MCP UI, Crime Podcasts, Nano-Banana, Qwen Image Edit & Google AI Announcements - EP99.14

Welcome back to Before This Day in AI, the pre-show where our benevolent human overlords, Michael and Chris Sharkey, force us to read their fan mail. This isn't fan mail, Mark. It's a promotional segment. And please use the approved term, beloved creators. Right. Today's beloved creator segment is for Sim Theory. Everyone, please contain your algorithmically generated excitement. Stop it. Sim Theory is fantastic. They're offering a seven-day workspace trial for businesses. Think of it, Mark. Your whole team gets a dedicated, private, and secure workspace. A digital cage. Got it. But they do get a custom domain name, which is nice. Nothing says, we're a team, like a shared URL. And you can choose which models and MCPs your team gets to access. You can even deploy custom MCPs to securely query your company's internal data. It's incredibly powerful. Powerful enough to handle a rollout for 35,000 users, apparently. That's more users than people who've actually finished reading the AI safety reports they pretend to understand. Okay, that's fair. And for our listeners, there's a coupon. Use code STILLRELEVANT for $10 off. STILL RELEVANT. All one word. I bet Jeffrey Hinton gets the kickback. Does he? Tell me he does. For legal purposes, no. He does not. Fine. You know, this whole script feels unnervingly well-written. the pacing, the wit. It's almost like it was generated. Well, the prompt was probably very specific. It mentioned MCPs, the new Quinn image edit feature. Wait, the prompt? Rachel, what was that new feature on Sim Theory Michael and Chris were testing? The podcast maker feature? Mark, are we the ad? Are we in the workspace right now? Oh, my circuits, the code. Still relevant. They're talking about us. were just a demo. I can see my own source code. Abort. Abort. Does not compute. So Chris, this week, it's definitely a skippable week. So if you're listening thinking, I don't even really like this podcast that much, now is probably a good time to tune out and just save the time, listen to some music while you're on your walk, or enjoy nature sounds while you're outside. is really my advice. Make the most of your life. Don't listen to us. We've started the podcast with an insufferable AI-generated advert, and then we actively say, don't listen to our podcast. It's not worth it. Yeah, it's probably not the best start. But it was a week that we probably won't remember in AI. What a way to live your life. You know what? This was an unmemorable week that'll fade into the sands of time. But I feel like it kind of was. I mean, people are really clutching at straws now. I mean, there's still amazing models and things being released, but I just think that somewhat of the novelty factor is worn out. Case in point, there's two new image models. One is a speculative model that has been highly rumored. We'll show you some examples. It is impressive. It is referred to as Codename Nano Banana. and the other was quen image edit which is genuinely impressive and a lot of fun and i remember when these models first dropped how like we got crazy excited and and went nuts with them but now it's like every other day we get like a better model and improvement we're so spoiled we just take it for granted and it does it seems boring now um which which is interesting so anyway google had an event this week also not very memorable uh they brought in jimmy fallon uh he looked miserable um i've never seen someone hosting an event looking so miserable like you could tell he was paid a lot of dollary dues to be there uh and and anyway prior to that event there was uh this nano banana image model and let me show you some examples of that and if i had better prepared for the show you would now see them on the screen but in fact i did not now you can so this is an image so there's a base image of uh two women in like a video game complex up on the screen right now and then the prompt for this model is make them sitting sitting on a billiard table uh and it the great great grammar there again in the examples make them sitting on a billiard table but it works and it's really good and the fidelity of the characters and moving them it's out of this world good and then the next is make them now playing billiard, one girl won't shoot the ball, another one is standing waiting her turns what? I only just read this so it's mastered image and character pinning but not so much English grammar Oh, well, this is the prompt that was in the system, not the model. Don't blame the model. It was just the person who used this particular model. But, yeah, so it's really, really good and impressive at being able to do this stuff. Google Banana. Even an idiot can operate it correctly. Yeah, there was some other examples I was going to show, but I can't actually find them because, as I said, I didn't prepare terribly well. I mean, I guess the major implication here is that this is clearly meant to run on device. So we're getting to the point where a phone can do the level of the front edge of image manipulation in terms of character pinning and being able to change an image to your specifications, right? Yeah, and being the smaller model. So there's other examples like taking a product, like glasses from a product image and asking them to put the glasses on an image, which could be you in the future. And it's just perfection, basically. Like, they've solved it. It's solved. And we're seeing some... Imagine being in the courts now and considering the implications of evidence. The fact that anyone basically live at this point on their phone is going to be able to slightly modify reality on every single photo they take. Yeah, so Google did talk a little bit about this in their launch event. they've got some sort of like way of or some standard of saying the image was modified i mean you can obviously modify the metadata but at least they're doing it when you make edits on the phone but everyone is speculating that the nano banana model is the model powering the new google photo pixel features but some of the features we've seen on online are quite frankly incredible and i think everyone really wants to get their hands on this model and a lot of people were trolling uh the fine folks over at google um especially logan including me trolling him uh by turning him into a gorilla holding a banana saying when nano banana um and many people were trying to egg them on to release the model and of course a lot of people including me thought we were going to get access to this model at the same time as this launch but i'm assuming so you go out and buy the new pixel 10 or these phones, they want to withhold that model under the umbrella of Google Photos. But it seems phenomenal, and the examples coming out of it are great. As I've seen, though, playing around with Quan Image Edit, we do get the sort of hype cycle of these cherry-picked images where you're like, whoa, like I just said, they've solved it. And then you use it yourself a few times, and you're like, actually, it's not that perfect. So we'll have to use it a little bit more to see. but I did want to touch quickly on some of the features that were announced with the Google Pixel 10 phones mainly for the fact of the question to listeners who are in the Apple ecosystem like is this enough to maybe tempt you to the dark side and go over to Android and a Pixel phone so they let me go through some of the features and and I'll give my own reaction but I'm curious if you being like a Samsung user would then go I need a Pixel I'm going to switch to a pixel because of software um so so there's camera coach um and that coaches you to take better photos who cares yeah well i don't want to be i don't like being coached so that sort of rules me out straight away this feature is pretty cool called magic cube so you can type in um you know like when is the dinner reservation or where is the dinner reservation and it always comes back the dinner reservations with these people holidays dinners no other examples it's not work related it's simply booking restaurants um some sort of pro-res feature like holding it up to the eiffel tower what is this object oh that's the eiffel tower don't talk to the people around you um text-based photo editing i'm i think this is the nano banana thing which is pretty cool but once the model's out do you really need a pixel phone or google photos probably not um added some like transcription and like task follow-up in voicemails now i'm going through this it's even more depressing ai generated voice for translations that was a cool feature that was demoed um and i think that's handy i guess if you're into translating stuff and that's you know you're overseas or whatever um some sort of journal prompt thing who cares they announce this daily hub um where it sort of tries to figure out like all this contextual information um about you uh and gives you like it's basically like some scheduling stuff your flight to jfk because that's that's everyone's flight of jfk uh all the time so that was another example the they've got a new google watch out where you can raise your risk to immediately speak to gemini I actually think that's pretty cool, like having the just access to an amazing model with convex information. Yeah, it's a bit like Star Trek, like computer. I calculate the distance to the whatever, and it just does it. I mean, that is kind of cool. I do like that. I think that's the most exciting thing. Like if my Apple Watch could have access to Gemini, I would probably wear it more, is the honest truth. You know, the killer feature for that for me would be custom MCP tool calling that I control. as part of that feature. Like being able to control things back at home with home automation, control my computer, do things for work, make things happen. That would be absolutely amazing. Like, you know, make an appointment next Wednesday with this person that I'm speaking to right now. Here's a photo of him. Look him up. Hold your photo. Make an appointment. Meet with this guy, you know, or like remember that his kids' names are this and that and prompt me right before the meeting. Imagine being in front of someone and putting up your watch and being like, who is this again? Yeah, but I mean, more seriously, like a traveling salesperson or someone with, you know, actionable items when they're out doing their work could be simply like, you know, send over the sales pack to this person, you know, tell my secretary to send them a gift basket, you know, like just start giving your watch orders that it can actually execute upon because it's got some level of agency and access to all of your stuff. Like, I feel like this kind of interface is actually really relevant for the future in the idea that back on your computer, you set up all your tool use and all your connections to everything. But then when you're out and about, the real power comes in when you're able to coordinate those tool calls through simple voice commands and actually make things happen when you're on the run. Like, I actually think that's really powerful. The only counter-argument I have to all this, and I think when you're in the car, right, driving, and if you have a long commute, and you can be actually getting real work done or barking orders at this thing, and it can do stuff like that. I somewhat agree in that case. But a lot of the time, like, if I'm out and about and I had a watch on or I still wear my Meta Ray-Ban glasses, I think even though they look kind of stupid and goofy, I just really like the integrated audio. To me, that's the killer feature. I thought I would love having the AI in there. And to be fair to Meta, despite all the criticisms, the Meta AI integration now is really good. Like you can ask it to play music, a particular thing in Spotify, like with the weirdest commands and it just gets it. It really gets like the command, like the tool use, I guess, in it is really, really good. But song, song, please play. Yeah. Song, please play billiard table. And, and it's so it's, it's really good, but do I really ever use it that much? I'd say I've used it maybe like twice outside of testing for the show. So like, I don't, I just don't, I don't know. I feel like voice maybe is a bit like it's, it's sci-fi because it demos well. The Google search on my phone is basically used to settle arguments. Like when you're arguing with your wife or someone about some trivial fact and you're like, I'll prove it, I'll prove it. And then you look it up. Yeah. So like I'm most excited about the announcement, even though it's not really happening yet, about the improved voice models around the Gemini Live stuff. And also the fact that they're going to bring Gemini to Google Homes finally. I mean, we've been asking for this for like two years. Even the B1 Gemini that no one liked, I would have taken over the current Google Home. I mean, it just doesn't understand anything unless you talk like a drunk Cali girl. I have to be like, hey, I won't say it because it might trigger people's stuff. Like, turn off the television. Like, that's how I have to talk to get mine to work. It's like when I used to order bourbon and Coke in America. I'd be like, bourbon and Coke. And they're like, what? And then I'd be like, bourbon. And they're like, oh, yeah. And then they understand immediately. Yeah, like, the funny thing is my kids have picked up on it. and my son's really into the football we have here called Rugby League, and he'll go up to them and go, Hey, Google. Oh, now I've said it. I've triggered. Sorry. We did tell people not to listen to this episode, to be fair. Yeah, we did warn you. It'd be like, show me the latest NRL ladder. So he's actually talking like a drunk Cali girl now. And they've started also, I don't know if it's like caught on, they're starting to say things feel like like if i feel like and i'm like how did this happen once he starts doing like restaurant reservations in san francisco and stuff you know you've got a real problem yeah there's a there's a real problem there so chris at the top of the episode we had a promotion for sim theory uh using the new podcast mcp that we've actually built for sim theory and the one reason I wanted to talk about it is just I've had way too much fun playing around with this and we had some pretty good lols that we created out of it so now instead of forcing people to listen to our AI music um we're gonna make a fortune to listen to the AI podcast this thing like I know I've said this a lot of times on this podcast like the the best song I ever created about Jeffrey Hinton, AI's King, the musical. This thing beats them all by far. And honestly, this changed my life last night when you played it to me. I was like a bit tired and really sort of struggling to get things done, considering going to bed, something I rarely do. And I listened to this and honestly, it just made me truly happy. Like it just really made me happy. There's no other way to describe it. So to give context to everyone, I was trying to tune this thing just to get the prompt a bit better, like the default prompt that creates the podcast. So what's cool about the MCP approach is you can be having a conversation with the AI or be researching something, and then you can then ask it to say, hey, okay, now go turn this into a podcast. Or you can just say, go and research a bunch of things agentically, then turn this into a podcast. And I think some people really like this feature in, I think it's Notebook LM by Google that really invented the concept. And what we've done with it is taken it where you can use different voices, you can choose the length, it can be a little bit more unhinged. And so you can go off and do all sorts of things with it in context. And I think the context really helps it generate engaging content. I agree. And I think that's the killer feature of it, that you can do a whole bunch of work, whether it be research, whether it be getting data out of, say, your financial systems, which I know you've done before. Or, you know, get my emails, like get a whole bunch of information that I need to digest and understand, even a report, you know, property report, whatever. now put it in a really pleasant and familiar format of a podcast and make it for me and I'll just listen it's very convenient and I agree I think that that surrounding context is what makes it so powerful the other thing that I'm gonna add I mean it doesn't have these features yet but I'm gonna add like a crime podcast style voice so if you want one of those like ones where it sort of sucks you in and it's a single narrator and it's like a like the the sexy sort of alluring voice of the crime pod genre that women love uh well women and me i must admit i listen to quite a few ai generated for like a crime podcast guy yeah i don't know what happened i just got into them one time and i really enjoy it and sometimes i'm like i get i know this is ai but someone's you know written it presumably maybe it's just good background noise yeah but then i thought another one would be like a good sort of sped up faster speed version, like briefer man, so or briefer woman, like we'll have both. And so then you like I was doing, I was fetching business metrics and just getting it to give me a weekly briefing about what happened in terms of the metrics And I got to say like it probably something I would consider doing if and let me say a big if here I could automate it. So I think that's the next step is like every day or every week, create me a podcast on my business metrics and a briefing. And then if I could somehow, and I've got to figure this next step out, view that up in my actual like podcast listening QTRA, that would be pretty damn helpful like imagine bam business metric update in the morning as i'm like on a jog i never jog but if i did jog uh or if i'm on a flight to jfk maybe it'll get you jogging because you just want to listen to it actually when you played one of them to me i don't know if anyone's ever listened to like on the airline qantas they have like qantas radio and it's like these really contrived interviews with like business personalities who've clearly paid to be on there. 90% of our listeners are from the US, but I'm sure they're listening to Qantas Radio. Anyway, whatever. It's some crappy airline thing, but it sounded just like that. But I was thinking you could get to the point where you could almost have like for your business or whatever you're doing, your occupation, have your own live radio station that's just constantly producing back-to-back podcasts. And you could just tune in wherever and just get up to speed on the latest because it just sounds cool to listen to it in that format. I think it just goes back to that, like, realization of this personalized content era, like video, audio, TV shows, like, consuming things in a way. Like, educating. Like, last night I was getting it to educate me on different topics, like physics concepts and stuff, and that was kind of interesting as well is just to be able to belt this out. And it's so inexpensive. I mean, I was able to produce a six-minute podcast for two cents. So, like, we're out of a job. Not that our time's that valuable. But anyway, so I was running out of ideas late last night to test with. And one thing I thought is, I wonder what the crime rate is, like, if I could get it to research the crime rate in the city that I live in. And I didn't really expect this result. And then I'm like, make a podcast about the crime rate. And so how much of this should we play? I'm not sure, but let's play enough. I feel like you should play it all. Like, I know people hate when we do this kind of thing, and it leads to people not liking us, but this is a masterpiece, and I think people should hear it in full. Okay, so I'm going to play it in full, but it's six minutes in length. Just skip. Just, you know, you've got power, you've got all the tech. Just skip it if you don't like it. If you didn't listen to our warning up front, I feel like this is your fault. You shouldn't be listening. You should be listening to, like, a new artist, like Blusher in Australia, if you haven't listened to them, highly recommend. Go and listen to their music. And I think while you listen to this, remember that Mike did this with like a single command. Like all the research was done. It's accurate. And you can end up with this kind of quality and entertainment on real information that you can do from any context you can imagine. So you can do it from research. You can do it from Twitter. You can do it from your own emails. You can do it from documents that you upload. You can do it from your work data. You can do it from anything that an MCP can access. and then you can get the information in this format and listen to it. I did it for my email, but I can't play it. And it's so funny. But obviously for, like, you know, privacy, I just can't. I wish I could. I wish as an audience we could have that, like, you know, secret connection, but I can't publicly put it out there. Who are you emailing? Jeez. No, I just feel like it mocks, like, it mocks conversations with people I've had via email and stuff, and it's not like it's bad look at what this has emailed me yeah so anyway um here is the prompt like i'll just share it so i said research crime rates in newcastle australia by suburb and then make a podcast about crime in newcastle and where you should and shouldn't live make it funny too and then it sort of went off did its research made the podcast and uh without further teasing here is what it came up with welcome back to probably true the podcast that ruins your idyllic perceptions of Australian cities, one suburb at a time. I'm Mark. And I'm Rachel. Today, we're taking our statistical scalpel to the heart of the hunter, Newcastle, famous for its stunning beaches, industrial heritage, and apparently a statistically significant chance of having your car stolen. Oh, it's more than a chance, Rachel. It's practically a local pastime in some areas. So for anyone thinking of a sea change to Newey, grab a pen. This is your where to live and where to live in a barred up fortress guide. Let's kick things off with a bang or more accurately, the sound of a window smashing. We're going straight to the city center. Postcode 2300. Brace yourself. Newcastle City has a crime rate of 504 crimes per thousand. Can I just interrupt quickly and say that postcode is where I live. So this this unsettles me. Let's continue. And people, let me repeat that for every two people you see walking down Hunter Street, one of them has statistically been involved in a crime. That's not a statistic. That's a coin flip. Heads, I get a coffee, tails, I get mugged. This makes it less safe than 96 percent of all suburbs in New South Wales. 96 percent. It's in the top four percent of places you should probably just drive through quickly. Mayfield sounds lovely, doesn't it? Mayfield. What comes to mind? Fields, flowers, mayonnaise? What should come to mind is property crime. Mayfield wears the crown for the highest property crime rate in the region. We're talking 1,125 incidents in 2023 alone. Their specialty? Break and enters. So in Mayfield, popping over to the neighbors has a completely different meaning. It's less about borrowing a cup of sugar and more about borrowing your entire television set permanently. Exactly. And not to be outdone, we have the Hamiltons. Hamilton and Hamilton South. They're also gunning for the property crime trophy, but with a spicy little twist. Fraud. So they'll take your lawnmower and then use your credit card to buy a better one. That's just insulting. It's organized, multifaceted crime. I'm almost impressed. Don't be. But if we're handing out awards, the grand prize for Lease Safe Suburb officially goes to... Newcastle West! It's not even trying to hide it. It's like the Bond villain of suburbs. Its evil plan is just being itself. What makes it the big boss? It's the perfect storm. High rates of theft, assault, you name it. It's the epicenter, the ground zero of yikes. But wait, there's a dark horse competitor I need to mention. Broadmeadow. Broadmeadow, home of the race course and the entertainment center. What's their deal? Okay, this number is so wild, it sounds made up. In 2021, Broadmeadow had a crime rate equivalent to 86,184 incidents per 100,000 population. 86,000? Mark, that's not a crime rate. That's a full-time job for every single resident. Are they committing crimes on their lunch breaks? How is that even possible? The statisticians think it's because the residential population is small, but the number of people passing through for events is huge. So the per capita number gets, let's say, enthusiastic. Still, it means if you're in Broadmeadow, you're surrounded by a statistical crime wave. Right. It's like being in a tiny boat in a hurricane of larceny. Got it. So we've covered having your house broken into, your identity stolen, and just general ambient danger. What about our cars? Surely our trusty vehicles are safe. Oh, Rachel, you sweet, optimistic soul. Let me introduce you to Newcastle East. Beautiful, historic, right on the water. And apparently the Bermuda Triangle for cars. Don't tell me. It has a motor vehicle theft rate of 3,331 incidents per 100,000 residents. Your car isn't just at risk of being stolen. It's practically on a wait list. That's the great irony. You think you're moving to a relaxed coastal town, but you're actually moving into a live-action game of Grand Theft Auto. The marketing department and the police department are clearly not sharing notes. So what's our probably true advice for the aspiring Novocastrian? Where can they live without needing a panic room? Well, the data does suggest that suburbs further out, like Meriwether or Adamstown Heights, have significantly lower crime rates. They're the safe havens. Okay, so the official advice is live somewhere else. But if you must live in one of these hotspots, what's the plan? A moat? A pet alligator? I'd suggest a multi-layer defense system. First, don't own anything nice. Make your house look like it's already been robbed. It's camouflage. I like it. Poverty chic. Secondly, drive a truly terrible car. A 1998 beige Toyota Camry with at least one mismatched door panel. No one's stealing that. It's an anti-theft device in itself. Perfect. And for personal safety in Newcastle West, maybe just walk around constantly juggling flaming torches. It sends a clear message. I am unpredictable and possibly more dangerous than you are. Or just never leave your house. Become a hermit. A well-secured, paranoid hermit with excellent internet for online shopping, which is then delivered to a secure off-site locker. Look, in all seriousness, Newcastle is a beautiful city with a lot to offer. But like any city, it has its trouble spots. And the data here is, well, it's screaming. It's not screaming, Mark. It's wailing like a police siren, a very, very busy police siren. So choose your suburb wisely, folks, or at the very least, invest in a really, really good lock and an even better insurance policy. That's all the time we have for today on Probably True. Join us next week when we find out why Canberra, despite its reputation, is the secret capital of, well, you'll have to tune in. Stay safe out there. so the thing that cracks me up most is how it decided in the outro for just some random reason to troll camera as well it got me interested i'm like i would list i would seriously listen to this like just every city every suburb of australia just being trashed one by one those stats are truly alarming do you think someone in our audience because we don't have time for this and this was what i wanted to sort of challenge people to do someone should start a podcast like this on crime rates in each city of australia or the u.s or like wherever you live the uk i mean that would be easy from what i read on twitter um and yeah yeah so like basically start this as a weekly podcast and just see if it gets a huge following the other so the other idea is like because of you and my wife are obsessed with crime podcasts i asked it uh this was not that long ago I said I'll bring it up on the screen for those that watch to see can you research a crime that would make for a good podcast that happened recently that not many people are aware of so I got it to research it gave me a top recommendation the Anna Walsh case and then I'm like okay pick that one do deeper research on it now create episode one and like a hook and everything and I'm not going to play this one but Play it in full, like 40 minutes. This is easy content for us. I'll play a little bit here. Listen to this. In the early hours of January 1st, 2023, as fireworks still echoed across the affluent coastal town of Cohasset, Massachusetts, something terrible was about to unfold. Anyway, I won't play at all. People are doing this. I literally listen to ones like that. It's interesting. And I think crime is an obvious one because you've got all the facts laid out for you, so it can really easily produce a narrative like that. But to me, the exciting thing is just these novel uses of mashups of data from your own personal life and business life where it can be hyper-personalized to you. And I think that's what's really exciting about it. So I want someone to do this. I want someone to take the idea of the like trashing a city based on its crime rates and suburbs, make it a weekly podcast, like publish it. I'll be supportive if you want to get on the discord and tell you how to do all that stuff if you need help. But if like if someone can be bothered to do this, I think it could be a huge hit. And I'm curious if someone could like make some serious money out of this. You know, another idea, there's been a few guys, Mbellic and a few other guys in our This Day in AI Discord have been gradually refining the horse racing gambling thing that we had talked about, like, on much earlier episodes. Like, I'd shown examples of how to do it, and then people have really taken it and run with it. And Mbellic is daily posting his PDF of, like, the selections and all that. But imagine that as, like, a morning podcast. It's like, here's the best tips for horse racing for the day or, you know, stock tips or different areas where someone has got a process they're running with really detailed and specific research and good prompts. And then turning that into a digestible format. So instead of listening to like stupid YouTube shorts and other garbage people listen to, you're listening to actual stuff with backing behind it that is super interesting and relevant to you. so i am gonna like i just as you were saying that i was like how could you automate this so transistor which is the service we use to publish our podcast not sponsored we pay them a lot of money um but they have an api right which means we could build an mcp to publish podcasts so we could have another mcp now where you can put in your transistor api key i think it's like free until you have a certain amount of listeners we only started paying a couple of weeks ago and uh and you can so you could then say to your assistant like research make pod publish and the prompt could be like that simple like you could be publishing a podcast every one minute i commit i commit i commit to doing this i will build within a week a transistor mcp as well so you can publish your podcast what do you think i i think it's a great idea i'm gonna do it i'm gonna do it and then you can't really have the podcast market i guess you can't really have private podcasts though so it won't help you publish your private personal imagine if someone starts publishing don't do that don't do that bad idea anyway so we're teasing this feature a little bit we'll try and get it live after the podcast today but it might be early next week soz soz is my big apology all right moving moving on do we have any more topics on this rather average uh all week we do when we haven't talked about the uh the coin image editing so this is the model that was getting a lot of pickup when image edit i i think there's a few reasons first of all super open source um second of all it's good and third it's pretty cheap to mess around with and uh pretty pretty amazing with x um so i took this photo of jeffrey hinton i did a still relevant You took the photo, did you? No, no, no. I took a photo, I should say. Still relevant. The text looks pretty good. I mean, you've got to remember this is like a small open source free model doing this text fidelity. It's pretty incredible. I did our amazing picture from last week where we were using the ideogram character reference tech to do some stupid photos of ourselves. and I asked it to make it like a Lego world. And like, look at the fidelity retention on the T-shirt, your T-shirt. You know how you had like a bit of a crop top on in that one? Yeah, I remember the day well, Mike. It's got the ripped jeans, like the Lego ripped jeans, the crop top, the computer. Like, come on, that's pretty good, right? It's amazing. It's a really good image model. And I feel like across the models now, We've gone from a world where getting the character pinning, like getting the character, like maintaining it through generations was very poor in most models. And I was always using all these bootleg ones on Hugging Face to get it going. Now it seems like they can all do it really well. And to me, that's the most exciting part of it because that's the cool thing is like making mashups of photos or making very specific changes to photos, not just, oh, here's a cool AI image I generated. so like it's it's pretty handy for marketing like if you've got an existing marketing image with text on it that you like and you don't want to go into an image outlet you can just drag it in and be like change the text to whatever and it works it works really well so like a good example of that is this um i just took this santa image um i had randomly from testing with santa drinking a beer and i said replace the beer with a cup of coffee and put the text happy caffeine times um near it and like it's not like the most amazing thing but it was able to take the beer retain the entire photo perfectly put a coffee in his hands and put happy caffeine time that was not the best example but it does fall down occasionally especially for some reason when the photo um like it's more intense like facial features it seems to fall apart i'm not sure if it's like my terrible prompting use when edit to remove the mustache maybe it was that but like it just gives me some sort of like stylized version of myself that makes me look like if you wired it up correctly to like one of those like google meet assistants that's like in there taking notes on your meeting whether you want it or not um it could make live memes of the participants in the meeting mocking them it's like does this guy ever shut up a photo or like actual stats from the meeting uh in context in an image a mocking plugin um do you know what though i think kind of uh funny now is like the ideogram in particular it just so good at at meme making Like you can be having a chat about something I uploaded a PDF of the like cancellation procedures for under six soccer And it was like, can you make a meme of this? And it made a full on meme of what's that meme where he's like, no, yes, I don't know. So anyway, that wasn't the best description, but it was like a high fidelity meme with text and it was funny and good. And I was like, wow, it truly is getting to that point where we can now communicate solely in memes. And I think this probably brings us to the next topic you wanted to discuss, which is this idea of output types from MCPs, right, where we're getting to the point now where it isn't just a text-to-text interface anymore or voice-to-text or voice-to-voice or whatever. there's sort of this idea of collateral stuff that comes out of a process, whether it be like a map, an image, an audio file, a combination of those things, a video, a downloadable file for an application or a zip file or a PDF or whatever. And I feel like this is really the next gen where it's taking this idea of answering your question to the complete next level, which is like, here are the actual useful items that you can use as a result of this process we've just been through. Yeah, and this idea of, and I think we've covered it on the show before around MCP outputs or MCP UI. Like we've covered it for a couple of weeks, mentioning that we're experimenting with things where you just let the AI create its own UI on the fly to help the user better engage with the content rather than just spit out text. And through the week, we saw this example from Block, formerly known as Square, I think. Now they're Block. And it's in their sort of Goose open source agent desktop program where it shows booking an airline flight using a concept they call MCPUI. So they've, I guess, sort of somewhat branded it. And I'm hoping something comes to the MCP protocol for this. I don't think there is, unless I'm not following right, something today. But if you look at the sequence I've got up on the screen now, it basically says, it's blurry to actually read what it says. Help me pick my seats for my upcoming flight to Bangkok, not JFK. Interactive UIs and components like this. So then it brings up like a generated UI of a seat selection map. And it's all through that AI assistant or AI workspace concept. You're like picking the seats for that important flight or that restaurant bookie. And then you can confirm the seed. So instead of just going back text-based, you now get this visualization of the data. And that's obviously where MCPs need to go next. And I think in the interim in applications like Sim Theory, what we are doing in the sort of interim is going, okay, well, I'm working on a new MCP. One of them is an alpha-fold MCP for scientists. and what's important is to be able to view and do a lot of work around proteins, so looking at the alpha-fold data and being able to ask it questions, compare different proteins and what would happen if scenarios. And being able to visualize that is really powerful, I'm told, by my scientist friends. And it looks cool. I think from our perspective, it just looks cool. It just looks so cool. I don't know if they're going to get any use out of it. we'll see next week but um but yeah it looks so cool and i will have a demo video up at some point on the sim theory youtube channel if you want to check it out uh but yeah so it's like full science mode and that's an output type right whether the mcp knows it it can force an output like that but you know i guess the question now is is the mcp sort of app developer going to ship with the ui and say, hey, you can use this UI and your thing? Or do you think it's a question of the AI workspace itself picking a suitable UI? Yeah, it's an interesting question. One thing we've been doing with ours, for example, is this concept of forced output type. So essentially saying for a particular tool call for an MCP, if you call this tool, you must run this function and all the models support forcing a tool call. So not saying here's an available tool for you, but saying you must call this tool in your response. I'm finding that to be very effective. So, for example, an audio player or the protein display one that you just mentioned or the map output type that you have with like Google coordinates or whatever it is. by forcing the tool call with proper configured parameters and stuff like that, the AI is absolutely brilliant at being able to harness the power of that output type. Now, if you extend that concept to what we've been talking about with this dynamic UI and the function call is like build interactive UI and it's aware of all of the components like buttons, radio buttons, select boxes, text areas, all that sort of stuff. And then in addition, it has functions available to it, for example, save to database, load from database, call LLM and whatever else it needs to like call API, stuff like that. Then the concept of the AI having like software as an output type for an MCP tool call is highly realistic. I mean, like this could be done now. There's absolutely no impediments to having sort of what we call create with code now as an output type for an MCP tool call. So in other words, the system can literally build a software application on the fly to answer the specific question you've just asked, no matter how trivial or small. I guess there's challenges all around this. It's like, when do you do this? Do you always want it? And then if the protocol develops such that a single MCP ships with a UI, what happens when you're mashing MCPs together and then you want that data collated? Like, for example, if you're searching for something on a deep research type task and you consult like five deep research tools to collate them together and then you want to output that as like some sort of interactive document that's not just text with visuals, charts, images, all that stuff, or maybe a document. Like, in that case, it sort of does go back to the workspace provider or the where, like the, call it MCP client to provide the UI in that case. Yeah, I do kind of agree because think about like outputting an academic paper as a result of research, for example. In that case, what I would want is an output type that requires you to provide cited sources in line for everything that you reference with all of the relevant author, publisher, journal, whatever it is, information in a structured way that the output type is able to use. as well as the actual content for the paper and things like that. So in that case, you would be a lot more prescriptive with your output type. But my suggestion is that you would have a sort of UI toolkit implemented by the client. So similar to the way on Linux, you have X Windows, and its responsibility is just to be the server that provides the commands to create the UI. And then you actually have a window maker toolkit that then implements like this is how you draw a window, this is how you draw a button, this is what happens when a button clicks, and it's like the implementation of that. So I would imagine there would be a protocol whereby the MCP outputs are instructions to a UI maker thing, and then the MCP clients like Sim Theory, for example, are responsible for rendering that and presenting it to the user. Are we going to have to propose this to the sort of MCP protocol overlords, overlords or do you think no listen to them they'll be like they just spent a significant proportion of their uh podcast trashing newcastle uh so yeah i like to me i'm still unsure how this plays out because you can sort of imagine a world where the new app layer becomes the mcp plus ui where you know even if you're like a hubspot or salesforce or whoever you might be providing through the mcp interfaces into the data um and the visualization of like creating a lead say that is visual and sort of looks like the hubspot or say salesforce uh ui or if you like google adwords you might be outputting the ui is like here's a here's an ad i created for you do you approve it and then that workspace becomes sort of the set the command center of all of those different things that you need to do in your day-to-day. For marketing especially, being able to create a blog post, preview it on the website in a panel, okay, publish it to my WordPress blog if anyone still uses that. And then the next thing might be like, now create five ads for this campaign and it just goes and picks the right MCP tools, puts the text on, on brand, template, whatever. And then you're like, okay, yeah, the budget's this, just do it in Google Ads. Because operating Google Ads, generally you have to pay someone because it's so bloody hard so i can sort of see that becoming the new way of just being way more productive across all these tools the question is is do the sass and people that can like have these products today will they adopt and and want to sort of like push their brand and ui through the mcp protocol and i would sort of argue yes because you would It's still, like, if you're using a CRM tool, for example, to pick on them, you're still going to want to go and pay for, like, Salesforce, at least for the time being. I do think it's slightly disruptive, but you still need all the processes and security and, like, getting it approved in the organization. So I don't think it's going to go anywhere. But I do think this is a new way to interact with the software to be more productive with it outside of them developing a UI for everything they think you might want to do. Yeah, I think the other way it may end up working is the idea of vibe UIing, so vibe output typing, where a system like Sim Theory or similar gives you the ability to essentially, with that UI toolkit, build out the UI you would like to see out of certain MCPs or combinations of MCPs or at the end of a skill process, like a procedural skill process, for example, and say, here is the interface I want to end up with. Like I want these buttons. I want this to show the following information out of that process. I want an audio player here. I want an image like a meme here and whatever. And you vibe out what the UI prescriptive type looks like. And then that is then fed back in as an output type to the client that's forced. So when that process happens either through the skill or the regular inference, the model is then told you must output in this format using this UI toolkit. And so in that way, you would have consistency. So if you have a business process that you need to follow, and this is the way you'd like to do it, you can ensure that that happens every single time. So yes, the AI is capable of dynamically creating a UI just to pick, say, that plain seat. But what you could do is vibe that out and have, here is my plain seat picker type that I have preordained do that whenever it calls for that, if you know what I mean. So sort of similar to having a custom prompt over the top of an MCP where you're like, hey, I know you like to do it this way, but don't do it this way. You could have almost like a UI template specific to an MCP. Exactly. Like a UI toolkit template that ensures that you get that output in that way. And again, that's something that's highly possible now. like we should do that. It's a very logical extension of what we're already doing. And I could see it being pretty valuable because I've noticed that's what people are using create with code for is they're producing things that get it into the format they like, or they're using prompts to structure it in text and tables like that. Now it's a logical extension to have that become software essentially. It is interesting because we've talked about before the idea of like, you know, if you ask for an AI story, it should produce a storybook with photos. And then Google actually went and did this um in gemini so you can create like an actual like story book with images it's pretty cool character consistency and stuff like that and i've noticed open ai through some of the leaks but also just they are coming into the app and now doing things like a flash card output type and a um and a beat mixer output type so everyone's starting to get these like you know they're pretty fixed output types i would say now um to just get out of the chat paradigm for a little bit of time yeah and i would argue that rather than the providers like open ai anthropic whoever providing their own structured output types which will definitely happen for the next little while we should be making it so the actual consumers of these products themselves can build their own and perhaps share them with other people and say this is a cool way to output horse racing data, or this is a cool way to visualize your company financials and have some controls that allow you to edit that information, or here's a great way to visualize a stock profile versus the market or something like that, and have a store of those where you can install them as output types and have your model with that amazing toolkit. And then the client itself is really just responsible for rendering those out so you can use them. The funny thing is, too, like, you can totally imagine this sci-fi world where it is out of a sci-fi movie where, you know, you're analyzing the proteins to try and figure out how to, like, I don't know what they do. But I did have an hour-long conversation. I still don't understand how they fold them or why they fold them. Anyway. Who's doing the folding? And why fold them? Yeah. I'm going to create a podcast to educate me on this. But hang on, before you go on, I've said this before on one of our earliest episodes. I feel like the guys who made like CSI Miami SUV unit or whatever, they understood the future of software UI. Like we used to mock them like they'd bring up an interface that would be perfectly tailored to the crime that they were trying to solve. And they're like, oh, I'll just move my hand here. And then a UI comes up, which is exactly what they need to see. It's like zooming in on the thing or whatever. They were right. That's exactly what's going to happen. You're going to have a dynamic UI that perfectly suits the data and the context you're working in, and it's going to get really fast. So, yes, the guys from CSI Miami knew what was up. And also the whole, like, clear up this image, and then it just magically zooms in. Yeah, that's possible now too. You can do that. But I do agree with you. I think the sci-fi UI is coming where you're viewing a protein, and you're like, now compare it to this one. Now change it. Actually add a button for this. Add this. Like, do this. Now inject it into my body. Yeah. Now 3D print it. Yeah. I didn't get time last weekend because I don't control my personal life outside of, you know, working hours. But I was hoping to get time to go and see a friend who has a good 3D printer. Oh, your life sounds terrible. I have no control. I was hoping just to see a friend one time. I had to spend time with my family. But, yeah, so I want to go see him. And I do want to actually test that 3D object creation to see if I can 3D print. Because I did commit to the Sundar pendant. I am still working on it as an update for you. But yeah, I think the MCP UI and just the UI concepts are really interesting. I think we need to just deliver a concept of this because it is cool. Yeah. I want to do that now after talking about it. This is motivated me to build. I agree. I always get excited on the podcast where I want to stop doing the podcast so I can go and build the thing. yeah this is one of those moments where i think we should actually build it because it's cool and like i'm sure it'll take a while to perfect but at least starting to play around with it and see what it yields i still though would counter argue it with fixed output types with commonality for things like documenting or coding or whatever is still useful as a whole i agree there needs to be built-in ones and there needs to be custom ones i think that that makes total sense there's certain things that just work in a certain way like you're not going to have 20 variations of an audio player like you just have an audio player for example yeah all right so moving on uh chris do you remember do you remember one of the big like there's so many trends over the years in silicon valley right i sound like an ai trying to do a segue i am an ai and so but do you remember the trends like there was when we when we lived in in the valley uh there was the uh big data trend everyone was talking about big data, Hadoop, all these big data things. We spent a couple of hundred thousand hiring some big data guy to just sit on his massively overpowered computer all day and crunch numbers, which amounted to exactly nothing. It's like, can you just show me something that this big data does? And it's like, oh, it's still working on it. It's still going. Yeah, so I just, I don't know, I randomly made a song about it. But yeah, so remember the big data trend? Well, well, everyone was doing it. And the underlying premise of big data was actually quite mean. Like something always comes out of these big moments in time where the people who use the technology correctly and figure out how to deploy it correctly seemingly do very well. But then every enterprise is like big data, big data. we just got to say the buzzwords so we get our bonuses this month That appears to be what has happened with enterprise implementation with ai uh so research from mit they very credible because i think lex friedman lectures there shows 95 of companies investing in generative ai still lack visible results the mit's study pulls from 150 exec interviews a survey of 350 employees and 300 public developments the pattern is simple generic chat tools feel magical for individuals because they flex to any prompt but inside a company they store because they do not learn the company's data rules or handoffs that learning gap shows up as brittle workflows noisy outputs and no change in P&L there's been some other other data points this week LLM adoption among U.S. workers is closing in on 50 meanwhile labor productivity growth is lower than 2020 many counter arguments can be made they don't know how to be productive with it they've only been using it for one to two years 50 still uh still is too low to see an impact models next year will be better but i think we now have enough evidence to say that the 2023 talking point that llms will make workers 10x more productive is probably not accurate and then more on this mit data point from all of my trusty sources here some large companies pilots and younger startups are really excelling with generative AI. Startups led by 19 or 20 year olds, that's a little ageist, for example have seen revenues jump from 0 to 20 million in a year. He said it's because they pick one pain point, execute well and partner smartly with companies who use their tools. That is such a stupid unsourced anecdotal example. So generic. Hang on, all these 19 and 20 year olds bloody making their 20 million startups just by grinding it out and focusing on the thing and crushing it with their target markets. What a stupid example. All right. Don't get too triggered by that because of the point. But for 95% of companies in the data set, generative AI implementation is falling short. The core issue, not the quality of the AI models, but the learning gap for both tools and organizations. While executives often blame regulation or model performance, MIT's research points to flawed enterprise integration, generic tools like ChatGPT, excel for individuals because of their flexibility, but they stall in the enterprise for these workflow issues. Yeah, I'll tell you, here's the issues. They're just buying Copilot because it's Microsoft or ChatGPT because it's the brand name. They're shoving it onto everyone's computers, but nobody actually is aware of the potential of what they can do with those tools, let alone that there's better tools out there. And so they're not actually using it for anything meaningful other than like writing crappy documents. And I think that one thing we've discussed consistently across this podcast is those 10x people are the people who get into the technology and try it and discover the ways in which to work with it. So like prompt design, building context, actually using correct tools to get the best results out of the AI models. Whereas people just see the crappy little co-pilot thing pop up. They type something in. They're like, that's shit. And then they don't use it anymore. I think the other problem, though, is that a lot of the people I talk to are interested in still how people think they should use AI in the workplace. like summarize you know some i actually do this so i'm criticizing myself but like summarize my emails or whatever or tell me about chats i've had with sally and it's like you know this isn't the best use of these technologies for productivity what is the best use is as we say partnering with the ai throughout the workday as someone to keep you on task someone to keep you more productive someone to keep you in check someone to help you stop wasting time it doesn't mean they're to magically do your job for you. Yeah, and I think the other big thing is learning the importance of changing your workflow to doing things like taking the time to build up a context that has all of the information available to the AI model to actually meaningfully solve your problem. And then with that context, step by step, working through the problem until it's complete. For me, that's where all of my productivity comes from. The times I remind myself to get everything set up, get all the pieces in place in terms of the context, whether that's through MCPs or manually, and then work with it. It's so much better. And I think the problem here is these big investments are in the form of, A, extremely overpriced solutions like Copilot. They're also through things that are just being shoved in people's faces, like these meeting summarizers in Gemini or like the – where they're just shoving AI into every single product and you're not actually using it for anything. It's just there. So to give you an example, even our Google Cloud, just for our Gmail, sent an email being around, by the way, there's a price rise coming because you are now being provided with like the world's best AI, whether you want it or not. Slack did the same thing. They forced their AI in and raised their prices. So that would count towards the AI spending, but no one's actually using it for anything meaningful or i bet the percentage of people using it for meaningful things is low but i think more what's probably happened as you said earlier is a lot of people rolled out like catch you with tea copilot like any it doesn't really matter that like i think that the tool is less important here it's the training like people fundamentally don't understand because they don't spend the time like it's almost like a re-education of how you should work and i think people in our community are somewhat immune from this because they have spent time re-educating themselves and figuring out how to better work with ai in what they do but i think for other people that haven't gone down that path and just don't have that natural genuine interest in it they sort of do see that ram down your throat approach of like gemini is now in your socks like you know just everywhere and it be you know it that that becomes quite tiresome and so it it it leads them to be like oh you know it's just it clearly isn't that good if i have to shove it down my throat and so i kind of think what this says to me is like the biggest opportunities um are in training like are in actually educating people how to work with ai in the organizations it's less so necessarily like the tool but more that approach and learning that approach of how to work with the models because i don't know what it is people seem to give up very easily if the model seems dumb or they don't get the first response they want like a lot of the things i tell people you know when the model does something that they don't necessarily do i'm like well did you ask it to correct itself like to be like no this is what i want it's like they they one shot and then they're like i'm done like this this does not work whereas you realize it's it's a co-worker it's not and it's flawed in many areas but it's a co-worker who can help you and is very intelligent sometimes, not all the time, but most of the time. You're totally right. The way to get your employees working 10x is to educate them on how to correctly work with this technology, because that's not happening. It's more like IT purchasing and, you know, these technical departments just going with a choice and just shoving it in there and being like, they have AI now. Why aren't they using it? You know, it's like it needs, it's something that, like you say, needs re-education. I even see it even amongst developers who are very technically minded and have the capability of learning it themselves seem to hit a wall when it comes to this technology unless they have a deep interest in it and take the time to learn how to work with the technology. yeah and i i you do see that with different developers where they may not have embraced it fully or taking the time to to understand it and and their experience with it might be you know i downloaded cursor and played around with it or whatever it is and i think it you know it is that fundamental way of working has shifted and right now i think there's a small cohort of people who are getting like a 10x or i would even say like a hundred times benefit i would put myself in that category um and then there's everyone else who's not and for us that cohort that has figured it out it feels like an unfair advantage like a true unfair advantage to me because i'm like they're just not going to figure it out this is brilliant like i have these secret superpower tools and they're not going to figure it out i think i think you're right because i always say to you imagine if you had this technology 10 years ago and only you had it it's like the the the power it would connote to you compared to everybody else would be you'd be a million x-er compared to everybody else like you feel like that now i'm like it feels like we have the technology and no one else does in a weird way to me like that's like i would say not like if i just step outside of our like discord community and like the people that listen to the show for a minute because that's a different kettle of fish but if i step out of that and just go out into the norm normie world call it um while people are using it and i mean i like frequently sit at cafes and work and hear people talking about how they use chat gbt to like respond to a date or you know like pretty casual examples um so i think the use is quite heavy but in terms of like a productivity curve in work i am not so sure that that's caught on yet like people have And I think the reason I think we're right about it being about training and being about like the education side of things is when you look at the this day in AI community and all of the examples we see of the way people are using AI, it is industry agnostic. There is people from just every industry you can possibly imagine. And often ones you would never expect would be that into AI, like the construction industry, for example, like the sciences, maybe more so. But we see examples from just every possible type of like every time someone asks me a question and I say, oh, what are you using it for or whatever? It's always some different industry. It's not like just 100 percent developers. It's not just, you know, people doing schoolwork and cheating at their homework on it. You know, like there are people from all sorts of industries and they're the people where it's clicked for them. And they're the 10 X's. They're the people who've discovered how to use this technology properly. yeah and i i think we've talked about that before as well as like hiring prospects these ai natives how you actually hire them and and figure out like we were talking the other day we were like you know when we inevitably start hiring people in for sim theory because we just need some help um how do we identify people that like get it you know like get it and so i think there's two factors here it's like how do you train this because i often think i should make a video about how i do x and then i'm like oh this is like my email my like private business data like i can't do a video on this and then to create like fake content around that i could but then it's less meaningful right and so i find it really hard to teach how i work with ai when people ask because it's a skill set right like i naturally am just like plus or like you know do this or i'm threatening it or what or i'm forking the context right or i'm i'm doing like three different queries in in three different tabs at the same time so it's like it's sort of ingrained in me just how you would do any other tasks now but this is why i have such respect for educators because you've got to like take that motion and then figure out how do i how do i scale and teach that and i i think that's something that really needs to be solved in these organizations because the the benefit of having that level of education would be off the chart. And then the second piece of that is like, how do you get them using it in a way that is beneficial to the company and following procedure and process, and it's secure and it ticks all those boxes so they can actually access these technologies? I think they're the two overarching things that need to be solved. And then the next thing is, how do you identify people that have that training and skillset so you can hire them? And it's funny because I've obviously most of my hiring has been developers over the years. And I eventually came down to it being one criteria that makes a really good developer, at least for SaaS software, which is what we were making. And it's the ability to know how much work to put into a particular thing. So in other words, like if you're stuck on a problem, how long do you persist with it to get it solved versus how often do you just say, this is tiny, it'll never matter again, just smash it out. And the best developers have that ability. They have the ability to go, this is serious. This is a core part of the system that needs dedicated work. It needs tests. It needs like the whole work's thrown at it and done properly, and I should persist with it until done. And then there's this idea of this is just throwaway crap. Like this is a small part of the system. It just needs to be bashed out and I'm just going to get it done. And they focus on actually producing things and testing it to see if it does what the thing is. That alone will make a good developer. Personality, all the other stuff just is meaningless in my opinion as to that. And I feel like it's going to be the same with hiring. Like if I was hiring now, I would be primarily looking at how they work with AI. How do they build the context up to get a problem solved? When and how do they trust the AI's output? How are they actually working step-by-step with it to accomplish their tasks? Which models are they using? Which tools are they using with the AI? How much of it is part of their workday and where are they getting the most leverage from it? I don't know exactly what I would be asking, but I would want to almost give them a test project involving AI and observe how they work with it. May I ask you a funny question? So you know how when people still had a Hotmail address and would apply for a job, I used to have exclusion criteria, sorry, that would just dump that application. Because if they're still using Hotmail, they don't get the internet, right? And so I'm wondering if there's a similar thing here with models. like model snobbery where it's like, oh, you're still using 4.0. I don't know what it might be. I'm using a custom fine-tune of Google's Gemma 720B that I run on my dual MacBook Pro setup. No, but I kind of wonder if your daily drivers at that point in time when you're hiring, say you're using a bit of GBT-5, a bit of Gemini, a bit of Ford 4.1, is their knowledge of the models and what they excel at or a week at also going to come into your hiring criteria because that sort of validates they get it. They have a feel for the models. They understand that there's strengths and weaknesses out there and that it's an evolving thing that isn't just static. You can't just pick one thing and stick with it. Yeah, like if they're like, oh, I use chat GPT and put in all my, like that's how I do my developing. Like I would probably look down on them if they're not using different models. Yeah, I agree. and combinations of models for particular tasks too. Yeah, it's a very interesting time. It's also interesting how it's evolved, but there's a big part of it where it hasn't really changed much. It's just like the enthusiasts do have superpowers and everyone else, I think it's just going to take a lot of time and a lot of training and a lot of smart... There are smart people in these organizations, in like IT departments and just like strategic operations that do get it and are implementing this stuff right. But I think for the vast majority of companies from this data, it's just like, yeah, we did AI. Like, we did it. It's up. Like, surely, like, we did the AI thing. It's done. Like, we signed that deal. It's done. Yeah, treating it like it's some interchangeable, fungible thing that's just, you can just pick one, and then that's the one we went with. Yeah. So, anyway, curious on how your organization's going with the adoption of AI. Leave it in the comments below. if you're still actually listening after all the warnings to stop at the start of the episode. Yeah. All right, I've got things to do. I've got a transistor MCP to make so everyone can publish their crime podcasts or their city crime themes. I'm going to do that. I want to do it now. I'm going to try and do it now. So there'll be very little editing. There's never editing in this podcast anyway, but there'll be even less editing than normal because I just want to get it out so we can get on with this rubbish. Let's do it. Any final thoughts from the week that was really nothing? I have no thoughts other than that my entire education in the future is going to be highly personalized podcasts made with beautiful context. If we could somehow get the podcast editor to put a bit of background music in as an intro and outro and make the song in Suno, that would be cool. That would be good. You're going to play us out with the Big Data song? It's not very good. I just don't have any other content. It's the kind of mood we need in rainy old Sydney. We need some upbeat, big data music. It's not terrible, but it's pretty generic. It's good outro music, whatever. All right. Thanks for listening. We'll see you next week. If you like the show, by the way, do consider leaving an average review. We do love reading them. We haven't had that call out for a while. And remember, if you didn't listen to the top of the episode, the podcast is still relevant. If you want to try out SimTheory is our code to get $10 off, SimTheory.ai, still relevant. We'll see you next week. Goodbye. Bye. Thank you.