1014: Anthropic Doesn’t Use AI  🔗 **Source:** [1014: Anthropic Doesn’t Use AI](https://share.snipd.com/episode/68e6128d-995f-4977-af52-e599f8cfda8a) ## Highlights - **Anthropic’s Fable Videos Were Handmade** - Anthropic’s Fable promotional videos were created by skilled animators, not generated by AI. - Wes and others confirmed the team hand-crafted assets from internal sources like the Brockus Encyclopedia and teamwork, showing high-end motion design labor behind the clips. Transcript: Wes Bos Things both related to the Claude 5 Fable launch, but not about Claude Fable. And's more um the the launch videos that went out so every time claude has something or anthropic launches something they have these like beautiful videos that are they're usually Like uh ui animations of like showing it and moving around but this one they had this like beautiful fives and nice animations, kind of stop motion feel to it. I just absolutely love the look of it. And I remember seeing that and being like, hmm, I wonder if those assets were programmatically generated or not. And somebody asked on Twitter, I'm curious how Anthropic makes the models for their promotional materials. Is this Claude collaging found images? It doesn't look like an image model. And Sam McAllister, who works there, he says, there was no AI involved in this project. And I previously had been DMing with somebody that works on the like After Effects ones that they put out. And they said, yeah, like we have some processes with AI in place. But all of the actual videos were totally done by hand by like extremely skilled people. He says much of the source material was made from the Brockus Encyclopedia and our internal teamwork. So some brilliant animators for the finished product. ([Time 0:02:16](https://share.snipd.com/snip/a9fedccb-8d72-4d9c-9a86-3c19154482c4)) - **Taste And Vision Still Drive Great Creative Work** - Human taste and vision remain essential for high-quality video and design outputs despite AI tooling that helps with tasks. - Scott and Wes emphasized iteration, vision, and revision are what make slides and motion graphics stand out versus one-shot AI outputs. Transcript: Wes Bos So let's talk about that. We edit a lot of video on our team, right? And our editor randy he is absolutely cracked and has all of this like claude tools that he's built out right um and i think that ai is fantastic for doing quick edits being able to figure Out what the false starts are being able to remove silences all of that stuff but the actual like storytelling and like the i hate I hate to say this, but the taste part of it has, same as Design, has still not been replicated. Scott Tolinski Yeah, I think that goes for a lot of things. I said the word taste so many times by accident at JS Nation and React Summit, where it was like, there is no other word that fits in to describe this. But again, yes, the professional hand of the storytelling and stuff. Been working on a tool to do motion graphics myself i know yeah like you said randy's been working on tool we we all are deep into this space so to me it is again no surprise that again the Uh that overall uh just creative overall like what do you would call that i don't want to say taste. I don't want to say taste. The overall creative vibe is something that is difficult to replicate because somebody asked me at the meetup or at the thing about my conference talks, and they were like, what prompt Did you use to create your slides? Because my slides were 3D. It was a solar system. It was in space. Like, there's no prompt that will do this. It is revision. It is overall vision. It is exactly describing to a T how you want it to look, not make me a thing go. Wes Bos The prompt seagulls, as I call them, where people are just like prompt, prompt, like what was a prompt, prompt, give me the prompt. Absolutely drive me nutty because like, use your brain for one more thing and figure out what you should ask for. You don't even have to do the work anymore. You simply just need to ask the box to make it. And the fact that you are so lazy that you don't want to ask it to do what you want, or you are unable to describe what it is that you want, but you're simply just like prompt, prompt, prompt, Prompt. That drives me crazy. ([Time 0:06:14](https://share.snipd.com/snip/4d7f6264-13a2-4faa-bfb3-eb8d324cd8bc)) - **Fable And Mythos Release Then Rapid Restriction** - Anthropic briefly released Fable and Mythos models, then access was restricted following government/export concerns and reported AWS involvement. - Scott and Wes described the models as powerful but noted the PR hype (therapist/sentience messaging) that complicated perception. Transcript: CJ Reynolds Let's do it. I got one more photo to show. A fantastic family photo of the Syntax crew. Because we were all there. That's one of the cool things uh nikki our our new editor sort of new i've been here a few months now uh october that's new yeah yeah uh but we were all in the same room so we got to take a fun Family photo on the couch that was super great um but scott you mentioned you used fable to prompt but how'd you do that because fable has been banned uh it was not banned while i did that Scott Tolinski Uh because it was banned what the the day after this or whatever it was banned maybe like the day after or the day after that but i did it um the day before my conference talk so it was not banned And in fact i was actually really looking forward to having time on the plane to really get deep into Fable. I'd actually tried it a little bit out to have it evaluate my graffiti library. The thing I heard that Fable does better than other models was specifically handles large code bases really well. So large, mature code bases, it has a good overview of them. And so I had asked it to evaluate some things in my CSS graffiti library. And it did come back with a really lovely detailed list of concerns, possibly, or things that I should be thinking about in a way that I felt was more complete than what Opus 4.8 had had Given me in the past because I'm often people, people are all the time, they're just asking, they're prompting this or that, whatever, but they're not asking enough things to consume Into their own brain back, right? Like, tell me what you think about this. What do you see? Whatever. So I'm not having the I'm not outsourcing some of the overall things to to the agent I'm asking the agent to distill larger concepts and then bring them forward to me help me make the decisions And things like that so I found it to be very nice with that and I am sad that I did not get to use it more I'll say that yeah as a canadian i am i'm sad that i am now banned from using it i we were Wes Bos At the conference so i didn't get to dip into it other than just a couple little prompts but not enough to like be able to say it's it's way better at x y and z in fact i was like like looking On twitter like what people are building with it and like all of the examples that people are building i'm just like that your, I'm like, I could have built that with like a year ago with A lot of this type of stuff. And like the fact that you one shot at it in a single prompt is not impressive to me either. Scott Tolinski The one shot benchmarks are not impressive to me. That's not something that I really am that excited to see. So for people who don't know, Anthropic released a new couple of new models, Fable and Mythos. Mythos is one that's been talked about for a long time here in terms of it being too dangerous. And then they were only out for just a few days. And even then, when they were out, you could use them in your CloudMax subscription plans, but they had announced that they would be making it API only at a certain date, which I don't Remember because I was like, oh, I'm going to get to try it out before then. So therefore, I'll have a good opinion on it. ([Time 0:30:55](https://share.snipd.com/snip/dbdaaaf3-6405-4c31-8d19-08ec809b05bb)) - **Browser Native HTML Streaming Is Landing** - Chrome added declarative partial HTML streaming (processing instructions) enabling native out-of-order streaming and island-style delivery. - Scott highlighted a polyfill exists and this could become the browser-native foundation for framework-level partial updates and Suspense-like behavior. Transcript: Scott Tolinski The HTML, the HTML is actually getting something called declarative partial updates. In fact, this is in Chrome right now behind a flag. And I will say, like many awesome things, there is a polyfill that you can use to have it working in other browsers today. And if you read this blog post on the Chrome Developers site, on the Chrome Developers blog, you will see that this is probably the future of most front-end frameworks and stuff. So what the heck is this thing? Well, it allows you to do out-of streaming of HTML directly in the DOM without a JavaScript framework. And this is going to allow for like partial HTML, like native HTML partials. It's going to allow for essentially framework style apps. And it's actually inspired a lot by the Astro Island architecture. So they are taking hints from the true modern web here and piggybacking on something called processing instructions. Have you seen processing instructions before, Wes? No. Explain that to me. Yes. So you'll see this this HTML tag that has a question mark and it says marker. You might be thinking are they adding new HTML tags well this is actually an old tag that has existed for a long time in XML and they're basically now piggybacking on it using the also the Template tag to be able to have as you can see even like loading states for streaming HTML in, it's really syncing cool. So it's basically like just an old XML tag that they were like, oh, we can now use this to do out of order HTML streaming. And my gosh, it's pretty dang cool. So it allows for, again, island architecture native in HTML. It allows for delivery of content when it's ready. So essentially like what we're doing with was a React suspense, right? And HTML can be delivered in the optimal order for page load performance. Pretty sick. Wes Bos So let me see if I understand this properly. So you have like a website and maybe you load the HTML immediately. However, there are parts of that site that may not be ready at the time that the browser sends a response, right? So, for example, you visit a website and immediately it starts streaming in. Maybe it streams in the body maybe it streams in like like the header the footer but maybe there's something in the sidebar maybe there's a piece of content in the the main content part That's not yet ready um previously not using streaming you would have to simply just sit there and wait for all of the parts to be ready before the entire website does come comes and renders Out. But with streaming, you're able to like ship the parts that are ready and sort of just wait for the rest of the parts that are being generated. Right. Maybe specifically a component that has a bit of a slow database look up or needs to be refreshed, something like that. So with this, we've we've seen that in like react suspense is a really good example of that um but with this you'll be able to put like you'll like native is that is that what that is like Scott Tolinski Placeholders for where these parts are yes yes including like uh like loading um placeholders um it's really stinking cool. And I love that this says this could enable things like client side includes natively in HTML, something that like, man, if you would have talked about client side includes in HTML so Many years ago, like that's like a kind of a holy grail feature, I think, for this type of thing. Batching preventing um what was it like revision so you could have a revision number and do versioning on content and have it update again there's a polyfill to get this working it's Pretty stinking cool one of the comments here it says i still feel like i'd use astro over this i'm not sold let me tell you i think astro would use this not yeah you would use this over astro Ultimately i think this is the type of thing that frameworks will implement and then it's just going to one make that experience better for everyone so whether or not the frameworks Already do it or not, and by the way, not all frameworks are doing this currently. And the ones that do, you have a substantial amount of buy-in to do so. So this is coming to the browser doesn't mean that you would necessarily be writing this stuff by hand, but it does, I think, change the world of what frameworks could be in the future. ([Time 0:46:01](https://share.snipd.com/snip/f83b953c-d929-49c4-9430-7ae1e94a8f5a)) - **Stream HTML Into Marked Slots Instead Of Replacing DOM** - Use native HTML streaming for partial server-rendered updates instead of heavy DOM orchestration. - Scott explained streaming inserts HTML into marked processing-instruction slots, avoiding full re-renders and complex diff logic. Transcript: Wes Bos Let me see if I understand this properly. So you have like a website and maybe you load the HTML immediately. However, there are parts of that site that may not be ready at the time that the browser sends a response, right? So, for example, you visit a website and immediately it starts streaming in. Maybe it streams in the body maybe it streams in like like the header the footer but maybe there's something in the sidebar maybe there's a piece of content in the the main content part That's not yet ready um previously not using streaming you would have to simply just sit there and wait for all of the parts to be ready before the entire website does come comes and renders Out. But with streaming, you're able to like ship the parts that are ready and sort of just wait for the rest of the parts that are being generated. Right. Maybe specifically a component that has a bit of a slow database look up or needs to be refreshed, something like that. So with this, we've we've seen that in like react suspense is a really good example of that um but with this you'll be able to put like you'll like native is that is that what that is like Scott Tolinski Placeholders for where these parts are yes yes including like uh like loading um placeholders um it's really stinking cool. And I love that this says this could enable things like client side includes natively in HTML, something that like, man, if you would have talked about client side includes in HTML so Many years ago, like that's like a kind of a holy grail feature, I think, for this type of thing. Batching preventing um what was it like revision so you could have a revision number and do versioning on content and have it update again there's a polyfill to get this working it's Pretty stinking cool one of the comments here it says i still feel like i'd use astro over this i'm not sold let me tell you i think astro would use this not yeah you would use this over astro Ultimately i think this is the type of thing that frameworks will implement and then it's just going to one make that experience better for everyone so whether or not the frameworks Already do it or not, and by the way, not all frameworks are doing this currently. And the ones that do, you have a substantial amount of buy-in to do so. So this is coming to the browser doesn't mean that you would necessarily be writing this stuff by hand, but it does, I think, change the world of what frameworks could be in the future. And it seems like the near future considering this is in Chrome under a flag and there is a working polyfill and from my understanding this is not a Chrome just yoloing this in and this Is actually something that the other browsers will be getting. Wes Bos Okay um I'm just reading a blog post on this because I was like look how is this any better than simply just wiring up a fetch request, streaming it in, and then dumping that HTML into the DOM, right? Or giving it to React or whatever, right? Substack, modernwebweekly.substack says there's some new streaming methods like set HTML, set HTML unsafe, and the inner HTML, outer HTML setters. So this enables a spa to stream HTML into the assertion point marked by processing instructions using JavaScript. This means that instead of using inner HTML to update your content, you can now just stream it in. Oh, that's great. Because if you were streaming in like a response from like a chat bot, every time you have like a new little bit of content, you have to do one of two things. You replace all of it, right? You just wipe the old HTML out and you just put what you have in so far and then the whole thing is re-rendered. Or you have some sort of mechanism to figure out what is already rendered, what needs to be updated, like does this paragraph have more text, and what is new elements that have to be added. And in that case, you need to reach for something like React because you're essentially now orchestrating the DOM. This will do all of that for you. You simply just stream HTML from the server into the client, and you just say, put it in here. As you get it, put it in here. Figure out the fastest way to do that. Scott Tolinski Yeah, just put it in. ([Time 0:48:20](https://share.snipd.com/snip/921e6207-f299-448d-a0ae-d610f2ade876)) - **Use An Image To ASCII Generator For Terminal Branding** - Try oxide.computer’s image-to-ASCII tool to create terminal-friendly brand assets quickly. - CJ showed options like character set, brightness, and color mapping to tune results for logos or console displays. Transcript: CJ Reynolds Give us a little comment as to why we should check it out but before that i have something to share an ascii art generator that was made by oxide computer this company i believe benjamin Leonard might be the main contributor or maybe it's henry wilkinson but it's an app where you can upload an image and turn it into ASCII art with tons of options for tweaking things. So for instance, let me get the syntax logo in here. So just by default converts an image into ASCII art. That's great. But you can play around with these settings. So this is like the brightness setting and that like increases the amount of characters used it's really fast and reactive too on this which is cool to see um there's also the white point Um which i don't know exactly what it does but it is adding or removing quite a few characters that's like at what point do you make it like it's it's either got to be like white or black right Okay um and at what point are you deciding if a pixel is going to be white or black i see so like a really high black point is going to just be like solid or i guess less maybe it's a opposite Like low black point will have more yeah more characters but it might be like a period or whatever your character set is uh this is fun because you can have you can choose from different Character sets so the standard character set is your classic ascii art with like at signs and exclamation marks but you could also do uh boxes so that's pretty cool um like the these are These are also ascii codes but they're ascii codes for like black and white boxes oh that's great and then the color mapping you can change to be based on hue. That completely blinks it out or saturation. But brightness seems to be the way to go for this particular logo. And then there's also a color set. That is what colors it uses to convert. So it's pretty fun. You can upload any image. I tried with like some actual images, like this is the photo of the whole Syntax crew on the couch. And if I adjust the settings here, it'll start to come through that this is like a photo of people and not just like a blob of text. Let's see. This is really should be used for logos. You can imagine the OpenCode logo, or even the Claude logo. A lot of these terminal UI apps have an ASCII-based logo, so this could be a way to design it and then turn it into ASCII. Yeah. Scott Tolinski Can you do multi-color in this thing? CJ Reynolds I couldn't find an option for it. Wes Bos I feel like you should be able to yeah yeah because it because the terminal might or the terminal supports like colored if you could do multi-color it would be a lot better for especially CJ Reynolds This kind of thing yeah i think you're you're totally right because we could use the binary boxes and then each one could be a certain color. Like what about could limit limit the color palette? Scott Tolinski There's a color set down there and it's set to yellow right now. What does that is that just change like main color of it? CJ Reynolds Yeah, it's just the hue overall. Doesn't seem to um background and foreground. That's all you got. Yeah. Scott Tolinski Boo. Wes Bos Give me full color that is cool i used to have a picture of myself in the view source of my website and like this was like 10 years ago and people thought that was absolutely nuts uh but now That everybody's living in twoies it's like it's sort of coming back you should also throw these in your console log as well. If someone views source on your website or opens up the console, we should have one of these in the syntax website. Let's add the syntax one. CJ Reynolds That's some ASCII art for sure. And the code's totally open source. ([Time 0:55:42](https://share.snipd.com/snip/f8e74bf9-183f-4954-8f8e-59ff36b20c20)) - **Check Replacements.fyi Before Adding Libraries** - Before adding a dependency, check replacements.fyi for modern, lighter alternatives. - CJ demoed searching for Axios and Express replacements like fetch, ofetch, H3, and Hono to reduce bundle size or simplify APIs. Transcript: CJ Reynolds This is a site I came across. It's called replacements.fyi. And the idea is you can type in a package name and it'll tell you what you could potentially replace it with. So for instance, if we type in Axios, it will tell you that you might just use fetch because it's available natively. Or you could use my preferred library called ofetch. I actually learned about this from the Nuxt community, but it's built on top of fetch, but it gives you all the niceties of Axios, like automatic error handling, an automatic parsing Of the response payloads, and automatic header setting, that kind of thing. But essentially, if there's a library you're looking to replace, or maybe you want to find alternatives for, you can throw it in here. Like another example is Express. If you search for Express, it'll tell you about H3. Also, from the Nux community, it's from Nitro, I believe. But then there's also Hono, which is what I prefer to use. So this is a nice site, especially if you're newer to the ecosystem or maybe you're getting tired of a package. You can plug it in here and it'll tell you what you could replace it with. Are there any packages you guys think I should type in here that maybe we could look for replacements for? Wes Bos The one that always gets me is Moment.js, it's like, what are you doing? Um, express is another one. CJ Reynolds So for moment, it says, yeah, yeah. Date functions, which is what I use and day.js. Um, yeah. In unrelated news, the, the newer version of date functions is finally going to ship built on top of the temporal API. Sick. So that's, that's going to happen. But yeah, date functions would be the way to go or even luxon which is like the lighter thing that came from the team from moment there's some others i could plug in here you got any scott Wes Bos No no okay um react.js i'm trying to think chat what are like the common packages that you you see people using like request js is a is another one request js it was like before axios and It is still one of the like top packages out there and i don't think people are using request js directly but it's just a low level dependency that maybe will never be um brought out chalk Let's see is there something for chalk chalk is good though well there is a built-in node js way of doing it now. CJ Reynolds Let's see. Yeah, it's not as good. Util.style text is what they're saying. But there's also Pico colors. I've actually used Pico colors. Yeah. And they don't have replacements for everything. Like if I type in react, that doesn't show up. So I think it's more so this is like a curated list. Yeah, that's nice. Yeah. Wes Bos Object assign. That's a good one. That one, everybody still uses that or a lot of the Lodash ones you see people still reaching for just because like that's how they know how to do things but most of the lodash functions Are no longer necessary because we have them built in uh via map sets whatever um or array methods or you can you can just reach for the individual one that you need totally so yeah check CJ Reynolds It out replacements.fyi Next time you're looking to install a package, maybe go here first and find an alternative or use a built-in. Wes Bos There should be a skill. Or like a... Is there like an API? Let's see. They have a repo. CJ Reynolds I wonder if there's just like a JSON file that has all the data inside of it. Wes Bos Yeah, because that would be a good skill is like, see if my package JSON includes any of these packages. And if so, let me know if I hit them all. It probably doesn't need to be a skill. This could be a script. CJ Reynolds So I'm just now looking into this, but apparently module replacements is a project that's community driven. And that's what this site is built on top of. So module replacements is a manifest of JS modules and they're more modern or active replacements. There we go. So this comes from the E18E. ([Time 1:01:33](https://share.snipd.com/snip/e562cf09-0fc0-4d28-9473-014a4b5f5ab2)) - **Beta Upgrades Surprised With Stability** - Scott installed macOS/iOS betas across all his devices ahead of the conference and found them surprisingly stable for a beta. - He reported smoother performance and fixes like consistent keyboard behavior across apps. Transcript: Scott Tolinski Or I have my phone, my laptop, my tablet. Unfortunately, I'm kind of outing myself here as being like an Apple guy, even though I don't really identify that way. You're outing yourself. We could tell, Scott. I got the vision pro that is right here i have it updated i have my watch update i got i got all the crap and i got it all updated and i gotta say folks it's good for a beta it's very solid i updated To the beta from my hotel room an hour before my flight to and knowing that I had to go on stage and give a talk with this OS and have it not goof up and it was perfect and honestly I feel like My computer functions better with it um there's some annoying little things with the new Siri but like ah the OS is good it's smooth I feel hopeful for the first time installing some Apple Software that it's like actually put some care into it. It doesn't feel like it is just, it doesn't feel like it's just a bunch of bugs tossed on more bugs tossed on more bugs. It feels like they did the right things in going back to some of the older styles of stuff. ([Time 1:06:05](https://share.snipd.com/snip/26c204fe-fb32-4510-ae7e-ff9816432707)) - **Local Quantized Models Make Offline Agents Practical** - New open-source and cheaper LLM models from companies like Xiaomi can run locally via Hugging Face with four-bit quantization at ~150–190GB. - CJ noted that local runs may yield slower token rates but enable offline or cheaper agent workflows. Transcript: CJ Reynolds I'll have to check these out. It looks like the four bit quantization is anywhere between 150 gigabytes and 190 gigabytes. So I might be able to run this even on my, my local AI setup. So, I mean, that would be crazy to get opus 4 6 level responses from just my local ai yeah so i'll have to try it out report back yeah patiently waiting for this new max studio uh it varies Wes Bos Um probably anywhere between 10 15 tokens per second and is that like what do you know what do you get when you like use like um like anthropic via the cloud probably getting closer to 30 Or 40 tokens per second when you're oh so like yeah like a three times slower you think yeah okay yeah but so that's probably not a big deal for like asynchronous agents hey like it is when You're sitting there waiting for the response but it's not if you're just running these things overnight yeah you're just letting it run and then coming back to check on the results Scott Tolinski Later yeah yeah not as important so yeah and a lot of the stuff i would use this stuff would for be like personal assistant stuff you know like stuff that i'm running whether hermes or whatever For so like coding you know different different, different tasks for me where like coding, I expect a much faster response than like, I really don't like running stuff overnight for Coding just because I feel like I want to have more control over it. But like, if I wanted to, you know, check my calendar every single morning or something, that would be a perfect use case for that type of thing to run locally for sure. ([Time 1:11:49](https://share.snipd.com/snip/8a78918d-111b-4451-ad16-dea4c523a5b4)) - **Supply Chain Risk Lives In Community Repos** - A recent Arch Linux supply-chain incident affected community (AUR) packages, not core repos, illustrating risk in community-maintained package sources. - Scott and CJ recommended auditing foreign packages and checking Pacman for unknown packages. Transcript: CJ Reynolds Arch by the way not arc um also i will say browser Yeah. Scott Tolinski Arc Linux. Arch Linux. Yes, and their logo is even like an Arch. Yes, I'm dumb. I'm sorry. I get it. No, no, it's okay. CJ Reynolds But what I will clarify, though, is it was not their main repository. Like, Arch itself and their core repos are fine. There are some user-contributing repositories, very similar to how of them. And so very similar to how NPM is kind of open to the world. Anybody can publish there. There is also like a community-managed repo that you can pull in packages from, and that's what was compromised. But a lot of people do install apps from there. Scott Tolinski Yeah, it's like the main, it's AUR, right? Isn't that the utility? CJ Reynolds But it is community packages. It's not the... Yeah, it would be like, so when you run a, I'm more familiar with like DBN or Ubuntu, but you do like an app get update, that's pulling from by default the core DBN repos, but you can also Add like the third party or proprietary repos. This in particular isn't the core arch ones. It's the community ones. Correct. Scott Tolinski Yeah, but it's a whole lot of them. So you can run Pacman hyphen capital Q, lowercase m. ([Time 1:21:56](https://share.snipd.com/snip/93e3484c-253c-4c80-b593-1c96d8b64e7f)) - **Generalist Models Keep Valuable Domain Knowledge** - Models trained solely on code would lose domain knowledge helpful for real-world apps; wide training data supports domain reasoning beyond syntax. - Wes and CJ argued generalist models can route tasks or call specialist models rather than replacing broad context with narrow code-only training. Transcript: CJ Reynolds So this message was from clarkio uh speaking of cost and speed i'd love for the frontier labs to provide more coding focused models that have no training in anything else outside that Scope um we haven't really talked about this before but like yeah so in the world of like trained models they train on massive amounts of data right not just code like websites and blogs And video transcripts and books as well as code and so when you're interacting with opus 4 7 or whatever else it's trained on all that other stuff too um so it is interesting to think about Like if there was a model that also i think we've talked about this in a past livestream, like, what if it was only trained on good code, too, like, good code examples? Examples of code that we like, like, clean code or good architecture patterns. Yeah, what do you all think of that? Scott Tolinski I would like a model that's trained on the code that I like and no other code. That's what I would like. Yeah, exactly. Wes Bos That's that's a good question like would a model that was only trade like there are models out there that are only trade trained on code and they don't seem to be particularly like i guess Like composer is is an example right it's significantly cheaper um although i have asked composer about like like love and things like that as well. You know, it seems to know what's up. I don't know. I think these mega models are it because like, I don't want to have to think, where am I going to do it? I think the day, like I often will simply just use Claude or ChatGPT or, or Cursor or whatever as my, I don't want to have to switch to a chat when I want to do something here and then switch To something else. I want to be able to be in my clod and be like, hey, what flights should I do? Or add these to my calendars, but also research where I should be going out for dinner. I think it should be doing it all. And the days of it being just like a coding tool, I don't know. I think it's got to do absolutely everything. CJ Reynolds Yeah, there's a comment from Frankie that says, without wide knowledge, they would probably suck at understanding your domain. If you're making an app about pizza delivery, it's good if the model knows what pizza is. And I think that's exactly to your point wes like it you you can have a conversation that says i want to build this type of app and it has domain knowledge like it knows about that type of Thing not just the code it knows about that the aspect of that thing like uh for instance i was building an app to analyze uh dna data and if i had a model that only knew hold on put it in that Continue okay but this is a domain that i'm not familiar with but like there's there's all this information in medical studies about if you find certain aspects um in someone's genome Then that means they might be more susceptible to getting certain illnesses or they might need to supplement certain vitamins in their diet or they various other things like that but This is all stuff that when i was prompting and planning out this app opus 4 6 was able to pull already had access to all of that knowledge even searching the web um and then it could use that Wes Bos To then create the algorithms for analyzing the dna files so yeah yeah like we're so much of building software is not about like where do the functions go and what database adapter should I use it's so much more about like i'm trying to build software for somebody in the world or something in the world. Therefore, you need to know about the world because I'm building software for it. I'm building traffic management software. You need to know how cars and traffic lights and all these things talk to each other. CJ Reynolds Yeah. Wes Bos So I don't think so. Scott Tolinski I don't think we're getting there. What if there was other – I mean, what of you're just reaching out to other models like explain to me what the hell pizza is and it's reaching out to a model that knows what pizza is right CJ Reynolds More like a model router of sorts like you could decide what what model knows about this or it could decide to search the web i honestly that might be how we get to a point of less hallucinations If it's choosing the right trained model for the right job yeah ([Time 1:23:47](https://share.snipd.com/snip/5a38d70e-6859-4208-945f-6b85d7a919bc))