BONUS: Our 2026 AI Predictions.... Who Wins, Who Loses, and What Changes Everything?

15:30Grant:Because how often when you think or write does the entire sentence come out fully formed?

15:36Corey:I don't know. If I'm honest, I've – well, yeah, usually I'm thinking – I don't know. usually i know what i'm writing some words ahead but i don't know how many uh i'm gonna have to sit down like with a pen or a keyboard later and see we should test that see what they'll feel like i'm really curious now you've you've sparked an interest well i've also heard that um some i've

16:01Grant:read a research paper where some believe that the thinking process or the way that thoughts come to us consciousness, let's say, is like quantum entanglement, like quantum. Yeah. Which I which I find compelling. I don't I don't know enough about that area to weigh in on it intelligently, but sounds right to me.

16:22Corey:Yeah. And that's closer to how Pathways model works, I believe. Yeah. I have another prediction here under model architecture and technical direction. and that is that brace yourself for the year of AI of things. I believe that this year we'll start seeing small language models live on board in every single thing you touch, from your refrigerator to your car, where it's probably already at, to embedded systems and all kinds of things. I think you should start expecting like, and by your appliances, I mean, like, you know, you go to Best Buy and you see the really expensive. Yeah, AI in your couch.

17:06Corey:Right, Cal? Cal, good to see you. What up?

17:08Grant:Yeah, Jordan and Cal. What's up, guys?

17:10Corey:Wow, I love it. I love it.

17:13Grant:This is great.

17:16Corey:Yeah, I think small language models and things like that. But think like, you know, you go to Best Buy and you see those like$6 ,000 refrigerators and stuff and a washer and dryer that costs, you know, months and months of salary. I really think that we'll see those start having a lot of onboard AI with small models, you know, things that can, you know, maybe a fridge that can suggest what you could make based on what's inside it. But AI is a fridge that can then prepare the food inside the fridge.

17:51Grant:Yes. And then it's ready for you when you show up. Yes. That's a ways out, I think. Yeah. From a robotic standpoint. But that would be a really cool home robot. It's like it's actually the robot just works inside your refrigerator and prepares your food for you.

18:04Corey:I want it to ding me at like, yeah, yeah, absolutely. Samsung will be all over that.

18:09Grant:Samsung is already all over that, I'm certain.

18:12Corey:I want my refrigerator to text me at 1 p.m. and be like, hey, here's a few options for dinner. Which one are you thinking? And I want to be able to be like, yeah, I want hamburger helper. Yeah. Or whatever the case is.

18:26Grant:No, that's great. um also so i've been reading a lot of headlines about this uh especially with ces this week that ai on your tv is it's like coming oh yeah if it's not already in your tv it's it's coming this year

18:39Corey:even in your streaming services i i think i mean you know and there's already a certain amount of ai in there and has been for years frankly that are you know what's making recommendations for shows to you. I mean, that's all machine learning. Like all of that. Netflix's algorithm. Prediction, stuff like that. That's kind of one of AI's earliest real world applications was Amazon. We think you'll like these. And, you know, related to your purchase. I mean, that's all essentially, you know, predictive stuff that was running through ML models. But I think you can expect to see a lot more of that and get used to and nauseated by the term AIOT?

19:24Grant:I think, I think, I'm interested. So we just wrote something in the Neuron today that was about Dell and Dell basically admitting like people don't want to buy computers just based on them being AI computers.

19:36Corey:Like they don't understand why they need it. They don't like it as a term.

19:39Grant:A lot of people are anti-AI, you know, obviously not in our audience, but in the broader community. So I wonder how they're going to approach AIOT because it's really good to talk about AI and your products for investors, but for regular people, we don't give a F about that. We care about what is the actual value, right?

20:01Corey:I agree. I agree. This is absolutely my favorite comment from Candice here. I need AI to maintain a specific level of something on Blanc in my fridge. For sure.

20:11Grant:You've got to keep those temperatures optimized. You've got to temp max it in here.

20:17Corey:Oh, my gosh. All right.

20:21Grant:Grant, do you have any thoughts on these before we go? Yeah, I think we would be remiss to not talk about world models on the new architecture front. Yeah, I think that's fair. Yeah, so I think that there's going to definitely be something. I think there's going to be a productized version of a world model in a really real way this year. So two things for more technical people. When I talk about world model, I'm talking about, like, a world prediction model where it's, like, generating an entire world, not from a conceptual world model.

20:49Corey:It's creating a world of Warcraft in front of you.

20:51Grant:Yes, exactly, that you can control and move. Google's Gemini is, or not Gemini, Genie is the closest to that, right?

20:58Corey:Yeah.

20:59Grant:Genie World 3.

21:01Corey:I don't think we're going to get access to Genie ever.

21:03Grant:No, no, I disagree. I think they're going to productize it this year. Yeah, there's a lot of incentive for them to productize it. I think probably what they have to do is they have to make it much more efficient to run. which having a ton of tpus available will probably help them with that um they're probably i imagine that's what they're working on right now and then you know google has a lot of ideas when it comes to product and a lot of cooks in the kitchen so who knows what form that'll take but i imagine it'll be something similar to uh the play the new playables uh feature on youtube have you seen that yeah or you can generate video games directly on youtube which is insane by the way i imagine it'll be something like that where you know you can make like a 30 minute movie that you can play inside like a movie game like we need a term for this what is a what do you call a movie game immersive video immersive video game i don't know yeah yeah i love that watching a movie and you

22:02Corey:know and it at like 32 minutes and it's like quick grab your controller no i think it'll be the

22:07Grant:entire movie zombie you walking around and like meeting people and talking to them yeah yeah Yeah, and you may be right on.

22:17Corey:I do think I hadn't even considered world models. I'm glad you brought that up, man. Yeah, I think...

22:24Grant:Here's the question. Is it similar to Sora?

22:27Corey:Do you think it will be the caliber that we saw in the demo when it gets released?

22:34Grant:Okay, what do you mean by that? Can you unpack that?

22:36Corey:Do you remember how when we first saw the Sora demo and then Sora came out several months later and it was like not quite what the demo was like it was it was a nerfed version because like yeah because when i watched the demo for it i very much get the vibe that like this is a thousand

22:53Grant:dollars per million tokens probably really expensive let's let's bring up this comment from cjw uh from cal at 9 21 and let's put this on the screen um knowing demis i i recently watched the Thinking Game, which is the Google documentary about DeepMind and how they kind of formed and eventually solved AlphaFold. And I really liked his thinking there. He loves video games, but he wants to solve intelligence. So he's going to try and find a way to make whatever they build useful for real world products and problems. So Cal said, I could see the world building being useful for companies storyboarding outlining video games and movies for cheaper i think probably they'll also find a way to offer this to industry so that they can do you know let's say like everything that we're using vr for now even better um you know you could do in simulation surgery that's like actually looks like reality you know instead of having to spend all this time creating these digital worlds it's like you can have the ai create the digital world for you that actually simulates reality and then you can focus on you know just making the surgery simulation actually accurate yeah um things like that like or you know how to train self-driving cars we've seen this is like a huge use case for world models and video generation models so i think they'll probably try to productize it for industry and and perhaps also as a fun demo for people to play but you know there has to be like a financial incentive there so unless they're going to launch a new streaming service or a new tier of youtube they might not do that yeah that's true too that's true i'm anxious to see where it goes though you know it's exciting

24:44Corey:all right is that all for model architecture and direction do we want to go to yeah i got we actually have a question in the chat we want to handle that first from kevin

24:56Grant:dearie you see that one as ai engines like perplexity and chat gbt become the new front door to the internet who are the losers in traditional digital marketing who fail to adopt this new discovery model do you have a prediction related to this

25:13Corey:not directly uh i think uh if if you're in traditional anything right now and and failing to adopt in any way, I think you're probably in trouble. And I would say that is not true across every industry because some industries have certain blockers. But, you know, they're essentially, and I think I mentioned this on LinkedIn last night, there are two approaches businesses take to AI when it comes to staff. One is touch it and you're fired. Two is more AI, more AI, more AI, more AI. Don't touch it and you're fired. don't touch it and you're fired and um the truth is we still need a human in the loop for most things in that 60 70 percent range but it takes a lot less people to do it and and i think the fact of the matter is that the companies who are taking the more ai more ai approach uh just by the fact of what it is and what it does sure they're going to stumble they're going to make some mistakes there'll be some embarrassing moments but i think they're going to steamroll those who are not.

26:21Grant:I think when you look at any industry,

26:24Corey:look at who's adapting heavy, who's already made the change, and look at who's still resisting and those who are still resisting. Honestly, as of January 2026, I think they're going to get steamrolled.

26:38Grant:Yeah, I also think, well, okay, my predictions are splitting off in three different directions. What you just said, we had a conversation with this with Alexander and Biriko from open AI. And I do think that there is a fast takeoff scenario where like the skills of working with AI compound on one another. And, you know, typically we talk about fast takeoff in the sense that like the AI gets better and better and it can create better AI and then the AI takes off. But I also think people who are using these tools, the more that they use the skills and the more that they get better at it, they have a skills takeoff.

27:12Grant:And I'm concerned about this personally, because we're writing about the news all the time and we're trying to keep up with what's going on, but we're not necessarily always in the tools working day in and day out. And I can feel behind a lot of the time. And I'm someone whose job is to learn about this stuff.

27:29Corey:It's literally our job to know what's happening all of the time and we both regularly feel behind.

27:36Grant:And it's really interesting.

27:40Corey:Hey, Ika, would you mind grabbing the questions as they come in and maybe throwing them into a dock for us? I'm thinking it might be easy to kind of run through what we have and then take some time and just walk through question by question everything that's there. Maybe it might keep us on track because I'm seeing quite a few come in that are really good, like defense and AI. And I have I have thoughts there that I would love to discuss in a few. but I'm afraid we'll miss them as we go through if we don't

28:13Grant:maybe let's try what do you think about trying to do a question then one of ours, then a question, then one of ours so we can keep the audience engaged

28:21Corey:that's a good idea by the way, watch this this is Ika, say hi Ika oh no, I'm on blast thanks for all you do Ika we appreciate you

28:33Grant:that's funny and I knew that would absolutely break her

28:36Corey:Well, let's talk frontier models Here's where I think this is going to get interesting My first, and definitely not the most controversial on my list Is that I think OpenAI drops a significantly better new frontier model by early Q2 I keep hearing Scuttle about it There have been some comments and rumors that people at OpenAI are kind of happy that people are underestimating a little bit right now. And the truth is 5.2's jump over 5.1 was probably the biggest jump we saw in 2025 in terms of capability benchmarks from one model to the next. And I expect what's next will be bigger than that. and that that's the reason why 5.2 was 5.2 and not 5.5.

29:33Corey:I think that they gave it a one-tenth of a point jump. Says a lot about where they think it is on the pecking order of what they have ready to go next.

29:46Grant:I think it might be worth also putting that in perspective with everything that we know that OpenAI is doing. I'm working on a piece related to this, but I haven't finalized it yet. Basically, we know that OpenAI is doing, they have their 100 fingers in 100 different pies on the product side, right?

30:05Corey:As they always do, yeah.

30:06Grant:Yeah, they just launched that.

30:07Corey:Part business.

30:08Grant:Yeah, yeah. Well, and also when you're selling like a whatever$1 trillion business, like when you're basically selling yourself as a big tech hyperscaler, you have to have hyperscale. So they have to have a lot of different product offerings that they're selling. But the two biggest things that we've heard about is voice, a new voice model and voice interface mentality coming into 2026.

30:32Corey:It's significantly better right now, too. Oh, yeah. I was screaming with it.

30:37Grant:Every day,

30:39Corey:like somewhere in the last month, month and a half, it's gotten an update because it stopped telling me, hey, thanks for reaching out. We're going to keep it chatty and clever, a little bit of humor, a dash of this. You know, it doesn't do that anymore. Now it talks.

30:52Grant:Right, right. Yeah, it's getting there. And I think the new version that they're working on is going to be even better. They also have a line of products, which maybe we'll see in 2026. Maybe we won't. It seems like we have a potential to see them this year because, you know, we actually got leaks on what they're going to be from the manufacturers who are going to be producing them. So in theory, if the manufacturers have a deal in place and it's actually in production, we're likely to see it this year not next year.

31:24Corey:I agree. People have been sleeping on voice. I'm telling you. It is such a great way to interact. I mean I even use Copilot on my computer and yell at it. Not yell at it but holler questions at it once in a while and it'll respond with what I need while I'm working and it's really handy. Yeah.

31:43Grant:Related to this, do you think voice is going to be as big as they say in the workplace and out in the world? It's hard because Because we kind of have to get into the mentality of I'm okay talking to myself all the time. And everyone has a preference. The truth is it may vary from person to person. Yeah.

32:00Corey:But would you want it to chime up in a meeting? Like you're in a meeting and it's like, oh, actually that number is this. What if it's just like a –

32:07Grant:Well, that's certainly coming. We talked to – We kind of do that. Yeah. Well, Otter CEO Sam Leong, we talked to him and they're working on that. They already have a voice assistant in meetings and it's going to be an active participant. But I digress. So I'm going back to the open AI point for a moment here. So we know that they're working on voice stuff. We know that they have something big in the pipeline for, you know, could be GPT 5.5, like Corey said. And we know they have devices.

32:38Corey:Or something completely different even. Yeah.

32:41Grant:And we also know that Jerry, I'm going to mispronounce his last name, but one of their main researchers just had a feud with Jacob Pachocchi. We don't know what it was about, but they had a disagreement about the research direction for OpenAI and he left. so whether that means they're doubling down on what they're doing now or they couldn't agree on which direction to go for something new um we know that there's internal disagreements there so that could mean number one maybe they do have a new direction maybe you know for for research and

33:13Corey:what comes next it's kind of interesting to think it is but putting all that in perspective i think

33:19Grant:a lot is going to be going towards this new device paradigm and a lot is going to be obviously staying relevant and and i think that i think that there's a there's a lot there's a lot more we can go there um but i i digress more into the interface side of the conversation so on the frontier yeah i think it's still going to be a three three-legged race between open ai uh google and anthropic and then xai is the fourth leg that could either come out and surpass everyone in a dramatic fashion or could just be the fourth leg that is constantly keeping up.

33:58Corey:Yeah. So I have a theory there. I have a theory there. And this is one of my predictions on my list. I think throughout the year, we can expect to see what we're used to, which is Google and OpenAI slapping each other every four weeks with something new to retake the top spot for the next six days or so as it seems to go every time. Right. And, but with that said, I think by the end of 26, the intelligence leaderboard looks a lot different. I think Google and OpenAI are still at or at least right near the top and batting back and forth. However, legacy tech is going to show up in 2026. I really, really believe that we're going to see legacy tech do some things like, for example, I think Microsoft's going to have a genuine contender for a frontier model and probably by Q3 at the latest.

34:59Corey:You know, I mean, when we spoke with Mustafa Suleiman earlier this year, you know, I mean, he even talked about their process was like, you know, yes, we're going to work with these models. We're going to work with these. We're going to build our own. We're going to try to do all of this Nokia AI. we're going to try to do all of that. And, you know, they've built out a really strong team that you don't hear a lot about. And they're a company with the dollars, with the existing infrastructure and data centers already in place. Like they don't have to go start from scratch. I think you will see them in the top five by the end of the year.

35:37Corey:Like I think it was, Was it MAI that cracked the top 10 briefly as an open source model?

35:42Grant:With either their image or their voice model.

35:46Corey:Okay.

35:46Grant:Not with their, well, maybe with their general MAI one. Yeah. Perhaps they did. And in addition to that, I want to add that it's not just Microsoft and legacy tech.

36:00Corey:I think you start to see Mark Zuckerberg's money paying off and Meta making some real moves. Meta bought Manus.

36:09Grant:See, I'm coming for you, Kevin. That's why we've been holding up.

36:13Corey:And they've bought Manus. It's pretty well known that they have a big model coming soon that is supposed to be significantly better than what they've done in the past and genuinely competitive.

36:29Corey:Jan Lacuna's gone, whose ideas were obviously at odds with...

36:34Grant:Gosh, and he torched some bridges on the map. Oh my gosh, he just keeps

36:37Corey:torching him doesn't he uh he is not a happy camper right now and we haven't covered that yet because i will i really wanted to like i really wanted to put it put it all all together

36:48Grant:um in one piece but we'll get to that basically for people who don't have context yen lacun has been talking smack about the new guy alexander wang and and you know what's

36:57Corey:happening over there now and i mean yeah it's uh i don't know it it felt a little unbecoming of a of of an academic well look he's a he's a startup founder now so gloves are off that's right that's right but the thing to know is meta has been building out you know solar fields to power data centers meta has been quietly building data centers around the country uh you know including a project or two that got killed right here in missouri where i'm at um they're not playing around you know and and if they can nail it they already have an ad business this that rivals google probably the closest competitor to google i would argue that's just me spitting off that's not based on numbers so it could be wrong there but facebook ads are huge man that's a big marketplace they're they're effective they're good they work uh they have Instagram they have WeChat you know is it WeChat that they have I could be wrong yes I think so yeah and no no

38:04Grant:WhatsApp WhatsApp WhatsApp thank you

38:07Corey:and I really feel like and and I heard it from a number of people in November and I heard it again in December keep your eye on Meta Meta's got some stuff cooking and I don't know what that means but they turned around and they bought Manus they've made a number of moves like there's there's Sam 3 their new 3d generation model is ridiculous like if you haven't Matthew Berman who's one of my favorite youtubers probably grants too uh and uh he did I think it may have even been a sponsored video with meta but demoing what it can do and it was really really cool uh you absolutely want to look at sam3 and and i think that plays a role in you know augmenting your lllm strategy i think they might also abandon open source but i think we see yeah i think that's pretty much foregone conclusion i mean they might still

39:01Grant:do it in the research realm um and perhaps they'll still do it while they're catching up but the the the thesis is you don't open source when you're ahead you open source when you're behind

39:13Corey:That's right. Yeah, you don't open source when you're in the lead. That's right.

39:17Grant:Unless you're open AI, but then you open source something that's not frontier level.

39:20Corey:It's not frontier level, yes. And, you know, that's what I really think. But I have one more that is going to be the abrasive one that people aren't going to like. Go for it. I think for the first time this year, a Chinese model tops the artificial intelligence leaderboard.

39:41Grant:I think that that's a foregone conclusion, but it might not be this year. It might not. I think if it's not Deep Seek v4, there's a chance it could be Moonshot. There's a chance it could be GLM. And there's a chance it could be Quinn because those are the three biggest ones.

40:04Corey:Don't underestimate the volume of money that Alibaba has.

40:08Grant:Yeah, no. Alibaba provides. It probably will come from Alibaba. world um but but we know that deep seek before is coming out and we know that it's competitive in coding as of the information's reporting um today so there's a chance that that could top the leaderboards but it could be for like a very brief moment it could yeah i i don't i don't think

40:28Corey:that they're going to top it and keep it it's really interesting because china has us just smashed in the power department they have the energy adversely we have more data centers and the best gpus yeah so there's this interesting divide and it's uh the truth is if you put those two things together god knows what level of ai we'd be looking at right now but uh it's interesting and it's an interesting dynamic uh that you know who knows where it's going to goes but i i think i think you see a legitimate contender out of china like in a larger sense

41:07Grant:than deep seek was so a report a report today from epoch just said that the um consistently since 2023 china has been about seven months behind the u.s yeah and that's been a consistent trend over the last we were worried that that uh rate was shrinking so like it was more like three months. It felt like more like three months last year. But according to Epoch today, they said, no, it's been consistently seven months behind the frontier. And we know that, you know, OpenAI has a new model cooking. We know that Google is probably working on something, though I haven't heard what the successor to Gemini 3 is going to be yet.

41:48Grant:But yeah, like Cal said in the comments, China can brute force anything into success. I mean, they just have the scale, right? So, So, you know, unless, you know, you really do need the most cutting edge NVIDIA GPUs to reach that next threshold. Yeah, I think perhaps they could they could break through and perhaps they could break through with bench maxing. You know, like maybe it's not technically superior, but, you know, on paper it is. And then people test it thoroughly and then they find, oh, it's actually not, you know, it's not good in these areas. It's not quite there.

42:21Corey:Yeah. Yeah.

42:23Grant:Allegedly, they were working on it over Christmas. So there could be a Claude, maybe not Claude 5, but a Claude 4.7.

42:33Corey:Anthropic would give us a 7, wouldn't they?

42:38Grant:At least give me the 7.5. Like, at least give me the halfway point, the official halfway point.

42:42Corey:Give me something that feels pleasing.

42:45Grant:Yeah. I do wonder when they're going to stop labeling them as you know, numbered releases. Perhaps never, but it would be nice.

42:56Corey:Yeah, you know, if it wasn't for the pace, the truth is we'd probably get into a cycle like the cell phone companies do. We're like, all right, new models drop every September from Apple. New models drop from so-and-so in May.

43:07Grant:Actually, speaking of Apple, so that's my controversial hot take in this area. I don't think that Apple is in the business of building frontier models. So I don't think they're actually going to have a frontier model. But on the UI side, I think they will finally show up in a big way this year. and it's going to change the game.

43:25Corey:That's kind of what I have on my sheet too, Grant, is that Apple will continue to not have anything for state-of-the-art,

43:37Corey:but we'll see small models begin to take a bigger role within their ecosystems.

43:42Grant:Yeah, I mean, well, we know that they're going to be using Google for some of the cloud-based processing. we also know that google has a series of small edge device models yes the name escapes me right now but you can access them through the edge gallery what's the what's the small one called do you remember oh god jimma yeah jimma yeah jimma yeah i remember jimma on my phone

44:11Corey:it worked on as old as an iphone 14 or 15 i could use it successfully yeah so so i you know

44:20Grant:perhaps they work perhaps part of their deal with apple is that they're you know run some gem models on device yeah um that that could be really cool um but but i think a lot of it will probably be you know offloading to gemini 3 um in the cloud i imagine and then you know we know open ai is working on a new voice ai presumably that's going to be used primarily for their devices since they're kind of competing with apple in the device care category though they say they're not they say it's complimentary to your iPhone, not to replace it. But perhaps their new voice model will be integrated in a more meaningful way on the Apple iPhone.

44:57Grant:I believe that's true.

44:59Corey:I think that's very possible. And it's absolutely related to, I'd just like to say, what is quite possibly my favorite meme of the year. And I'm trying to figure out a good way to show it.

45:13Grant:You can present your screen, maybe.

45:15Corey:I don't know. I was wondering if there's not an easy way in here to just upload a picture and show it. You can present a tab.

45:24Grant:It's probably the fastest way. While you're doing that, I want to address a couple of comments that people had where they were talking about topics. They asked about topics that we already covered. So a recent interview with Zuzana was really interesting. How do you think Pathway will develop this year? I think Pathway will definitely launch something, and they're going to launch something with AWS. So if we don't hear from them sooner, we will definitely be hearing from them around reInvent next year. So run the clock 11 months from now, we will definitely hear from them by then. Oh, yeah.

46:06Grant:So this was Apple paying Google$1 billion for Gemini, Google paying Apple$20 billion to stay default search engine.

46:13Corey:That was one of my absolute favorites this year that just cracked me up. I was like, yes, we'll give you$1 billion. Okay, that's it. I just wanted to share. By the way, I know how to do it now. When you go to screen share, there's an option to just share an image file.

46:29Grant:Oh, sweet. Which is really handy to know. Yeah.

46:34Corey:We're learning the tools, people. Bear with us.

46:39Grant:But yeah, I think Apple will definitely. that was one of my hot takes for this year is that Apple is going to show up in a meaningful way. And all of the memes that we made fun of them over the years, I'll always get one comment from someone. Anytime we make fun of Apple where they're like, well, you guys, you think you're so clever, but be nice to them. Like they're working on it. Maybe they'll do something cool this year. You know,

47:02Corey:something other than just publish research papers that dump on the rest of the industry.

47:08Grant:Yeah, exactly.

47:09Corey:Which they've done a couple of times.

47:11Grant:One more comment that we already addressed. So, Solarpunk asked our take on Yon LeCun and the world models. I think they will have a big relevance. I think what he's building, it depends on if he can raise the money that he's trying to raise. Some people are skeptical that his path is the right one. Without getting too in the weeds, his path revolves around JIPA and energy-based models and some of this other stuff that I am only slowly learning about and understanding. I think if he raises some money, I'll spend more time investigating them.

47:44Corey:I'll understand it better as soon as it's a little more real with some dollars. Yes, exactly.

47:50Grant:We've talked about it a bit before. There's lots of different versions and such. But yeah, we'll see if he's successful in raising the money as far as what he's doing. But we invited him on the show as well. So if you want to come on and give us your pitch, please do. Come see us. But again, excuse me. But anyway, yeah, world models themselves, I think, will show up in, number one, an industrial application. So how can you basically replace VR simulations with a world model that's lifelike and realistic? And number two, in a more entertaining fashion, which is perhaps generative video games where you can generate your own 30-minute playable movie.

48:34Yeah, I love that.

48:37Corey:How about if we take a bop down to agents and practical autonomy, Grant?

48:44Grant:Let's do it. I have a lot of thoughts here.

48:46Corey:I do too. I have three. How many do you have?

48:51Grant:I don't know if I have them listed like that.

48:53Corey:Okay. I'll ramble.

48:55Grant:You go first. I'll react to you. All right. I'll get my take.

48:57Corey:All right. My first one, and the one that I think is the biggest impact, is that I think agent building

49:03Grant:becomes a natural language process.

49:04Corey:I think you'll go in and you, and this, this kind of maybe ties to Manus too, because there are tools like Manus, like Opal, who have this to at least some extent, but I think it gets more manageable and easier to understand where, you know, you set up an account and it, it, maybe it walks you through how to go get an API key. Maybe it walks you through how to set up your tools really simply. And from there, you just tell it what you want an agent to do, and it starts building it. And as it's setting things up, it would be more complex. And like in an 8N situation, it just comes back and asks you, which of these options do you want?

49:39Corey:And explains what they mean. Makes it simple to where just about anybody could do it. And honestly, maybe some of those decisions, it just makes itself because it knows, okay, for this, it needs to be that. For this, it needs to be that. But where it would be a kind of a back and forth thing. But I really think it becomes a natural language thing and anyone can build an agent who wants to.

50:00Grant:And I'd argue it's close to that now. So I agree. But with the caveat that if that happens, it will have to be because there's an agent builder agent under the hood. And let me explain what I mean by that. So I think what this, I was talking to some people yesterday and this solidified my thinking on this. The harness is the most important part of the AI experience right now. And by harness, I mean all of the software that helps direct the models to do what they're trying to do. So when we're creating agents as individual users, are we creating our own harnesses or are we creating agents? And I think that there's really good harnesses.

50:44Grant:Cloud Code is having a moment right now. It's going viral for everything other than coding and coding. Like it's really, you know, if you haven't, if anyone hasn't tried Claude code yet, please go out and test it and play with it. Even if you're not an engineer, just use the ghosty terminal app. It's really easy to use there. Use factory. Factory has a web app. You can chat with Claude there. Really test it out because it is a, we talked to Eno from factory recently, and he solidified this as well, which is that a general purpose agent by necessity has to be the best coding agent because why are humans so great in our environment?

51:32Grant:He says it's because we use tools. Well, a software engineering agent in a software environment, the computer, is a creature that is perfectly designed and tuned to use tools in its environment. it. So if we're going to have a really great agent, that agent has to be really good at coding, which is why CloudCode is so good at everything that you can do on a computer, not just coding. So the reason that CloudCode is so good is the harness. It's whatever Anthropic did under the hood to set it up and make it really easy for you to use it. So the harness will be the most important part of agent building.

52:12Grant:So when you're creating your own agent, the best tool will have a really good agent harness to help you create agents really simply. That's my belief.

52:22Corey:I like that. I think it's a good call.

52:25Grant:And I think it's more important than models because you could have the best model in the world, but if you can't harness it and use it effectively, all of that capability is lost in the sauce. It's lost in the numbers and the parameters and the weights. You can't get it out. You can't get the information out of it.

52:45Corey:I have something else I think gets a moment this year or that I hope does. And that is computer use agents, which to me are the closest thing to real agentic AI that exists right now, honestly. Like, I call out ChatGPT's agent mode, but Claude has one too. ChatGPT's is quite good. Claude's has improved a lot, and I'm sure will be great with time. But people really sleep on these, and that is such a killer function. Like, I've done some neat stuff with letting that go run my computer this year. Like I had it scoured DMVs in the state of Missouri and find one with an appointment for 9 a.m. on a date because we couldn't make a trip if we didn't.

53:35Corey:And like it went, it found it, it popped up, let me book the appointment. And it turns out it was 12 minutes from my house. But it went searching things like that while I sat back and watched Stranger Things, which I think was delightful. And nobody uses it, which is crazy to me. And I don't know if it's because, you know, there's a 40 use limit per month or something unless you're in the pro plan. Because it is compute expensive. And I think that's maybe why it doesn't get the promotion that it should.

54:06Grant:It's also slow, which is a bigger problem. But it's like the best way to get around like a lot of websites that block.

54:13Corey:Yeah, it's faster than it used to be. And it's improved along the way. But I feel like, you know, it's not going out and doing it at lightning speed. It's almost doing it. It's doing it at a speed probably a little slower than you could. But if it's I mean, I don't want to go to the DMV so bad. I don't even want to go scour their websites like, you know, so it's like go do this. So I also every morning I have it go search the news. Go go find AI news today. And it finds different articles than Google News. Google News finds it. So we like we use a bunch of different sources to find news and know what's going on in the AI space.

54:52Corey:And every approach we use seems to always come back with different things. And that's one of those two. And I think, you know, maybe we get beyond these kind of brittle workflows that are slower and find a way to make that sort of thing more efficient. Because like I would love for it to work like that. I mean, deep research will do some of the same thing, but it's different. And it feels like, to me, it's not agentic AI until I can say, hey, go order me a pizza. And I don't have to sit down and program a bot to do it. I need to be able to say, you know, agentic AI can go and do tasks on command.

55:33Grant:So this is where I'm wondering about the harness versus the agent versus the agent builder. because if you, Corey, wanted to make your own agent to go do something, do you want to control how that agent behaves? Do you want to set up all of the parameters of what it can do and what it can't do, all the guardrails? How much control do you want over it, or do you just want the end result of what you want to do?

56:01Corey:See, now, in business applications, they're absolutely used for that kind of agent. But to me, agentic AI was never, let's go build a specific workflow that will do this thing.

56:11Grant:And then it knows it's, you know, it's hyper specific goals.

56:14Corey:Agentic AI should be able to just do what I ask it in a very general way. And that means if what I need it to do is gather information for a report, cool. It'll go gather information for a report. If that's order my lunch, that's order my lunch. If that's, you know, help me plan out a new diet or a new something or other, it knows what it knows about me because it has access to my account memory. And then it just goes and it does that. And it's like, hey, you needed this. You needed that. We should probably order you one of these whoops or something, you know. And here it is. If you're down, go for it.

56:53Corey:And but you don't have to build it for every function you want. Like what we're doing right now, whatever you want it to do, you've got to go build it. And to me, that's not the same.

57:06Grant:Yeah. I mean, it's interesting because, you know, as in our work environments, we do want to be able to build agents to go do things for us to help us do work more efficiently or help us increase what we can do. But, yeah, for like the for the consumer use case, yeah, you just want the you just want the thing to work. I want it to work. So you don't want to build it. You actually just want one that already works. That makes me think that continual learning and memory is the trick to having an agent actually do what you want.

57:36Corey:Yeah, and the truth is until marketing teams got a hold of the term AI agent in early 2024, that's what agentic AI was.

57:45Grant:They made agents feel like a scam, basically.

57:49Corey:Yeah, yeah. They basically made agents into a scam. And in my opinion, a lot of what we're doing now is, you know, yeah, it can make decisions, but it's very guardrailed. It's very engineered. It's very, you know, practical, broadly useful agentic AI needs to be general. Yeah. Like, you know, consumers don't have a reason to – let me rephrase that. They probably have reasons they could justify using automations the way we are now. But, frankly, we're building automations and calling them agents in a lot of ways. Thank you.

58:23Grant:Um, what's your, somebody's going to stab me for that. I can feel it. What's your third one in this category? Because there's so many things I want to react to and what you said, but yeah.

58:33Corey:Yeah. Oh, here's a question on that too. Somebody's like finding DMV appointments or new articles different. How much do you want an agent to use your bank account or personal information?

58:41Grant:Honestly, I want it to be able to, I genuinely want it to have, you know, access to maybe it's an Apple.

58:48Corey:Well, I didn't want to put my credit card information into a chat, But I do think that there's going to have to be a way for some sort of secured wallet system to work within an agent in a way that is safe, that is passable from the agent to a website, just like it is when we use Apple Wallet and click Apple Pay on Little Caesars, for example. They're definitely very different things. But I want to be able to – I mean, if I tell it to go buy something, I want it to be able to go buy something. Like I know what pizza costs If I say order me a pizza Like hey get me a thick crust sausage pizza Delivered to the house I want it there about seven I want to be able to say those words And then go back about my business And do what I'm doing And then get a ding when my doorbell rings And maybe that's expecting a lot I don't know Would I let an agent do my taxes?

59:42Corey:Only if it had loose morals Cal

59:48Grant:only if it can get me the biggest discount possible only if i'm doing my taxes you mean it's gonna help me pay zero taxes yeah yeah actually i'll be honest i i kind of look forward to the idea of ai helping with the taxes a little more it's gotta match the boss i uh you know i well just because that's what cory wants nine design doesn't mean that that's possible today or what cory does like no that's how it should work right it should be trustworthy enough that you can do that and speaking for myself i don't feel that it's trustworthy enough where i trusted to do that maybe once i see not today many people doing it uh a lot then i might trust it or if they or if when they show it off and they release it they make it as easy to understand but this is an active problem that they're working on.

1:00:41Grant:Like there was a report in the information today that said that OpenAI, Stripe, and Shopify are really struggling with getting the shopping side of this to work the way that they're marketing it because it's a complicated problem. There's a lot of reasons for that. Getting all the vendor information to show up correctly, all of the privacy issues we talked about. I don't know if they're struggling with that. I think they focused on that first and they do have some benefits there. But personally, I wouldn't want a cloud-based agent with my credit card. I would rather that be a local phone agent.

1:01:16Corey:I would like to add that as a guy who remembers the early internet, I remember this exact discussion that there is no way in hell I'm putting my credit card on a website. Have you lost your mind? I am never buying anything over the internet and now it's the only way you buy yeah and to web's to web's point yeah you do you

1:01:40Grant:do i do trust it to go gather documents now now i i had this conversation yesterday the reason that i don't or i i think of it in the same way i think of security and web research in the same way where i can trust it to go out and find something but i don't at this point trust the ai to go out and and search everything and come back with the right thing always, which is why you always have to do your own research. But even if you're doing your own research, you're still only looking at the top page of Google. Yeah. But at least you can be discerning and say, okay, if it's on the top page of Google, I can trust this or a source I trust this up-to-date information.

1:02:18Grant:You have a little bit more agency in the process so you can make the call. But yeah, I want a web research agent and I would want a shopping agent to do the same thing where I know that it's looked at, gosh dang, everything before it makes a decision. Like I want it to like, cause I can't do that. So I want it to be better than me at making shopping decisions. Look at everything, come back to me, give me the best answer. Yeah. At this point, it's not efficient for them to do that.

1:02:47Corey:And to be clear, I'm talking about the difference in what I expect to be the case at some point and what's today. Like what's today? I, you know, I'm not typing a credit card number into a chat interface to be clear, but, you know, I mean, assuming there's a good enough secure process, I'm down.

1:03:03Grant:You know, for example, if that can do a tool call to your Stripe account where you're. Yeah.

1:03:08Corey:Or your demo or whatever. I mean, a simple tool call to those. And that's quite secure already. Your credit card number is not getting anywhere near the inside of their database or anything they have access to. It's connecting directly to the vendor. uh now with that said i want to by god know what it's buying and i want to trust it i don't want it going out there and you know going like you know claude vending machine on me where it ordered like what did it order that was so ridiculous that was so funny oh oh um i i don't remember off the top of my head but yeah sold well and it ordered like 5 000 of something i can't remember what it was one of your members drop it in the chat because it was it was it was funny yeah

1:03:42Grant:the vending bench is what we're talking about where where it failed pretty hard yeah yeah because Because it ordered items for the vending machine, but something nobody wanted. So it ordered like 5 ,000 of those.

1:03:54Corey:Yeah,

1:03:56Grant:exactly. But that's where I think continual learning and memory is going to be a big part of this. And I don't know. Do we have a section on this, Corey? If not, I'll share my thoughts on that now.

1:04:05Corey:On learning and such?

1:04:06Grant:Learning and memory. Do you have a prediction on that?

1:04:08Corey:I do. Okay. Well, I'll save my thoughts for that. I say that, but I'm not seeing it in here. Okay. I will say that as far as learning and memory go, I think that's – oh, I do. I do. Right here. AI memory improves substantially, removing one of the biggest bottlenecks in agentic usefulness. One major lab cracks it first. The others all adapt and do the same thing, making persistent context and long-term continuity far more reliable.

1:04:44Grant:I've heard rumors that Anthropic has a solution here. They technically introduced memory, but as we talked about with Zuzana, when the model companies right now do memory, they're adding it on top of the prompts that you're sending, or it's in the system instructions, or it's strategically deployed. opposed to connections to other data yeah yeah it's not true memory and i'm there's a lot of companies that have plug-in solutions where you know you're creating documents and markdown files that like effectively is rag or effectively is uh a memory solution but again i think that's a that's a short-term solution memory has to be solved at the model info at the model architecture level um and right now they're solving at the harness level so i don't know if anthropic solution is going to be at the model level or at the harness level.

1:05:37Grant:But I think they have something. And I think one of these labs is going to produce something that's closer or better.

1:05:45Corey:Like, judgmentally, cross-account, like, full-account memory works very well for me. Like, it's very good at referencing, you know, my preferences and things from the past that I've discussed. The one thing I hate is that everything I say might go into it And it keeps things in memory that I would really rather not be there. So maybe the answer is like, you know, you need to manage the memory in chat GPT, for example. Otherwise, it's kind of like, you know, when you think of algorithms, how you say something out loud and then, you know, you're getting videos for it later. Like, oh, I don't want to search for that.

1:06:22Corey:It's going to be in my algorithm. Now I'm going to see videos for this stupid thing all the time. And I think that that may be a similar concern with current memory in AI is that if I ask this question, it's going to always want to tell me, include this context. I mean, maybe the answer is getting in the habit of going and managing them, but I really think there will be much.

1:06:44Grant:It's in context memory. So it's when I am dealing with this specific task in this specific scenario, use this specific information. At this point, it doesn't work that way. It's like a catch-all where it either is never used and never relevant or it's like overused and made over relevant.

1:07:03Corey:Yeah. Yeah, I think you're right. It's, you know, like I said, it's like all things. It's not quite there. But it feels like we're in the realm of solvable.

1:07:14Grant:Yeah, I think there will be a solution for this. And whether it's a harness, I think there will be a harness solution that probably feels much better and much closer to solve than today. And I think there could be a model solution this year as well. We've certainly seen ideas there.

1:07:31Corey:And Pathway's solution.

1:07:33Grant:I think Pathway's solution is really clever. We'll see if it scales well.

1:07:39Corey:Let's talk interfaces. Yes. experience i have i have strong opinions here um i think the ai interface is the battleground of 2026 i think finding ways for companies to differentiate from one another whether that is something like memory context control speed integration with your other workflows user experience more than just intelligence like they're all pretty good now like the fact of the matter is we have preferences. So-and-so likes Claw. So-and-so likes Chachubiti. So-and-so likes Gemini. But 95 % of what you want is just fine in any of those if we're being real with each other. I mean, like if you're a developer, you're probably going to have better luck with one.

1:08:25Corey:If you're a writer, you're going to have better luck with another. So, I mean, there are preferences in areas where they'd be best, but they're all good. And I think really starting to differentiate the workspace is going to be important. And right now, and this is going to be really unpopular with some folks, the Copilot MS 365 interface is really cool. All of your software is right there. Copilot is right there. You log in to Copilot as the front. But your Word's right there. Your Excel's right there. Everything is right here and just a click away uh whether you don't like it at all is fine but the truth is uh yeah yeah i done broke cal copilot is the take you're pro copilot and and you know you know the truth of the matter is when it comes to copilot now it's it's chat gpt and claude i mean you know the same models you love are there it's just in an interface that's pretty easy to use easy to understand yeah uh and uh but i really believe that that their interface is is good i'm gonna take the under on this okay copilot is cooked in 2026 maybe here's why here's why it goes to my it goes to

1:09:54Grant:two of my other predictions so one of my predictions is that generative ui shows up in a big way in in 2026 generative ui yeah i've been talking about generative ai with chat gpt uh since 2023 and i'm sure sam altman read my chats at one point because he started talking about generative ui this year last year um but basically generative ui is the idea that um you know you don't have a fixed interface like we're watching this on youtube youtube looks a certain way because a bunch of people at google and youtube designed it to look the way that it is generative ui is youtube everything is a back end everything is just data um and api calls which is for people don't know api call is like connecting two pieces of software together to do a thing um so in a generative ui world you could say like hey i want to watch youtube but i want it to all look retro 50s and I want to add these features to it and you being able to actually edit the YouTube interface and it adds those things.

1:10:55Grant:And not only does it change the look of it, because you can change the HTML on your page right now to make it look however you want, but adding functions and features to the tool as well. So I think that is going to be a bigger thing. I don't know who is going to roll it out. I don't know when they're going to roll it out, but we're at the point basically where we

1:11:14Corey:could do that like a widget based interface kind of not not you know where you could like there'll

1:11:20Grant:be guardrails you can't add adult content to the website or whatever but like whatever the whatever the guardrails are you know they'll be guardrails like hey you can't do this or that thing but like i think that that is going to happen and if that happens um i think that the need for co-pilot starts to go down and the reason for that is you know you have word you have excel you have um presentations uh whatever else you can do all of those things in other platforms you just need yeah you just need like one pane of glass like i mean copilot is trying to do this and to your point this is what copilot does better than google right now which is you have one pane of glass where all of your stuff is and you can just work in the where google has 179 websites 460

1:12:09Corey:62 products and none of them, none of them integrate with each other.

1:12:14Grant:Yeah. Yeah. It's a problem. It's a problem of scale. Yeah. Yeah.

1:12:19Corey:You know, I would, something I think though is almost essentially what I feel like you're talking about is an agentic operating system, you know, almost an agentic OS and, and that's another thing where Apple should be discussed because the fact of the matter is we're an awful lot closer to a fully agentic windows than we are to a mac os right now yeah you know i mean like like if you're dealing with like early beta versions of windows right now it's it's it's it's it's all intertwined together in a in a really interesting uh way and and i think it's only going to get better but apple's behind there and i've been i was a A guy who was an Apple guy for many, many, many, many, many, many years.

1:13:08Grant:Because as a journalist, that was the way of the world.

1:13:11Corey:And I still love them. I just don't currently have one. Okay.

1:13:16Grant:Ankit, tell me how I'm pronouncing that correctly, just suggested something really, really smart, which is UI predictability is really important for user experience. This is where the phrase, the devils and the defaults, comes into play. Right? I think the default will still be the default experience, but I think the ability to change your default experience when needed will probably play a bigger role in how we experience websites. And I don't know, like people are going to change kicking and screaming because you've obviously optimized the default to work a certain way. And for a lot of companies, that's to try and maximize engagement with the tool, right?

1:13:59Grant:For the user's benefit, it's like, let's make the tool as useful as possible, but also for the company's benefit. Let's, you know, create lock-in. Let's make sure people are addicted to our tools and using them every day. And I do think that we expect a certain level of predictability. When I go to Google Docs, I know it's going to look a certain way. It's going to have certain features. Same with any other tool that you use. So I do agree with that. And I do think that there will still be a default interface. But we're seeing more and more, like take Cloud Code for example. Cloud Code for people who use it, it has all of the functions that they need inherently in it to use anything on your computer eventually.

1:14:43Grant:The one thing Cloud Code doesn't have is an interface. so I saw and this is going viral on x somebody adding a canvas window to quad code so when you're booking a plane it shows you it creates an image on the right hand column that shows you like all of the seats that are available with like little green boxes and little red boxes or if it's drafting an email for you it shows the email and an interface so if you have a tool that can do all of the functionality that you need and you can create interfaces on demand i think that you're gonna your default behavior is going to go to the most powerful tool and create interfaces and even save those interfaces to your default preferences you can even say like for example you could say make this exactly like microsoft word and then it will create something that's you know maybe not 100 % feature complete for IP reasons, although it'd be pretty ironic if any company in big tech tried to sue someone over IP usage these days.

1:15:44Grant:But you could set your default preferences that way. And that's kind of how I imagine generative UI working. You could change the defaults on the websites. I find that's probably going to be a kicking and screaming kind of component where you're going to have to drag people away from their default preferences. But I think you'll be able to recreate interfaces that you love in the tool that's the most powerful.

1:16:10Corey:I love that. I think that's really interesting. And oh, I have another one on this, on interfaces and product experience. And I touched on it earlier, but I want to drop it in. Go for it. AI features become deeply embedded in consumer platforms. So earlier we talked about hardware, but I'm talking about streaming services, podcast discovery, music recommendations in your games in a lot of ways that maybe aren't yet. Like right now, when it comes to AI integration in like various streaming platforms, that area still feels like it's a little bit in the, what's the phrase here? A little bit in the let's bolt a chatbot on it era.

1:17:03Corey:that SaaS is finally starting to pass, thank God. And I think it's going to be incremental, but I think by the end of this year, what we're doing with AI inside of, let's say, Netflix, inside of your banking app, for example, are going to be very different from what they are today.

1:17:26Grant:Yes, it is Luddite, But also a let I take every now and then is totally valid. Absolutely. Absolutely. Like, yeah, I don't know. I mean, this goes back to my idea about the marketing, right? Like, I don't think people want to hear AI, AI, AI all the time. And I know that's ironic from a brand that is all about AI. But I think like people just want things that work better and are good.

1:17:52Corey:Yeah.

1:17:53Grant:So I think putting the AI in the background makes a lot of sense as long as it's reliable.

1:18:00Corey:Yeah. Yeah, I think so too. I think so too.

1:18:06Grant:Yeah, AI fatigue is definitely a thing.

1:18:09Corey:Let's see. Let's see. Some of these, because we've gotten so excited and a little ahead of ourselves a couple times, I'm looking through my list and being like, oh, I kind of already talked about that. I don't want to bring that up separately.

1:18:19Grant:I want to talk about another one that's related to this, if you don't mind. Yeah, please, please go ahead. Okay, so in the same spirit of generative UI and in the same spirit of not bolting a chatbot into everything, I don't know exactly how this works, but we're going to see the rest. I've heard this prediction multiple places, so I can't take full credit for this, but we're going to see the rest of the business world have its cloud code moment with a general purpose agent that user experience-wise feels like working with CloudCode, but for whatever business tasks you need. That's sort of also why I'm like under on Copilot is because I think if there's a company that does this really well, you're going to see companies be like, oh my gosh, we don't even need this anymore.

1:19:09They'll still use it for, you know,

1:19:13Grant:obviously all the lock-in with Teams and email and whatever else companies need it for, Azure, et cetera. but yeah I think that I think that we're gonna see a general purpose agent and this is to the Manus point I don't know if our Manus our Manus question is still in the chat but but Facebook bought Manus for a reason Manus was like a really good agent harness that can do a lot of things and so they're going to try and build that kind of functionality into all of the Facebook tools that you use so what what does the alternate user experience look like that's not a chat bot version but but works agentically like cloud code because cloud code is still a chat bot but then it's also going off and doing things for you it's interesting uh favorite meme number two let's see it let's see oh this one's a little blurry it's not a great picture but i'm gonna share it anyway

1:20:23Corey:image file bam this

1:20:30Corey:mark zuckerberg and alexander wang man this is pretty good can you buy it for me on it done did you buy pro or premium i thought that was really funny yeah that's amazing

1:20:40Grant:That's great. So related to this, I was watching Dan Shipper's interview with Reid Hoffman recently. It's really great. I'm going to include that. It's part of an AI predictions roundup that we're going to do after we finish this episode. Dan was talking about how every new piece of software he's creating this year is basically cloud code in a trench coat, where it's the same features of cloud code, the same kind of experience but it's you in your chat and or it's like in their tools right and so i do think that the user experience could be similar to that it's like the same experience of chatting with agents having the agents doing everything for you happening in the background you're checking in on their work you can see everything they're doing you know maybe you're giving permission for things maybe you're not maybe you walk away from the computer for five minutes and come back and see what it's created for you.

1:21:36Grant:Maybe you leave it for an hour or two hours and see what it's created for you. So part of that experience is trusting an AI to be able to go off and do something for you autonomously. I think that that business experience is going to happen in a big way for other people in other industries like finance or imagine if you could just walk away from your computer for like 15 minutes and all of your spreadsheet analysis is done. You can kind of do that with chat dbt uh agent now but but not really um and then you know for the existing products we know they will feel more agentic where it'll feel more like hey the thing's going to go off and do the thing for you and then it's going to bring you in when you want to be brought in but uh it's not like a chat bot that you like are going back and forth with it's more like a here's the task i'm trying to do and then you can watch it work or you can walk away um and just seeing that experience explode everywhere, I think will, will undoubtedly happen this year because the success of cloud code has just been so monumental that people are going to copy it.

1:22:41Corey:It has, it has, uh, cloud code has been big and my gosh, I don't know what's happened in the last three weeks, but just suddenly it's all I hear about. I don't know what that is. Cal mentioned it earlier too, that his algorithm is frigging cloud code propaganda. Uh,

1:23:01Grant:is what, what, What's what propaganda?

1:23:03Corey:Cal said, my algorithm keeps throwing me cloud code propaganda.

1:23:07Grant:It's because people are waking up to it. And then there's a viral moment where, you know, obviously people want to talk about it because it's trending. But there is something to it. And if you haven't tried it yet, you definitely should. Not to continue the cloud code psyop. But yeah. Do I have another take on this? No, I think I covered my...

1:23:32Corey:my angles there yeah okay yeah most people did just discover it recently i agree um

1:23:39Grant:yeah general gpt users yeah and then when that happens that then like it's like oh my god it's new um someone had an interesting question about nano banana earlier can we bring that up i did have thoughts on this oh okay

1:23:57Grant:where was that

1:24:00Corey:I have thoughts on it too

1:24:01Grant:maybe let's go to yours and then I'll share mine it's related to GenUI which is why I'm thinking of it now

1:24:09Corey:oh wait I don't know what you're talking about now

1:24:12Grant:okay so somebody asked if Nano Banana Pro's image model or vision model is going to get better well sorry image model not vision model is going to get better this year and I've been thinking about this a lot like every time there's a new image model that shows up pretty much every image model is like so good now I I can't imagine what would actually make the quality of the images better for me it's all about the control how controllable is it like do I feel like I'm actually I can actually create and ask for and get what I want and from an actual image quality standpoint, I think without question, the current level is there.

1:24:51Grant:I find like, what are you going to do? Make it like 10K resolution?

1:24:56Corey:Well, I can show you something interesting. Yesterday morning, I was trying to make an infographic and I went to, I developed kind of a full list

1:25:08Grant:of what I wanted and kind of like a design work order.

1:25:13Corey:and I took it to both Nano Banana and to ChatGPT Image. And let me show you each of these real quick the other day because they both fell pathetic, not pathetically short. They both got really close but didn't get there. Here is the ChatGPT Image, which looked really cool. you know i'm stuck to the width it gave me the image models don't do a great job of like make it more narrow none of them do that really well yet in my opinion okay so better dimensions durability right yes they need better dimensions durability really bad uh most of this is exactly what i asked for it even went and found the tech republic logo this is going with my ai adoption in the enterprise report that just came out this week we can also drop that in the chat um And it did really good.

1:26:12Corey:But when you get down to around here, there's weird things. There's like a misspelled word. There was something here that didn't make any sense. And the small text went nuts. And I went back and I tried to edit it and it didn't work for squat. It didn't like here. It was I can't remember what the word is right there. And it's too small for me to see on my screen. but it wouldn't make a change here. It wouldn't make a change. Actually, it did make a change here for me later, but, and this, I got about 75 % better, but never good enough. I was like, I would publish this. And now I did. Yeah. Every one of those sources are real.

1:26:50Corey:I provided the sources. I saw Cal ask that real quick. Now, after that, I also went and did the same thing in nano banana. and it did this. For the life of me, I could not get it to present it in any format other than this. I couldn't get it to remove the smaller copy in each of these cards. I could not get it to make this a vertical image. I could not get it to change the text, make the header more bold. they both failed on all of this. With that said, they both got me to like 90%. If this was able to be downloaded as like a Photoshop document, as a PSD, I could have made either one of these good to go in 10 minutes.

1:27:48Grant:Yeah, I mean, I guess that's bullish for Adobe, right? Because if you could use Nano Banana inside Adobe and then you could break out all the layers. I really liked the Quen image layered came out with this. I've been talking about this for years that they need to make it so that each aspect of the image is like its own layer so you can edit it. And Adobe's been making this possible with some of the UI experiences that they've changed. But yeah, basically, yeah, the built-in functionality, I guess this is where image models need a harness, right? Like you can create the raw image file fine, but you need the harness to be able to like layer it and edit it directly in the app that you're using.

1:28:32Grant:Because otherwise then you have to be combining multiple tools together and it's really clunky. Exactly.

1:28:37Corey:Exactly. Big shout out here for our good friend from data science with Dennis.

1:28:42Grant:AI powered flying chickens. Yeah. If we're talking about drones, we do have some predictions on that.

1:28:51Corey:Dennis is a good friend and you should absolutely go check out his channel. Yeah.

1:28:54Grant:Yeah, check out Data Science with Dennis.

1:28:56Corey:Yes, he covers all things data science and BI. He's very, very knowledgeable. He's worked in the industry for years, has all the experience. And one of these days here in a few weeks, probably, we were talking last night about having him on for an episode where we could talk about things like predictive analytics. And the truth is there's a big intersection between data science BI and AI. like they kind of go together like peas and carrots in a lot of ways. So we thought there's definitely a good discussion to have there. And plus, Dennis is just a fun guy. We met Dennis at Microsoft Ignite back in November and hung out and had a lot of fun together.

1:29:39Grant:We might not be going next year after I called the under of co-pilot.

1:29:43Corey:Yeah, after you called in. Well, maybe it balances out since I gave it the over.

1:29:46Grant:Yeah, exactly. Exactly. Corey gets to go. I'll stay home.

1:29:51Corey:I get to go and they don't invite Grant.

1:29:52Grant:No, I say that, though, because, like, look, all of these legacy tech companies are on notice, right? Sam Altman called it. This is the era of fast fashion SaaS. What that means is that as these coding models get even better than they are now, I mean, even where they are today, you're going to be able to create anything you want with software. Like, it's going to happen. And if it's not possible now, it'll be possible probably by the end of this year, I'm imagining. Now, will it be really good? Will it be really robust? Will it be at the level of something that people have worked on for 20 years?

1:30:26Grant:Probably not. But can you create essentially the same thing for your needs? You probably can. Yeah. So then what what do these companies do when that's the case? They have to create their experience that much better and they have to make it that much easier to use, that much cheaper, potentially. Maybe not. Maybe they don't lower their prices, but, you know, they have to add that much more on top of what they already offer so that they're staying ahead of this capability that people are creating.

1:30:54Corey:Yeah. Yeah, I agree. I agree. You know, one caveat we also ought to add here to every single one of these predictions, especially the ones where we've said by the end of this year, you know, the interesting thing about the AI space is that it moves fast and weird sometimes. So if any of these happen to be like Tuesday, don't be shocked.

1:31:17Grant:Yeah, but end of this year could be by the end of January, man.

1:31:21Corey:Or 2030. That's the other option. Some of these could just as easily be way out. But I really don't think they are. And I think most of these we've hit are pretty solid. I have one I was going to lead into right after we talked cloud code. And that's that I think Anthropic further solidifies its position as the developer engineer model of choice. I really think that they've found their niche there. and that community has dug in around them quite well and will support just about whatever they do. I'm not saying that to say it's not worth that. I'm just saying that I feel like that's a niche where they have really, really dominated and I think will continue to.

1:32:07Corey:I think Codex will continue to get better. I think the other companies will also get better and that there will be people who use those but that Anthropic will still likely maintain its lead in that space for a long time to come.

1:32:22Grant:Yeah, it depends on how much of a lead they actually have. Because I hear from engineers that this was like a step bump can change better, Claude Opus, than everything that's out there. And that was weeks, a week after Gemini 3 came out and everyone was like, okay, Gemini 3 is actually possible, like is at the level of Frontier. It's better than OpenAI. And I released something right after that kind of put that in contention. But it depends on how strong that leap actually is. I would love to talk to Anthropic and actually learn what they did differently.

1:32:57Grant:Love that. I don't know how to pronounce that, but it's hilarious. Yeah, people have made many versions of TARS, by the way. It's pretty cool. That is cool.

1:33:09Corey:do you think anthropic kind of holds on to that lead for a while or do you think

1:33:14Grant:do you think if one of the others comes out with a tool that is 10 better that people abandon it jump ship go to what's new and flash all of this comment from onket like that like this is that i

1:33:30Corey:think it's this that's anthropics biggest problem yes i think it's i think they're gonna get

1:33:36Grant:financial pressure especially as they start to go public to uh either raise their prices or um you know i i already saw today you know rumblings that they're they want to prevent people from just offering cloud code uh in a trench coat right like because it's the the 200 max plan is very expensive for for them to run if you use it um theirs is more than 200 isn't it uh i think they might have like a$300 version now. I thought that was thinking that

1:34:05Corey:Jagupt came out with Pro at 2 and I was thinking Anthropic came out with either$250 or$3. I couldn't remember.

1:34:11Grant:No, Gemini has$250, maybe even$350 at the max or the ultra tier. But Anthropic has a tiered system where it's$100 max plan and then you can do a$200 max plan and perhaps there's even a$300 one. I don't know. Okay. I thought I heard somebody launched a new price here recently, but I never saw any links about it. but I could see them going away with that and completely killing it once the adoption's there. And if that happens, it's just too expensive. It would be like thousands of dollars a month you'd be paying for this coding capability. It's not going to be effective. So you can have the best coding model in the world but if you have the best cheapest model like the Pareto Frontier on cost, that's going to win because engineers are nothing if not practical.

1:34:58Grant:and you know cost will be the biggest factor in that case an open source Chinese model will win or XAI will win because XAI can burn more money than Anthropic can

1:35:15Corey:yeah that's true the one advantage to being the richest man in the world I like that like there's one they're all the advantage all of it is the advantage yeah but uh is that yeah you can hemorrhage money for a good long while yeah yeah you know and uh and in the interest of you know building i don't i don't mean that like even carelessly hemorrhaging but i mean oh yeah yeah i mean you could argue with x perhaps yeah yeah and i think uh i think it's really worth keeping an eye on them because as much as as much as they've been divisive models they're not bad they have concerns uh but they're not um it is decent you know i think they're the ones to let me put it this way i would say

1:36:11Grant:anthropic is the one to be for the general purpose agent i think chat to be the agent but that has not that does not have the same adoption and somebody asked me recently what the difference why one is better than the other. And I think because using collage code, even though it does a lot more for you, it feels much more active. It's like you get these little snippets, these little tiny dopamine hits of like, ooh, it's doing a thing. Ooh, it's doing another thing. Ooh, where you're almost addicted to watching it and you can't leave. But then - And it's like, why am I not going

1:36:43Corey:and doing something productive? That passes with time. As a person who uses it every day, after the first week or so, I just leave it up open on one of my other monitors and occasionally glance over to see, has it popped up a window for me to click something or does it need anything? Because it'll come back with needs, which I think is exactly what they should.

1:37:04Grant:Yeah, but I think the experience of using it and the fact that it can manipulate files on your computer and you can create this whole memory ecosystem that's like local. I think that is the true power there. And I think that they're going to have an edge there for a while, even if they don't have the best or cheapest or Pareto Frontier coding model. I did see that. Do you want to pull up that comment?

1:37:26Corey:Yeah, I have some thoughts on this. I was holding my finger over it to bring it back.

1:37:30Grant:I was going to say I cannot comment.

1:37:33Corey:I can. I sat in on a couple of panels during AWS where these things were discussed. AI, you know, what are your insights for AI and banking? What I would say is that the thing that has blocked banking is the same thing that has blocked healthcare, that has been an issue for personally identifying information. And it is, you know, this idea that you need a few things in certain fields. Like banking is a field where precision counts. Like, you know, in areas, other areas, we can have a little bit of an error, a little fudge room, but these are fields where exact accuracy is vital. And there are lots of enterprise AI apps, right, or companies right now, as well as some others who are heavily, heavily, heavily focused on building what they're calling trustworthy AI.

1:38:31Corey:The idea being it's going through, it's grounded in your data somehow, it's grounded by search. It's grounded in other things. It's going through multiple verification steps that are even hardware-based in some instances, I believe. I really think that those industries where that precision is necessary are about to have their moment because it's less even about

1:38:59Grant:this need for

1:39:04Grant:the models themselves to get better. What they're doing is they're building

1:39:09Corey:the ecosystem around it to force the accuracy is essentially what's happening. And all in all, they seem to be having a lot of luck with it. I keep hearing about companies adopting things like that now. It's very much still early, I would say. But in 2026, I think you'll see a lot of big banks bring it on.

1:39:32Grant:Okay, we're having a sidebar here. I'm going to bring this in. So Cal said, 90 % of business and markets are held together by banking and banking processes. Ergo, the precision and accuracy needed applies to business as a whole. And I said, solve banking, solve business. So if they can make it verifiable enough and trustworthy enough that banks will do it, you basically have captured the whole market, the whole global economy.

1:40:00Corey:Yeah, if banks will use it, they're absolutely the last.

1:40:04Grant:Yeah. Aristotle and Axiom Math are another company that I thought of as you were talking, Corey, that would help with this. these are companies that are using lean as a programming language actually i don't know if aristotle is using lean but both of them are using verifiable math proofs essentially or

1:40:25Corey:math of formulas in order to to make ai that is extremely trustworthy with calculations so

1:40:32Grant:i can see that being like yeah it's maybe you don't trust chat gpt with your bank but maybe you trust Aristotle or Axiom or some of these more purpose-built models for it.

1:40:44Corey:So I don't know if it's like, no, no, I was just, I was just going to reiterate. I don't know if it's going to be Chattu Ruti, Google or, or, or Anthropic that wins there

1:40:53Grant:could be something more purpose-built. That leads perfectly into my next prediction, which is that industry incumbents will increasingly become the AI leaders in their own spaces.

1:41:06Corey:companies that already dominate workflows and we've seen this with adobe we've seen this with

1:41:11Grant:thompson reuters uh will leverage the domain expertise they have their existing distribution

1:41:17Corey:within a space and become the ai provider in that space um and i think you'll see maybe that's city bank of america maybe that's uh in the banking space in you know health care maybe that's pay

1:41:34Grant:Pay attention to what AI companies the banks are acquiring. That might give you an insight here.

1:41:40Corey:It really, really would. And I think we'll be seeing a lot more of that. And I think Adobe is a good example of someone who kind of really led that way. And I say that despite the fact that Canva creates problems for them and other tools have as well. But they're already on every graphic designer and illustrator and artist's computer now. So what they've done is they went and found ways to implement AI that would be appealing to the people in that community. I think of the object selection tool in Adobe Premiere, which I used the other day, and it's amazing. Like trying to rotoscope someone out over every frame for, you know, a 30 minute video is insane.

1:42:25Corey:But with this, it's two or three clicks and it takes 20 minutes to maybe go run and do its thing. But it's not 20 minutes of you click, click, click, click, click, click, click, click, outlining things. It does it all on its own. And can I add three ideas to this? Yeah.

1:42:40Grant:So that because we noticed when we did the Adobe content that we covered, a lot of people were using Canvas or Figma as an alternative, Canvas as an alternative. A lot of people really don't like Adobe for its business practices. So I think that there could be a lot of these like really like cutthroat fights between like the leaders in each industry and how well they implement AI is going to determine, you know, which one takes off there. And I don't think it's just like adding more AI. I think it's how well do you actually improve the product? Not just like saying like, we have these AI features now.

1:43:23Corey:Yes. Yeah. Yeah, and I think you watch those leaders as well as the niche pop-ups in each industry and kind of get an idea of who those are going to be. And whoever they are, they're already doing it at least to some extent or they're not the one.

1:43:39Grant:Yeah, I'm going to add another thing. I'm going to add another element to this. So I think there's going to be a battle between the general purpose agent and the industry specific agent. so if you look at cloud code right people are realizing that you can use cloud code for more than coding so how long is it going to take anthropic to launch like a general purpose agent like cloud for business right like they cloud code for business um but then all these under other industries are going to launch cloud code for x cloud code for this cloud code for that cloud code for you know aerospace cloud code for banking cloud code for whatever and a truly general purpose agent should be able to do all of those things you don't need you shouldn't need a niche specific one but it's a marketing problem and who who is who is the best at marketing the best agent for their niche like they're going to win it's a who's going to make the money problem too okay so let's explain more on that i like that i think i think it's it's equally uh

1:44:42Corey:you know, everybody wants their piece of the AI pie right now. And, and you want to be the one that provides that specifically to your niche instead of them using that. But it's a really double-edged sword because like, yep, you can absolutely go fine tune and train models, build your own models on, on truckloads of data. Like Nvidia just did with their new auto model, which actually was just 1700 hours of AI created video that they trained it on for this autonomous car thing.

1:45:19But I don't know.

1:45:20Grant:Yeah, yeah. I got you.

1:45:22Corey:You can do that, but the problem is.

1:45:24Grant:This brings me to something else too.

1:45:26Corey:These companies are going to steamroll you. The fact is by the time you have that project done, the latest and greatest model is probably going to be better than whatever you made.

1:45:36Grant:Well, this is why if you're building an agent for your niche, like if you're building cloud code for X, like you have to, you have to make it. So you're betting on the models, the cloud model getting better. You have to bet like, and that's a hard thing to do. Like, what can we build if the models increasingly get better? You can't just be adding little additions to it and trying to improve what it's currently doing. You have to be betting on it getting even better than it is now. And if it's even better than it is now, what can you do with it? So you have to be kind of like projecting your ideas into the future, your product ideas into the future.

1:46:20Grant:And saying like, okay, maybe we can't do this today, but let's develop assuming we can do it in three months.

1:46:25Corey:or there's a great Sam Altman quote on that talking about you know if you're building a tool on the idea that it's not going to get better uh we're if we just do our fundamental job right some things we'll feel really good about some things that's not the one where is it but he's like you know if you're not planning on the model itself getting better as part of

1:46:46Grant:your strategy you're in trouble because I really I really want to find a good resource that says how to do that. Like, like what is, what does that really mean? Like stepping you through it? Because I, you know, I struggle. We should do a thing on that.

1:47:00Corey:Let's, let's do a, let's you and I do a project on that, on what that is, what that looks like. The difference in, in building something versus build, that a model's just going to have as a standard, that a model's just going to do like, you know, versus something that's really unique and improves as the model improves because those are very

1:47:24Grant:different things i'll give you an example that i can think of i think gen ui is is one that i think uh is is relevant like the more that the more that these models get better the more features that you can add kind of intuitively on top where you let the where you let the users decide how they want to use your tool and and you just give them the tools to decide that and like change the interface for what works for them so that's one that's one um aspect another aspect switching gears slightly besides giving you more controllability over image models i imagine that image models need to improve to become real time uh in the sense that you know we have a real time voice api that where we can chat back and forth with um with uh chat gbt for example what if that what if um i saw this really amazing product last year called mirage that basically can change your and your screen um in real time so like if you and i were chatting here we have to get access to this so you and i could be chatting and all of a sudden i turn into the orange neuron cat or like or we could completely change the entire background around me and and it's generating uh frame by frame um like like predicting the next frame like we need the image model equivalent of that where you can be like drawing on the screen and it's generating an image as you draw it and generating around you and generating like that's the type of thing where i can imagine the image models getting better in that way you could create so many cool tools with that but that's where i imagine them to get better not just like you know 10k resolution you know ultra hd quality you know

1:49:01Corey:yeah something else i'd like to dive into because i don't think we can just can have an ai prediction show without discussing the 4.5 trillion dollar elephant in the room i know i wanted to get to this too i have i have nvidia thoughts and they have when i when i mentioned at the beginning of this that there are are a few of these predictions that i've gone back and updated a number of times over the last four weeks getting ready for this uh this is one of those uh and here we go nvidia it faces growing competition from customers and partners developing custom chips uh but they're going to continue to dominate the gpu space in my opinion i following the vera rubin announcement this week and the insane differences like you know uh and and i don't want to quote these because they're wrong but the difference in efficiency and cost and speed and uh sustainability uh all of these things are so so dramatic and this is not even what's out there now blackwell's still shipping like the the h300s are still going out right now you know and and microsoft built i think

1:50:20Grant:microsoft has the biggest concentration of them right now specifically yeah uh but you know they're

1:50:25Corey:going out daily and to think the next level is that much of a step better in terms of speed, accuracy, reliability, efficiency, that's insane. So like I think sure, some other players are going to build some cool chips and other approaches and they'll be good but I mean like you know tipping them over means for starters knocking over a$4.5 trillion monster and simultaneously needing to uproot CUDA in the systems of everything

1:51:02Grant:that exists right now to make that

1:51:04Corey:happen. I don't think that's easy or possible.

1:51:08Grant:So, yeah, I really don't feel equipped to add much to that or much to this conversation except three things that I think are worth paying attention to. I love this comment. I wore my Chicago apartment with NVIDIA.

1:51:22Corey:That's what we need to do. Data Center heat needs to be rooted into old steam vents in cities and used to heat houses.

1:51:29Grant:I think basically you saw NVIDIA have this like$20 million aqua hire situation with Grok, right? Where they basically brought the executives in from there. And fun fact, the executives at Grok, from my understanding, and this is Grok with a Q, not K, were the guys that helped build the TPU at Google. So that was very strategic. So it is definitely going to be a knockdown drag out fight between TPUs and GPUs this year. And we're going to see if you really need the cutting edge or if you can have something just below the cutting edge and just scale it more cost effectively for inference, for example, if that will cover your needs.

1:52:13Grant:It depends on what new model architecture and ideas come out there. The other thing I'll add is there's an even bigger element to all of this that we should cover, which is number one, the ad bubble conversation, and number two, the Taiwan conversation, which is we've read reports that China is gearing up to potentially take Taiwan by the end of this year. now that was like very speculative and we saw reports um at another company i was working at back in 2024 or 2021 that said it would be 2024 so who knows if they're going to do it but you know things things going on in the real world right now people are kind of taking taking things that don't necessarily belong to them or maybe they do we don't know um so could could there be a a geopolitical situation that jeopardizes chip production in taiwan where all the chips are

1:53:04Corey:created possibly we don't know boy wouldn't that be a doozy yeah oh my god um and that could throw

1:53:12Grant:a wrench in all of this because i think a lot of people are saying like hey look the the debate around the the speed at which all this is happening like let me really phrase this differently we are about to experience like incredible rapid takeoff if we keep off if we keep up at the same pace we're at right now and there's three things that could potentially slow that down it could be an ai bubble crash it could be political backlash to ai and demanding that things slow down and who knows what form that'll take or it could be um you know like now we can't make chips anymore you We saw there's a RAM crisis happening right now where the cost of RAM is skyrocketing.

1:54:03Grant:So that's market forces impacting, potentially slowing things down. There's a lot going on that could slow it down. That's not at this, are the models, can the models get better? Yes, the models can get better, but everything else around the models could slow it down. Yeah, that's interesting.

1:54:22Corey:That's interesting. And I want to save my comments on the bubble for last. Okay, let's save that. Because I do have some ideas there.

1:54:33Grant:Oh, and the other thing I wanted to say on the chip front, if you want the most cutting-edge information on chips, read Semi-Analysis. Semi-Analysis is the single best source on AI chips. We don't cover it in the Neuron, really, unless it's more broadly applicable. but some but semi-analysis they are the goats in that category so you want chip takes semi-analysis excellent excellent well let's uh

1:55:00Corey:uh I have a theory that I want to drop here. I guess it's a prediction. I think perplexity is acquired this year by someone else. And I think it should be Apple. If Apple was smart. It would be a real kick in Google's search problems for sure. but I feel like

1:55:31Grant:they're not going to do that

1:55:32Corey:the value of perplexity right now is different than it was early on like I don't feel like there's I don't feel like there is anything they're providing me I can't get elsewhere now

1:55:48Grant:where there really really was

1:55:51Corey:at one point and I say that as a guy who loved the tool and loved Arvind Srinivas and everything they stand for and building in public, and I love that. But I feel like their relevance is shrinking every day right now. And they're mentioned less, they're discussed less, they're used less. I think the best move for them would be to be brought into someone else's ecosystem and be part of their tool. And I mention Apple because it's a tool that feels a little Apple-y and because Apple's not having a lot of luck on their own, at least not publicly.

1:56:32Grant:Apple has Apple News too, which I feel like the Discover feed is the best part of perplexity for me. Yes, same. Which, by the way, if anyone from Perplexity is listening, probably not. Add a freaking search tool to the Discover feed. I've asked this multiple times.

1:56:47Corey:Just give me a search so I can search previous discovers. The truth is the Discover feed was what kept my perplexity subscription paid for for as long as it did. I really love the Discover feed. It was a great way to get news. I loved being able to hit read out loud while I was driving my car or like while I'm paused and be able to just listen to a news article that's crafted for me essentially in a lot of ways. And really, really targeted me in a way that Google Discover does not do, frankly. or at least better.

1:57:23Grant:Who is the most valuable media corporation? Like is any media corporation worth more than$32 billion? Cause if I were a big media conglomerate, I would, you know, they're currently suing the crap out of perplexity right now. I would get them to the point where they're on their knees in the lawsuit. And then I would buy them and use their tech.

1:57:45Corey:Yes.

1:57:46Grant:Like, like I don't think an AI, I don't think a big tech company will buy Perplexity. I think it'll be maybe not a media publisher because I don't think they have that much cash flow to buy it unless the valuation of Perplexity just dies.

1:58:01Corey:And it includes Comet, which is – I like Comet. I think Comet's a good browser, despite the fact that its memory footprint is about quadruple Google Chrome. The sheer volume of Ramideats is a lot. I want to share this because I think it's fair. Chaotic neutral.

1:58:22Grant:I'm chaotic neutral. Thank you.

1:58:26Corey:Grant is evil. That's a T-shirt. That's a T-shirt. One of these days we're going to do some merch with fun stuff. I have had suspicions. I don't really have a stance on this. But I want to at least address it. And it's a discussion. Grant and I have had a number of times over the last year about Anthropic, and that is that perhaps someone might acquire them.

1:58:56Grant:I think that ship sailed, actually. I think it may have.

1:58:59Corey:They've raised a lot of money this last four months.

1:59:03Grant:They're raising$10 billion right now, and they are the one to beat on the enterprise side. So I think they're actually good to go. And they're planning for an IPO this year, in case people miss that.

1:59:13Corey:When you say enterprise side, what are you thinking? not enterprise adoption as companies on the api yeah yeah they're winning the enterprise they're

1:59:20Grant:winning the api enterprise uh or the api what what i don't actually know what i would say that's tied

1:59:26Corey:to their developer engineer yeah community thing is that those are the people that are that are

1:59:32Grant:making those decisions but i mean just today they had a deal that was announced with the lions which is like a giant german um insurance company yeah um so they are winning at least in terms of momentum on the enterprise adoption side. Yeah.

1:59:46Corey:On the enterprise backend adoption side, not in the two-year companies. I don't know the exact numbers,

1:59:53Grant:but I think they might also be winning on the enterprise plan side too.

1:59:57Corey:You think so?

1:59:58Grant:I think that's what everyone is saying. Everyone is saying this.

2:00:01Corey:I haven't heard that.

2:00:02Grant:I mean, again, I don't have my sources ducks in a row on that, but in terms of momentum, they have the heat right now and momentum is the momentum potentially. So they're the one to beat in that area. ChatGPT also came out and said, you know, the enterprise is where we're focusing our attention this year. Because they're, you know, in theory, you could assume they're losing in that category.

2:00:25Corey:And Microsoft already have most of the world.

2:00:28Grant:And Microsoft uses Anthropic models. So if you want to use Anthropic and you already have Copilot, you can do it directly. Either.

2:00:36Corey:That's cool. That's one of the things about Copilot that makes me not understand always why some people don't like it. is that you're essentially just using ChatGPT or Claude and you got both of them right there in one little package. I mean, you lose some of the features like I use projects a lot but generally they're just called something different there. Yeah, that's a good call out.

2:00:59Grant:Yeah, so I don't know. I think that they're the one

2:01:01Corey:I think you're right that the ship has sailed on Anthropic being bought.

2:01:05Grant:I thought it was going to be Amazon who bought them but I think they Anthropic got too much too much heat. Amazon can't buy them anymore.

2:01:15Corey:I would say no, I don't secretly work at Microsoft. I just I've spent a lot of time researching their products this year, more than I have in the past.

2:01:26Grant:Because full disclosure, our company uses Google products. Google Workspace. Google Workspace. But we use both Copilot and Google Workspace both for personal use and enterprise use. But to your point, Corey, the reason why you're saying Copilot is so good is because it's far better than Google's workspace solution. Like Google's workspace solution is like, here's a bunch of random tools split up everywhere. You know, like, good luck. Hope you can find what you need. Microsoft has a much better solution compared to that. Is it better than just using Anthropic or ChatTBT out of the box? I don't know.

2:02:04Grant:I use a combination of everything because I have to, but if I had one thing that did everything I needed to do, I would use that.

2:02:12Corey:And I don't want to share the comment I made the other day about Google workspace and, and who uses it, but I think I have to, you know, we were talking about the other day and, and to Cal's question, I use Excel for everything when I'm not at work. I have a 365 account. that's mine, not tied to what I do at the day job. And I use it pretty consistently as much as I use, not as much as I use Sheets because I do a lot of spreadsheeting during the day. That's fair. But I genuinely prefer it still over that. And I always felt like people who use, my joke the other day was that the people who use Google Workspace do it because Microsoft's too expensive.

2:03:07Corey:I mean, you're not wrong in my personal experience,

2:03:10Grant:but that tracks with my lived experience, and no one can take that away from me.

2:03:15Corey:Yeah. The truth is they're very comparable. The differences are pretty minimal. Yeah. But what I will say significantly is that the AI provided inside Microsoft's ecosystem is significantly better. Significantly better than the AI provided to than the AI provided in Google Word. It could be a scale issue, right?

2:03:45Grant:Because we don't know, maybe it's written somewhere, but we don't know what the model is that Google lets you, lets everybody, all of their 3 billion users use inside Google Docs for free. It's probably their cheapest model.

2:03:57Corey:When they added us, when they force-added everyone to Gemini last year, because that was how Google pushed their numbers up, was by adding everyone in the world to Workspace. So if you were a Workspace client, you got a price increase attached to, here's Gemini, which was going to happen, obviously. But the problem is these enterprise companies are paying thousands and thousands of dollars a year to Google, and they provided them with 1.5 Flash as the model that was already dated. It was already dated when we were force added to... Yeah, it's just not smart enough. I spend most of my time... But its ability to work in tools sucks.

2:04:42Grant:Yeah, I spend most of my time in AI Studio and other tools, Anthropic, ChatGPT, trying to recreate a good version of an agentic Google Doc. Like the fact that it's not, like the Google Doc is not a good place to write using ai shows you that like they're losing that battle um because it's not their focus they're focusing on the consumer which makes total sense and i think they've they've made a ton of ground there clearly they put chat chibut on the back foot um but chat chibut is still way ahead um but yeah i just they've they've sort of let the the workspace adopter fall by the wayside for that while giving us other good stuff that's that's what i would say is my criticism is that the

2:05:23Corey:workspace user has kind of fallen by the wayside a little bit.

2:05:26Grant:But we still get access to Gemini and you can do a lot of pretty good stuff.

2:05:29Corey:Their models are great. I don't have any complaints about their models. I love AI Studio. I love a number of tools they've created. I love Nano Banana and all of that. But I

2:05:39Grant:feel like

2:05:41Corey:while they were chasing consumers to keep the ad business propped up, because that's what it boils down to, is that Google to keep its ad revenue, which is the lion's Bread and butter.

2:05:53Grant:The reason they can do all this really amazing stuff with DeepMind.

2:05:55Corey:They need the consumers, and that's why they're competing with ChatGPT instead of Microsoft. And as a result, though, an awful lot of customers that spend an awful lot of money every year have... It's improved a little along the way, but it's still... Like, in my opinion, as recently as the end of Q3, it should not have been rolled out publicly as it worked in Workspace. Yeah.

2:06:27Grant:I think, you know, one wildcard prediction could be that, like, the rise of Claude Code really supercharges Google and Microsoft to get their act together. Maybe. And I think they're probably already thinking about that. and they're just lucky that Anthropic accidentally called their tool QuadCode and not CloudAgent. So we'll see. We'll see what happens there. Do you want to touch on what else do you have, Corey?

2:06:56Corey:I think I have two last things I think we need to address. I'd like to hit a couple comments I have on evaluation and benchmarks and I'd like to end on ye olde economy. on evaluation and benchmarks i have two predictions that i'm going to bundle into one here and i think that's that humanity's last exam and arc agi2 both fall this year

2:07:22Grant:uh they're both really really really close to falling already like in the 80 percentage

2:07:31Corey:range. I feel like what we get is Arc AGI 3 as we continue to say, well, it's not quite AGI because a number of mundane tasks still just remain out of reach.

2:07:48Corey:I think we just continue pushing it a little farther and pushing it a little farther and we get a third when we reach the top of this. It's AGI benchmark. and that we still haven't accepted. And I'm not saying what we have is, but I think if we use today's tools in mid-2023, we would at minimum believe it was close, very, very close. If two years ago we used Gemini 3 and 5.2 and Opus 4.5, I think we would be shocked with what we're using today, but because it happens so incrementally and, and because it happens every two weeks, like it's always happening. That's the problem. And as a result, I think people don't notice.

2:08:41And I, that's what I think.

2:08:44Corey:I think humanity's last exam and RKGI two both crash. I think they both get beaten soon. Okay.

2:08:52Grant:Let me add to that. I don't know what model architecture or technique will solve this, but maybe it's just all the RL reinforcement learning that they're doing in these environments, which environments can basically be a data set. It could be an actual staged environment where you have a user interface that AI is going around and testing in. But somehow the action efficiency, which is how quickly the AI model learns and picks up what it's doing, will start to really surprise us. And that area of development will increase. So what that means is AI can basically start to do things and figure out and orient itself in an environment faster.

2:09:52Grant:Because that's what Arc AGI 3 is going to be testing. It's going to be testing, can you efficiently be put into an environment, figure out the rules of that environment, and solve that solution quickly? Because that's one of their thesis on the true test of intelligence is how quickly can you learn? And I think we're going to start to see some real progress there as ArcGIS 3 gets rolled out, and that becomes the new standard. Yeah. Yeah.

2:10:20Corey:Yeah.

2:10:22Grant:Yeah, someone said RL is kind of garbage. Yeah. Andre Karpathy says it's drinking through a straw. So we'll see what becomes the dominant paradigm if people continue with RL.

2:10:34Corey:But, you know, it's some type of continual learning, though, is is what will allow people to accept that it's AGI. Like like to me, you know, I always hear these arguments that are like, well, why can't it, you know, detect I'm going to use an old one. ours and strawberry and it's like who cares like some of that little stuff we have little things we suck at too yeah like like there are things that we just don't do well like balancing a basketball on a finger seems like a pretty minor task considering we can do advanced physics but we can't do that why is it that you know we can't do a variety of things and i think I think the idea to call it or not call it AGI is totally stemming from the idea of how we expect it to be like humans, when in reality, I don't think there's any reason for it to be like human.

2:11:36Corey:I think we should be building something that's better. And if that means that maybe it can't recognize purple from blue, but it can predict stock trends with 98 % accuracy, I think we shut up about purple and blue.

2:11:54Grant:Maybe that's wrong. This goes to something else I've been wrestling with, and I wasn't sure if I should make it a prediction or not, but the rise of small niche models versus one full AGI model. and we've kind of talked about this where it's like maybe the secret to agi is like a really good harness and like just strapping a bunch of smaller models together to like do different

2:12:14Corey:components really well like like one of those monsters on transformers yeah yeah exactly just

2:12:21Grant:strap a bunch of like strap a bunch of cars together and make this monstrosity but you know our brain has a lot of different cells that specialize into different areas to do different functions and it's sort of it it seems silly to try and just build one gigantic thing that costs trillions of dollars um that can do everything versus just making hundreds of specialized things kind of like anthropics um you know skills thesis where you can instead of making this giant mcp server with all of these um different tools and docs and how to use it you may make a bunch of little tiny skills that you can all lego brick together and then um yeah like like that makes up all the components that you need to actually function the same multi-agent hive brain exactly and i and i've heard uh dorkash and other folks in the industry say that's that's basically where this is going is it's going to be you know like a bunch of systems that can copy themselves and and you know once one thing learns something it adds it to the database and then they all know it and that's definitely where it's going.

2:13:26Corey:AGI is a computer, not a model at that point, essentially.

2:13:29Grant:Yeah, which is why Reid Hoffman said a lot of what this year is going to be going into the end of this year in 2027 will be about orchestration, which is it's no longer you and a single agent chatting back and forth, but it's you and a swarm of agents that you're controlling and orchestrating. Yeah. And I don't know what the dominant user experience for that is, but it'll be interesting.

2:13:51Corey:I like it.

2:13:53Grant:software is just dropping like mad knowledge in the chat here. Look at this.

2:13:57Corey:Yeah. I was reading, I was reading this and, and,

2:14:02Grant:and if I'm honest, you should be getting a meta paycheck. You should be going to meta and trying to get them to hire you. Be standing outside with a,

2:14:10Corey:where's my a hundred million dollars mark. Yeah. Bring it.

2:14:13Grant:Like, yeah. I, I do think that there is a conflict though, between making a truly thinking, uh, AI and like just one that's really good for business. I don't think they're actually the same thing. I don't think you want the, the truly thinking AI to be the thing that you have running your company.

2:14:34Corey:That's fair. I think that's a really good call out. Actually.

2:14:37Grant:I think you just want it to be the most controllable, smartest thing possible. And perhaps that is a, an over on the giant model or, or, you know, Google, Google says they're building an Omni model, which is something we haven't talked about yet um what does an omni model actually look like it's something that can do all of these things um you know like it can do vision it can do voice it can do world models maybe that's a bunch of models put together that um it you know just has a harness that's controlling it or it has like a cortex model that's like saying like i'm going to use this one and for this and this one for that and this one for this but like the user experience as we go to it and it's a single model that can do all of that.

2:15:16Grant:I don't know. But that's also something that's in development. So we'll see.

2:15:23Corey:So Grant, do we want to talk a little economy before we go?

2:15:26Grant:Let's do it. Let's chat financial bubble talk. Paper, rock, scissors for who goes first? I'll go first because I'm not going to really share my own opinions. Who cares what I think? I'm just going to share two points that I think were quite interesting. So the first point was what Zuzana shared on our chat with her this week, which was essentially at this point, we've only spent about 0.7 percent of GDP on the AI data centers that we're creating. And at the height of the telecom boom, we were spending something like 2 percent of GDP on telecoms. And that was just infrastructure for phones or internet.

2:16:19Grant:It's not as fundamental, as she says, as AI is, where if you have true intelligence that you can scale and apply to any problem anywhere in the world, that's a lot more fundamental than just, can you connect to the internet now it's like can you scale intelligence like it's it's fundamentally probably it's probably worth five percent of gdp before all is said and done um yeah if not more than that yeah um but is the money being deployed efficiently which is the problem that i think we have now where you have one company nvidia that is just like the king um and they can charge kind of whatever they want um you know they obviously that they've been taking down a peg with google and now there's real competition.

2:17:02Corey:Otherwise known as the legs of the US economy.

2:17:05Grant:Yeah, right, right. Yeah, exactly. So they, you know, now that it's, you know, perhaps there'll be some price dynamics that I'm working with Grok. Maybe they release a cheaper, faster chip, you know, like Grok can do to compete in that area. Maybe that brings the overall cost down. As Jeff Bezos said, it's an industrial bubble, you know, not necessarily a financial bubble. um those are different things they're different things yeah and i think we've taken some air out of the balloon by talking about is there an ai bubble um this year but uh i have a i have a kind of hot take on this 2026 if there is going to be a financial bubble it could be the year that uh open ai dies wow that's a moment i is the most at risk in a financial bubble burst personally

2:17:56Corey:I think that's possible. I mean, I don't necessarily agree that that outcome is necessarily likely, but I see where you're headed. And I think in 2026, once again, we hear a lot of people yap about a bubble and we just continue to grow. I really think that there's still an awful lot of leeway here. There's data center construction. There's the build-out of everything that goes inside these data centers. There's all of these companies scaling up their AI efforts across all industries right now, trying new things, some of which will make money, some won't. Some people will fail. Some people will definitely lose money, but I do not in any way expect an AI bubble to explode this year and all of the horrible things that people expect could follow that to happen.

2:18:53Corey:And I, there's so much more room still like, you know, these companies are already working into the, you know, they're, they're talking 2030s and, and, and working on things. And there are, there are deals, there's stuff being built. There is, I, I just, I don't think it's there. I do think that there is absolutely, there will be startups that fail and investors will lose their money on some

2:19:19Grant:of those, but I think those are less tied to, I think those are less tied to, you know,

2:19:28Corey:our stock market and our overall economy more so than just, you know, the situation with investing in risky startups, you know, and maybe it's more than those than normal. I'm trying to read your conversation.

2:19:42Grant:Well, sorry, I'm sidebarring. I'll bring it into the main chat so um uh let's see onkit said uh google has such deep pockets that they could survive it and i said basically every big tech company playing in this could survive it um no problem nvidia will survive microsoft will survive google will survive i mean one of the best financial takes was essentially like honestly the big tech companies based on their cash flows could could raise like could spend like five times the amount of debt or something to that effect oh yeah if they wanted to actually and be totally fine and totally justified doing it so really it's not as big of an issue like the actual spending is is it seems astronomical but relatively it's not that big of a deal yeah

2:20:25Corey:which i would say any ai company that collapses right now is going to be bought before it hits the ground that's what i that and that was there's still people dying to get their teeth into it do

2:20:36Grant:mad it's actually impossible for me to imagine open ai actually going down which is why i say that kind of ingests because like it's actually a really bad take like there is no way first of all they've made themselves vital to this entire industry they are the band-aid of ai like there's no way that they're actually going to fail but if they do fail um they will certainly be swept up i mean i feel like microsoft would have first dibs to buy them based on how much they own of the

2:21:03Corey:company and absolutely should have it frankly you know they were investing in them early you know i I mean, it wasn't a given then. And, you know, I mean, the fact of the matter is, ChatGPT is now a productized name. If you talk to people who aren't in the know or staring into the little microcosm of AI like us and probably everyone watching this do, it's like Kleenex, you know? I mean, they don't even talk about AI. They talk about, you know, ChatGPT. And it doesn't matter what tool they're using. It's what they call it. and there's a lot of power to that. And when you think of the sheer volume of paid users, I mean, even if they're operating at a loss, first off, Elon would buy them.

2:21:53Grant:Yeah, that's true, actually. That would actually be hilarious.

2:21:56Corey:Elon would fight Microsoft over buying OpenAI, and God, can you imagine what they'd have to sell for? Even borderline financially collapsed. It would be insane.

2:22:07Grant:I just don't think that's going to happen. I think the only... The arguments and the reason why I even brought this up to begin with is that they just stretch themselves so thin with everything that they're trying to do, which you have to do when you're at hyperscale. Like I said earlier, they've also made themselves kind of like integral to this whole industry. Their API, everyone can use it easily. But they just don't have the momentum at their back. And it seems like other than the consumer, which costs them a lot of money, actually. Some reports I saw recently said that a majority of OpenAI's usage is the free tier.

2:22:47Grant:And it's not necessarily free users that advertisers would want to advertise to. Maybe. So even if they launch ads, which they will do in 2026 and will be good for them, I think. perhaps it's not enough to justify it.

2:23:05Corey:I think they'll also, I think when ads come, they've been very, very vocal about it not being the results of your prompt.

2:23:15Grant:I saw some reports recently that were...

2:23:18Corey:No, I mean like there may be ads alongside it in your sidebar in chat GPT, but that your answer remains pure. And they've said that really publicly a number of times I want to say, Kevin said it, Greg said it, Sam said it I say that's right, I could be wrong And, you know, the truth is, I get it's a business It's got to be paid for, this ain't cheap Nothing of this is cheap I mean, you know what$150.90 costs Let alone a data center full of H300s

2:23:53Grant:This is why I almost wonder if like the average person right now can't tell the difference at the frontier of intelligence. Like we can't really tell the difference between Gemini's best model, Google's best model, and OpenAI's best model in terms of just asking questions. Obviously coding, you can tell the difference. Obviously some of these more niche use cases, you can tell which one works the best for your specific use but you can't i mean who can really say other than the benchmarks and you know there's a lot of benchmarks going on which one's actually the smartest um i've heard people say gemini is the smartest for science stuff but then you know uh open ai is the smartest for you know just general like it can answer most things i need it to do um and you know claude i just i find the experience of talking to claude more fun than talking to the other two so just i talk to it

2:24:49Corey:more often um so i don't know well i don't know if they actually need to push the frontier of

2:24:55Grant:quote-unquote intelligence in any meaningful way they just have to make the tools more useful and then they and then they'll probably focus on making them cheaper to run yeah yeah i mean especially if there's any sort of downturn or financial conversations. Just do that and then just turn them into money printers. I think they could absolutely do that.

2:25:16Corey:When I talked to Scott Guthrie, something we discussed, Scott Guthrie is the EVP of AI and cloud or something like that. He's a big deal at Microsoft. He had an interview when we were at Ignite. One of the things we discussed was that all of the incentive is there for companies to ensure that This is done as efficiently as possible. The fact is it costs a fortune. It's a fortune to serve it. The hardware is a fortune. Anything they can do to make this more and more efficient, sustainable.

2:25:55Grant:And communities are demanding it if you want to build a data center.

2:25:58Corey:That's the other thing. I mean, all of the incentive is there for this to become more efficient over time. And that is absolutely the case with how it works in all tech. Like, you know, you build the thing, you make it strong, and then you start worrying about how can we make it run a little faster? How can we make it a little smoother? How can we make it cheaper? You know, you innovate and refine, you know. And I think that, I think we're kind of watching a little bit of that go on side by side. And I think, honestly, NVIDIA is kind of leading the way in the refining portion of that. When you look at, like, Blackwell over Ada and Vera Rubin over Blackwell, you know, there's a lot of focus on keeping those efficient.

2:26:39Corey:there, I think, and requiring less power because we don't have the power. That's the problem is, you know, you know, they're building out solar and wind farms and all of the, all of the renewable resources, not, not out of sustainability concerns, but because it can't get enough raw power to, you know, I don't know if he'll ever feel the pressure enough, but, you know,

2:27:04Grant:President Donald Trump revoked some of the solar projects that were in the pipeline. And then perhaps maybe he'll feel the pressure to put it back on for the AI companies, which would be kind of funny. I think there's two things there. Someone brought up, Softwire brought up, efficiency would improve if they improved electrical transmission. That's very expensive, but that project is ongoing. They're replacing a lot of the transmission lines right now. And then also storage.

2:27:31Corey:Get them underground, for God's sakes. We don't need them dangling in the sky.

2:27:37Grant:Yeah. Surely better battery tech is coming. Yeah. And, you know, it's just scaling it.

2:27:43Corey:We might have a guest on that coming soon too. Just an FYI. We've got somebody we've been talking to that had a really interesting approach that intrigued us.

2:27:54Grant:And I don't think it's a done deal yet, but we're working on it.

2:27:57Corey:We hope to have an awful lot more cool ones this year.

2:28:00Grant:Yeah. Definitely. Definitely. I got my money on it.

2:28:03Corey:I want everyone to help me. I want to get Altman on here. And I think we need to make that happen. And I don't know what the answer is other than maybe everyone helping me badger on Twitter to make it happen.

2:28:14Grant:Well, maybe if we, I don't know, maybe counterintuitively, if we say 2026 prediction, OpenAI might collapse. Maybe he will come on to dispute that. To dispute that. Or they'll just not answer our emails anymore. Could be out there.

2:28:30Corey:I've been cooking with them right now too, I think. Yeah, we'll still see after my prediction. But I say this all for content, folks.

2:28:42Grant:I am just thinking, I try to think realistically, but there's so much going on in this space. And any of these companies could release a model tomorrow that is AGI. I don't think so, but they could. And that just changes the game for everything. for all of this. So I'm just trying to think about what could happen, not necessarily what is going to happen, because nobody knows what is going to happen. I will say, how do I say this?

2:29:18Corey:I have spoken personally to someone who has touched the next OpenAI model, and they are pretty floored. I am absolutely certain Google has another one very similar probably. Waiting in the wings. I mean, like I assume these guys are all working three generations into the future. But from what I was told, it looks like there's something really cool coming. I don't know that it's that, but we'll see what it is. Either way, it's curious and exciting. Nothing beats new model day. Grant and I get all excited and flappy. But yeah, I've had a literal personal conversation with someone who has touched a thing.

2:30:09Corey:And it sounds really cool.

2:30:15Grant:Did GLM-5 just come out? Can you pull up that comment, Corey?

2:30:20Corey:Yeah, I have yet to run any of the GLM models. um there's a new cool video last one i saw personally was 4.7 i just i just linked to it

2:30:30Grant:today but to be fair i feel a little behind at the moment that's okay you took a day off now you don't know what's happening in the world yeah it's the way completely behind trying to catch up on everything you know don't look at the news for one day you have basically missed a week's worth of content okay so glm5 is not out yet well they just went public um that company so i imagine with all that newfound money you're gonna buy themselves some nvidia gpus which is now possible because of the deals between the two countries and uh they're they're gonna be up and up and cooking i think i would expect glm5 soon the the believe it or not the open source chinese model ship faster than the like Frontier Labs do.

2:31:17Grant:I was really impressed

2:31:18Corey:with the Quinn models. I've used them quite a lot. Yeah, they're big. And there's another interesting model came out yesterday. I don't know how much you know about it yet. I say yesterday. Which one is that? LTX2 from Litrix.

2:31:32Grant:Oh yeah. That just went out.

2:31:34Corey:Yeah, they're actually I think they may have advertised with us now that I think I was doing, I was setting ads up in future issues yesterday. Yeah, yeah.

2:31:41Grant:Yeah, no free ads, but they did sponsor today's newsletter. Yeah. So people know.

2:31:47Corey:Yeah. Yeah, and this is totally unrelated to that. But that is how I knew it happened and went and did some research and checked it out last night. Of course, I don't have anything I can play with it on yet, except for maybe hugging face spaces. But it was interesting. And I'm trying to remember what it was that was so unique about what it does. I want to say seamless video generation. I know you can give first and last frames, bring stills to live text video.

2:32:18Grant:I'm going to have to find it.

2:32:19Corey:We'll report back. It was an interesting model. It was a couple of things it was doing that others weren't. And I read it late, late, late, late, late, late, late last night.

2:32:31Grant:Someone mentioned electrical engineering earlier. Darby said, what certification should you be looking into right now? If you're open to a career change, I would look into electrical engineering.

2:32:43Corey:Electrical engineering, no joke.

2:32:45Grant:Yeah, I feel like that is actually a legit good career path at the moment, helping plug in all these data centers and pay for all this stuff in energy infrastructure. Oh, cool. The Stanford course. Yeah, that's a good one. Yeah. Yeah, there's a lot of – all of the big elite universities have published these free courses online. I don't remember if that one's free or not, but a lot of them are on YouTube, and they're really, really good. So you should definitely check them out.

2:33:21Corey:And I'll be honest. I wouldn't even worry about specific certifications so much right now as I would create things. I would build things. I would save your most impressive prompts. I would build a portfolio of what you've done because the truth is that people running these companies, they don't give a damn where you went to school right now. What they want is to know, can you do the thing? Because talent is thin. Talent is not all there. They're willing to pay good money. Just be able to prove what you can do, I think, is what matters. And it doesn't have to be building software. Like, I think there are plenty of options.

2:33:57Grant:Yeah, YouTube for Harvard and MIT lectures.

2:34:01Corey:I've taken, I've run through a bunch of those. Those are excellent. You should absolutely watch some of those videos, take notes, treat it like it's a real class, learn the stuff.

2:34:10Grant:You know, that's what I would say is I was going to caveat the thing about, uh, certification and courses. They're really good. If you need to put your money down to actually complete them. So that's what I would say is like, if you have disposable income and you could, and you need that structure to guide you, I would 100 % recommend doing so. And we include courses that we find are helpful all the time. DeepLearning.ai is a great resource for this. I'll include a link to them in the chat.

2:34:41Corey:And if you don't know where to start, go to Coursera or edX. Just find something. Find

2:34:47Grant:something 100 level. Can't stress that enough. You know, something that's very, you know, intro.

2:34:52Corey:take it and then by the time you're done with that first course you'll have interest areas and then you'll kind of think maybe i'm going to go take this one and that one and when you take them on corsair you do get like a certificate kind of a badge uh that you could throw on your linkedin profile if you're worried about certifications because we all know we love certifications they

2:35:11Grant:make us feel good and accomplished and yeah and it might it might appeal more to people who need some signal um and don't understand the industry themselves like if it's a company who's hiring someone with AI experience but isn't an AI company, a certification might really help them realize, oh, this is someone who we should pay attention to here because, you know.

2:35:32Corey:If you're a non-technical, you should go do that today, honestly. Like, go and it'll teach you enough of the fundamentals of, like, you know, how to communicate with one that, if nothing else, you will be leaps and bounds above most of your non-technical co-workers.

2:35:48Grant:yeah i took an mit uh course uh through their um edx i think is what it is i can't remember exactly their their platform yeah their their online platform i took one of those um and that was really helpful like that helped me understand a lot of the more technical terms that i just had no no experience in when i first started this newsletter and i was like okay what's a gan um you know what what is the difference between gan and diffusion like what what's a i think it was van or something like yeah like there were some of these terms that i just that were relevant at the time oh vae uh what were these terms that were relevant at the time that i just kept seeing everywhere but didn't really know what they meant it really helped with that and then it gave you gave you practical hands-on um experience like building stuff and i think they've improved it a lot since uh since i did it um but but that's just one you know there's there's there's any anything where it's like they try to keep the content as relevant as possible um you You don't want something to someone's point in the comments that's from the 60s and 70s because the field is moving so much faster than that now.

2:36:56Grant:They're all relevant situationally. Okay, that's fair. That's fair.

2:36:59Corey:Yeah. Well, guys, I think we've gone way over what we normally do, but guys and gals, excuse me, I really appreciate it. This has been so much fun. Before we drop, just a quick reminder, if you haven't yet, please like and subscribe to the channel. Make sure you go check out the newsletter at the neuron.ai.

2:37:20Grant:Join 600 and I think it was 32 ,000 this morning.

2:37:25Corey:Others who receive it each morning and get their AI news. We really appreciate all of you and the conversation you bring, the insight you bring. Yes, I'd love, software. Yes, I'd love to check out what you're talking about. Send it to, is it team at the neurondaily.com?

2:37:42Grant:Yeah, you can send.

2:37:44Corey:We can probably put that on the screen somewhere, too, if I had to guess. We should learn these tools better. Thanks, Anga.

2:37:52Grant:Dude, thank you so much for hanging out and sending so many good comments and questions. This was great.

2:37:58Corey:And thanks, everyone, for everyone else tuning in. And if you email and we don't get back to you, please know that email gets hundreds a day. We need a better solution. you know we've debated a variety of things maybe even a discord or something that would it would be fun and an easy place to back and forth and chat for things like that

2:38:20Grant:i'll also be a personal shout out i'll also be uh writing up a recap of this in the email or sorry on the website and then i have a list of i put together an omnibus of all the other predictions that i've seen people make uh and so i'll release that as well so you'll have both of those to look forward to. Hopefully I'll get it done today, but if not, it'll be in your inbox sometime next week.

2:38:43Corey:Yeah, and if you email and we don't see it, please, please, please, please don't hesitate to email again and maybe again, because it's so many that it's hard to keep up with them, and sometimes they all get funneled together. But we'll be keeping a lookout, and we really appreciate every one of you. And hopefully here soon, we're going to continue to ramp these up, do more. We got a cool interview with David Cox from IBM dropping Tuesday. You'll want to watch. And if you haven't yet, go check out the Zuzanna Stamoroska

2:39:11Grant:episode from Pathway.

2:39:13Corey:It's really cool talking about Baby Dragon Hatchling. Have a great weekend.

2:39:18Grant:It answers a lot of questions as far as continual learning and knowledge. One pathway, potentially, to do it. You don't know if that'll be the ultimate way, but it's very interesting. It is. We don't have an email agent. That's a corporate bureaucracy problem. Not a technical problem.

2:39:37Corey:agent that's like, is this a real human? If this is a real person, is this a real thing? We do respond to a good number of them every week and try our best to make sure, but most of them aren't actual questions and things. Anytime we see one that is, one of the two of us bounces on it. Sometimes both of us do.

2:39:55Grant:One other caveat I'll just give, probably should have given you this at the top. This is all speculative. We have no inside knowledge to any of this stuff. We're totally guessing and and just trying to entertain y 'all with our, our best guesses at what could happen.

2:40:10Corey:Yeah. So, Hey, yeah. Yeah. Don't, don't, don't go make investments or anything based on it. Not financial advice.

2:40:17Grant:Not financial advice. Absolutely not financial advice. If you already traded on this.

2:40:20Corey:You shouldn't take it from me anyway.

2:40:23Grant:Not, not a financial advisor. Okay. Anyway, got that out of the way.

2:40:27Corey:Thanks everyone. We'll see you later. Bye. Farewell humans.