BONUS: The AI Starter Kit: What to Try...and What to Ignore

34:34So what I've done is I use this, it stores, it's a simple lightweight application. Let me say that. You can use it on your phone. You can use it on a Mac. You can use it on a PC. You can use it on the website. You got access to it everywhere you go and you can connect it to your chat GPT, to your Claude, to your OpenClaw, to your everywhere and can serve as like kind of a almost a central file server, and you can tell it, like, save this to Obsidian. And it'll go in there, and it'll create a very tiny, very simple little markdown file. Markdown is fancy for text file. It edits core, like, you know, I mean, but what they are is they're really lightweight, so an AI can read them really fast.

35:21They don't take up a lot of space. they're very clear because you don't have like when you give an AI a word doc there's a lot of trash in a word doc before it ever gets to your words and when you're thinking of how a machine is going to read it so these are super easy they're super fun I use them for organizing ideas I have logs of our podcast guests I have logs of guest appearances I've done on other shows I have sources. I save papers there when I come across a research paper I like. They've got a nifty Obsidian Web Clipper Chrome extension you can use, and you just click it and it sends it, and it'll save an article over there for you as Markdown as well.

36:05Yeah, it's really handy, and it's a good way to manage files on your computer. So if anyone's ever used Codex or Cowork, which we can talk about a little bit, You can think of these as kind of like phase two of the chat bots from OpenAI and Anthropic. Phase one was, you know, chat GPT in the chat window, you know, all of the tools that they've added on top of that. Same with Claude in the UI, the web UI, and then they made a desktop app. And Codex and Co-Work are phase two, where you're giving the AI more agentic power. What do I mean by that? I mean, it's able to go off and use tools so it can run software to it can access other other applications with connectors and which is what they call MCP, which you might have seen some people in the chat talking about MCP.

36:59that's what that is like the companies call it uh connectors and uh it can access other applications and when you run it on your computer either in your terminal or through the app you can then use it uh to actually use other applications on your folder use any basically any command that you could do on your computer the agent can do and so that's what's like phase two of of the these tools is that the agent can actually go off and take actions on your path it's not stuck in the chat window and it's way cool and it is exactly as downy said here a gateway drug uh yeah i uh yeah i went i went down the claw rabbit hole for a good while i'm tinkering with the various alternatives now but i haven't really formed a good opinion on on what i love and what I don't love.

37:54What's your take on that in terms of what you would avoid? Would you say Open Claw is one you would avoid at this stage? If you were a beginner, I would avoid Open Claw. Okay. And I say that not because I don't love it. I do love it, but it just takes – you've got to be willing to kind of learn your way through it. There are going to be errors you've got to deal with occasionally. There are going to be a variety of little problems. And with that said, when it works, it's magic. Yeah. It's absolute magic when it works right. That's fair. But I don't want to get too lost in the more difficult things.

38:33Yeah, there you go. That's what you were looking for. Yeah. So if you see here, this is essentially how you could use Obsidian. And we have these instructions. I will also share this in the chat. i'll say how to create a local wiki file with obsidian follow this ai skill of the day and this was a four-line chat that allegedly andre carpathy used to create this um it says dissect dissect this raw note into atomic obsidian markdown files each file equals one concept and then use wiki links between them uh and you know you can follow this it's it makes a cool knowledge graph of everything you say than there yeah so that's a cool way to do it so if you are if you are wanting to be call it chatbot agnostic meaning you don't want to be stuck in any one interface that's the way that you would do it but that is a little more complicated than the ai starter kit so i want to make sure that we don't get too i was just thinking that oh we've done it grant we went too deep yeah but um but uh so we talked about projects i think projects are something that i would definitely um learn if if you're just getting started with this stuff so projects context memory yeah projects context and memory i would say memory is less important i i still find that it's not it's not as effective as i want it to be at this point i think context is all you need at this point personally but i 100 disagree with that and i think this is a difference in that we have both used different tools grant grant traditionally uses a tool that has a larger context window and i traditionally use a tool that's got better memory so i think that's maybe why we uh um why we disagree there there could also be a lot of ways in which that i'm not giving memory enough credit like if we look at my claw like if i say hey i'm working on a thing for the neuron it knew what that means right and without saying go to your memory and read this you know well you can set that up with instructions but then for example on clawed clawed has memory on top of it and it has a lot of information about me so i'm taking a lot for granted with this memory up in the corner just because i never touch it and never edit it that's so there there is a lot that that memory is helpful probably doing more than you realize it's doing is what i would say i think that's fair yeah i think that's fair i bet it is attention is all you need yeah well now it might be sub-quadratic attention is all you need i don't know more on that you know potentially later today in the near term future yeah um but yeah so what else what else is something that people need to know if they're uh totally starting today okay so i don't think we need to go as basic as a prompt if anybody disagrees let me know but i think i think how to use projects understanding what context is and how context can help you get better results are very important and the other thing is understand that it doesn't have to be a one-shot world you know ai is about iteration you know get what you can get on shot one and then talk to it work back and forth with it make it better treat it like talking to another person just understand that just like another person it's not going to have context you don't give it and those three things work together really well to help you really better get a handle on better responses?

42:21I think another thing that I would say is I've found the longer that I go on that I rely less and less on its like, quote unquote, memory that it has in the model. Like, I just don't trust its training data and its memory to be enough. so i'm always giving it the exact context that i want it to work off of uh or having it search the web and come back with results or connecting it to other applications like i think the the the built-in um what do you call it like it's model weights which not to get too technical but that determine how it responds and what it knows um i find is good for helping it use the context you give it but if you're going to get the most out of these tools give it all the context that you that you know it needs to to do the job like don't rely on it to just like if you say like like explain um quadratic um equations or or whatever like explain algebra to me something like that um it could do a decent job but anything beyond is just like something that like a really smart person would just kind of know um off the top of their head each variable narrows it down and narrows that scope yeah yeah just give it give it the attention and let it use all those model weights to sorry give it the context and let it use all of its model weights too another thing i want to mention that could fall in a things to know and things to ignore is most of the tools are really good now.

44:04Like for the vast majority of knowledge work, if you work in the standard fields that every company has, I'm not talking about super technical fields. I'm not talking about, you know, highly specialist things. But when it comes to more general fields like project management, human resources, all of these are going to do that well. Don't get too hung up on which one's the greatest right this minute. Because the fact of the matter is, most of what you're going to do, you're probably not going to notice a big difference. If you're dealing with like, you know, research, if you're dealing with things that really require that leading edge, that extra 2 % the latest model has, it matters.

44:48But if what you're doing is, you know, standard office work and getting help with meetings and organization and things like that, But go with what you like. Go with what you're happy with, what you can afford. But the flip side of that is that you're going to see a billion different models along the way. And I'm going to let Grant lead into that here because I can tell he's got it up and ready to explain away. This is a suggestion from M. Downey 14, and it was a really good one. It's a good one. They suggested maybe choosing which model to use. Some are more expensive than others and what thinking and effort level means.

45:25So this is really important because a lot of people don't even realize this exists. So there's this slide. Sorry. Say that again. Back in two. Oh, yeah, that's fine. Yeah. So essentially, I think this is now, I feel like they always change it. It was in the top left hand corner here and now it's here. But essentially, every time you use a chat interface, there is, you know, the interface is like this portal here. But then there is a model that's running under the hood. And that model is deciding basically how to respond. And all of the big chat companies have essentially narrowed down the way that they present these models to you into three formats.

46:12So the three formats are Instant, Thinking, and Pro. So you can see here, there's a little dropdown. I'll zoom in. Not that far. So right here, Instant, Thinking, and Pro. And I believe if you just leave this on, it will default to Thinking. Seems now that you can also configure it. So that's probably where I'm getting this from. And there's also the amount of effort that you can apply, but we'll get into that in a minute. But it's important to know that if you have instant on, it will give you a really fast answer. So I'll say, explain the speed of light. And it's going to give me an extremely fast answer, right?

46:57But it has a lot of emojis in it. It's giving me like a very basic, a very like AI slop kind of answer. So me personally, I never use this version. I just never think it's worth it, in my opinion. I don't care how fast it is. it's just not worth it um other people have different opinions on that and as m downey shared in the chat uh the the higher up on this scale you go there's this thing called rate limits and rate limits basically determine how frequently you can use this so um for example if i have thinking on here i'm gonna ask the same question i'm gonna say explain the speed of light and now what it's going to do is it's going to be thinking and basically what that does under the hood is it takes a second to reason through the prop through the answer where it's almost like writing out the answer ahead of time and then it it it says okay this is good and then it gives it to you that's like a very oversimplification but that's essentially what's happening and so it says okay the speed of light and then it explains it and it as you notice it has a lot less emojis it has a lot less like slop in it in my opinion um it seems like more like the answer i was looking for right it's more much more reasonable now let me do it again only this time i'm going to use pro and not every account type even has access to pro but it's going to show you a good example here um okay so now i'm going to ask explain the speed of light so now it's doing Which is a three-week request for pro, probably.

48:34Look, we're trying to show the same process. Across all three levels? Yeah. So one thing you'll notice right away is that this is much slower, right? And that's because it's using a lot more, I guess you could say, the model weights are, it's a much bigger model. And so it's going to take longer to answer, essentially. So in this case, I thought for 27 seconds and it gave me a slightly longer answer. I don't know if I would, you know, I'd have to read this to say, OK, was this really worth the pro request or not? But why would you use pro versus thinking? Well, pro is for situations where you're providing a ton of context when you need to do something.

49:21Basically, you need to use the AI at the top of its intelligence. Like the smartest version of this AI that you could think of because you're doing a problem. You're trying to solve a problem that's extremely complicated. Think of it as a complexity guide. When you look at the three options there, the more simple it is. If it's super simple and you just want to know, how long do I cook a baked potato because I can't remember in a microwave? Ask it on instant. It'll be fine. As they get a little more complicated, you go up. And like Pro, I would say Grant and I use Pro for, suppose we want to know how the newsletter is doing.

50:04We'll go and pull down like the last six months of numbers and be like, what's happening here? Why are we seeing an upward trend or a downward trend? What do we want to, maybe we want to do more of this or less of that. Ken Little Coyote has a great answer. The difference between asking a smart person and a nerd.

50:25yeah um and uh you know if if anyone has anything else to add to that i'm trying to give like the very basic most easy to understand explanation between the differences here but um obviously we can go much more in the weeds on that one other thing i want to explain is under configure so you can set what model is your default there's also this thing called model up here and 5.5 is the most is the newest one there's also all of these old uh older legacy models here so So there may be use cases when you go back and test them, but I think we can kind of say for sure that this is a step change difference between 5.5 and any of these other ones.

51:05So 5.5 was a jump. Yeah, I think you should just roll with this for now, but you can set your default. So here I usually have it on thinking. As you can see, there's, you can toggle this when you have instant on that says auto switch to thinking. So this is an interesting feature. So here, chat GPT can automatically switch from instant to thinking when you ask a complex question. And if you think it sucks from when GPT-5 released, I would recommend giving it a fresh try because it's not bad. It's come a long way, the autorouter. Was 5 when they released it the first time, I think? Yes. Because a lot of people freaked out about it and were like, I want to pick my model, which is fair.

51:42Yeah. Yeah. But now you can set a default, right? So this is a default. And here, when you set your default to thinking, you can also set the effort from standard to extended. And the easiest way to understand this is just it's the time it takes to think. Extended is going to think longer. And that doesn't mean that the answer it gives you will be better necessarily, but usually it is because it's thought of it. Think of it in terms of how you think of a problem. Like if a problem is something you know you would spend more time on, assume the AI will do better with more time to spend on it as well is probably a good way of thinking.

52:18Like if you're doing something that's involving a report or some kind of research, I absolutely would. There's also something in here I'd like to show real quick. We had a comment from Chemical Coyote here about bouncing things off both Claude and GPT. And it's interesting how one catches things. I want to show what is one of my – the feature that's why I still talk about Copilot quite often. Sure. Take it away. Because it is absolutely the sickest thing Microsoft's released in the AI space, in my opinion. Okay. So when you go in here, we're going to say explain the speed of light.

53:02And we'll go right over here. And as you get down here, you'll see there are some options. Oh, not this one. Up here. Where are we at? Not auto. Where did it go? Where did it go? Hang on. There are tools. There's a tool called council. I wonder if you got to do it on web.

53:29Where did it go? It was literally here days ago. You want to take a second and look into that? Yeah, I do. because now I'm annoyed. Okay. This always happens with Corey, unfortunately. Corey has this great thing. He's tested it ahead of time, and then the second we stream, it's like something... It falls apart in my face. Well, I'll show you... Okay, so there's a really cool feature that he's talking about, which we can share, especially if people use Copilot. But if you don't use Copilot, there's another tool that I'll share while he's doing that, and it's called Open Router. So... Yeah, Open Router will do it too.

54:09That's right. Yeah, so one sec. Okay, so Open Router is basically, you can think of it as the unified interface for LLMs. But basically what it is, is it's where you can use one key. And, you know, I don't want to get too lost in the weeds on this as far as keys go, but a key lets you access a model outside of the chat interface. So you might be comfortable, you know, for the people who are level ones or call it a sub level five. If you've never used AI outside of the chat window, the way that you do that is with what's called an API key. And here you can use one API key to use any model that's listed on Open Router.

54:55So you can basically plug this into any system that you use and you can use it from here. Now, what you can do is you can also do a chat on OpenRouter, and you can actually add models and test them out. So here I could do Gemini 3.1 Flashlight. I could do, let's say, there's a lot of models on here. GPT 5.5. And I can add Claude. Let's say, Sonnet 4.6. And now I can ask one chat. I could say explain the speed of light and it will send it to all three of them at the same time. Now, I don't have I haven't paid for it, so I can't run it right now. But there are free ones on here that you can use.

55:42A lot of times that'll be for promotional. But yeah, what you do is you just add like$10,$20, whatever it is here, whatever you would pay, you know, whatever is your testing budget. And then you can ask the same question to all three of them here. I know what I did now when you're ready. Yeah, go for it. So if you use Copilot, which probably a lot more people use Copilot than use Open Router at this point, this is how you do it in there. And this was an ID10T problem. It was very much me that was the issue. It's in their researcher agents. So Copilot has these agents stuck through here that you can use.

56:20You can go build your own and all. But I don't want to get into that. But when you go to their researcher agent, this is where you have abilities that you do not have elsewhere. so you can use computer use you can come down here you can add context you've got sources but up here that didn't do it here we go critique these are when you when you send a request gpt responds claude refines model council this is where you ask a question and both gpt and claude go out and answer it and come back and then decide on here's where you are you'll it's really cool to watch like let's say okay let's do it this way explain the speed of light and it will it will do this with both of them and what you're going to see is two windows pop up Oh

57:17I say that. Okay, lining things up. Give it a second. Explanation. What level should I aim for? Go ahead. Just use your own judgment. And it should now pop up with a window for each. Living AI labs. That's hilarious. Co-spyware pilot. Yeah. Here we go. Check this out. This is for people who use Copilot at work. Yeah, if you use Copilot at work or whatever the case is. The truth is, this is a cool feature that, you know, was previously, you know, for harnesses developers had built to do it themselves. But they made it really simple and put it on 70 % of the world's computers. So it is there and one of my favorite things in here.

58:09So they're both going to answer. You never know which one will be fastest. It varies. and they're generally something different in each of them almost every time. And then you pick the one you want or you can go up here and tell it to do other things with it. You can start a group chat with this as well, which is kind of cool. Why was my message blocked? Might be the Google. Oh, I responded. Yeah, that's probably just the auto mod on YouTube. Also, for some reason, if you share a URL in chats, you're not allowed to do that. So just keep that in mind. It just disappears. Yeah. Like it looks like it's sent, but it doesn't actually go through.

58:49Okay. And we don't have to sit here and watch this. I just wanted to show you that it is an option to have them both in one place really easily in a simple tool. The researcher agent is pretty sick. I use it quite a bit. Let's take a poll for people in the chat. What do you use as your daily driver? because we can demonstrate stuff in Claude. We can demonstrate it in Copilot. We can demonstrate it in ChatGPT. But let us know what the majority of y 'all are using, and we can show a bit more in those tools. Yes. Absolutely. But, yeah, this is really cool. I think more of them need to add these features.

59:35I think so, too. Critique is also nothing to shake a stick at.

59:43Like, I like the way this works a lot because it'll go through and you're going to get the chat GPT response first. Go ahead. Stop asking me questions. Oh, here we go. Hang on. Short, one to five pages. Long, five plus. Let's go short. I like that. That's nice. That's a really cool feature, isn't it? I actually hadn't noticed that before. Stan, you must be like level eight, man, if you're using Claude Code in the terminal. Let's go. That terminal life. Love it. That's me, dude. When I first Claude. Codex is good. I like Codex. Claude. A lot of Claude, but RBT uses Claude and Chachabute, so that's good.

1:00:32It really pays to use multiple tools. I think it's really good to know what else is out there and not get too sucked into your own universe where you don't even know what's happening. And it depends on what you're doing, as we've mentioned here earlier. Like if you just are getting pressure to use AI at work and you just want to understand what you need to understand to be able to do that, grab one of these and go. You will be just fine. And you can always come here. You can always email us and ask. If you've got something you're trying to figure out, you think you ought to be able to do it and you just can't get it, we always want to hear about them.

1:01:09We keep saying one of these days we're going to do a mailbag thing. But the truth is the email gets too many emails, and we haven't figured out a good way to extract them yet. But you would think the AI guys could figure that out, wouldn't you? Yeah. Well, there are perhaps bureaucratic reasons we can't do that. There may be barriers, yes. Yeah. Okay. Okay, there's a lot of good stuff. There's a lot of good stuff. It's pretty mixed. So I think we'll kind of keep showing things off in different tools. By a show of hands, who has done anything with agents or automating workflows or anything like that?

1:01:51Has anyone experimented with that sort of stuff?

1:01:58And by a show of hands, I guess I mean comment in the chat. hands hands yeah yeah put a hand emoji in the chat oh it'll block it probably yeah open code and open work both good yeah i haven't i haven't actually tested open work surprise surprise i've used open code but not open work yeah that's for but i have a feeling that i'm not familiar with open work but i have a feeling that hermes i haven't i haven't spent the time I need to with Hermes. I have it. I have set it up. I got distracted, had to leave town for a few days and I've just not gotten back to it. So that's that's on my list. I keep hearing good things, though.

1:02:40Yeah. Yeah. I've heard really amazing things about Hermes. And we we have I have tested it to the extent that I was like, yes, we should put this in the newsletter. But I need to spend like a weekend on it. You know, that's the that's the problem with all this AI stuff is you really need like a full weekend to understand like what it can do for you just to switch to it right like that's like a lot of times you'll you'll see people get um call it ai pilled or agi pilled after uh like a winter break or a spring break because it actually had time to go in and play with the things instead of just like the constant barrage of new launches that come out every week um you need a little bit of time to actually test it something worth showing here real quick if you're a newbie that is worth knowing is that in models like this where you're dealing with here it's researcher but if you're using thinking mode and one of the other models other tools as well where you see you'll see this little carrot here and i'll talk about what all it did copilot lays this out in a really simple way in my opinion like instead of it reading like it gets a little crazy on the back end of chat gpt this lays it out really clear as to what it's doing starting research pulling together sources and it'll let you know what it's thinking it'll let you know where maybe it's gotten sidetracked so like if you get an answer that you're like that's not right you can go back through these and be like ah that's where it went wrong right here and uh and kind of figure out you know a better way it's a nifty thing to be able to do especially if it's something really important you know if you're working on something for your boss your ceo your you know where where accuracy is vital and you want to make sure that it went through the process as you wanted it to go through the process now check those it's handy actually this is a really short answer funny enough yeah it's funny i think it says everything that matters you want it's funny it took so much time we should have asked it like um could can humans ever be able to travel at the speed of light?

1:04:45The answer is no. The answer is no. But could we get, like, within spinning distance? Yeah, potentially.

1:04:57Does a scenario exist where humans could one day exceed the speed? Oh! Exceed? Absolutely not.

1:05:09Interesting. It'd be funny if it just comes back and is like, No.

1:05:17That was quick. Oh. No, it's asking the questions. A lot of times they'll do this where it will ask you some follow-up questions before it kicks off. And don't do what I'm doing here and just blow past them. Take the time to answer them. You'll get a much better response most of the time. That's good context you're providing. Yeah. Oh, we are already at the speed of light. Whoa. that would be wild okay so I want to show a couple other things in the UIs that most people use I didn't see much people saying copilot was their use case here so while you're doing that I'm going to pull up I saw a lot of Claude so I want to show a couple things in Claude well I'll return with the cliffhanger answer to whether we will one day exceed the speed of light.

1:06:13I really want to know. I really want to know. It will be interesting. Downey, thanks so much. Appreciate you being so active and having good conversations. Yeah, thank you. Appreciate it. All right. To everyone here real quick, if you haven't already, please take just a moment to like the video, subscribe to the channel. It really does a lot to help us continue to justify what we're doing and bring these great things to you. And it means a lot to us. Also, pop by the Neuron.ai and sign up for our daily newsletter where all of the – where Grant, every morning – how do we say this? Compresses the wide world of AI into 100 kilobytes or so and sends it out to the world.

1:06:56So I'd love for you to be one of them. It's difficult. You have to be able to code to do agents. This is a great question. I want to talk about this. Can you put that up on the screen? Yep. No. That is a great question. do you have to be able to code to do agents? So there's an answer that's yes, and there's an answer that's no. And I'm going to say the no answer first. So there are versions of agents that you can build inside the chat apps or the desktop apps, depending on what version you're using, where no, you don't have to, but it depends on your definition of agent. So I think it's like a big question, but I think the way that I would put it is I would say there are ways that you can build agentic workflows with no code required.

1:07:52And there's a lot of different paradigms or there's a lot of different ways in which people have been trying to conceptualize this to make it as easy as possible. and some of the easier ones, some platforms make it easier than others. For example, you could look at what OpenAI just launched with the workspace agents as a really easy way to build agents. And so I'll pull that up real quick. We did a whole stream on this last week, so I won't go too deep into this, but they have this tool here called Create an Agent. And what you can do is you can, you know, pick one of their planners here or you can ask them to create it for you.

1:08:32and essentially I'm going to hit skip to builder and you can connect it to what the agent needs in order to work. So in order for an agent to work, it needs contacts, it needs skills and it needs to be able to run tools in a loop. And so this, I don't want to skip over skills to talk about agents. I think we should probably talk about them before we talk about agents. But as you can see here, you can add different channels. You can add different apps that it'll be connected to. And then you can add skills, which are repeatable processes that it can do. You can upload files. You can give it memory.

1:09:09And you can give it instructions. And all of that is doable inside the chat GPT app. So you don't need to be a coder in order to work with that, right? Now, can you create a much more powerful agent if you use code? Yes, you can. and in fact you you don't even need to code yourself to be able to do that you can work with some people have mentioned that there's a terminal agent that they use or there's like cloud code cursor some of these other options codex somebody has mentioned you can work with an agent an agent that's built into these tools and have it help you build agents and write the code for you and then you can run it locally on your computer you can create what's called a harness which is your own sort of instructions and your own interface for working with the agents yourself.

1:09:59And you can write that yourself. And there's a open source tool called Pi that lets you do that. But that's definitely not a starter. I was like, Grant went off to deep end. Yeah, that's definitely not a starter tool. But like, it's as easy as asking your AI to help you do that, right? And it's as easy to say, hey, I want to make an agent, you know, use web search and help me look up the ways to do that. these are the tools I use. This is what I want to do. And it can, I am not a developer. Tell it that, tell it, but it knows your experience level, uh, because that matters. I mean, you know, I always throw in an Eli five, like explain like I'm five.

1:10:36And, uh, quite often that, that helps me learn something I didn't know. Yeah. But I, I do think like to chemical coyotes point, you should try and learn as much as possible because asking the right questions or suggestions is key a lot of times. Like, that's a great quote. Like, just using the AI to go back and forth and ask a question and say, hey, I don't get this. Can you explain this to me? Like, can you walk me through this step-by-step and don't move on until I've got it? That's a really, really helpful resource. And that will probably take you 90 % of the way of where you want to go, right?

1:11:15And you'll learn as you go. Yeah. Yeah. Now, one of the things I mentioned is called skills. Corey, would you say that skills should be in the AI starter kit or no? What are your thoughts on this these days? I would say yes, with the exception that I don't know if they hit a broad enough part of the population yet. You know, we're still kind of with Claude and ChatGPT's business plan. if you're using either of those it's it's a big unlock and if you have not used them before if somebody wants us to show skills hit us in the chat we'll absolutely run through them i think they're really powerful and they're probably the the best way to use these this stuff in you know 2026 i would say i want to want to respond to coyote here too who mentioned that you know we were joking about vibe coding and, and any mentions you should try or, and learn as much as possible though, because asking the right questions or suggestions is key.

1:12:18A lot of times that is absolutely so true. And I've very much taken like my, my journey into vibe coding as a way to, to learn, you know, I quite often will make it explain things to me. So I understand the code that I'm dealing with, you know, I'll pull it out and go throw it into another app and be like, what is that? Show me. Yeah, that's a good question. All right. The people have spoken. They want to see skills. All right. We got three yeses to skills, four yeses. Yes, let's do it. Okay, so I'm going to do an example and I'm going to say I need to create a skill to. What's a good example?

1:13:06Sound off in the chat. If you have something you've already done well that you could go in and be like, Oh, yeah, yeah. I have a lot of stuff I can show, but I want to show like the process of actually working with one. So give me an example in the chat, folks, of something you want to create a skill for. The best way to think of a skill is when you go to skills, they are. Let's see. What's one? That's one. Let's let me show one of the Anthropic ones. So here's the skill creator skill from Anthropic. And you can see it has a markdown file that has a series of instructions. And then it has a series of subfolders with markdowns in them.

1:13:43And it also has some code files as well. So you don't have to worry about any of this because they have this skill creator skill. So in this case, there's a skill to make skills. So I'm going to say, please help me create a skill to show off skills for the folks for the lovely folks on this live stream and then what i'm gonna do is i'm gonna hit at nope that's codex version i'm gonna do backslash skill creator and as you can see i have a library i'm gonna zoom in here a library of all of the skills i've created that helped me every day on what i does what make you tag the skill or does it nope you you don't even have to do this i'm just showing this for demonstration purposes i didn't think so that's why i was asking yeah The only thing I would say is that, let me see, please create a skill.

1:14:34Let's see if it's going to call the skill appropriately. You know that you have the skill is working. If once we have this little guy here, it says thinking it needs to switch to the scroll. So yeah, and it'll eventually say read the skills create skill creator skill. Exactly. So this is how you know it's using the skill. Now, like I said earlier, in my opinion, it's just better to force it because then you never have to give it up. Yeah, you don't have to worry. Also, notice that it says skill.md there. Remember we were talking about markdown files earlier? That's literally all skills are is it's a markdown file that the AI pulls up quickly and goes.

1:15:16Yeah. Love this. Metaskill for the live stream. Before I draft it, a quick gut check on the angles since show off skills can land differently. and then now it's going to give me the same thing that copilot did earlier which is um uh yeah to skill that teaches what skills are while demoing one is that meta enough uh yes i i like that let's do that full meta yes i go full meta i mean why go half meta right Yeah. So it asks you a series of questions in order to clarify any uncertainties, which is really useful. Now, notice that I'm using here, I'm using Opus 4.7, which Anthropic hasn't fully given up to the Instant Thinking and Pro naming convention, but their models follow that convention.

1:16:10So Sonnet is kind of like their basic thinking version. Haiku is their instant. And Opus is their pro model. So you can think of it that way. And instead of having an effort toggle, they have this new thing called adaptive thinking, which lets Claude decide how long it should think. It's okay. It's okay. Is it not great yet? I mean, it's good. It's gotten better since the first week. It clearly takes a little tuning to do that well once you release one. yeah yeah um but yeah so what it'll do is now it'll go and it'll create this skill for me which means that anytime i'm in a chat i can either type like neuron ai or neuron newsletter writer or neuron podcast email editor or uh good taste or job search cover letters which is the one that we created in the uh uh the ai for total beginner stream that we shared in the chat i have a vibe coding questions to always ask skill that like anytime I'm, I'm making something, I run that skill and it asks like, are there any of these exposed?

1:17:14Are there any of those? Is this a problem? How is this? And it kind of runs through and does that. I call it the idiot check for me to make sure I haven't done anything careless or reckless. Yeah. Now to make, to actually create skills, you might have to have a certain setting turned on. Let me double check that. Under capabilities, this is on Claude. Yeah, you have to... Yeah, skills have moved to customize. Okay, so I don't think you have to turn them on. I think they're just automatically on with a pro account, which is good. And as you can see here... They were kind of in a preview for a while, weren't they?

1:17:57Yeah, exactly. Yeah, so we'll let that cook. I'll show you how to do this on ChatGPT as well while that's happening. Also, there's a good mention here that you can edit them as well. You can do that in both. On ChatGPT, these only appear on business-level accounts right now because it is in research preview. But they've been there a couple of months. I imagine we'll see them rolled out on a broader scale in the very near future. So the way that you get them is, you know, you click on your account in the bottom left corner and you see skills. skills you click on that and you can see all of them listed here and then you can create a new skill create with chat you can create with a skill editor which is just basically like you set the name the description you write the instructions you upload files so you can get really granular with it if you want I would recommend just doing in a new chat at skill creator and then say hey let's create skills to do xyz now that's probably the easiest way to do that now if you don't have chat gpt this is a little different than my approach too my approach is when i go through and i do a thing manually and i get it dialed in to where i want and then at the end of the conversation i'll say create a skill that will give me that and it will just instantly pull up the skill creator skill then it will create the skill and then anytime you use it if you notice you know what this isn't so great it's giving me something I don't want just tell it say hey edit this skill to not include 15 of these things or something and it will just make the change and it'll have you I think you click a there's an install button but there's also an update button when that happens that you hit oh i did not mean to click this what did you do design design yeah this is a whole other thing uh let's let's not get into that now

1:20:03okay i'm gonna skip out of claw design well it's very cool that's a whole other cool feature that we can talk about if people are interested okay so why would you want a skill well skills basically help it help you so that when like if you're following cory's process you never have to repeat yourself again in terms of how you want something done like you don't have to save the prompt yeah you don't have to save the prompt you don't have to set a dedicated project folder because this is how i was using skills back in uh in in uh 2025 yield 2025 eons ago um is that I actually had project folders for all of my previous things that I would use skills for.

1:20:47So if I had a prompt saved, like let's say I wanted to like condense the bullets for around the horn, I would save that prompt as instructions. And then I would just start a new chat and I would, you know, paste all the tools in and then it would just run that prompt. Now, that's a skill, right? Around the horn is a skill that I can call. So I don't have to have a dedicated project folder for it. I can have a project folder that's just like the neuron general purpose. And I can say like around the horn, you know, right around the horn. Make this an around the horn. You know, and then I can include the tools, right?

1:21:23So it helps you where you basically have essentially a prompt library. Oh, did I stop sharing? Oh, yeah, you did. I thought that was intentional. Oh, no, I meant to share everything I was just doing. that's funny. Let me show it again because I think it'll be helpful. Take two. Okay. So previously, previously, I had different project folders for every recurring task I had to do. And this was a great way to use project folders because basically you could set, and I'll zoom out a little here. I could set the instructions, which is just a pre-saved prompt and I could save that. And then anytime I have a chat in here, all I would have to do is share the context and it would run this prompt and that works perfectly but now i can just have one project folder that's just like hey this is everything that i do for the neuron and i can run the around the horn skill and i don't even have to go into a separate project folder for it i can just say like turn these into around the horn bullets yeah so this is like a really really awesome way to automate your work without needing an agent now yeah and it's it's like i'd like to think of it as a step back from full automation like it's much faster but you're still in control yeah and then i mean you are with an agent too you're just more wearing a catcher's mitt and less guiding it down the road yeah so nick nick mezzoni uh forgive me if i'm not pronouncing that correctly uh cory could you pull up that that question so are skills just reusable custom instructions you can reference there um there that and more plus tools plus yeah um let's go back here so under the customized tab on claude we're going to go to skills and we're going to look at because i love this as a reference the skill creator skill so so first of all you can see the project format of the, sorry, the folder structure of the skill.

1:23:33So there's the skill MD, which is the basically like the instructions for the whole thing. And it explains how you create a skill, capture intent, you interview and research, you write the skill.md. Then it shows you the anatomy of a skill. And this is actually really helpful because this shows you exactly, you know how how skills work you know it's got the skill md which is required then it has the yaml front matter which that's technically don't need to worry about that and markdown instructions then bundled resources and you can actually include scripts here so this is where you can use skills in coding sessions and you know in automations that you run there's a lot of different powerful ways that you can combine scripts and instructions in order to create skills And then you can also add references and assets.

1:24:24So it can read docs that are loaded into context or files that you need in outputs like templates, icons, fonts, et cetera. So they use a three-level loading system. There's metadata, which is the name and description. The skill MD body, which is in context whenever the skill triggers. And that's when you see the little scroll comes up. That's what it's reading. And then bundled resources, as we just said. So, and then you can look in here and you can look at all these folders and examine, you know, there's some Python code, there's some HTML, there's some schemas, et cetera, et cetera. So, but you don't have to worry about all of this because you have the skill creator skill.

1:25:03So all you need to say is you need to say, skill creator skill, skill creator skill, skill creator skill, make me a skill. You know, boom. It's like, that's, it's, they're, they were really genius in, in, in creating that. And it's really, really helpful. And the most important part, right, is like, let's say you use Claude and Codex or and ChatGPT. When you go to customize. You can actually download any skill that you created. You can't. We tested this last week. You can't download anthropic skills that they created. But like, for instance, I have a subject line generator skill. I can click this and I can download it.

1:25:44And then now I can take this. Yeah. Yeah. Now I can take this into Codex. I can take this into the terminal. I can use this with open router. I can use this anywhere. Right. So it's very helpful. Yeah. It's funny, too, because like the ones you can't download or like they're not anything fancy. Like it's like super proprietary. And like we did last week, you could technically just read them all and then recreate them. All right. You can literally copy paste the text out. Yeah. Yeah. Now I want to show Codex because Codex is the one way if you don't have a business account, you can actually use this.

1:26:22Do you want to share your screen for Codex or I can do it? Yeah, I'll share my Codex. One moment, please. By a show of hands, I saw somebody, or show of hands, by a show of comments in the chat, who uses Codex? I know I saw at least one person, maybe two. Chemical Coyote, let us know if you use it. I'm curious since you're a power commenter in this chat. which appreciate you, dude. Okay. Let me make sure I'm big enough. That was not a short joke, Grant. Let me... Also, for folks who, you know, this is the first time you're hearing about skills, if you have any clarifying questions about them, do ask in the chat.

1:27:08Yeah, absolutely. And where they come into play here, this is Codex. and Codex is I love Codex so speaking of which Grant have you seen Pets yet? Yes we shared it on Monday in the newsletter. Oh okay. My favorite is the Dario pet. The angry Dario yeah and they made the little Claude Code guy and Tebow commented on it was like oh four hours I lost a bet like obviously that was coming really quick That's hilarious. I don't actually know how you make them, though. Do you know how to make them? I saw, like, the directory of them, but I haven't made one yet. Watch this. I think it's pet, yeah. Oh, wow.

1:27:56And it's supposed to... We should make one for Ollie, the neuron cat. Hang on. Settings. Okay, so Chemical Coyote says, I have not used codex yet, been meaning to. I've been caught up in the whole new AI laws awareness. Oh, yeah. That's a whole other can of worms. Yeah. Yeah. Yeah. Are we talking about the one I wrote about the other day, Grant? I think so. Yep. That was terrifying. Look, freedom of speech. Gotta use it. Hey, that's right. That's right. That's what it's there for. Okay, so like in Codex here, Codex is essentially it was designed for coders, but I use it for everything I do. I use it for writing.

1:28:40I use it for anything I want is the truth of the matter. And it's really intuitive. It's easy. You can change a lot of things. It'll control your computer if you want it to. You can adjust what your permissions are. Like this is my work computer, so it has default permissions. My home computer has full access, which is a different ballgame. I don't recommend it, but, man, it's more fun that way. And, you know, you can adjust these levels a little bit differently in here. So the nice thing about codecs is you have low, medium, high, and extra high for whichever model you're using. So you can use 5.

1:29:21You can use GPT-5.4. You can go back here. 5.3 codecs was a hell of a model, too. I keep thinking we'll get a 5.5 codec specific one at some point as well. It was great to write with because it doesn't fall into all of the AI writing traps that a normal chatbot does. And you also have speed. So you can run these on fast mode, which is lightning fast. This is running through the Cerebris chips, which is crazy. It's also exponentially more expensive, if I recall correctly. Yeah. So you'll blow through your limits pretty quick. I've used this for a long time with the$20 account, and it's been okay.

1:30:06I now have upgraded to ProLite because they have the— Which is the$100 one. Yeah. Coyote gets it. My claw has complete access. I hear you. That's funny. Well, here is interesting, Corey, because you can do plugins and skills from here. Plugins are essentially chat GPT apps or MCP connections for the most part. And they have a billion of them. And there are other ones you can add as well, built by OpenAI. And I don't think they've let outsiders into them yet, other than like through official channels. Like they have Little Caesars. they have. Yeah. That one was a little disappointing last week.

1:30:53We tried to order a pizza last week on the live and it didn't work. That's like, I always joke. That's my AGI test. Little Caesars needs some agent automators over there or something. Yeah. And here under, after plugins, you'll see another tab over here in Codex. It's your skills. and you can come down here and it has a number of them built in that you can use it has you know here's your skill creator skill and just like with i like to have this option where you can try and chat sorry to cut you off but at the bottom yeah yeah like without installing it you can just try it and see if you like it you can also uninstall skills if you choose turn them on and off they got an easy toggle uh the difference between disabling and uninstalling by the way is like when you uninstall it, that removes it.

1:31:43If you toggle it off, that just means it won't randomly trigger. I mean, it won't randomly trigger, but if you, for example, didn't want it to use, you know, XYZ browser use or whatever for every task, you could toggle that off. Whoa, it went over and opened that in VS Code. It downloaded their created skill into VS Code. I've got a, is it just a plain old markdown file, Grant? Yeah. Their skill creator. Well, that's nice. You can download that. But you can take that wherever you want, I guess. That's a good check mark. There we go. There's also Plugin Creator, which is really nice. I'm going to go shove it into my open source models.

1:32:18The Plugin Creator would basically let you create your own app connector. So that's something to keep in mind. If there's a piece of software that you use on your computer that there isn't a connector for, you can use that Plugin Creator to try and create one. Neuron Podcast Notes, this just helps us kind of starts a research doc. before we interview someone is the idea is like grab their latest info, current title, financing, all of that information for us. So we can sit down and instead of spending an hour and a half Googling those things, have it all ready. And right there in front of us to, to plan how we want to, how we want to tackle a conversation.

1:32:57We have a cool one this afternoon. You will not, if you've been following AI in the last, let's say three days, you will not want to miss next Wednesday's episode. So that's all we can say. We can't say who it is yet, but it's going to be interesting. So this is an interesting comment from Chemical Coyote. I think Codex is a good place to start with beginners since OpenAI does restrict GPT more for safety than Claude does. I 100 % agree with that. That's kind of my approach. And I like it because, you know, it used to be a little slower, but it's not so much slower now. It's fast now. but I find that that extra bit of time resulted in a, in a better hit ratio for someone who is not an experienced developer.

1:33:44Um, and it's, it's really good about walking you back and explaining things. Like I'll put in my, my personalization notes, like, Hey, I'm not a developer, but I'm going to mess around and we're going to play with things. So make sure I'm learning as we go, explain what is happening. Ensure if you ask me to do something that you say it clearly and that I have a step-by-step guide and, you know, make sure I'm learning. So you can do that and work through them well. The truth is they're both great. You know, if you use this or Cloud Code, you're going to be happy. I would just expect that for every three months, another one will be the best for three months.

1:34:22They'll just continuously flip quarterly as they have a tendency to do. well rhymes with rodriguez no no but that's funny fair guess yeah so i don't think it rhymes with anything in retrospect no that's not guess um so there's a couple things on here that i wanted to highlight so obviously i think anyone can use this with the 20 plan is that right cory do you have to have a specific codex plan i know that they added a codex plan you can use it with the 20 plan okay yeah so you can also run open claw with their 20 plan if you want open claw i'm going through it pretty quick because open claw likes the tokens man so this is the way that you use skills if you're not on the business account right you can see here cory can use all of these skills in codex it's just a simple download if you use chat to bt you can use codex um and it's not that intimidating um cory why don't you show them how to make uh projects on here where you can organize your chats so okay that's a good idea that's a good idea so let's go over here you'll see i've got a project called cory but uh let's make a project start from scratch we're gonna call this uh let's name it live stream demo because we're original like that yeah yeah yeah we didn't disconnection you're losing connection or i am no no no no somebody in the chat asked if we lost connection we did not if at any point you um actually i'll just put this in the comments go ahead yeah so you know you start a new a new thing here what you should know is it's going to show you when you hover over it the location of that folder it can't go beyond the walls of that folder unless you specifically instruct it to.

1:36:17So like if you say you're wanting to build a simple app of sorts, if you build an app with a folder with what you want to call it, it's going to be as simple as that. You know, it will work within the confines of that folder unless you tell it otherwise, which is the smart and responsible way to use tools like this. I have to say that, but I don't always practice it. And so, yeah, you've got your permissions. And you could come in here and you could say, build a small timer app that records, oh, wait, that only sets for five minutes. So, you know, we could do, I should have put it on fast. Since we're on live, smart and responsible.

1:37:13That's right. Going to take a quick look at the workspace. It's going to run through things. It's going to check its skills to see what it has there. And in a minute, it'll build a thing and we'll play with it. Tiny standalone web app, one HTML file plus iOS. Anybody can do this. I could literally teach. I don't know. Anybody. Anybody could do this. The catch is you could also break a lot of things. So it's important to just do it slow, be deliberate, have another AI as your like verification tool. You know, don't hesitate to like grab what you see here and go say, hey, is this normal? Is this right?

1:37:51And, you know, you can work back and forth with them and it will it'll be very helpful for you.

1:37:59So we had an interesting comment in the chat that was, is there such a thing as too much context where there's diminishing returns in context you give to the AI? yes there is um and that's a problem with the current models um a lot of the companies will offer up to what's called 1 million uh tokens context window but if you get over 150 000 it sucks it's like 200 000 to 300 000 depending on the model i find like gemini you know can do well well up to 300 000 um claude i think is like 200 000 although i will say that claude code in the terminal is pretty good. And with the 1 million contacts, it's not, I feel it more at 700 ,000 contacts, like with Cloud Code, where I'm like, okay, but it charges you more.

1:38:50So the more contacts you use, the more expensive it gets. And there's reasons for that. But I would say like, once you get to about$200 ,000 worth of token,$200 ,000, sorry, 200 ,000 tokens, above that point, that's when you start to get diminishing returns. Yeah.

1:39:13Okay, what are you doing, Corey? I'm making the corniest app ever. What is it, a five-minute timer? Oh, this is a five-minute timer, lonely time for five minutes. But I just wanted to show how simple that was for anyone who thinks, you know what, I'd like to build something. Like, you can do it. Pause, reset, it'll do the thing. Now I came back and I told it, that's great. Add a slider that will let me change the amount of time And let's change the name to Unlimited Timer. So I'm going to share a prompt that I thought was really cool. Visual comments for codecs with a single click or drag to select an area.

1:39:44Check this out. So, Grant, you can be like, use a different font here.

1:39:57And boom, it's got a comment right here on the title where it will go back and change the font. You could come down here and say, I don't like these buttons. Make them rounder. Whatever you want. It's a lot of fun. That's awesome. Oh, compaction is my friend. I overuse it relentlessly. Do you trigger it or what do you mean by that? It automatically compacts when you're in between and it's close to the context window in Codex. It just goes compacting and it takes about a minute and goes. and you'll lose some elements uh but i mean i've done it in excessive 10 times in the same conversation easily so i just shared in the chat something we talked about earlier which is um from every's stream yesterday which was really great we shared it in the neuron this morning you should definitely go watch that video um subscribe to every they're amazing um but they they shared a chat where Austin basically said he uses codecs now for all of his tasks and he sets up different folders like we showed you for everything that he's working on so for example Corey's got this live stream demo folder here for any side project you have you can create a project folder and work on it that way but he recommends he he recommends sending this prompt that he shared in the in the stream yesterday which is go take a look at the things I use the most and think of some automation that would help me with my work.

1:41:28And so this is a great way, like if you don't know how to come up with automations or even skills, like you could ask, like, what are some skills I could create for the work that I'm trying to do often? You can adapt it and use it based on what we've learned so far. That's a great tool for helping you. If you don't know what you should automate or you don't know what skills to create, have it look at the resources, like the apps that you use and tell it about yourself and have it give you suggestions. I think that's a really helpful tip that they shared. And, you know, that's why we shared it, because we thought it was great.

1:42:03We thought it was great. I love it. I love it. We don't have to sit here and watch me make this stupid timer. I really was just kind of illustrating how simple it was to play around in here. Like, you could absolutely create a single project, have a folder, and be like, build me something. You know, build me a silly web game, you know. And if you can't figure out how to open it, tell it I can't figure out how to open it. Give me a give me a direct link or something, you know, and it'll it's pretty good about finding ways to make that better. Yeah. And, you know, this doesn't have to be this is not limited to applications.

1:42:43You can create spreadsheets in here. You can create presentations. You could create pretty much anything you can think of inside of Codex. It's a really powerful app. It'll do Google Docs, Word Docs. Well, Word Docs. And yeah. Last year, before compaction. Last year. Back in the day before compaction. Maybe we could cover very briefly automations. Do you use any automations, Corey? Do you recommend people start with this? Do you recommend they ignore this? I have a couple that I have set up in here that I have not had good luck with. But there are a number of things you can do. The ones that are pre-built into codecs are very much centered around coding and summarizing pull requests and stuff like that.

1:43:26But I do believe they have a couple in the pre-built. But you can go up here to New Automation in the corner and tell it, you know, I want you to check the weather at 6 a.m. Central Standard Time every weekday. and report it to me. So this is where we get into the agent territory. Like someone asked you, you need to know how to code to use agents. Well, with automations, it's not quite an agent, but it's like the next closest thing. And it's an agentic workflow and you can connect it to your skills with that button. Or sorry, that's the model. Can you connect it to your skills actually? Yeah, you can just tell it to...

1:44:14Yeah, you tag them in the instructions. Yeah, you would just say, you know, use the neuron blog article skill to write about one of them. Yeah, so once you have the skills created. Oh, wait, not the weather. I thought I had it checking the news. That's so stupid. Oh, God. No, that's fine. But it shows you how to use the skill, right? Right. So basically, once you have skills that you know you work, you know, you've created them, you've tested them, you can then put them on an automated cadence here. And you can either set that on a timer. I think there's more powerful ways where you can set it on a trigger.

1:44:57And that's how you can have an agent or an agentic workflow in this case run for you automatically. Codex itself in this app is basically an agent. So if you're working with Codex or Co-Work, which is the Cloud equivalent, that's essentially working with an agent. And so you've essentially, you know, like you can have that agent be your like orchestrator agent, and then it can create sub agents for you to go off and do things. And you can literally ask it to do exactly that and it will come up with it. But it's better to have a documented process for the workflow that you want it to do. and in my opinion, and I think Corey would agree with this, Corey, you can let me know, the best way to do that is by trying to do the task yourself manually in the chat, writing down all the instructions, giving it all the context you need for it in order to do the task.

1:45:45And then once you've figured out that it can do it, you know, you've connected all the connectors, you've managed to get the result that you want out of it, then you can turn that into a skill. And then it only then after you have the workflow set as a skill, you can then automate it to run for you automatically. Yeah, and I wouldn't hesitate to dial it in a little bit before you do it. I mean, it's only going to work better. I mean, you know, when you get to a finished result, like I regularly use, you know, create a skill that will give me that finished result when asked. And it usually works great.

1:46:21It's really not that complex. It's more intimidating to think about than it is to actually do. If I had to – I would say skills are probably top two unlocks since 01 released. Yeah, I agree with that. From a productivity standpoint for context here, if you run a business, something you should know is three weeks ago, I had a company leadership retreat where we went with all of the company's managers. It's like 60 people. And while there, I taught a three-hour course on how to build skills. It was what's a skill. We built one together. And then I brought someone up and built one with them so everybody could see how it thought.

1:47:08And then set everybody loose to go build their own. And God knows how many man hours it's saved just spending a couple hours doing it that day. I highly, highly, highly recommend understanding these because it'll make your life easier. yeah it's just like this is i think this is the path right you go from from you know manually chatting figuring out what you need to know what you don't know in order to get it to do what you need it to do adding the connectors um which are you know the external applications that it needs to access so that's adding the context and then converting that into a skill and then once you have all these skills that you can use on a rotating basis you know now you're off to the races and you can actually start automating things.

1:47:52It gets much easier. Yeah. Then the step beyond that is what Stan does in the chat, which is you use this to actually write code for you. Leroy Jenkins. Yeah, exactly. Yeah. Especially if you're ADHD. Yes, I can say with speaking for myself, you know, I relate to that. So it's incredibly, incredibly helpful for keeping, you know yourself on track with things that you need to do and speeding up your time um to do things and we really kind of put the duh in adhd if i'm being honest yeah i think cory and i can both relate to that yeah that's why that's why i'm always telling people like it's okay if you need to tell us to focus you know get on point yeah so i forget so um is there anything we haven't covered yet that we that folks want to talk about i know we talked about creative tools at the beginning.

1:48:48I don't know if the person who asked about that is still here. You know, we could go a little bit more in depth on that. What do people want to know? We can go deeper on agents. This is for complete beginners for starters. Teleportation. Let's talk teleportation. I'm just kidding. Teleportation. Yeah, don't ask. I got a text message that mentioned it, which is weird enough in itself. By the way, quick reminder, if you're here and you're still here, We're grateful for you. Please take just a moment to like and subscribe above. Helps the channel grow. Helps us keep doing it. And it means a lot to us.

1:49:24Also, pop by the Neuron.ai and read the Neuron newsletter every morning of the week except for most Saturdays. And I guess that's the way to say it. Yeah, for now. And it very much, you know. Hey, Paul. Oh, Paul's still here. That's awesome. Yeah. Yeah. Okay, so creating 3D models from images. There's a tool that I wanted to mention that's sort of an alternative to Adobe Firefly, which Corey mentioned at the beginning, which is Comfy UI. And I think Comfy UI is probably where you want to do that. Yeah. So I'm going to share in the chat. I've not done true ones, but, like, I once took a picture that I drew, like, with a pen.

1:50:07And then I had Flux make it into an image. And then I used another thing. I used Flux again to make it like a three-dimensional Pixar-style character. So, like, it was really wild to watch. Then I took it to a video model and made it move. So, like, it went literally from pencil to paper to an image to a 3D-ish character to video. And it was the wildest thing. We should do that sometime on a stream. That'd be fun. Yeah. I'm going to share my screen here for a second. And I'm going to share. So I also want to get to Jay Bormel's question, which is about the limits of LLMs. Yes. Yeah, that's probably a good place for us to wrap up today because we've got another thing we've got to be at.

1:50:53So this is Comfy UI's website called Comfy.org. and basically it's a an open source product which means you can actually just go to it on github you can download it for local you can do it on the cloud and basically you the way that it works is you add models and you can run those models and you can you know let's say you start with this asset and then you run a workflow and you combine these two sub models together sub images together And then, you know, you've got the canyon edge here and the depth map, and then it creates a image. So the way that you would do this for a 3D model is you take a 2D image and then you convert that into a 3D model with a workflow.

1:51:37And Comfy actually has a lot of workflows that are already, like, established that you can download. It's very community oriented. So, for example, there's all these ones here. Let me see if they have a 3D one specifically. you can see all the popular models that are on here um node graphs might be under there let's see if i was to search just like 3d yeah so now you can see the different yeah now you can see the different 3d models so there's image to 3d which is a new one uh i think this is from Tencent. This is a popular one. So HY3D. But in addition to, oh, Meshi, that's another model. In addition to this, there's also a website called Spline.

1:52:29By the way, we did a full demo of that if you're curious and wanted to know how to get Comfy UI up and running. Have you seen Comfy make a 3D model for 3D printers? Yeah, I think it can do that. You might want something more, a little more powerful though. Or want to use it on the cloud versus one's own computer depending on what there's a caveat i'm going to give to this as well so spline's another one where you can create from a 3d design and it's a little bit more powerful and dnd minis yeah they have a lot more different options like there's mock-ups there's logos like they're they're all dedicated to 3d work here which is really useful so i would save this website and check this out as well.

1:53:14But then there's also, let's see, Claude for creative work. Claude also offered some new connectors last week. And so I assume that you could probably run these connectors in GPT as well, but don't quote me on that. But you can actually use a connector for like say autodesk right directly inside clod so this allows designers to basically make 3d models in autocloud directly with conversations with clod um same for blender so now you can work with clod to help you create blender um tools and the way that you would do this is you would uh sorry 3d assets with blender the way that you would do this is you would do this on your desktop you would have the desktop model open and you can use what's called computer use and you just tell Claude like, Hey, you know, I, you know, I'm creating an asset and blender, you know, for a 3d printer.

1:54:11I think Claude's answer to apps. Yes, exactly. Okay. Good context. Yeah. Yeah. So these are different connectors that are like apps. So for example, I don't know if I click this, if it'll take me to, yeah. So you can just go, I'll share this link because I think this is helpful. So when I clicked that tab, it took me directly to the Claude connector here. And you can see, okay, here's the blender. But I have to use it with the desktop. Yeah. But yeah, it shows you like you just download this and you attach it. And you can call it the same way that you call skills that we talked about earlier. And install it like anything else, I'm sure.

1:54:53Yeah. So another one, just to wrap up on this, is Meshi. We mentioned Meshi. Meshi is like a spline competitor. So we'll just look at that real fast. And then we'll cover the last question from Bormel. um and so you can see like they have different workflows like if you want to do like product design if you want to do film production game development 3d printing they have 3d printing right here yeah so they have workflows for that and that are custom around that and i would say like the the hardest thing here is uh top uh topological topography i forget exactly what it is but it's basically like once you've created the mesh right like you also need to do the the the creative work on top of it in order to like do all of this coloring and make it manipulatable and these companies have been working hard to solve that problem i don't know if it's fully solved yet i imagine it's not but uh they get you a lot of the way there so where it fills it with like sample colors or something yeah and then you could go back and say well i want to change this one and this one instead of having to select every color and every shade of everything to go throughout the whole thing i guess yeah even if it's just like to help you like get started and then you take it into blender and you edit it yourself yeah with that now accepting 3d printer donations yeah yes exactly i will take one as well well to everyone here thank you so much for hanging on it's been a it's been a long one but it's been good and uh i hope i hope you last question cory let's not forget the last question the llm uh the limits of LLMs.

1:56:29Let's see the Bormel question. Oh, yeah. Yeah. Where was it? Where was it? Where was it? Advice for advice on the limits of LLMs. So before we fully wrap up, let's make sure we cover this. And I think what we should do is we should just go around the horn. You answer and then I answer. If you want me to go first, I can go first. Go first. Okay. So I would say the limits of LLMs, it's a hard thing to fully explain in three minutes. But one of the things that Bormel called out, and forgive me if I'm pronouncing that incorrectly, at the beginning was a lot of times when you ask a question, the way that you ask a question biases the answer you're going to get.

1:57:15So if you say like, hey, is this a good idea? Like, this is my business plan. Is this good? It will try to help you. And so it will tell you all the ways that it's good. But if you look at like Mark Andreessen's prompt that went viral a couple of days ago, we shared it in the neuron. He had all of this language around saying like, you know, do not preference my ideas, like push back, like don't be afraid of hurting my feelings. It was very much like like bullying the model into like being mean to you and bullying him, bullying him back. Yeah. Yeah, yeah, yeah. And trying to weed out a lot of that, you know, what some people call sycophancy, but basically it comes down to, you know, like it's trying to help you.

1:57:58And so it's going to tell you your ideas are great. It's going to try and help you accomplish your goals. so the way around that is you know there's a lot of tricks you can do to try and get it to not self-preference to the way that you phrase it and that has to do with you know you can say push back against my idea you could say like i'm considering two options a and b um what do you think of both um give me the pros and cons of both there's a lot of language that we share in our prompt skill of the day to help with that bias um but but that's just something important to understand. And there are more limits as well, but I wanted to call that out because I know Jay Pormel mentioned it specifically.

1:58:36On the Lumen AI Lab's comment here about semantics and framing impose and shape the reflection, Grant and I dealt with a deal like this very, very recently on a problem where there was a specific thing causing an issue. I'm not going to get too into the specifics. There was a specific thing causing it. We were having an issue and Grant was pretty sure it was around this one thing and started working on it like, Like, how do we optimize that? How do we optimize that? And didn't really see the results we wanted to see from it. And then we went back and took a much more open approach. Like, what's the problem with this?

1:59:10And just let it. go look at all of it and make up its own mind, which is a thing that actually got us both results neither one of us ever would have considered. And we got the same result from both Claude and ChatGPT pretty much with a couple of little variations, but enough that we were like, okay, so there's something to this. Yeah. And on that note, like something to keep in mind, which we didn't cover in this at all, because it's a one layer deeper than like the chat interface for a starter kit is the way that these AI models work is that they are probabilistic, not deterministic, meaning that they have a distribution of possible ways they can answer.

1:59:50they don't have a set way that they're going to answer every time. So the way that you ask the question, to Lumen's point, can totally fundamentally change the answer you get. Even changing one token could give you a completely different answer. Even asking at different times can cause a completely different answer. I think they've solved this by now. I'm not 100 % sure. But there was even an issue where however many people are asking questions at the same time you ask it, you could get a dramatically worse answer because of the batch size. I think understanding bias is very helpful. Understand bias, understand self-bias and accept that we all have some of it in one way or another.

2:00:27And if you can learn to correct for that bias in some ways, you will find it. I'm not talking about political bias or any of that. I'm just talking about bias in the way we ask questions. Perspective and ensuring that you're stepping back to take a good macro look at things. I believe that you'll find you get better answers. And that's, yes, that's why context and memory are so important. Yeah, because there's ways you can add these automatically where you don't have to remember to prompt it a certain way every time. Like you create a recurring skill that basically goes and checks your results for biases.

2:01:06Like Corey mentioned, like you have a vibe coding skill that you run that checks your code for bugs. I mean, it's essentially the same thing. You could create a skill that that, you know, checks for the 10 most common errors that vibe coders make. Yeah. So you can do like check for like check this for like the 10 most common logical fallacies and logical errors in my assumptions here. You know, you know, check my assumptions, research my push back against my beliefs. Like, why am I wrong? One of the things that Andreessen said in his prompt was kind of good. is like, don't change your answer unless I give you a better, like, unless I give you new information that actually like changes it.

2:01:45Like, don't just like, like, give me a different answer just because I push you, push back on you. So I like that as well. I'll share the link to that before, before we drop off here. Well, everyone, I want to say thanks again. It's been, it's been a blast and we're still going to keep doing these every week, but not always a beginner one. We're going to try to, since turnout seems to be good and they seem to interest people, we'll try to do a beginner one once a month or something, especially when we can update on new things that might impact beginners and what they mean. Next time, maybe we'll take that approach.

2:02:17Like, what's the latest that you care about? Oh, I didn't even, we never even got, Grant, we never even got to things you should ignore. Benchmarks. If you're a normal person, ignore benchmarks. You really don't need them in my opinion yeah that's why when we test stuff we're just asking it questions like for fun we're not running sophisticated i think evals are important but that's when you're running them in production so that's not a ai starter kit thing um so you know put put a pin in evals and we'll have a whole we need to bring in someone who's like an eval expert and how yeah we've done we've done that a couple times we brought in artificial analysis and we brought in the other by the way Don't forget to check out next Wednesday's edition of the podcast.

2:03:00There's going to be a really cool one coming out that if you're keeping up, you're going to want to know about. Also, we may very well have an interesting announcement to follow that in a few days after. So stay tuned. We'll be back and see you next week.