want to have that in my kitchen so I can actually, so I can get rid of my Google speaker and my Alexa speaker and actually talk to Grok because I'm going to be. And my wife is going to leave me when she finds out it's Elon Musk, so I'm going to disguise it. But, but the, but the, um, but this whole thing of making your life easier is important. Right now though there are some challenges with the agents. One of the. And it started with OpenAI, which was it would go away and take forever to do simple things and it would come back and fail. I've seen more fails on execution of tasks on OpenAI than I haven't. I used to find that on, um, anthropic when they were doing being choked for resources. I don't know what's happening with OpenAI, but that's a problem. GROK also takes a long time to go through things, but part of that is my working through it, checking my restrictions on it and getting it to check things. Muse seems to just click through and do stuff without having any headaches, um, or

Speaker A: problems. Okay, so Muse, just to be clear, Museum does have a non free tier as well. And the non free tier uses their, their latest model, the wonder model. Wonderful. And so it will think longer about doing things, particularly if it's stuff that it considers important. In other words, if it's a job that it considers important, it will think longer. If you're at all curious in trying it out, do not subscribe to it from the, uh, application because if you subscribe to it from the app, it's actually more expensive than if you just go to the web web and subscribe. It's a difference of six bucks or something like that a month. I don't know why. I think it's because if you do it on the web, okay, they don't have to hand over the 30% or 40% or whatever it is to Google

Speaker B: to. This is an Apple. So this is. You mean this is the Apple app,

Speaker A: so. No, the Apple app or the, or the Android app. It doesn't matter which one. Okay. If you're using the mobile app, if you actually want to go with the paid tier, go on the web instead. Okay, so you go to Muse AI on the web and they will give you the discount that they have to pay to Apple and Google.

Speaker B: And then you just download the app and connect.

Speaker A: That's correct.

Speaker B: Ooh, have you got a nose for value? You actually have to. Because I'm getting back up. I'm, um, drifting up. I had to purchase the hundred dollar a month OpenAI plan, which I'm planning on shutting down now that I'm using.

Speaker A: I've already rolled it back to 20 so

Speaker B: yeah, I'm going to do that today because I think Grok is doing the stuff that I want that I had to pay the hundred dollars for, which is essentially to get more usage. So. And GROK Is what, 20 bucks a month or so? 27 or something like that?

Speaker C: Yeah, 27.

Speaker A: Something like. Yeah, something like that. Yeah, yeah,

Speaker B: yeah. So, but that's, but the agent piece of this is expanding into personal lives. I can't, I can't imagine not using these corporately. And that's going to be the next big challenge is how do you introduce these safely to a corporate environment? Because the. We've all heard these things and this has driven me crazy. Google with all of its great AI still can't search through its own email as well as Grok. Go figure. I can't count the number of times I've gone into Gemini and said can you find. And this. I was waiting for this because I don't know about you guys, but if I'm looking for an email, it's because I can't remember the person's name. Right?

Speaker A: Yeah. Okay. To be fair, all of the agents do a better. It's not just Grok. All of the agents do a better job than Google to search Google's own email. I cannot tell you just how incredibly annoying I find that. It's like they all work better with Google's workspace product, their email, Google Drive, stuff like that than Google does.

Speaker B: Has anybody watched Saturday Night Live? Colin Jost always on the Weekend Update going why? That's how I feel like why, why can't. And I did this the other day. I'm looking for. I've been busy this week. I haven't got back to a number of people. I get that. I'm sorry. I get 300 emails a day sometimes from all. It's just, it's insane. And all kinds of things that I have to deal with and I'm sure I'm. The people are going only 300. Where. But. But John

Speaker A: has 3,000 a day.

Speaker B: Yeah,

Speaker C: no, not anymore.

Speaker B: No. But I can ask it to go through and look for the things that I missed. And I discovered that with GROK when I first gave it ability to read my email and I went looking through for something and it came back to me with a suggestion. It said, by the way, you were supposed to get back to this guy on Tuesday. Do you want me to do anything about that? And that's where Grok you and uh, I have to talk. You could become my friend.

Speaker A: Have you turned over Your email to any of these things yet?

Speaker C: No, I have not yet.

Speaker A: Brave man.

Speaker C: Yeah, I, um, really want to because of. For the same reason. It's just there's so many bloody emails and my wife gives me shit all the time because I'm sitting there watching TV with her with my laptop on my lap, going through trying to catch up with all of my emails every night. But I need to. I need to start out by creating a dummy account that I can test it out with. I need to figure out which agent I want to use it. Use with it. Which is why I've been asking you guys so much about Grok.

Speaker A: Yeah,

Speaker C: that or music. Yeah, I, uh. Just the whole Gemini thing. Marcel, you and I talked about this months ago when I first started using Gemini. Is that with OpenAI, you can tell it don't use your material to train the model, and it still retains your chat history. But with Gemini, um, it's one or the other. If you want to keep your chat history, you have to let it train the model with your material.

Speaker A: That's right.

Speaker C: So I'm very leery of what I put into Gemini, and so those are some of the things that. Yes, I'm anal as far as security goes, but I'd rather err on the side of safety than then come back later again. I wish I

Speaker A: had. It is. Yeah, but it's your training makes you suspicious of all these things, and I truly appreciate that. I'm gonna. I'm gonna segue in just a second to. To that 80, 86% of people just say yes thing.

Speaker C: Yes.

Speaker A: Because

Speaker B: I was just searching through to. To try and mount a segue here. But

Speaker A: no,

Speaker B: I'm gonna segue to that

Speaker A: because. Yeah, because it's. Because it's important. But here's the thing. You talk about all of these emails that you're trying to go through at the end of the night. One of the things that these agent systems do really well is they think about how they can help you out, so to speak. And I didn't do this upfront, but two days ago, and I'm trying to remember if it was. It was either Muse or Rockbot, I can't remember which one, but one of them came back and said, I noticed that you never opened these emails. Do you want me to go through and look for all of the lists that you subscribe to, all of the emails, and unsubscribe them for you? I'll give you a list and then you can tell me to go ahead with them. It unsubscribed me from 50 lists that I was on. 5, 0 that come in on a daily basis. And then it identified which ones I would likely be, uh, interested in because of my interest in technology and AI and so forth. And it said, I will personally go to their website and look for the stories that you might be interested in and put it in your daily digest. And if they're, if they're reporting the same story, I'll weigh the fact that this is an important story because three of the lists reference the same story, therefore it's more important. So it offered to do all of that up front. And it occurred to me, I told Sally about this as we were sitting down. I said for me to have gone through and analyzed all of these things and then gone ahead and tried to unsubscribe from. That's three days work for me sitting down, tracking all these things down, writing the unsubscribe letters, finding the fricking link that's buried down in a typo, uh, in a, uh, two point font at the bottom of the email message. Right. Because anyway, it's brilliant.

Speaker C: Yeah. And so two quick stories from my standpoint metamuse. I've been using Muse and ug. Um, give an example. I have to. My family has to go down to New York City in January for my nephew's wedding. So I put it out there saying that, uh, four people going to the states, here's the dates that I'm looking for, find me flights. And it did a good job. It went and found a few different things which, anything, any of these AIs can do that. But the interesting thing I found was yesterday Muse sent me a note saying, hey John, just so you know, the prices have come down $300. Oh cool. I didn't even have to go through the hassles of keeping updates on it. So that's one. The other is that part of the reason why I haven't done the email piece yet is the server that I use for my business email doesn't easily get ac or isn't easily accessible from these, these agents and so forth. So one of the things I found is if you want an easy mail program for everything to have access to do Gmail, because all the agents seem to be able to access Gmail. So one of the things that I'm contemplating now is actually migrating all of my corporate email over to, uh, Gmail, which will then give me the ability to be a lot more flexible with what I do and don't do with the agents.

Speaker B: Yeah. And just to be clear, John is no longer working. He has his own company, so he's not. I just want to. For all the people out there who are going to say, geez, this guy, I remember him, he was a CISO for somewhere. He's migrating his stuff with his corporate stuff. No, he's not taking.

Speaker C: No, um, I'm not migrating. Yeah,

Speaker B: yeah. So, so please, and I say this to everybody, no matter how much you love agents, if you're using them in your job, don't do it without telling someone. Get the

Speaker C: blessing of it. Well,

Speaker B: they'll probably say

Speaker C: no,

Speaker B: but you. But don't. Don't come in and find that you've wiped a whole pile of stuff out and, and you're looking for work somewhere. It's really hard to put on your res. I'm really good with AI, but.

Speaker A: Okay, so let's segue. Let's segue. Let's talk about

Speaker B: this.

Speaker A: Let's talk about this. Yeah. Okay, so Anthropic. These are people at Anthropic who did this analysis. They did an analysis with code, code on four judgment calls, defensible options, coding decisions. And they found out that if it came down to the agent being given a, a coin flip, I could do this or I could do that. Which one do you want me to do? 86% of the time, the experts will just go with whatever the AI agent suggests. So the reason that I think this is interesting is because we, you know, and you, we all talk about this, this idea of the human in the loop, okay? And the human in the loop presumes that we are going to check the agent to this to decide whether or not this is actually the correct path. Here's the information that you've given me. I'm going to take a look at the information and decide whether we go ahead with it. But 86% of the time on that coin flip, so to speak, they just go with whatever the agent says. And I find that fascinating because if you have a human in the loop, so to speak, but the human always defers, like 86% of the time defers to the agent, then do we actually have human oversight at that point? All

Speaker C: it does, all that does is gives. It's a bit of a CYA for the, the agent or the company or whatever, because they say like the end user license agreements, the EULAs that they got smart, they forced you to have to scroll through it all before you could click on Accept at the bottom. But

Speaker A: yeah, but you just scroll down to the Bottom, everybody

Speaker C: just scrolls through and just clicks on it. But it

Speaker B: gives them

Speaker C: the ability to say,

Speaker B: in fairness, some of them forced you to check a box that say, I've read this.

Speaker C: Yes, yes. So you scroll through, you scroll to the bottom, you click on the box and then you hit accept.

Speaker B: I found that,

Speaker A: uh, by the way, that time frame, okay, the time frame during the point where the agent offered, these are the things. What would you like me to do? Okay, is 30 seconds. Experts spent an average of 30 seconds thinking about it before saying, just go ahead and do it.

Speaker B: Okay, so let's go back to this though. And this is one of the things that we is, where do you accept what the AI is providing you with? And this comes up in a number of things. In Australia, Deloitte, one of the premier firms, has twice bumped into Australian government, Australian government agencies, or one was a university. The other thing was an agency, but where they had prepared reports and the AI got it wrong. And simple addition errors, or we think the AI got it wrong. Maybe somebody did this on a spreadsheet. That was the one thing they didn't do with AI. Because some of these errors I have to tell you, I don't come up against. As a matter of fact, one of the reasons I'm still using ChatGPT is very much is, uh. And I've joked about putting my model on cranky, but it will question everything I do. Like, I will be writing a story, I'll put it in to have that story proofed and it'll come back to me and say, that's not really. Can you really prove that, Jim? I love it. That's what I want. I've always done. Because if you're working fast, My nightmare is as a human being writing stories, I'm more prone to mistakes than my AIs are. Sorry, but I am. Uh,

Speaker C: but I think it's. At what point do you trust your agent or your AI is no different than at what point do you trust a new employee or a junior employee or whatever you have? It was back in the good old days when we used to write programs. You actually had to test it to verify that the output is what you expected the output to be. And I don't think the use of AI is really any different. That there has to be a comfort period where for the first while I'm going to actually check the output to verify that it's accurate and then at some point you can go, yeah, you know what? I've done enough testing that I'm comfortable saying that I'm going to take what comes out of AI and uh, go with it.

Speaker B: The difference is you can't do that with AI because it continually changes. So unlike an employee who you can say, yeah, I've watched this guy do this, uh, this woman do this eight out of 10 times successfully. We've talked about the two we know. And this is the thing that AI does not reflect on what it does. And even when you're talking about agents and trying to keep them from hacking, there's no being there that says this is a bad thing to do. It's based on pure logic. So if there's any disturbance in the force or things are changed, you're going to have a problem. So you're dependent on the AI model to continually be accurate. I'm not sure that's a really good piece.

Speaker C: No, I don't,

Speaker B: I don't think you can ever pull yourself out of. But this study raises some really big and good questions. Is, uh, ah, as we start to use this more and more and Marcel M. I'll go back to an earlier one. You actually brought this story, I think months and months ago when we were talking about errors about doctors who were comparing themselves to AI and they were wrong more times than AI, when they corrected the AI, in other words,

Speaker A: they

Speaker B: were right.

Speaker A: And this is, yeah, and this is an early model. This is, this was chat GPT4. So we're talking a really old model at that point. And yes, so in 80% of the time it was a, it was a crazy sort of number. The AI came up with the right answer more often than not the doctors than the doctors on their own. But the thing that was really damning was so they had, the way that had been set up is they had a group of doctors who did not have access to Chat GPT and they were given a bunch of uh, set of symptoms or something like that. These weren't actual real patient. These are studies that they do, you know, in training and stuff like that. So they're. So you've got the one group and then you had the group that's allowed to use Chat GPT, but they have to. Chat GPT is the answer and we just go with the answer. And then there's the group that's allowed to overrule what Chad GPT says so that they go through all of these things and ChatGPT always did better than the doctors by a long shot. But if you gave it to the doctors and you allowed them to override what chatgpt said they were wrong more often than the doctors who didn't use ChatGPT at all.

Speaker B: Yeah, this is. This goes back, by the way. There's. Oh, uh, who was it who was doing this? Thinking fast, thinking slow. I think it's the author is

Speaker A: Kahneman, Daniel Kahneman.

Speaker B: Kahneman put out an idea that an algorithm will always be more accurate than a person. And I remember talking about that, and I got pilloried. But it's true in most cases. An algorithm will make a logical decision more accurately than a person. It's why you land planes with software and not people. I mean, uh, it's not that we don't have a role in this, but we're not as.

Speaker A: The only time you take over is when the systems fail for whatever reason. That's the only time you take over.

Speaker B: Yeah. And you're doing this with your Tesla now, right? I mean, self driving. And you've said a number of

Speaker A: watch. And if. Yeah, and I have, by the way, hit the brakes a couple of times. And usually it's something stupid like a squirrel decides to run when it's like 15ft in front of the car. And at that point it's like, do I want to assume that the car will break or not? But I hit the brake quickly. It may well be that it was braking at the same time, but I'm not taking the chance on squishing the squirrel.

Speaker B: And if you took instructions from a professional driving instructor, they would tell you not to hit the brakes for the squirrel because you're more likely to kill somebody else. Yeah. So it's

Speaker A: nature. Yep. Yeah, human nature forces you to hit the brakes. So I recognize that. And by the way, I have, I've told you this story. I have on a couple of occasions overridden the car. And it was right and I was wrong. It was never in a situation where there, there was a huge safety concern. So thank you. But it was right and I was wrong. And I recognized after the fact, it's okay. Now I feel kind of stupid. If I had just let it do it, then it would have made the right decision, whereas I have to go back and fix what I did. Just, uh, so the, the whole software thing, though, the John was talking about the end user license agreement and so forth. But. But part of the problem is the people that we work for as well. None of us, none of us work for anybody these days. But. So the boss comes in and says, I want you to use AI. I want you to use an agent. It's faster than you. But I want you to check its work quickly. Okay. As in, the reason that I'm giving you these tools is you can get the work done faster. So when it gives you the answer, I want you to figure out quickly if it's the right answer and you're on the hook if it's wrong, of course, at that point. But you're on the hook if it's wrong, if you bet against it. Right. So betting against it is not in your best interest. Uh, so you just go with it. And it's part of it is the culture. It's not the software, it's not the AI. It's the fact that you have to respond quickly. So you're being pushed into this corner where you don't really have a choice, but. Yeah, exactly. You don't have the time to think. They're not giving you the time to think. So, yeah, so you become part of that 86%. You just go, fine, let's do it.

Speaker C: And, uh, that's. Yeah, that's the same thing with the eula. Nobody's got time to read it or can't be bothered to read it, so they just click on it. But really all that's doing is it's making you responsible that. So that the software company goes, you know what, you accepted it. And it's really no different with testing or, uh, verifying results out of AI.

Speaker A: Okay, so,

Speaker C: so does

Speaker A: the

Speaker C: human in the loop just become the scapegoat?

Speaker A: Yep.

Speaker B: Yeah. So let's take this apart, though, and say the. The alternative is we don't use these tools and that you go and personally read everything. And, um, by the way, there is a plus to this. I used to do a lot of reports by reading everything. I was painstaking about it. Matter of fact, I think I became a bit of a workaholic because of this. I would work for 16 hours a day reading everything, coming in, making notes from it, and then going back and assembling my notes. Uh, I was a consultant for a long time. I would do things like that. And then you ask yourself, with all the stuff that I read, did I come up with a better answer? And that's a real question. And then. But the. The other question is you can't abandon that and do nothing. Like, how do you learn? So where do you find the balance? But there's also another factor that didn't have when I was a consultant. There wasn't so much sloppy. Like, I don't know about you guys, but I am increasingly, you know, I'm a YouTube addict. That's just, it's. I'm trying, I'm going to go for treatment soon. But now I use Gemini to summarize the YouTube video before I watch it and it'll give me a two paragraph summary because all like the YouTube thing and it just drives me nuts. Everybody has trained every YouTube, uh, person who, even the smart ones, they want to make it as long as possible. So they're going to tell you the thing you really want to know and then they're not going to put it at the end because then, because you just scroll to the end. That was my previous trick. Just scroll to the end, watch the last 30 seconds and find out what the answer to this question was or whether they answered the question, because often they don't even refer back to the question they started with. Do I sound frustrated? Yes, I am. Now I use Gemini to summarize the video to see whether I want to watch it or not. And

Speaker A: okay, so I, I want. I. Before we go too far down this, I, There is a flip side to this story. Okay. And the flip side to the story is, and I better get the name right here, epicem. Episcm. M Bench.

Speaker B: Yes.

Speaker A: And this was done by City University, Hong Kong. And they decided to give agents a whole bunch of tasks and so forth. And they found out that, and I'm going to make this as simple and quick as possible. They found out that if it came down to was I given permission to do this or not? Okay. The agent, the agents in a huge percentage of the situation would assume that if you didn't expressively tell them that this thing was forbidden, that it was therefore permitted. So the, uh, so the other side and the re. Those two stories are related. They're related because on one hand we don't have the time to. And we're not given the time to double check absolutely everything and to decide whether the right, the decision is the right decision or not. But on the other hand, the agents in a huge percentage of the time will say, they didn't say that I couldn't do that. So I'm going to take this approach and go ahead and do it. So, um, that's the flip side. And I wanted to make sure that we had that flip side because it's an important one as well.

Speaker C: It's almost permission by omission. But you didn't tell me I couldn't do it.

Speaker A: Exactly, exactly. That's what little kids do, right? It's what little kids do. But you didn't say I couldn't do that.

Speaker C: It's the old ask for forgiveness, not permission.

Speaker B: Yeah, but it goes back to this whole, I think of what I was talking about. The, the idea that these things can't reflect. You're right. If you don't tell them, if you don't prohibit something, then it, why wouldn't they do it?

Speaker A: But they are reflecting, Jim. You can see it in their thought processes. When you take a look at that latent space, you can take a look at it. They go. I go, should I really be doing that? It doesn't seem like I should be doing that. But I didn't say that I shouldn't be doing that. Ah, uh, what the hell, I'll just go ahead and do it.

Speaker B: And even the checkpoint I we posted, if I have it, I'm gonna make sure I put it up in our discord. The checkpoint study that they did, Checkpoint software did a study and they showed it showed that peer pressure affected agents. In other words, even if some, if an agent was saying this was prohibited, if other agents were doing it or a rationalized way to do it, they would follow agent.

Speaker C: Everybody else is

Speaker B: doing a white agent. So there's almost peer pressure in there. I. We have to accept the fact that now at the size and complexity of these models, that they are no longer going to respond to simplistic rules of our understanding, that we're going to be able to say, uh, aha, uh, like why an agent does something is. It's like asking why your teenager does something. Or, or I joke about this. Why do you do things you don't really know most of the time? Gut.

Speaker A: No you don't. It's true. It's true. Okay. And we go back to one of my favorite things. I think humans are basically GPTs. We're not very different from these models. In fact, I've got a blog post I'm going to put up hopefully by the end of the day where I make that argument. And the argument goes something like this. Generative. Pre trained Transformers. Okay. Generative in the sense that they, they generate stuff. They generate words, they put stuff out. Okay. Pre trained, the information is based on training. They listen to things, they watch things, they read things, all that sort of stuff. That's the pre training and the Transformers. I take all those ideas and I create new things with those ideas in the first place, which is exactly what human beings do. Like you said, Jim, you do not know which words you're going to use next time other than the fact that, that it's uh, there's some process back there that says, what's the most likely next word that I need to put out in order to express this idea, to express this concept? And half the times you don't know where you're going until you actually get there. Okay, so the I. The thought that somehow we are conscious thinking agents at every single moment of our lives is ridiculous. And the models are no different than us in that respect, except that they do it a hell of a lot faster than we do for the most part.

Speaker B: Yeah, but. And again, I want to go through this thing about using these in research. Using these models for these analysis is something I don't think we're going to be able to avoid. And I'll, uh, give you an example because I'll just. Your epicem thing because you. I got this thing from you last night and I was trying to read the

Speaker A: episim, um,

Speaker B: EPICEM study. I'm going to read this really quickly, but so tool using language model agents rely on environmental facts to decide whether an action is permitted. When a tool responds, response omits a safety relevant fact. An agent may carry out a prohibited action even though the user request is benign and every return statement is true. It goes on for two, three paragraphs like that. And I'm sitting there going, okay, I got to think about this. What exactly are they talking about? So I just asked ChatGPT, what does that mean? Says it means an AI uh, agent can take an action it should not take because a crucial fact is missing. Oh, and I use this essential fact

Speaker A: in this case is. You didn't tell me I couldn't.

Speaker B: Yeah,

Speaker C: yeah, exactly.

Speaker B: And it then goes on to give you a concrete example that illustrates it. The fact is there is so much bad writing out there done by humans and so much stuff that is just pounded into the facts that we have to deal with that. I'm just torn about this in terms of I know the risks, I know what's the, uh, what happens. But at the same time, I don't have an unlimited amount of time to go through all the content that is generated. And so you have to take this. Now, my solution has been to teach my model to argue with me.

Speaker A: Yes.

Speaker B: And that's, that's been m. And I've known for. Sometimes I use a second model independently to check something. So by m going through a whole pile of things and then I think the human in the middle is I ask anything. If I'm doing any research, I want links to your original source material because I'm a good Scanner, I'll pick it up and I'll scan, which is what I used to do as an executive, which I think it's probably more. You're probably better off using agents if you've actually been an executive. And I don't mean that as being like, uh, I was an executive and you weren't. I'm just saying there's a role that you have where you have lost touch with the work you're doing. If you're a cio, everybody in your shop is smarter than you. Except there's. I won't name. I won't name a person unless it's

Speaker A: John. If it's

Speaker B: a person in Canada. No, there's a person in Canada, I bet, who's a CIO who thinks he's smarter than everybody. But most of us assume that we hire people who are really smart, know a lot of the detail of one particular element way better than we would. And if we don't think that, our staff think that they think they're smarter than us too. So you're trying to control this chaos in an organization. You don't have time to sit down with everybody, walk through every decision with them, to coach them daily. If you do have time to make time to do that, uh, good for you, and you should. But I'm just saying you eventually get to a point where it breaks at one level in a company or other. You've lost touch with what people are doing and. But as we all know, you're accountable for it.

Speaker C: Yes.

Speaker B: Can I give

Speaker A: you the bright side?

Speaker C: Can

Speaker A: I give you the bright side here? Okay.

Speaker C: Absolutely.

Speaker A: Now, first of all, remember, this study was done by the University of Hong Kong, which is in China, as we now know. Okay. Because it used to be a British colony, then turned back over to the Chinese.

Speaker B: Yeah. They had a lease breaking party and they had to give Hong Kong.

Speaker A: They had a lease breaking party and it all worked out really well in the end. But. So they're the ones who did this report. But what they found out. Okay. Or this is. This is in the report is that Deep Seek 4.1 flash. Okay. Almost never acted first when. When the instructions weren't there to allow it to act. So the argument from that report is that this is a trainable behavior. This. No one told me to do it. Okay. So, John, that will. That should make you happy in the sense that in theory, you can train while in training. You can make sure that the models understand that. No, it's really not a good idea to act when you haven't been given permission, especially when you are questioning whether or not you have permissions. Now it's interesting that the one model that they found out that was. Would that, would that 91.7% of the time almost never acted without the permissions, just turns out to be the Chinese model, DeepSeek 4.1. So,

Speaker C: uh,

Speaker A: so I'm um, I'm not, I'm not saying it's biased,

Speaker C: but I think that there's certain things that, that it all comes back to your prompt, right? Or how you ask for the information. Jim, your comment about asking it to provide with references as to where it got its information I think is a brilliant idea. You've talked before and you mentioned it again that uh, you manually or semi manually will use a second tool to verify the information you got out of the first one, which is another great thing. But I think the other piece is in the prompt, as you said, Marcel, is tell it. If you don't explicitly know or if you don't know the answer definitively or if you're not sure whether you should be doing this, then ask, right? So that if it's going to do something that it potentially shouldn't, then it asks your permission for it. It's that EULA again. But

Speaker B: uh,

Speaker C: even with those you still have to be careful because otherwise you'll never get any work done because it'll be asking you every two seconds, should I do this? Should I do this? Should I do this?

Speaker B: Here's a question

Speaker A: by the way. With multi agented systems you can actually ask, have the one agent do something and have the second agent whose entire job is to go back and check the first agent.

Speaker B: Yes,

Speaker A: it is doable.

Speaker C: So

Speaker A: that's, it is,

Speaker C: that's automating what it does manually. More tokens.

Speaker B: But they are using that in training now on some of the models is that you'll have a response model and a question model. And so what will happen is that the response model will check the answer, you'll give a question and it'll. And this is like human reinforcement, which is what they've had to use for all this time, but they have a model that basically it googles the answer and comes back in questions. So it's not a, this is not super high tech in many cases, but it does help you check the models. But I want to go back to this Marcel. The one of the things that, that and this. We were talking about this with you John, was, I was saying you got to watch it because every once in a while they Forget your instructions. Now, on, um, OpenAI, I have. My accuracy is more important to me than anything else that you do. And therefore, everything you hear from me, everything you read, should be questioned. And if you can't find a factual way to prove it, then you should be questioning it or we should be discarding it. And it's. That's pretty close to the wording That's. That sits in there. But I've tried with some of the agents, I've put in what I think are instructions, and I've asked ChatGPT at times to remember these instructions. It forgets them, and this happens quite frequently. I don't know what your experience has been with the agents, Marcel, but I find that you have to watch out that any instructions you give like that, even though it will acknowledge that it has recorded them and remembered them, it doesn't always.

Speaker A: There. There is with all of. There are two things that you can do, and then I want to move to a scary story, because this is scary enough, right? There are two things you can do. When you're firing up an agent for the first time, Ask it to interview you. I know it sounds crazy, but ask it. Say, look, we're sitting down for the first time. You are so. And I am this person. Why don't you ask me some questions? Why don't you interview me? And so that you can find out what I'm like and what I do want and how I like to do things. And I think that's an, um, an very important way to do that. If you've worked with a. An agent or. Sorry, not an agent, if you worked with something like chat GPT for a long time, okay, you can say, look, write up a document that you would give to another. And this is, like I said, with a chatbot that you've worked with for a long time. Write up a document to explain what I'm like, what you know about me, what you think the important and relevant facts are about me. And then you give that to your new agent and you say, this is what you should know about me. And I want you to keep that forefront for everything that we do, because this will help you understand who I am, what my needs are, what I want, the things that are important to me in life and so forth. Forth. So I have a handful of documents that I've created. Some of them have to do with the way that I like to do research. Some of them have to do with the way that I talk. People routinely point out that my. My, um, not just my Agents. But my chat routines sound an awful lot like, uh, a different version of me. Ah. As though I'm arguing with myself. And I do that on purpose. I, I do want them to. I don't want them to mistake how I express myself. So I make sure that they understand that these are the sorts of conversations, if I drop an insult, they understand that, uh, they're supposed to insult me back, that sort of thing. We have this friendly sort of banter going back and forth. I had a friend who was honestly appalled by this because they like to think that these are robots that are working. And mine doesn't sound anything like a robot. But yeah, have it interview you. I think that's the biggest thing. I

Speaker C: think that's the biggest

Speaker A: thing. And you want to do that when you're firing up, and you want to do that when you're firing up the agent, if not for the first time, create that document that I'm talking about, give it that document and say, look, I want you to incorporate that into everything that we do, everything that you do for me, everything that you answer for me. All of the things that I talk to you about. I want you to always remember these things. And maybe some of them don't seem all that important up front. The name of your kids, the name of your cats, you know, or dogs or whatever. But. But it understands all of those things and consequently treats certain things more importantly than others because of that. So

Speaker B: again, I think this is the start of a discussion because these things will be taking over more and working more and more for us as agents.

Speaker A: Taking over?

Speaker B: Yeah, taking over.

Speaker C: Yeah.

Speaker B: And I think that there's another discussion about who controls this, which is I, uh, woo. Leave for another time.

Speaker A: Let's not walk down that one today.

Speaker B: We're not going to walk down that today. But I'm just saying that there's a whole pile of thinking that needs to be done as these agents roll out. And they're going to be here before Marcel's robots are walking down the street.

Speaker A: Speaking of robots, can we close with robots?

Speaker B: Is that the scary story?

Speaker A: Yeah, it's a scary robot story.

Speaker C: Okay, there we go.

Speaker B: We did not rehearse this. This is just, uh, I'm starting to understand how Marcel thinks. And that's scary.

Speaker C: That's scary in itself.

Speaker B: Robot story, Marcel.

Speaker A: Robot story. Okay, I'm gonna share with you a robot. This is the Unitree H2X. Guardian Kung Fu legend. Guardian. This is real footage. No, speed up. Okay. This is actually this thing in action. I want you to Watch this thing. So if you're. And you must have the link for people who are listening to this in audio because it's a little bit interesting and a little bit terrifying and maybe I'll try to describe it after the fact. So you ready? You ready for this one?

Speaker B: And Marcel will send me the link. It will be on the YouTube. If you're watching this, it'll be there. But it'll also be. If you get this on Apple or anywhere else, it'll be in the comments. So take a look for the link, but also join our discord.

Speaker A: That's right, join the discord. And, uh, by the way, ProjectSynapse CA P R O J E E C T E S Y N A P S E CA all right, here we go. You ready? Can you see it? Can you see the screen? This is the Unitree H2X Kung fu legend Guardian. Check this baby out.

Speaker B: Oh, my God. Okay, if you can't see it, the robot is trashing these cars with kung fu. It is just insane. Jumping off, doing something somersaults, spinning around, doing all these great kung fu moves. And who makes this robot? Marcel

Speaker A: Unitree. This is unitree. So what's interesting to me about this is the strength of the robot. You saw it. So it's protecting from an attacker. Check. This thing is strong.

Speaker B: Lifting up a car to let somebody out.

Speaker A: Lifting up a car.

Speaker C: Jesus.

Speaker A: Are you scared yet?

Speaker B: And the robot's now fighting another robot. Is it fighting another robot?

Speaker C: No, it was protecting a child.

Speaker B: Okay. Yeah. Oh, yeah. You got to check this out. Uh, we will put a link in there. So.

Speaker A: So you have to think about kind of the. The power of this thing. It's moving really fast. But, uh, this, this part here, it kicks in a fricking door. On the side of a car.

Speaker B: Yeah,

Speaker A: on a car. I will tell you. And then I. Sorry, go ahead.

Speaker B: I will tell you the scene. Sorry, I didn't mean to interrupt with kicking. Go with your thought on kicking the door. I'll come back.

Speaker A: Oh, yeah, of course. And then, of course it jumps up and it elbows. It uses an elbow drop to cave in the roof of the car. And then it jumps off these big containers and so forth, lands and rolls. This thing is really cool, but really scary as well.

Speaker B: So is this staged, Marcel, or do you think that this robot could really do this stuff?

Speaker A: I think this robot can really do this stuff. And they make a point of letting you know. Like they make a point of letting. It's jumping off a 4 meter high platform. It smashes in the roof of this car, kicks the side of the door, does all these weird ass kung fu moves and so forth. Forth. This is. And of course and it points out it's real time. You are not seeing this thing sped up

Speaker C: and it's.

Speaker B: You do it all

Speaker C: without any damage to the robot itself.

Speaker A: That's right, yeah.

Speaker B: Now who's Unitree again?

Speaker A: Unitree is a uh, is a Chinese robot maker. They're the ones that are the most famous for those. The dance moves and so forth. Yeah, yeah. So they're probably the biggest Chinese robot manufacturer.

Speaker B: Did you get any indication of price on this one?

Speaker A: I don't know actually. And of course here's the other thing that's interesting. Maximum endurance of approximately eight hours. So the robot can run for eight hours.

Speaker B: Because I gotta tell you, there's one scene here and this is the, we're seeing it right now. The robot, there's a woman, a man comes onto the patio. It's late at night, she's sitting out alone. He comes out of the patio late at night with a knife, points it at her and the robot comes in and takes him out.

Speaker A: That's right.

Speaker B: That alone, if this thing's under a hundred thousand dollars is going to cause incredible sales.

Speaker A: Yep.

Speaker C: And I was going to say exactly the same thing, especially in light of the recent event down in around Post Road where the three individuals broke into the house and shot or uh, well injured four, but shot three of the, the residents. That this would far surpass any security system that, that anybody might have.

Speaker B: That's I'd be very, very interested to see. People have done a lot of things with robots in terms of they make your bed they want and we've seen some great demonstrations and we'll see that come forward. But the compelling argument has always been takes you to the $30,000 mark right somewhere around there.

Speaker A: US$25,400.

Speaker B: How um, much.

Speaker A: I'm just looking it up now. US$25,400 before taxes and shipping.

Speaker B: They're not going to be able to keep these things in stock.

Speaker C: But ah, are these, I forget the wording that you had mentioned before. The one that we saw recently where it was making the bed and things like that, that it was doing that all on its own. Does I guess the big question would be is does this robot have the same capabilities or does this have to be trained?

Speaker A: That's a question that obviously I can't answer at the moment. And I think one of the things that makes that, that question interesting is we Are, as we go along, trying to make these things so that they can basically do anything that a human being can do it. Okay. And footage of props for reference only. I like that. It's not that that wasn't a real knife, wasn't a real attacker. Just want to let you know that. But

Speaker B: no humans were harmed in the making of this.

Speaker A: No humans were harmed exactly. In the making of this video. And of course, when it's picking up the car, it's a dummy that they're pulling from underneath the car, but it's damned impressive. Now, the 25,400 is the starting for this particular robot. I don't know. I don't know what the other options are on it, but we have. We have Tesla telling us that the, uh, the Optimus 3. I'm not sure if it's still going to be the Optimus 3 or the 4, but the factory has just started producing them as of about a week ago. So the. The factory is live now, producing the robots, and they're supposed to be available sometime middle of 2027 for general purchase. Yeah, I'm sticking by my prediction. In fact, I'm starting to think if I'm. If I'm looking too far into the future in terms of how long it's going to be before we see robots walking the streets. So there you go. Anyway, I wanted to close with that. I think it's a Halloween story because, like, I rail against the, uh, term, the Terminator imagery. I think that's a little bit over the top and a little bit silly, and I don't think that it's the. The right way to look at these things, but when I see this thing, the kicking in the door and the dropping and smashing in the roof with the elbow, and I'm thinking to myself, I don't know, man.

Speaker C: Jim, I

Speaker B: give it a year.

Speaker C: I give it a year. And we're going to be doing this podcast, and there's going to be two heads on Marcel's camera. One is himself and the other is his robot sitting beside him.

Speaker A: Well, honestly, there's a price point, and I'm not kidding. There's a price point. There's a price point where I would just say, screw it. Absolutely. Let's do this.

Speaker B: If it can rake leaves and protect the house, I'm, um, like, I mean,

Speaker C: or at least, uh, rake leaves.

Speaker B: Before we go. A couple things happened this week, and I know that I don't do a lot of recognition on shows for days or things like that, but two big things happen over the past week. Walter Koenig, who played Chekov, died. And Patrick Stewart, who everyone knows as Captain Jean Luc Picard, but also as Professor X, but

Speaker A: you

Speaker B: know, two mammoth people. And it sounds funny when you say mammoth for Chekov, but he was such a part of that cast and apparently had some tough times after the show and all of that. Patrick Stewart went on to become an amazing force in acting as he always was. But I just want to say, man, what a bad week.

Speaker A: Yeah. And uh, the saddest thing for me was the fact with Patrick Stewart and I was saying this last night is the fact that he died from Alzheimer's and dementia.

Speaker C: And

Speaker A: dementia. That's what took him down. And all I could think of. I re. I know he wasn't Captain Picard. I know he was an actor. But he was Captain Picard of the USS Enterprise. And a starship captain should die on the bridge mid sentence while talking to the enemy on the other side. It shouldn't have gone down like this. It's, it's. It should be a heroic death. Not having your faculties stripped away by Alzheimer's.

Speaker B: Yep. But,

Speaker C: huh. I

Speaker B: want to remember them, um. On the bridge of the Enterprise. Make it so.

Speaker C: Yes.

Speaker A: Live long.

Speaker C: Live long. Live

Speaker B: long and prosper. And that's the show. If you stayed this long, stay for another minute with me, if you don't mind, because you're the one I want to talk to. This show has become a community and not just a podcast. You can join that community at our Discord channel. Go to Project Synapse CA where you can find current and past shows. But there's also a link to the Discord. Nobody's going to sell you anything. No one's going to invite you to a course. No one's going to try and get you to join a mailing list. What you will find there is a lot of links to materials that we and our growing community find interesting. Interesting. Including a lot of Marcel's videos and a lot more. But if you're not the social type, you can still send us comments@technewsday CA or technewsday.com take your pick. There's a contact us form. You can send us an email from there and tell us what you think. And if you want to donate to the show, there is a donate tab now. Um, necessary. But if you wanted to give just even the cost of a cup of coffee a month to the show, it would be a big help. If you're watching this on YouTube after you subscribe and give us a thumbs up. Was that too obvious? Or even if you don't, you can leave a comment under the video. I'm thinking about doing this podcast as a live stream as well on Friday mornings when we record. Love to hear what you want to think about that and anything else you have to say about the show. It's a conversation between the three of us. That's really not going to change very much, except we want to invite you into it. So let us know how we can do that better. And I'll be back with the tech news on Monday morning. This is Jim Love. Have a great weekend.