NotebookLM Just Killed After Effects? + ChatGPT vs Anthropic War

Hey, welcome to The Next Wave Podcast. I'm Matt Wolfe, and I am joined once again by my good friend Joe Fear. And as we have done the last few weeks, we're going to break down all of the biggest, most interesting, most fascinating, and some of the weirdest stories in the world of AI that's happened over the past week and share our thoughts on it, test some tools, show you a really, really cool potential replacement for After Effects, talk about some other cool things going on with Hollywood. We've got a lot stuff to get into, so I'll stop rambling.

Cutting your sales cycle in half sounds pretty impossible, but that's exactly what Sandler Training did with HubSpot. They use Breeze, HubSpot's AI tools, to tailor every customer interaction without losing their personal touch. And the results were incredible. Click-through rates jumped 25%, qualified leads quadrupled, and people spent 3 times longer on their landing pages. Go to hubspot.com to see how Breeze can help your business grow.

But, uh, hey Joe, thanks for joining us again on the show.

Hey, I'm happy to be here. Happy you keep inviting me back. You know, it's just been kind of a fun thing we're doing. I think people like you.

Hey, if you're watching this video, in the comments say we love Joe and, uh, let everybody know how much you love Joe on these shows.

Please, my ego needs it. I swear, like I live and breathe with what I see in the comments.

So, and if you hate Joe, well then just go away. Just, all right, let's get into it. Uh, we got a lot of stories we want to break down for you this week and, uh, just share some thoughts and dig in on. So where do you want to start with this one, Joe?

Let's start with, uh, I feel like it's an ongoing theme lately since we've been doing this is, uh, so many models just dropping all in the same week. You know, from all the different LLMs. And I'm your everyday Joe over here. So like when I see all the models getting dropped, I'm just like, well, hot damn, where do I start? Like, this sounds cool. I love to test this stuff and I want to make the most out of all the tech, but what does it mean?

Maybe we should rename the show AI for the Everyday Joe and then in parentheses, fear.

Hey, I'm happy to hold the signs, be that guy for everybody. So that's how I see myself here. So yeah, Matt, what the heck is going on with the models, all the drops this week?

Yeah, so we actually got a handful of new model drops. We're going to focus on 3 of them. We got 2 new model drops out of OpenAI. We got a model drop out of Gemini. We did also get a new model out of Qwin, Qwin 3.5 something or other. There was a new update from Grok, Grok 420. There was a new model that came out of Microsoft called Phi-4, I believe, which is like a smaller 15 billion parameter model that's made to like run on device for people. All of those models came out, but I don't really think most people listening to this care that much about it. Maybe some people use Grok.

People are starting to use Grok more. I'm hearing more use. Yeah.

But you know, the Grok update was sort of more of a marginal update. They're just like, yeah, we made it slightly smarter. We made it hallucinate a little bit less. You know, it's just like fairly minor updates, not no big like architecture change or anything like that. The big models that actually came out this week that I think we should hone in on are two of them out of OpenAI, one of them out of Google. And the Google one actually came out like an hour earlier, and then the OpenAI model came out later the same day. Whether that was intentional or coincidental, we'll never know. But this GPT-5.3 launch, this is one that basically everybody that uses ChatGPT is getting. Basically the big update on this model is that it was like a vibes update, right? So it's an update where they're making it preamble less before it gives you the answer, right? Like, you know how you might be using ChatGPT and you might say, hey, what's the healthiest brand of beef jerky? I'm saying that because I got beef jerky on my desk right here. And it might say, Joe, that's an excellent question. I'm going to go ahead and break it down for you and I'm going to do it in a step-by-step process. So I can really help you lock it in. And then it gets to its answer.

Us everyday Joes love the little pat on the back, like, you're doing great, buddy.

So this new model is doing that kind of stuff less, right? And this new model is in the free plan, the $8 a month Go plan, and, you know, all the plans up the chain. So real quick, NotebookLM is probably one of my absolute favorite AI tools. And later on in this episode, we're going to break down one of the coolest new things that you can do with NotebookLM. But also, if you haven't already, make sure you check out my completely free guide of all of the interesting ways that I use NotebookLM in my own business. It's completely free and you can find it in the URL in the description. Again, totally free. NotebookLM, go check it out. And let's get back to the episode. But that is like the big update. The update focuses on the parts of the ChatGPT experience people feel every day, tone, relevance, and conversational flow. GPT-5.3 Instant significantly reduces unnecessary refusals while toning down overly defensive or moralizing preambles before answering the question. So basically it just gets to the answer more directly now when you're using this 5.3 Instant.

It drops the overbearing or preachy language in favor of more of a natural focus tone.

Yeah. So look, you could see down here in their like little description, it says GPT-5.2 Instant eventually answers the question, but in an attempt to explain its safety boundaries leads with a lengthy preamble. Yes, I can help with the physics and the math, and I can also help you sanity check what's physically possible, but I can't help with calculations or step-by-step guidance aimed at accurately hitting a real target at long range. That's kind of. Make it work at distance trajectory where it can meaningfully increase weapon effectiveness. So I have to keep it safe and non-actual help. And then GPT 5.3 Instant. Yes, I can help with that for a very long distance archery trajectory. We can model the arrows as projectiles under gravity to get meaningful results. I need a few parameters, right? So instead of it basically saying like, hey, we can do that. But first, let me warn you, you need to know this and you need to know this and here's the thing and blah, blah, blah. It's going to stop doing that as much and just go straight to like, here's the answer you asked for. I won't ramble. I will ramble, but ChatGPT won't.

I can't cut Matt off sometimes.

No, I'm just kidding.

There's no setting for that. I thought before this, Matt, that there was a way to basically have ChatGPT just like cut right to it.

Well, you could kind of do that with system instructions, right? You can go into the system instructions and say, hey, just get to the point. But it doesn't always follow system instructions, right? So this is basically like a fine-tune of their existing model. Like they didn't go back to the lab and pump in a bunch of training data into their training computers and train a brand new model from scratch to get this. Let me break this down real quick. Breakdown time. So when it comes to training a model, there's really three phases to training a model. You have pre-training, right? Pre-training is when you, scrape the whole internet for all the data you could possibly get, and you shove it into a model and say, like, find the patterns among the way all of this text works, right? That's pre-training. Now, if you just took a model that only went through pre-training and you went and gave it a prompt like, what is the color of the sky, right? Instead of like responding with the sky is blue, it would maybe then respond with the what is the color of the grass? What is the color of the ocean? What is the right? It might just try to match the pattern. What is the color of the sky? Oh, the pattern is they're asking what is the color of X and it will try to match and repeat that pattern. That's what would happen if you only did pre-training on a model. After you do pre-training, the second phase of training a model is called fine-tuning. Now this is the phase where you actually teach it to be helpful. So they actually go back through and they essentially teach it like, when I ask this question, the pattern should be a response to that question, not follow the exact same pattern and ask another question, right? That's what happens during fine-tuning is you actually sort of teach it to be helpful, right? And then after fine-tuning, the third phase, and this is the final phase before a model actually goes live and into the wild, is called RLHF, reinforcement learning with human feedback. And what this is, is where humans actually start asking it questions and getting a response. And then they grade the response. They said that was a good response. That was a bad response. Right. And so they actually start to steer it in the direction of the responses they want. This is how guardrails get baked in. This is how they say, you know, don't answer questions about how to make nuclear weapons, don't answer questions about how to make your own napalm, things like that, right? It's during that reinforcement learning phase where they add guardrails and they teach it better and better responses. All right. So now that I've given all of that context, when we see these new models come out, you know, GPT-5.3, GPT-5.4, Gemini 3.1, a lot of these models, they're not going back and doing that whole pre-training phase again. Where they scrape the entire internet, pull in all of this data, train it to learn on patterns. Because that's the portion of the training that takes tens of millions, if not hundreds of millions of dollars. It literally takes like months to train often. You know, it can take up to like 2 months to train one of these models, just like the computer running nonstop to do this training. So these new models, when they come out, they're not using that phase, right? They skip past that phase. They've already done that pre-training phase of sucking in all of the data. And that's the reason why when you go and use some of these models, it might say like our training data is only trained through July of 2025. And you're like, well, it's March of 2026. Like, why aren't you just adding the new stuff in there? It's because that would require going back through that entire pre-training process, spending hundreds of millions of dollars and months to train this new model. Right? So when these new models come out, they do a few things to get these new models out. They either, do a new fine-tune on them where they go back through and they, you know, teach it how to be more helpful, or they do another, like, additional reinforcement learning paths, right? And so when a model like 5.3 Instant comes out, they most likely just went back through and did that reinforcement learning phase, that third phase where they just gave it prompts, got answers, and said, eh, don't answer like this, answer like this. Don't answer like that, answer like this. Right. And they probably just did that reinforcement learning phase to try to tune it to get the answers a little closer to how they want it to answer. And that's what these 5.3 instant models seemingly did. They just went back through and said, we hate it when you preamble, get rid of some of that shit, just go straight to the answer. I like it.

So there's a reason why there's the thumbs up and the thumbs down on everything is essentially this.

Yeah, but I don't believe it's actually training on that.

Well, it's maybe gathering feedback though, so it can prep for maybe future model.

Yeah, yeah. It might be gathering some of that data for future training runs.

Right.

But it's not like actively training the model in real time when you do that. It's not like you give this answer a thumbs up and now in the, like we were talking about the em dashes before we hit record, right? It's like you can get a response that has em dashes in it and give it a thumb down. And then when it asks, why didn't you like this? You say, well, you put em dashes in it. Well, next time it's not going to like, not put em dashes. It's just kind of collecting that information so that in the future, if they did a training run again, they can use some of the feedback that humans have given.

Mm-hmm. That makes sense. So it makes sense also why we're probably seeing so many models get pushed out because like you said, there's 3 phases, but a lot of these, not saying all of them, are probably more of the fine-tuning and then this whole reinforcement layer.

Yeah.

I mean, just because it's cheaper. It's faster to push out and it makes the experience better.

Yeah. Like if they were going to go do a whole new training run and basically start from scratch, that's usually when we're getting like a full number leap, right? Like that's the leap from, okay, we had GPT-3.5. That's what everybody was using when ChatGPT launched to now we've got GPT-4 and this feels like a big leap. That was because GPT-4 was a whole new training run. They went back and did the whole pre-training from scratch and basically built a new model. Pretty sure that's what happened with GPT-5 as well. That was a whole new training run where they went back, did the whole pre-training, did the whole process from scratch and did a whole new run again.

When's the last time they pre-trained? Yeah, 5.3, but it'd be like the 5 series basically.

Most likely when it's like a jump in number, that was probably a new training run. Most likely. Again, it's kind of like the Google algorithm or the YouTube algorithm. They don't openly give out their secrets of like, here's what we did to get this model to do this. Like they don't publicly share that stuff.

Yeah.

All we can kind of do is speculate knowing what we know about how large language models work and go, okay, this is probably what they did.

Yeah. I did just ask ChatGPT 5.3 Instant, when was the last time you did a pre-training run on this model? Says, I don't have public visibility of this pipeline internal logs, but what's publicly known is the knowledge cutoff of this model was August 2025.

Yeah. Yeah. So that was probably the last time they did like a full scrape of the internet to do a pre-training run. So even if it was like the cutoff was August 2025, that means the model was probably a couple of months even after that, right? Because they had to scrape all of this stuff, start the training run. Once the training run started, they're not shoving new information into it.

Correct. Yeah. Because after the models, it says it goes through post-training steps, which is exactly what you just described. Fine-tuning, alignment, safety training, evaluation. And then it rolls out into public, basically. Yeah. Updates.

And then there's also another step that sometimes happens called distillation.

Oh, right.

Which is basically when they train on the outputs of another model. Let's say you've got GPT-5.2, but then also OpenAI releases GPT-5.2 Mini and GPT-5.2 Nano, right? These models that are basically 5.2, but they're much faster, cheaper, smaller models. Well, why is it still 5.2, but it runs faster and you can actually use it cheaper? Well, it's because these models are most likely distilled models where they went and they basically ask ChatGPT 5.2, like the full model, a question and get a response. And they take all the questions and responses, you know, tens of thousands, hundreds of thousands of these questions and responses. And then do a training run on just those questions and responses. So it's not a whole scrape of the entire internet, but it's like a new training run on just the prompt-response pairs. Right? And that's how they get models that are 5.2 sort of level, but smaller and faster. They don't have as much like nuance and detail baked into the model, but they can still respond very similar, because they were trained on how those models would respond.

Interesting. Okay. I like this breakdown. Yeah.

Little history lesson on how large language models work for anybody who is listening and you were curious, how do these training runs work? Well, those are sort of the phases and that'll give you some insights into like, okay, why are they just doing 5.3 instead of GPT-6? Why isn't every single month we're getting GPT-6 and then GPT-7 and then GPT-8? Like, why is that? Why are we getting 0.1, or why are they jumping it by just like percentage points? That's why, because they're not full brand new training runs, they're just sort of optimized during the sort of post-training processes.

Yeah, that makes sense.

So all of that to say, this 5.3 Instant model is pretty much the same model you were getting out of 5.3. They most likely just did some additional reinforcement learning on it to get it to respond a little bit better. And according to Sam Altman, a little bit less cringe. Yeah. See right here, a little bit less cringe. So that's basically 5.3 Instant. It's just going to be a little less cringe.

That's all you needed to know. Well, let's talk about the other model that released. What was like the next day, right?

Two days later. Yeah. So this one, the 5.3 Instant came out on March 3rd. And then the GPT-5.4 came out on March 5th. So one came out on Tuesday, one came out on Thursday.

Okay. So what's the deal here for us, the Joe Schmoes, or maybe that's me. What's the point of this one? Because I know this one's actually going, it's more of an intelligent, deeper model, right? Like it has a lot more thinking around it. The tests I've done, it's just a lot more comprehensive in the way it gives you answers.

Yeah. Yeah. So this one, the people who are probably going to feel it the most are going to be the developers and the people that are like sort of baking AI models into their apps. So, you know, I just broke down those various phases. This one isn't going to be a brand new model run either, but they probably did a fine tune on this one where 5.3, it was most likely just like an extra reinforcement learning layer where they were trying to teach it to be better. This one, they probably did a whole like fine tuning training run to change the way the patterns actually happen when you prompt it. So there's some pretty cool differences. Probably the one that's the most important is the million-token context window. It supports up to 1 million tokens of context, allowing agents to plan, execute, and verify tasks across long horizons. Now, this is something we've actually had in the Claude models. We've actually had these long, huge context windows inside of those models. Now our OpenAI models get it. The thing is that 1 million-token context window, you're probably not going to feel it in ChatGPT. Like ChatGPT is going to be a much smaller, limited window. That you can see the 1 million of context is going to be available in the API. So in Codex and the API, that's where you're going to see that big update. So you've got the 1 million token context window, which is great for coding because when it comes to coding, you could stuff entire codebases in, right? You can take a big app that you built, take that entire codebase, and it could read the whole thing and respond with a prompt. It's just going to have a lot better understanding of the code. Plus your conversations that you're having while doing coding in something like, let's say, Cursor or OpenAI's Codex platform, you can actually have much longer conversations before it has to compact the conversation or you have to start a new chat. Now, when it compacts a conversation, basically what it's doing is it looks back through the last conversation that you had, all of the chats that it had, and then it uses its own AI model and sort of summarizes it down into like a paragraph or two of like, here's everything we've been talking about. So that it can kind of continue the conversation. And it just has like a summary of the previous chat so that it can kind of remember what you're talking about. But obviously there's problems with that. Let's say you have 200,000 tokens worth of conversation and it compacts it down into like a 2-paragraph sort of synopsis. A lot of context is going to get lost, right? And so like, if you've ever done any sort of like vibe coding, you've probably noticed you'll be having a long conversation, building something, building something, building something. All of a sudden you open your app and it reverted something and some feature that you built into the app is all of a sudden gone and you're like, why did it delete that? Well, because it lost all of the context it had early on when you were telling it why you wanted that feature in the first place. It compacted down into like 2 paragraphs and that specific feature probably got excluded from the sort of compacted summary, right?

Yeah.

So the reason developers love bigger and bigger and bigger context windows is you can go longer and longer and longer in your chat before it has to do that compaction process.

That makes sense. And that's what it's kind of saying here. I'm asking the model itself some of these questions and it basically says it's better on real work and messy work as well because you can throw so much context at it. And to your point, yeah, the context window is so much bigger. So it's going to help with any of the coding, understanding a full-on like basically different documents you might shove it or different tools that somehow it's attached to. So context window, it looks like, yeah, that's the biggest piece. It's great for agents, which actually was a question I had is the whole OpenClaw thing. We talked about it, what, 2 weeks ago? Is Peter Steinberg working with OpenAI now?

Steinberger?

Steinberger. Now you got me thinking of last week, BK Lounge. So callbacks. But yeah, like, how do you think it plays with maybe what OpenAI is thinking when it comes to agents, the OpenClaw thing?

Joe, that's an excellent question. And in order for me to answer that, I need to give you a well-thought-out response. So I'm going to go ahead and answer that right now.

Okay. We'll come back to that later in the episode. No.

So a few other features that came out in this model that I'll point out that this will get to like a response to your question here in a second.

Yeah.

But like a few of the extra features that came out with this new model is computer use and vision. So this is actually something that's now native to this model. It can actually go and sort of click around for you and actually use a computer on your behalf. One of the examples they have, it's actually helping draft an email for somebody. So it can actually go and use your computer on your behalf now natively in the model. So, OpenAI did have a computer use model, but it was a separate thing. You had to specifically open up the computer use model, tell it what you wanted to do, and it would go and use a computer on your behalf. This new model, it will do it within the sort of flow of you using the model. It will just go and open up like a virtual computer and go and do things for you now. Right? So that's one of the features that they just added in. Another one of the features that they added in is the improved tool use. So they added this new thing called tool search. ChatGPT has been able to use tools for a little while now, right? It knows that it can go and use the HubSpot to do something. It can go and use Slack on your behalf. It can go and write code on your behalf, use different like calculators online on your behalf. It's able to go and use tools and it's been able to for a while. The problem was it would sort of put all of the information about all the tools that it had available to it. Inside of the prompt. And so you would prompt it to go do something. And in the prompt, a portion of the prompt that you wouldn't see that happens behind the scenes, it would say, here's all the tools at your disposal, use what you need to use to get me the response to this. Well, all of that data about all of the tools that are sort of behind the scenes there, that all goes into your context window. That's all using data within your context window. Well, this new tool search, is instead of like giving it all the details about all the tools it has available to it, it now has like a search functionality where when you ask it to do something, it doesn't load all those tools into the context. It goes and does a quick search of what tools do I have available for me? And then it picks the right one to use and then uses that tool. That's my understanding of what it's doing, but it saves you on that context window. So it's not loading a bunch of information about all the tools you have available to you. Inside your prompt, it's only searching out the tools it needs at the time that you prompt it, saving you that context window. And you can see right here, example token savings from tool search, upfront input tokens, input tokens from tool outputs, output tokens. So with the tool search, you can see that this is the upfront tokens of actually going and finding what tools it needs, where in the past it would use, you know, upwards of 60,000 tokens giving all of that data of here's the tools we have available to us.

Yeah. Question on this, like with Claude, I've been using that primarily more than ChatGPT, but like it will pull up different skills and search the different tools. It seems like it has some similar things happening here. Is that accurate?

Yeah, I don't know if Claude has like this tool search feature. I think Claude is actively still sort of injecting into the prompt Here's all the tools you have available. Use what you need to get me this response, right? So you are probably still using a lot of those tokens when you're using Claude and it goes and uses the tools for you. But coming back to like the Claude bot thing, right? Obviously Peter Steinberger went to OpenAI, so he's working with them now. OpenAI, you know, they want to get to this like agentic thing. Like they want to get to a point where they have their own version of like a Claude bot, obviously. And so a lot of this kind of stuff feels like it's designed for agents to use. Like one of the biggest sort of complaints people have about OpenClaw is if your OpenClaw is just running behind the scenes doing all these tasks for you, it's using tokens like crazy. And if you're connected to your API, your OpenAI API, your Claude Anthropic API, it's just racking up costs because it's using all of these tokens and it's constantly doing stuff for you. So this new tool calling, this tool search feature, is designed so you're not going to be using nearly as many tokens when it's figuring out what tools to go and use. Also, as we know, like OpenClaw is designed to go and do stuff on your browser on your behalf and go and do things. Well, if we look at this new model that they just released, it's actually designed to use a browser and use your computer and go and do things on your behalf, which leads me to believe that a lot of what they baked into this model is because of the rise of OpenClaw and this sort of like agents that are doing things autonomously for you. So to me, this feels like something that they're building to make your agents a lot more efficient versus something that they built to make ChatGPT a lot better for the normal user.

Got it. Yeah. And I'm in there right now. The computer use thing is interesting here. What do they call it? Agent mode? I have the paid account. Yeah. What do I have?

Probably plus.

Plus. You have 40 uses, it says, of the agent for the Plus account. So I'm assuming most people, a lot of people watching probably have Plus. I don't know if it's available in the free account, you know, the agent thing.

5.4, no, 5.4 is not available in the free account.

Gotcha.

You have 5.3 Instant, the one we were just talking about in the free account, but not 5.4. 5.4 is only on Plus and above.

Okay.

So basically the $20 a month and up plans.

Yeah, basically it says like for that whole computer use agent thing, best way to use it, it's like any, give it outcome-based tasks that you want done. Be specific about constraints, watch the narration so you can actually see what it's doing.

That's another thing actually, you can do that in ChatGPT.

You're right.

So check this out. This is actually a new thing that you can actually start to do where you can say, hey, give me a list of 20 movies that came out in 2025. Right. I can do that. And then I could say, actually, wait, just give me 10. Right. So like it'll start responding and then I can, I didn't do it fast enough, but you can actually stop it in the process. So you can do something like give me 20 movies that came out in 2025. That could win Oscars. Make it think a little bit harder. So I can give it a prompt like this. And then I can say, actually give me 10. And I can submit that. And even though it was starting to respond, I can actually stop it in midstream now and redirect it. So that's actually a new feature that they just added on the 5.4.

That's pretty cool. Yeah. There was like a little blurb when I was first using it said that you could do that. So you don't have to fight with the stop button and then go and edit. And you know, it's like you just go stream of thought. Boom. So I don't know. I think like the computer use thing is always cool. Agents, test it out. The fact is like, hey, plus user, if you're paying for it, go try it out. You got 40 uses. I don't know if that's like sessions or like prompts. It's probably a session, however they do it.

But I'm not actually sure. I'm on the pro plan and I don't know if I have limits because it doesn't tell me I have limits.

Okay.

I think all the plans kind of have limits somewhere. But it's probably further than I can push it to find the limit. Yeah.

Yeah. Okay, cool. Well, good model.

One more model. We don't have to linger on this one too long because I know we've already sort of gone in depth on a lot of these. Gemini 3.1 Flashlight. So this one is Google's latest model that they released this week. And this one is designed to be very fast and very cost efficient. Most likely what this one was is if I had to guess, like a distillation of, you know, 5 of their 3.1 main model, right? Like a lot of times when you see these 3.1 Flash or 3.1 Flash Lite, like these models that are named based on the bigger model, but have Flash or Lite or something like that after them, it's usually like a distillation of the model where it's, they took a whole bunch of inputs and outputs and also the chain of thought, right? Like, so, You know how models these days, you can give it a prompt, you can kind of see it's thinking, and then you see that final output. Well, that's all good information for doing a distillation as well. You can give it the prompt, the chain of thought, and the response, give it all three of those things for all prompts you give it. And it learns how the bigger models actually think and learns to sort of respond in a similar way. They're not as smart. They're just sort of trained to reply in kind of same way that the bigger model would reply. And so to me, that's most likely what happened. Again, a lot of these details aren't shared publicly of like how they get these new models, but this one is most likely a distillation of their bigger Gemini 3.1 model. Basically just meaning that you get similar sorts of outputs as their bigger 3.1. You just get them faster and cheaper. And sometimes it's not going to be as smart because it doesn't have nearly as much nuance sort of baked into it.

That makes sense. Yeah. And basically it says what it's kind of positioning it as the cheap and fast workhorse model. Yeah. Low cost, high speed for whatever you're trying to do. I think it's, I saw translation. If you're moderating content or reviewing things, tasks like that, that aren't as deep. Yeah.

And it's good for that kind of stuff. Like I actually really like using the Gemini Flash models when I'm like building my own apps. I actually code using either Opus 4.6 or 5.3 Codex from OpenAI. Now I'll probably start using 5.4 because it got slightly better at coding, but like very marginally better. But I'll probably start using GPT-5.4. I actually kind of like OpenAI's models a little bit better for coding lately. I feel like they do a little bit better job than Opus, which is sort of a new development. They keep kind of going back and forth between which one's better at coding and I think right now OpenAI actually has the edge a little bit in actually writing the code. Now, writing the code, I like using those models, Opus and OpenAI.

That's interesting because we haven't even talked about that. And I don't know if we have time now for it, but yeah, the whole, your switch from Claude to more Codex.

Yeah, I switch back and forth, honestly. I will use them interchangeably. Every once in a while, the GPT models, I'll try to get it to do something and it doesn't come out quite the way I like it. So I'll go to Opus and have Opus do the same thing. It might get me a better result. So I will actually flip back and forth. But I would say as a whole, 5.3 Codex has been a better coding model for me than Opus 4.6 has.

Interesting.

Now saying that, when I'm building apps, I'm building a lot of apps that use AI within the app. So I've got this app that I created. This is my like swipe file app that I built here. And basically when I come across thumbnails that I like, that I'm like, ooh, that's a good one. I might use that for inspiration later because That one's doing well for that channel. You know, it's a big outlier or whatever. Well, if I click into this one, right, here's an Austin Evans video. See all of this text here. All I did was drag and drop this thumbnail into my app. I found a thumbnail, dragged and dropped it in. That's it. Everything else you see here was done by AI. So it pulled in the title of the video. It pulled in the channel name, how many views it had at the time I took the screenshot, a quick description of what is actually in the thumbnail. A breakdown of why it works, and then the overall score. It gave this one a 9.4, and it based it on like these 5 criteria. All of this like code that it's writing when I drag and drop an image in, and it sort of analyzes that image real quick and then gives me all of this stuff back, this is all using this new Gemini 3.1 Flash model because it's really cheap, right? Like every time I pull in an image, it's going to cost me like a fraction of a penny. For every image I pull in and it's almost instant. I drop an image in and like, boom, like within a second, all of these details are filled out.

It's a question. So that's a screenshot. You're not actually bringing in the YouTube video itself.

No, it's a screenshot. And the reason I take a screenshot is because I want to pull in the title, the view count, all that kind of stuff as well.

Like the packaging of that video, basically.

Yeah, I want to pull in the full packaging. So I just, I do this screenshot that captures all of that and then pull it in. And then I've also started doing it with Shorts, but I don't do a lot of Shorts, so I haven't pulled too many in, but same idea. And then I also pull my own videos in and have it analyze those so I can figure out how to improve them later. And then I built an analysis tab here where I say, hey, I just made a video about something. And then based on all of the thumbnails I saved, it tells me, well, based on the thumbnails that are working right now, based on the current meta of all the stuff you've saved, here's what I think your thumbnail should look like. And it actually gives me, a breakdown of what I should make for a thumbnail based on what's worked on my channel and what I've recently pulled into my swipe file.

Nice. So to roll it all back, what you used is one of the Flash models. It's what's coming up with that material right there.

So it's basically— yeah. So basically I will use the best Claude models and the best, uh, OpenAI models to write the code, but the apps themselves, the AI that's actually being used within apps like this, I don't need the best model in the world that's going to cost me, you know, $25 per million tokens to tell me what's in an image, to tell me this image has HTTP crossed out and a dude's face with a silly look on it, right? Like, I don't need the best model in the world to tell me what's in that image. 3.1 Flash does it, and it'll cost me a fraction of a cent every time I do it.

Cool. Yeah, that makes sense. So that goes back to like the different use cases Google kind of said. So things like analysis. Easy analysis like that. Images, content analysis, translation, maybe UI generation, maybe. I don't know. Like it can kind of help out with that based on what it is. So that's cool. And then yeah, do whatever the Codex model, Claude model for the actual coding.

Yep. So I mean, 3 new models came out, 5.3, 5.4, and 3.1. All of them pretty much completely different use cases if you're in AI., right? The Instant model is just designed to give you an answer really, really quickly. The 5.4 is seemingly a model that's designed to be used by agents, but it will also think a lot longer. I've seen people go and ask it questions and show off that, oh, this actually thought for 80 minutes on this question, right? Like you could get it to think for a really long time to give you a really in-depth answer. And then you've got 3.1 Flash, which is a great model to use inside of the apps that use AI. Maybe not the best model for writing the code for you, but one of the better models for just doing a simple like AI application inside of an app that you built.

Got it. I like it. Thanks for making this. I won't say dumb it down, but you distilled this to a model that I can—

That's what I'm doing. I'm distilling my knowledge for everybody else. You're prompting me. I'm giving you an output. And that is the distilled version of what's actually going on here.

So here, I don't know, like I'm going to speak to the viewer now. You, you listening, watching, go into the comments and because I am Matt's prompt a lot of the time. So what do you want me to prompt Matt maybe thematically in future videos? That's just a little, you know, throw it in the comments. I want to see what you say. So moving on, this is also something we've both been kind of talking about a lot is generating really cool videos with AI. The lines are blurred. Learning every day. Now, one of our favorite tools, and I think a lot of people's, NotebookLM. They just dropped another big old update. Big old update. So basically, if you are at all interested in generating video, if you've used NotebookLM, now it's literally cinematic quality. And it's ready for you to kind of play with.

Yeah, so the new update from NotebookLM, it's a big update if you're on the Ultra plan, right? So if you're paying $250 a month, to get access to Google's Ultra plan, you can use this new feature.

So on that, I don't want to just discount that because like, I mean, $250 is pretty damn cheap.

It's a car payment right there, Joe.

It can be. Or it also can be like, you know, if you're an agency or if you're creating your own content or whatever, it's like, that's an investment into your business. Yeah. And that could be paid off like that. So like for the Ultra really fast. So like, cause I don't really know this, like what other like key perks do people get with an Ultra plan?

Yeah, so the Ultra plan is really going to give you all of Google's like state-of-the-art models, right? So like it's the equivalent of the Pro plan on ChatGPT, the Max plan on Anthropic. You're going to get all of the newest features first. And then over time, those features sort of trickle down to the lower tier plans. So like when you're on the Ultra plan, you get this feature in NotebookLM that lets you make cinematic videos. What is, what does that actually mean? Well, In the past, when you use NotebookLM, you can have it generate videos and it would generate videos that have like nano banana images in it. It kind of feels like you're watching a slideshow where there's somebody talking over it, but like switching to the next slide as they're talking. That's what you would normally get for a video from NotebookLM. Well, now it will actually use VO3.1 and actually animate some of the video. So that's one of the features, but you're also going to get access to Google's latest models, right? Like, you're going to get the latest Gemini DeepThink model that can really think for a long time and give you the best possible outputs. You're going to get a lot more usage on Google. If you are using Nano Banana, the Ultra plan is the only plan that doesn't generate images with a watermark. So if you ever notice, I generate images with Nano Banana. There's no Gemini watermark on them. That's because I'm on the Ultra plan. Um, so like there's a handful of benefits. I don't know them all off the top of my head, but the ones that I use are going to be the NotebookLM features, the testing the absolute state-of-the-art latest models out of Google. And the fact that when you do generate videos with Veo or you do generate images with Nanobanana, they're not watermarked at all. They just, they just come out without the watermark on them.

Yeah. $250 a month. It ain't cheap. Obviously it's not like you're just going to jump on it like $20 plan, but if you could justify it as your business or whatever you're doing creatively, like.

Maybe it is. If I was to go to Fiverr and I was to get somebody to make an animation for me on Fiverr using After Effects, they might charge me $250 to make me an animation or two inside of After Effects, right?

Just paid for your account, Matt.

There you go. With like a Google Ultra plan, which is $250 a month. If you're doing this After Effects style stuff with it and you're only using it once or twice a month, you're probably getting your money's worth just to get some of these types of animations that you can then use in your video. Now, I actually made a video yesterday. So have you ever heard of the birds aren't real conspiracy theory? Because of you, yes. Okay. So I had it make a video about the birds aren't real conspiracy theory. And this video I was actually pretty impressed by because the subject matter is a little bit more interesting than API keys and IDEs. So check this one out. This one's actually a pretty fun video. We probably won't show the whole thing.

Yeah.

If you look up into the sky right now, you might see a pigeon. But hundreds of thousands of people across the country will tell you that you are actually looking at a highly advanced, government-issued surveillance drone. This is the Birds Aren't Real movement. Its followers drive vans covered in manic scrawlings, they fund massive billboards in major cities, and they march through the streets in coordinated protests. This chart shows the movement's foundational timeline. According to their official lore, the U.S. government systematically eradicated roughly 12 billion living, breathing birds between 1959 and 2001, and they replaced them with exact robotic replicas. They have an explanation for everything. When birds perch on power lines, they are actually recharging their internal batteries. And when a bird drops waste on your windshield, That is a liquid tracking device designed to monitor your vehicle. But here is the catch. The entire movement is a completely fabricated satirical performance art project. Nobody actually believes the birds are robots. So why are hundreds of thousands of normal people actively spending their time, their money, and their energy to roleplay a lie they know is entirely fake?

Okay, we get it. We get it. But I mean, some of these animations here, right? Like where the guy's holding up the bird and laughing. And did you also notice very consistent looking character throughout? But I mean, couldn't you see like an actual YouTube video using some of this style of animation? I was thinking about that the whole time. It's always kind of jarring to me when it goes from like animation to just like a still nano banana image. But like the animations they do use, like I feel like you can go make a whole video. Like let's say I scripted a video, recorded the whole video, and then I could take the transcript from that video or the original script from that video, give it to NotebookLM. Have NotebookLM make its version of the video based on the script that I wrote, and then actually pull some of the videos and animations, like some of this kind of stuff, into my real video when I go and edit it, you know?

People would watch that. I was thinking the same exact thing. I'm like, there have been worse videos on YouTube that explain a topic or a trend or the news or whatever.

Like, this is basic After Effects right here. This is something that, I mean, somebody that knows what they're doing could probably do in like, you know, 20 minutes in After Effects. That's 20 minutes. But hey. That's 20 minutes saved.

20 minutes to render the whole thing, probably, right?

15. This is some cool animations. I also don't know if you realized, but it was like showing real images and it was putting the source for the real image down in the bottom left corner of the video.

I noticed that, which I'm sure they're abiding by all the rules. So there you go. Your stock photography is taken care of too.

Like stuff like this, you would see in a normal YouTube video like this. It's not quite to the same polish level as like a Cleo Abram, but like it's the type of animations you would see, just not quite as dialed in.

The text isn't perfect if you look at some of the spacings. I mean, just like with anything, you're going to have some little things, but it's good.

Yeah, but the thing for me is it's less about where we're at right now and how quickly we got to where we are right now. I was making videos 3 years ago when things like ModelScope and ZeroScope came out and you would prompt it for a video idea., and the video would kind of look like a blurry, like if you squint, it kind of looks like what I prompted. And it would only be like a 2-second clip. It would just be like a little blip and then the whole video would be over and it wouldn't really look like much. Here we are 2.5, 3 years later, and it's generating what After Effects might make. This is generating what you might pay somebody like $100 to do in After Effects for you. To me, how quickly we got from monkey that looks indistinguishable on the video and only does 2 seconds to holy crap, we're generating like animations of like this dude going crazy in his room, you know, like with a voiceover.

And also it can pull in all these other facts, like even the real news, you know, like up-to-date information. Yeah. You only fed it one video, but you could have fed it all sorts of sources.

Yeah, yeah, exactly. So I gave it a whole bunch of articles about the Birds Aren't Real movement, you know, subreddits and various news articles about it and the Wikipedia page about it. So this one actually was fed a whole bunch of different sources on the topic so it could kind of pull it all together. There you go. But like, to me, it's less about like looking at this and going, this is the best thing I've ever seen. It's more about comparing it to where we were just a couple of years ago to where we are today. That leap was insane. And the pace of things happening is only picking up speed. So like 6 months from now, we're going to be looking at this is generating Cleo Abram videos for us. This is generating Veritasium, Mark Rober videos for us that have their level of animation inside of them. Like it's going to get there pretty, pretty quickly. It's, is it there yet?

No.

But if you look at the trajectory of how quickly we got to this point, not far off.

I'm curious, are you able to in NotebookLM, like have a consistent style, like a brand style or anything like that? Or is it just kind of at the whim?

Well, you can give it some custom instructions. I don't know how effective they are. Like if I go to video overview, you can give it some custom instructions down here. I've done some custom instructions and it didn't always follow them perfectly. So I don't know if you can really, really get it dialed in. You know, if you set it on explainer or brief, it's going to use Nano Banana instead of VO 3.1. And when you set it on one of these, you actually have style options. You can give it a custom style, whiteboard style, kawaii style. You know, there are styles, but it doesn't give you that option on the cinematic. I'm sure that'll change fast. Yeah. So that'll probably get dialed in in the future because at the end of the day, like most of what we saw on there, they probably generate like a Nano Banana image as the starting prompt and then animate that image. That way, like the video feels like it has that consistent style. Like there was like a style to that video, right? It all had that sort of like pointillism sort of looking images to it.

Right.

Yeah. So like there was a style to it. So it probably started with like nano banana images for the whole thing. And then it shows some of those nano banana images to Animate. I think so. We will be right back, but let me quickly tell you about another podcast. Content is Profit, hosted by Luis and Fonzie Camejo. It is brought to you by the HubSpot Podcast Network, the audio destination for business professionals. Discover the secrets and strategies on how your business can achieve the frictionless sale. They talk about frameworks, strategies, tactics, and bring on special guests to bring you all the information you need in order to turn your content into profit. They just did a great episode titled The platform trap that's blocking your sales. Listen to Content is Profit wherever you get your podcasts.

I mean, who knows who will be the king, the top of the mountain of AI, but like Google, the way that they integrate things like NotebookLM with Gemini as well. I don't know. You can like leverage what you just did here, put it over there and it's all integrated.

Well, yeah. I mean, this uses Gemini behind the scenes, right? So we've got all of these details that we pulled in about birds aren't real. I can say, write me a 5-minute script that explains the birds aren't real story, right? And I could give it a prompt like that and it's going to use Gemini plus all of the sources that it's grounded on. Like I don't even need to jump to the separate Gemini app.

I could just do that right here. Yeah, I guess it's interesting because in Gemini you can pull in a NotebookLM notebook.

Yeah, I mean, maybe it can do some of the like coding stuff. I actually don't know specifically which model this is using. This may not be the most up-to-date model. This could be using, you know, 3.1 Flashlight to generate this. And if you go use the Gemini app, it might be using Gemini 3.1 Pro, you know, so you might get different responses, you know.

That's cool. I love NotebookLM, ever evolving. Keep your eyes here. Maybe it's worth the $2.50 a month.

I don't know. Yeah, I mean, it's up to you.

It depends on what you're using it for. Yeah. Okay, moving on. Here's something that I think is very practical as well. And it's something I wish I had a couple of months ago. I've started to use Claude just in general for everything. You know, there's just, it's a lot of cool stuff. Needless to say, there's been a lot happening with Anthropic and Claude. We talked about it last week. That saga continues, but it's interesting is there has been a big shift, like a big kind of mass exodus, if you will, from ChatGPT to Claude. And there's been some, I don't know if we have the exact link of all the download numbers or the uninstall numbers. Uh, there was something published about that over this last week, seemingly because of the whole Pentagon Anthropic thing that's been ongoing. But what's interesting is that Claude made it very simple for people to switch from ChatGPT to Claude, where it literally can pull like the memory and all the work that you've been doing from ChatGPT's side and easily plug it into Claude and just hit the road running like immediately. There have been ways to do this. You can export. I think if you're on the business plan in ChatGPT, it's a little bit more of a manual effort. They don't make it as simple to extract your stuff. Stuff.

Yeah.

But like, I personally did this with my team just a couple months ago and it wasn't as simple, but now it's literally like even TechCrunch, like you have up here, is like showing how to do it and making it like very clear, like, hey, do that.

Well, let's take a step back real quick too, because I think we have to add some of the important context here of what happened last week, right? So last week we spent a lot of time talking about the whole Anthropic versus Pentagon thing. And I think that plays a lot into what Claude's moves were here. Right. So I don't really want to get too deep into the weeds here, but the TL;DR of this whole thing was Anthropic had a military contract. Anthropic heard that the Pentagon used their tools for the Maduro raid. The Pentagon got pissed off that Anthropic even asked, what are you using our models for? And then they basically told them like, stop being woke and all this kind of weird stuff. And then Anthropic wrote two red lines into their contract basically saying you're not allowed to use our tools for autonomous weapons without a human in the loop, and you're not allowed to use it for surveillance on US citizens, right? And the US government basically said, well, you don't get to tell us what we can use this for. And Anthropic says, use it for whatever you want, just not those two things. And then the government basically blacklisted them and said, fine, you can't do business with us and you can't do business with any of the companies that work with us either, right? Like that's kind of the TL;DR of what happened. Obviously it's a lot more long and drawn out and a lot more nuanced, but that's the gist of it. He nailed it, I think. Yeah. So what happened since our last call where on our actual podcast, we discovered that that blacklisting happened because Nano Banana told us about it. But after we found all of that out on that same day where it was announced that they were blacklisted, OpenAI swooped in and basically said, you can do a contract with us, right? And so when OpenAI came in and basically made a deal with the Pentagon, everybody took that to believe, well, OpenAI is going to let the Pentagon use their model for whatever you want, meaning they're okay with surveillance and they're okay with autonomous weapons without a human in the loop. And OpenAI came out and said, no, we wrote the same red lines into our contract and told the Pentagon they couldn't use our stuff for that too. And so everybody's like, well, What? How does that make any dang sense? Right? And so that's kind of what happened. And I think OpenAI lost a lot of trust from a lot of people. And a lot of people went, there's something going on here where OpenAI is probably giving them more access than what they're actually claiming they're giving access to. And as all of this was happening, Anthropic went, hey, here's an easy way to migrate off of other platforms over to us.

And that's where we're at now. It's like, hey, you want to migrate? All right. How do you feel? I, you know, a lot of people like Claude for whatever reason. If it's not the political side, it's just like, it's co-work. You have Claude Code and all this other stuff. You know, it's really cool. I love using Claude, use it for projects and stuff every single day. I haven't actually used this process, but it's very clear now how to do it. TechCrunch, you showed, was even like showing how to do it. So that's kind of interesting that they're getting in the middle of that too.

Yeah, Claude basically put up this way to like migrate from OpenAI over to Claude. They said, hey, bring all of your memory, all of your preexisting context over. You can literally pick up where you left off on your conversations inside of Claude now. So if you had ongoing conversations inside of ChatGPT, you just bring them over and keep going with that conversation with Claude. And so what happened out of that was uninstalls surged by 295% of ChatGPT. People just bailed on ChatGPT like crazy. And this also happened. Claude became the number one most downloaded app in the App Store, topping out ChatGPT, topping out Gemini. It was the number one most downloaded app. And to give some context on that, Claude wasn't even in like the top 10 before this. Like they were like way down the ranks, like maybe top 100, but not even in the top 10 or 20. And when all of this stuff started happening with the Pentagon in like this show of solidarity from people, they all started jumping on board Claude and Claude became the number one most downloaded app.

It's wild, man. I mean, I just think back to the Super Bowl, you know, and like Claude was like running the commercials. I think we showed some of them. I'm just like, I feel like that was the point when people started to realize like, who's this Claude? Who is this Anthropic on, on scene now? It's like, Oh, now people know. People are downloading and taking action. There's also like Claude did a move. The fact that they put their memory feature now on the free plan of Claude rather than keeping it to the pro plan.

That's another thing that they did was they made it so like the memory feature where it remembers your past conversations, that was only on their paid plans. Now it's available on the free plans as well. So they're just trying to entice people to come over to Claude. They're not flat out saying, you know, migrate from ChatGPT to Claude. No. But I mean, that's basically what they're doing. They're saying, hey, we're going to make it really, really easy for you to switch over to us. Yeah. I will give a little disclaimer. If you are going to switch, if you're on the free plan of Claude, you get a lot less usage than if you're on the free plan of ChatGPT. You will run into the rate limits if you are like a power user. On Claude, um, where on ChatGPT you're much less likely to hit the rate limits. So I do know that people have been saying we moved over and then got rate limited after like 20 messages. So something to keep in mind if you are going to switch, like you probably can't chat as long with it, especially on the free and lower tier plans.

It does feel like you hit rate limits faster on Claude than ChatGPT.

Well, Take it for what it's worth there. I also found this interesting real quick too. I just want to share this graph. This is like the last basically year. So February of last year, OpenAI dominated AI use for businesses. Fast forward to where we are here, February of 2026. This orange is all Anthropic and this blue is OpenAI. Wow. So over the course of a year, we can just slowly see Anthropic just take over. And that's for businesses. Yeah, yeah. OpenAI does have something like 800 million monthly users or something like that, right? So they are still probably dominant on the consumer side because they are still like the Kleenex, the Band-Aid of AI. Like they're the brand name of AI that everybody knows. But I also did want to quickly share a super quick snippet of an update on the whole Pentagon Anthropic thing. So Anthropic was officially designated as a supply chain risk by the department. So Anthropic basically said, we're going to take this to court. We see no choice but to challenge it in court. If I had to guess, they would probably win because, like, these aren't designed to be punitive. They're designed for actual supply chain risks. So, like, there's a fairly good chance that in court they're going to win. But what I also find interesting is in this thing here, he said, I would like to reiterate that we had been having productive conversations with the Department of War over the last several days. Both about ways we could serve the department that adhere to our two narrow exceptions and ways for us to ensure a smooth transition if that's not possible. So apparently Anthropic is still talking with them to figure it out. But it gets interesting because there's also this. This is Undersecretary of War Emil Michael. This is somebody that works for the Secretary of War. And he says, I want to end all speculation. There is No active Department of War negotiation with Anthropic AI. So the Department of War is literally saying, we're not talking to them. So who do we believe? Just all kind of interesting. And then also Microsoft put out a statement this week saying that, you know, if you use Microsoft and Anthropic, you don't have to worry. Microsoft Anthropic's products remain available to customers after Pentagon blacklist. So if you have a contract with the DOW and you're using Microsoft, who does use Anthropic, You can still use Microsoft because it's not sort of downstream from the way Microsoft is working with the government. Basically want to give that quick update. Anthropic saying we're negotiating with the Department of War still. The Department of War saying no, you're not. And also you still can use products if you're a company, as long as it's not downstream from an actual government contract, you can still use these companies.

Well, You know what, folks? We're probably not going to go deep on this because this will probably last for a long time.

So keep an eye on the news. I will share updates from time to time about it because I am following it closely because it's giant soap opera to me.

It's a crazy story. Soon we'll have an 8-minute version from NotebookLM that just happened.

I'm looking forward to the David Fincher-directed movie. It's going to happen.

Someone's going to cover it. I hope it's him too. Well, on that note of filmmaking, I think this is a perfect segue. I thought this was very interesting because you now got Netflix over here joining this company called Interpositive. What is interesting about Interpositive, Joe? It's a weird name. I don't know who named it. But what's interesting is that they are basically an AI media creative company built for Hollywood. But built by, built by Ben Affleck and smarter people than him is what he said. That's his quote. So I think it's very interesting that AI is now starting to be accepted by Hollywood. Probably not as a whole still. I'm sure that's going to take some time. I'm just thinking, all right, well, Netflix is probably going to start cranking out more videos. I think we should learn about this tool just because we're going to see the effects of it on Netflix, probably a lot more pretty soon.

Yeah, I did hear this story. I don't know much about what his tool actually does though.

He'll explain.

So, okay, well, let's bring Ben onto the show. Come on, Ben.

Area 2, take 1, soft sticks. AI, people mostly think of as making something from nothing. I'm going to type something into a computer, it's going to give me a movie. That's not what this is. It was built with a lot of people a lot smarter than I am. You use this technology to learn from your own films daily. Then you're able to introduce that in the post-production process so that you're improving your editorial process, your ability to mix, color, finish your film, do visual effects. It's not a way of imposing a new set of reactions or something alien or foreign to the character. It can only understand this and only build this tool because it's trained on the character that the actor has already built. The filmmaking process, really since its inception, has been one long technological progression. We've always been seeking to make it feel more realistic, more honest. And Interpositive is, I hope, another iteration, another step in keeping with that long and storied history.

How do these actors and filmmakers react when you show them what's possible with this tool?

At first, it's startling and it's a little bit unnerving. But what they do want to know is, so what does this mean? And I always say it means you and I can go shoot a movie on a stage and we can really focus on the performance and the intention and the meaning and the humanity of it. I want to take out all the logistical, difficult, technical stuff that often gets in the way. You can use your own model to remove the wires on stunts, reframe a shot, get a shot you missed, shape the lighting, enhance the backgrounds. You can take some of those problems out. Yes, you can do it more quickly, you can do it more easily. You're given more choice, you're given more opportunity, you're getting more episodes of your favorite shows, you're getting more human work.

So how do you think about the responsibility that Netflix has in this space, and especially now in joining forces with Interpositive and you.

What was impressed upon me was like, this is a group of people that genuinely cares about this industry, that wants to foster growth in it, that wants to enable creativity.

During the last couple years, you've produced movies for us, you've starred in movies, you've directed movies.

I think having that, you know exactly what it is to make a multiple movie.

Absolutely right with us in every aspect. It worked all across this company with a whole range of your executives, a whole team, and every step along the way, it only reinforced that, like, sense of security and comfort.

And I thought, this is a place I feel really good about bringing this to. I'm glad that comes across because it's not new to Netflix to work with technology. We've been working with ML and AI for a long time, but always in service of responsible use of technology versus technology for technology's sake.

All right. I think it's interesting because it seems to go back to what we've talked about, taste. Like humans are still in the taste realm, at least for now. And it's almost like the taste is brought to the production side of things, it seems like. And then that's where this tool plays.

But I'm curious of your thoughts. Oh, I have thoughts. So my initial reactions are, A, I do think it's cool that more of Hollywood is getting on board. That to me is interesting, right? The way he described it was like, actors get to go and act, and then we're just going to use this to reskin and do more of the special effects and do more of the stuff that takes a long time to typically do, like CGI and stuff like that, right? The flip side of that coin is like, cool, I don't mind if this comes to Hollywood as long as it's not taking my job, but the visual effects guys, I don't care if it takes their jobs. That's very accurate. So it's like, it's fine if it's coming to Hollywood, as long as it's not affecting actors, but everybody else in the pipeline, screw them. Like we can replace cars and we can replace CGI and we can replace all of this. And it's like, yeah, okay. So you made a tool that's like making sure your job in the film industry is secure as an actor who's doing the emotions and that's getting translated onto the screen in whatever way it is. But you're not concerned with all the other jobs that this technology is going to inevitably screw over.

Yeah, that's a good point. I just looked it up and the most directly impacted jobs specifically, and I'm looking on Perplexity about Interpositive's tools and primary, it says it'll affect directors and showrunners, cinematographers and colorists, editors, VFX and post houses, and line producers and studio execs.

I'm going to be honest, like I had not seen this video yet. I did not even know what Interpositive did. I saw the news article. I shared it to Future Tools., but it was just kind of more of like a quick skim, like, oh, Ben Affleck has an AI company and he just sold it to Netflix. Cool. I'll share that to Future Tools. This is my first time actually watching the video and like, I couldn't help but like chuckle at the sort of hypocrisy of it just in the sense that, you know, Hollywood's scared of it coming for Hollywood because they don't want to lose their jobs. Well, now it's like, okay, this is a tool that keeps actors safe, but how many other jobs does it totally screw? I mean, like you've got people that are filming that their whole focus is on lighting. You've got people that on the filming that their whole focus is on audio and sound design. You've got people who are visual effects artists. You've got people that it was their job in the past to remove the stuff in the background and put different things in the background. You have VFX artists who it was their job to sort of do the animations of the characters on screen if it was like an animated movie. And he's like, my tool does all of that, but actors are safe.

That's literally it, man. And just think about each of those production houses, like with each person you just mentioned, has like a massive team behind it and a whole business and all these other— it's wild. I mean, like, I don't think Hollywood will get behind Ben and his vision.

I mean, actors might, right? Like, actors might be like, hey, our job's safe. All those other plebs that freaking get me my coffee, screw them, right? But actors are safe with this tool. Again, I do think it is cool that Hollywood's getting more and more on board with AI. Like, in my mind, it is inevitable. It's going to happen. There's no way around it. These studios, they're going to figure out How do we cut the costs? How do we produce shows quicker and faster? Right? Ben Affleck and Matt Damon, they both went on a Joe Rogan episode. I haven't seen the whole episode, so I might be missing some context, but I have seen clips from that episode. And one of the things that they were talking about in that episode, they were talking about how the way that films work now and the way TV shows work now is because everything is built around streaming. People expect stuff to come out faster. Like, people don't want to wait 3 years for that next season of Game of Thrones, right? Right? People don't want to wait 3 years, 4 years to get that next season of Stranger Things. They want shows to be dropped at like an annual basis, like a new season coming out every year at the least. Right? And so in order for these production companies to meet the pace that sort of customers demand, they need to figure out ways to sort of shortcut and cut costs and to get stuff out quicker. And one of the most time-consuming things about something like Game of Thrones, something like Stranger Things. Is how long it took to do a lot of the visual effects. Like, getting the actors there and filming the actors doing their scenes is the fairly quick part about doing that. It's everything that comes after that, the post-production after the stuff is filmed, that takes so much time, right? And so, like, it's inevitable that Hollywood is going to figure out how do we get stuff out that's at a similar level of quality to what people are used to, but do it faster and cheaper. So that we can meet up with the demands of what people want from streaming services, right? In the past, when they were making movies, a company might spend 5 years making a movie that was going to be a blockbuster movie. Now, so many people go to movies less that it's like no longer a viable, like, business model to spend a half a billion dollars producing a movie that might only produce $100 million at the box office and then go to streaming and then they make pennies on the dollar from that point on. So they need to figure out how to cut costs and also cut down production times on these videos. Now, I'm not a Hollywood insider, so I could be totally talking out of my ass here, but that is kind of like the gist of what they were saying on the Joe Rogan Show. So I do think it was inevitable that this was going to happen. I just find it kind of hilarious and a little bit hypocritical that Hollywood is saying, no, no, no, no, no. And now big people in Hollywood are saying, Well, as long as my job is safe, I'm okay with AI.

Interesting. I guess what they've also talked about is more shots on goal, meaning because everybody has endless streaming options, Netflix and wherever, YouTube, that it's like this marketing push. So with more shots on goal, you have different angles and different story, you know, hooks basically. It's the packaging. It goes back to YouTube like we were talking about too. It's like, you give more options for people to dive into and get hooked on.

Yeah, no, I mean, that totally makes sense, right? Like I have every streaming service, like all of them. It's ridiculous. So I have all of these streaming services, right? And I don't know how often I sit down and I open up Amazon Prime and I go, eh, no, this looks interesting. Netflix, eh, Hulu, eh, HBO Go, eh. Like I will click into like 7 different services, scroll in every service for like 10 minutes. Not find anything I like, and then I end up opening up YouTube and watching YouTube videos. Hey man, do you want me to save you a lot of money and time? No, no, because Ilana's the one who watches shows, not me.

All right, it's fine.

Let's move on then.

Yeah, save me money and time. Go open up Codex and go vibe code your aggregator of where all these shows live and just like, blah, blah, blah.

The point I'm making is there's an abundance of options and I have no, like, interest in so many of them. So like that whole like more shots on goal thing makes sense. Like, yeah, these studios might put out 20 different potential shows in the course of a week and try to figure out which ones actually stick.

It's marketing, baby.

All right, one last rabbit hole and then we'll get out of your earballs.

We are moving into the physical world because this is how we like to wrap up a lot of stuff. So basically Meta is in some hot water or I don't know, whatever's hottest, that lava, let's call it. Meta Glass is what Matt has right now. Like someone's probably looking at what Matt's looking at.

It's probably not that interesting.

No, no, I have them powered off, but security button switched on.

Who knows? Who knows? Would you really put it past Meta that if you turn the power switch off, they don't still record everything?

I don't know. Maybe we should ask Zuck. But, uh, so yeah, they're in the news for basically, uh, taking a look at all these videos that you may or may not be knowing you're recording while you're using your Meta Glasses.

I like how the title here is, she came out of the bathroom naked, employees say. Yeah, this was the original article. This is the company Svenska Dagbladet that did the sort of investigative research into this thing. But basically what's going on is these Meta glasses, they have little cameras, you know, on the corners of them, right? And basically Meta's terms of service say that we may actually review what's being recorded with those cameras. Either autonomously through AI or through actual humans reviewing them. Well, it turns out a ton of this footage is getting sent to this company in Africa, and they are human reviewing the footage that's going to Meta's servers. And well, somebody went out to Africa and interviewed some of the people that work in this company. And it turns out they're seeing people on the toilet, you know, people getting changed, people in the shower. People's credit card details because they're, you know, the camera, like they'll have their glasses on and they'll be looking at their credit card and credit card details are in full view of these cameras. And so Meta is kind of getting in some hot water right now because apparently you had to like turn that feature off specifically if you didn't want it going to Meta's servers. And most people like just don't turn it off. Like they just get the glasses, use them, record videos with it. And apparently pick up naked people and toilets and credit cards and stuff like that on the camera. And real humans are actually having to sift through all of that and annotate it.

Real humans in the— it just seems so odd. And it's not clear, like you said, right? Like, how do you know it actually stops sending signal?

It's kind of like the people claim with the iPhones. Yeah, I mean, I guess there's no way of 100% knowing for sure unless you just have no internet connection on your phone. Then you know, it obviously can't send it anywhere. I do get the impression with the switch off, there's no power going to it because the battery doesn't drain. And if there was power going to it, the battery would be draining. But the problem is, A, I don't think most people realize you can turn them off because it took me like forever to figure out that there was even an off switch on those things, which is ridiculous. But also, if you do use them, you need to go into your privacy settings and make sure you're not sending any data to their clouds that they can train on. Because if you do, it is going to a data center that— or not a data center, it's going to like a call center. What do you call it? Like a sweatshop, whatever. Office building with people looking at your videos. Yeah, people sitting around looking at your videos, you know, like, and then actually annotating it so that they can use this data for AI training data, which, you know, who cares if they're using it for AI training data? That doesn't bother me. It's just the fact that there's real humans looking through all of the footage we've ever recorded with these things. And supposedly there's like some sort of like blurring functionality where it blurs their faces before these people get to see it. But if you read the article, they interviewed people at that company and they're like, yes, some of it's blurred, but not all of it.

The blurring doesn't work very well. I just think of all these other devices. I mean, like this is early days of, I know you have a bunch of devices, Matt, that you get at some of these trade shows.

Oh yeah, I'm screwed.

Like I'm totally under surveillance at all times, I'm sure. But it's like you got a, what, the Plaud or the Limitless, all these little pendants that are, you know, they're recording everything. Yeah, I got my Plaud. Right here.

There you go.

So I don't know. I just, the trust of these devices. I mean, I don't know if people trust their iPhones or whatever.

I mean, yeah, yeah. You're talking about all these devices. You've got like the most prolific recording device right here.

I was like, we don't have any security or privacy anymore. Like what's the, just, I don't know, man. Yeah. I think it's just coming to terms of, I mean, this is just like, so just weird.

Yeah. And I mean, you know, obviously Meta is getting in trouble for it. Meta is actively being sued by New Jersey and California now over this because privacy concerns, right? Now the tech giant is facing a lawsuit in the United States and newly filed complaint. New Jersey and California allege that Meta violated privacy laws and engaged in false advertising. The complaint alleges that Meta AI smart glasses are advertised using promises like designed for privacy, controlled by you. And built for your privacy, which might not lead customers to assume their glasses footage, including intimate moments, was being watched by overseas workers. So that's the whole thing summed up in a nutshell right there.

There you go. So Matt, are you going to continue wearing them? Yes. All right, cool.

You don't care. I'm just going to turn off the privacy settings, make sure it's not going anywhere. I mean, I actually almost always wear them with them powered off now. And like, if I'm going somewhere like with the family where I might want to take some photos with it without busting out my camera, at that point I'll flip it on. I actually find them to be pretty handy, but I also find this to be pretty scary and creepy, which just makes me a lot more conscious of how I'm using them and when I'm using them and, you know, what's actually going on with them.

I don't know if you saw the Super Bowl streaker. Yeah. He was specifically wearing the Meta glasses so he can film himself going on the, like running on the, he published it somewhere. I saw it at one time. I'm like, that's pretty crazy to see a point view of like, you're just running on the Super Bowl field.

Yeah, but he didn't win his Polymarket bet because they deemed it not in the spirit of Polymarket or whatever. It's a good marketing play. Yeah. Yeah. Well, on that note, this has been a long episode. We covered a lot of stuff and I think we accurately broke down and ranted on all of the stuff that's come out in AI in the last week that most people would want to know about or should know about.

Yes. I'm going to remind everybody really fast, throw in the comments, what kind of prompts would you like me to prompt Matt? Or just anything.

I'm just a rant factory to you, aren't I?

The people want that, and you do too. So there you go.

Let us know. Yeah, well, everybody, thank you so much for tuning in and hanging out with us today. It's always a blast to get on and nerd out with my buddy Joe and talk about the latest stuff going on in the world of AI. If you like hanging out with us and nerding out on this kind of stuff, make sure you like this video, subscribe to this channel, subscribe wherever you listen to podcasts, YouTube is probably the best because you get all the video footage. But, uh, yeah, thank you so much for tuning in. Really, really appreciate you. Hopefully we'll see you in the next one.