WEBVTT - The Missing Piece Is Ideal State 0:00:19.800 --> 0:00:22.400 Hey. What's up? So Theo's got me triggered right now. 0:00:22.440 --> 0:00:25.520 He basically put out this video here talking about how. Yeah, 0:00:25.560 --> 0:00:27.520 turn off cloud code's memory. That's what the title of 0:00:27.520 --> 0:00:30.160 the video is. I want to let you hear the 0:00:30.160 --> 0:00:32.080 opening piece that kind of triggered me here. Made me 0:00:32.120 --> 0:00:34.520 want to do this video. So let's listen here real quick. 0:00:35.400 --> 0:00:36.680 Work effectively in a code base. It needs to know 0:00:36.680 --> 0:00:37.840 what the code base is, where things are, and how 0:00:37.840 --> 0:00:39.040 to get stuff done. I found that a lot of 0:00:39.040 --> 0:00:40.800 people want to store this information in some special way 0:00:40.800 --> 0:00:42.280 that AI agents will operate with. Well, I've seen so 0:00:42.280 --> 0:00:44.319 many different systems and libraries, plugins and features and stuff 0:00:44.320 --> 0:00:45.480 that I have built over the last few years to 0:00:45.480 --> 0:00:47.080 try and automatically encode what they're doing in the code 0:00:47.080 --> 0:00:48.840 base in some special way. That'll make the agent always 0:00:48.840 --> 0:00:50.080 do what they want. I've even noticed that in cloud 0:00:50.080 --> 0:00:51.880 code recently, it seems to use its memory a whole bunch, 0:00:51.880 --> 0:00:53.199 where it will save things in some magic hidden file 0:00:53.200 --> 0:00:54.760 somewhere on the computer after you ask it to do something, 0:00:54.760 --> 0:00:56.400 and now it will keep randomly doing that forever. You 0:00:56.400 --> 0:00:57.760 can't tell from how I've been spending this so far. 0:00:57.760 --> 0:00:59.280 I'm not a fan of these memory systems. I don't 0:00:59.280 --> 0:01:00.560 think memory is the right place to keep track of 0:01:00.560 --> 0:01:01.720 how things should be done in your code base. That 0:01:01.720 --> 0:01:02.120 worked okay. 0:01:03.480 --> 0:01:07.440 I don't think memory is a good place to keep 0:01:07.480 --> 0:01:11.440 track of how things should be done in your code base. 0:01:13.360 --> 0:01:18.660 That's what he said. Okay, so here is my response 0:01:18.660 --> 0:01:22.780 to this. The primary cause of AI systems not making 0:01:22.780 --> 0:01:26.700 us happy, is it not knowing what we want? Okay, 0:01:26.940 --> 0:01:29.940 now he is objecting to a specific thing in a 0:01:29.940 --> 0:01:33.899 memory system, which is this automatic writing that cloud code 0:01:33.900 --> 0:01:36.180 is doing. So basically cloud code has a feature. If 0:01:36.180 --> 0:01:38.060 you turn it on where it will write to this 0:01:38.060 --> 0:01:41.219 memory system. I think it's literally memory.md unless they changed 0:01:41.220 --> 0:01:44.660 it and that memory MD would be consumed every time 0:01:44.660 --> 0:01:47.340 you start a session. So it's a cool idea, right? 0:01:47.380 --> 0:01:49.260 But he's saying it decays over time. He's going to 0:01:49.260 --> 0:01:51.340 show you how it decays over time. And he's actually 0:01:51.340 --> 0:01:53.900 right about that. But he is still throwing out sort 0:01:53.900 --> 0:01:57.660 of this baby with the bathwater, basically still saying that 0:01:57.660 --> 0:02:00.980 memory systems are bad. I'm going to present an alternative view, 0:02:01.220 --> 0:02:04.660 which goes against both Theo and Matt Pocock on this, 0:02:04.660 --> 0:02:06.580 and kind of takes us in this direction of talking 0:02:06.580 --> 0:02:09.140 about bitter pill engineering, and what are we actually trying 0:02:09.140 --> 0:02:11.340 to do with a harness? So let me let me 0:02:11.340 --> 0:02:14.500 show you the architecture that I use, which is open 0:02:14.500 --> 0:02:16.780 source project. It's called life OS. This is the overall 0:02:16.780 --> 0:02:20.829 system that I use and I think it is superior 0:02:21.150 --> 0:02:24.550 to what Matt Pocock is doing and to what Theo 0:02:24.590 --> 0:02:27.269 are doing. Okay. And to be clear, they are both way, 0:02:27.310 --> 0:02:29.870 way better developers than me. Ten years ago, like I 0:02:29.870 --> 0:02:31.710 was not a developer, I was a programmer. And they 0:02:31.710 --> 0:02:35.590 were both hardcore real developers. So in terms of development skill, 0:02:35.590 --> 0:02:39.990 I think they are just way better. Okay, no question there. Unfortunately, 0:02:39.989 --> 0:02:44.230 I think they're still in that world of building development skills, specifically. 0:02:44.269 --> 0:02:46.350 Matt Pocock. And I think Theo looked at his skills 0:02:46.350 --> 0:02:48.910 recently and wasn't a huge fan. But anyway, Matt Pocock 0:02:48.950 --> 0:02:51.550 is doing amazing stuff on the development side. Theo is 0:02:51.549 --> 0:02:53.990 obviously doing amazing stuff. He's actually my favorite AI creator 0:02:53.989 --> 0:02:56.030 right now. I watch like every video that he puts out, 0:02:56.549 --> 0:02:58.470 and that's why I was specifically triggered by this because 0:02:58.470 --> 0:03:00.710 I think he's getting this wrong. So the way that 0:03:00.710 --> 0:03:04.829 I approach this in life OS is I am specifically 0:03:04.829 --> 0:03:09.190 addressing the problem universally across the entire harness of one problem, 0:03:09.549 --> 0:03:14.510 the AI, not knowing exactly what would create euphoric surprise 0:03:14.510 --> 0:03:18.010 in my brain. That is the ultimate mission. It's actually 0:03:18.010 --> 0:03:22.769 in our system prompt is to pursue what would make 0:03:23.730 --> 0:03:26.570 euphoric surprise in my brain. Okay, that is the meta 0:03:26.609 --> 0:03:30.329 goal of all the entire harness. Okay, so in order 0:03:30.329 --> 0:03:33.010 to do that, guess what you need, you need extremely 0:03:33.010 --> 0:03:36.570 detailed knowledge of exactly what I want. Okay. And this 0:03:36.570 --> 0:03:39.410 is where I think Theo is getting confused and he's 0:03:39.410 --> 0:03:43.130 conflating multiple things because his video is basically saying memory 0:03:43.130 --> 0:03:46.170 systems are bad. Okay? That's ultimately what he's talking about, 0:03:46.370 --> 0:03:48.770 but he's kind of throwing out the whole thing with 0:03:48.770 --> 0:03:51.610 this specific problem that he has, which is automatic systems 0:03:51.610 --> 0:03:55.170 that write crappy memory files are bad. Well, of course 0:03:55.170 --> 0:03:57.130 we agree with that, right? So let me show you 0:03:57.130 --> 0:04:00.890 an alternative. Okay. Have the goal of the harness be 0:04:00.890 --> 0:04:06.010 to produce euphoric surprise in you, the user, the principal. Okay. Second, 0:04:06.650 --> 0:04:08.930 I believe and this is a strong claim, I believe 0:04:08.930 --> 0:04:11.570 that the way that we've been doing development before, and 0:04:11.570 --> 0:04:13.689 the way that Matt Pocock is currently doing it with 0:04:13.690 --> 0:04:15.490 all these extra skills, and you've got to call the 0:04:15.490 --> 0:04:18.820 extra skill and you're producing these test documents you're producing, 0:04:18.820 --> 0:04:21.780 these plans and these specs and then a PRD, and 0:04:21.780 --> 0:04:23.580 then you've got big documents and you turn it into 0:04:23.580 --> 0:04:26.580 small documents. And then the Ralph Loop actually, where you 0:04:26.580 --> 0:04:29.219 take one agent and you work on one particular feature. 0:04:29.780 --> 0:04:33.940 In my opinion, these fail for one specific reason. They 0:04:33.940 --> 0:04:36.340 don't have the plot. When you break things apart, you 0:04:36.340 --> 0:04:40.660 tear them apart. They cannot see the plot. The plot is, 0:04:40.860 --> 0:04:44.900 in my opinion, for all of AI, the ideal state. Okay. 0:04:44.940 --> 0:04:49.220 You cannot achieve euphoric surprise. Your harness cannot achieve euphoric 0:04:49.220 --> 0:04:52.260 surprise for you if it doesn't know what you want. Okay. 0:04:52.300 --> 0:04:54.700 And here's the trick. I could be asking for a 0:04:54.700 --> 0:04:59.140 beautiful picture of a watercolor butterfly. Or I could be 0:04:59.140 --> 0:05:02.060 asking for come up with an amazing business plan and 0:05:02.060 --> 0:05:03.820 idea that's going to make me tons of money. Or 0:05:03.820 --> 0:05:06.340 I could be asking for build me a beautiful website 0:05:06.340 --> 0:05:10.020 that talks about typography. All of those things can be 0:05:10.020 --> 0:05:14.380 articulated in one sentence or two books. Okay. How much 0:05:14.380 --> 0:05:18.280 detail is in my mind for what a perfect typography 0:05:18.320 --> 0:05:21.880 site looks like, how much detail is in my mind, 0:05:21.880 --> 0:05:24.080 and how much have I given the AI. The gap 0:05:24.080 --> 0:05:26.600 between these two is the problem with all AI harnesses. 0:05:26.600 --> 0:05:28.600 It is the problem with all agentic work. It is 0:05:28.600 --> 0:05:32.800 the problem with all AI, prompt engineering, intent, engineering, graph, 0:05:32.840 --> 0:05:36.320 all this stuff. It's all floating around. It's all orbiting 0:05:36.560 --> 0:05:39.240 this core concept. This is my argument to you. Doesn't 0:05:39.240 --> 0:05:40.880 mean I'm right, doesn't mean you have to agree with me. 0:05:40.880 --> 0:05:43.200 This is my argument. I think I am right, okay? 0:05:43.720 --> 0:05:46.760 It is the ideal state. What we do in life 0:05:46.760 --> 0:05:50.360 OS is the ideal state artifact one document. Now, if 0:05:50.360 --> 0:05:52.360 you have some giant system, you're building an entire role 0:05:52.400 --> 0:05:55.960 playing game. You can have one Isa that links to 0:05:56.000 --> 0:05:59.360 other ISAs and they're all cross-linked. Like if you've got 0:05:59.360 --> 0:06:01.840 a full geography system and a language system over here 0:06:02.080 --> 0:06:04.680 and like creating NPCs over here, like you could break 0:06:04.680 --> 0:06:06.840 it apart if it's big enough. But for all intents 0:06:06.840 --> 0:06:11.360 and purposes, that is still one giant Isa document. Now, 0:06:11.400 --> 0:06:14.000 what a single document allows you to do is what 0:06:14.000 --> 0:06:15.720 we have in our document. And I'm going to go 0:06:15.720 --> 0:06:17.650 into detail in another video where I take you through 0:06:17.650 --> 0:06:18.930 the code version of this, I want to show you 0:06:18.930 --> 0:06:23.450 the architecture first. So this architecture basically says, here is 0:06:23.450 --> 0:06:26.250 the stated goal. Here's the stated ideal state. We also 0:06:26.250 --> 0:06:28.730 have a section where you can just dump a full 0:06:28.730 --> 0:06:31.170 text conversation. I can actually call Chi and say, hey, 0:06:31.170 --> 0:06:32.930 I want to make this thing. I literally call him 0:06:32.930 --> 0:06:34.570 on the phone and he answers and we talk back 0:06:34.570 --> 0:06:37.930 and forth. It's absolutely insane. But you could also use 0:06:37.930 --> 0:06:40.890 the discuss skill, which is similar to Matt Pocock's really 0:06:40.890 --> 0:06:44.970 excellent Wayfinder and grill me skills. So you're gathering information, 0:06:44.970 --> 0:06:46.930 but what he's doing is he's kind of putting it 0:06:46.930 --> 0:06:49.570 in all these various places and he's breaking it into 0:06:49.570 --> 0:06:52.850 pieces and recombining it. It's like it's not one unified 0:06:52.850 --> 0:06:55.370 thing that he is enhancing with this skill. And that's 0:06:55.370 --> 0:06:58.090 the difference. That's the difference. When I do discuss or 0:06:58.090 --> 0:07:01.010 I call Chi or we text back and forth or whatever, 0:07:01.250 --> 0:07:04.370 whatever the interface is, we are updating, we are filling in, 0:07:04.410 --> 0:07:09.090 we are enhancing, we are improving the quality of the 0:07:09.090 --> 0:07:12.130 description of what perfect state looks like, of what ideal 0:07:12.130 --> 0:07:15.130 state looks like for this website, or this local application, 0:07:15.130 --> 0:07:16.990 or this mobile app, or whatever it is, or a 0:07:16.990 --> 0:07:21.390 picture of a butterfly inside of here. In the ideal state, 0:07:21.390 --> 0:07:26.230 we have these. These are individual testable claims. Boolean. Is 0:07:26.230 --> 0:07:28.870 this true or not? Along with a testing function that 0:07:28.870 --> 0:07:31.350 can make sure it's real. So this is the beauty 0:07:31.350 --> 0:07:34.030 of this thing. The ideal state describes what you're trying 0:07:34.030 --> 0:07:38.190 to build. It also has the specific criteria for building it. Okay, 0:07:38.270 --> 0:07:40.150 here's exactly how you build this. It must do this. 0:07:40.150 --> 0:07:42.270 This form must do this. It must be secure. It 0:07:42.270 --> 0:07:44.670 must submit to the following location. It must only accept 0:07:44.670 --> 0:07:48.910 these inputs. It then updates this database. Right. So very specific, 0:07:48.910 --> 0:07:51.470 very clear things just like we've been doing with spec 0:07:51.470 --> 0:07:54.510 driven development and test driven development for all this time. Right? 0:07:54.550 --> 0:07:57.630 We're bringing it over into this larger sort of master concept. 0:07:57.630 --> 0:08:01.670 So it is full of ideal state criteria which are 0:08:01.670 --> 0:08:05.110 inside the ideal state artifact, the ideal state criteria. This 0:08:05.110 --> 0:08:09.230 is insane. They are the build criteria. They are also 0:08:09.270 --> 0:08:13.150 the testing harness. The ideal state artifact is the testing 0:08:13.150 --> 0:08:16.110 harness as well. When I have a production app, I 0:08:16.110 --> 0:08:20.800 have surface the surface.ai. I have human 3.ai. Both two 0:08:20.800 --> 0:08:22.840 products that I have that have users that are out 0:08:22.840 --> 0:08:25.800 there doing real things. In the private repo, there is 0:08:25.800 --> 0:08:29.520 the ISO document. If I get rid of my entire 0:08:29.520 --> 0:08:31.840 system and come back and say, okay, I've been whatever 0:08:32.000 --> 0:08:33.800 meditating for six months and I come back and I've 0:08:33.800 --> 0:08:35.920 got a brand new MacBook, I set up life OS 0:08:35.920 --> 0:08:39.000 or whatever. I say, pull up the Isa for surface. 0:08:39.000 --> 0:08:41.800 I don't say, hey, go find every single session where 0:08:41.800 --> 0:08:43.840 we talked about surface. I don't say, hey, go find 0:08:43.840 --> 0:08:47.000 all these different random documents spread all over the place, 0:08:47.280 --> 0:08:50.559 plan file spec files, every conversation we ever had, all 0:08:50.560 --> 0:08:53.920 the places where I corrected you and fixed previous errors 0:08:53.920 --> 0:08:56.480 that we had in the Isa and went back and forth. 0:08:56.760 --> 0:08:58.640 I don't have to say go and collect all of that. 0:08:58.640 --> 0:09:02.200 It is in one place. Now you might be thinking, okay, yeah, 0:09:02.200 --> 0:09:04.240 but this document is going to get really huge. Well 0:09:04.280 --> 0:09:07.320 not really. That is getting easier and easier to pass. 0:09:07.640 --> 0:09:10.280 First of all, for AI, as the context windows get 0:09:10.280 --> 0:09:13.640 better as we get more efficient, but also the name 0:09:13.640 --> 0:09:16.120 of the game for life OS is rip grep and 0:09:16.120 --> 0:09:18.780 JQ right. We find the things we need in the document. 0:09:18.780 --> 0:09:20.300 We don't have to parse the whole thing every single 0:09:20.300 --> 0:09:22.300 time we have sections in there, you can go and 0:09:22.300 --> 0:09:25.060 pull out the appropriate section. And most importantly, the work 0:09:25.060 --> 0:09:28.140 being done by all these agents. It's all versioned, it's 0:09:28.140 --> 0:09:31.100 all in GitHub, it's all committed. You can roll back 0:09:31.260 --> 0:09:34.700 and you're also updating the changelog, the list of changes 0:09:34.700 --> 0:09:37.380 that have happened to the OS or to the ISO 0:09:37.380 --> 0:09:40.500 document for this particular task. And it's all being updated. 0:09:40.500 --> 0:09:43.180 The history of decisions is all there. It is one 0:09:43.180 --> 0:09:45.100 single document. So I could just pull this up for 0:09:45.100 --> 0:09:47.660 a production app, and I have the full thing to 0:09:47.700 --> 0:09:50.300 continue working on to add new things to. If I 0:09:50.300 --> 0:09:51.860 want to add a feature, guess what? It gets added 0:09:51.860 --> 0:09:55.540 as a feature. We add additional ideal state criteria. And 0:09:55.540 --> 0:09:57.220 those are the things that get worked on. And those 0:09:57.220 --> 0:09:58.819 are the things that get tested to make sure it's 0:09:58.860 --> 0:10:03.420 actually functioning. So that is the ideal state system. My 0:10:03.420 --> 0:10:05.660 argument is that it is the way things are going 0:10:05.700 --> 0:10:08.300 to go. I said late last year that this is 0:10:08.300 --> 0:10:09.940 the way things are going to go. It hasn't happened yet. 0:10:09.940 --> 0:10:11.260 I thought it was going to happen soon. I thought 0:10:11.260 --> 0:10:14.660 people like Theo and people like Matt and other people 0:10:14.700 --> 0:10:17.400 building stuff like me on YouTube would already be talking 0:10:17.400 --> 0:10:19.559 about this and the fact that we have to unify, 0:10:19.560 --> 0:10:22.400 and the fact that all of AI's problem is the 0:10:22.440 --> 0:10:25.560 gap between what's in our brain and what we want 0:10:25.600 --> 0:10:28.280 versus what the AI knows, right? Another way to think 0:10:28.280 --> 0:10:32.000 about this. It's called writer's block or writer's blindness or 0:10:32.000 --> 0:10:34.480 expert blindness, where you try to explain something, you try 0:10:34.480 --> 0:10:35.800 to write a book or something, and it's just a 0:10:35.800 --> 0:10:38.800 garbage book because you haven't explained actually the stuff that 0:10:38.800 --> 0:10:41.439 you know. Right? So that's a good reason to have 0:10:41.440 --> 0:10:43.360 an editor because they can actually pull that out and say, 0:10:43.360 --> 0:10:46.320 you need to explain this more. This is not obvious, right? Well, 0:10:46.320 --> 0:10:49.920 the ideal state document and the discuss skill is designed 0:10:49.920 --> 0:10:52.439 to fill that in. Right. And it actually triggers now 0:10:52.440 --> 0:10:54.360 at this point, if you don't have enough detail, it 0:10:54.360 --> 0:10:56.640 actually triggers discuss and has a conversation with you back 0:10:56.640 --> 0:10:59.080 and forth, which then fills in the ice more and more. 0:10:59.440 --> 0:11:04.360 So central concept, you have to synchronize your brain and 0:11:04.360 --> 0:11:09.040 your AI harness across the entire system. Okay. And let's 0:11:09.040 --> 0:11:11.440 talk about skills next because that's also where this is. 0:11:11.560 --> 0:11:13.880 So I have a skill system. It's a custom skill system. 0:11:13.880 --> 0:11:17.090 It's based on cloud code, but it's a specific skill 0:11:17.090 --> 0:11:20.770 system that has three main components. So the skill.md is 0:11:20.770 --> 0:11:23.370 basically the overview, and it's got a router in there 0:11:23.370 --> 0:11:26.329 similar to our cloud.md. It's got a router for workflows. 0:11:26.370 --> 0:11:30.210 Workflows are specific prompts, specific tasks that get done within 0:11:30.210 --> 0:11:33.570 the skill. So the workflows are write the blog post, 0:11:33.610 --> 0:11:36.530 edit the blog post, publish the blog post, scan the 0:11:36.530 --> 0:11:40.330 blog post for errors. I have one specifically, it's called enrich. 0:11:40.450 --> 0:11:42.570 So that goes and takes all my previous content, finds 0:11:42.570 --> 0:11:44.690 out if it's relevant to this area, and it adds 0:11:44.690 --> 0:11:46.970 a link for where I said those three words. It 0:11:46.970 --> 0:11:49.970 links to a previous thing. So I probably have like 0:11:49.970 --> 0:11:53.090 15 different workflows inside my blogging skill. So the concept 0:11:53.090 --> 0:11:56.010 is I want to blog workflows are these 15 things. 0:11:56.010 --> 0:11:58.730 And then the third piece, which is pretty unique here 0:11:58.850 --> 0:12:02.170 is as much as possible is turned into CLI tools, 0:12:02.410 --> 0:12:06.330 deterministic code, direct API calls. I use very few MC 0:12:06.690 --> 0:12:10.010 because I've reverse engineered the MC into CLI tools. So 0:12:10.170 --> 0:12:13.370 the workflows call the code. The code executes, hopefully doing 0:12:13.370 --> 0:12:16.830 as much as possible, and the concept calls the workflows, right? 0:12:16.870 --> 0:12:18.590 So I just say, hey, I want to blog about 0:12:18.590 --> 0:12:20.550 this and it goes and builds a blog. Now let 0:12:20.550 --> 0:12:24.750 me show you something. This is my site here. Sorry 0:12:24.750 --> 0:12:26.309 about the flash bang. It's not quite a flash bang 0:12:26.310 --> 0:12:29.070 because it's not exactly white, it's sepia. But this here 0:12:29.070 --> 0:12:31.590 is my blog and I want to scroll down here. 0:12:31.990 --> 0:12:34.350 Look at this latest content. Do you notice anything about 0:12:34.350 --> 0:12:41.710 these images? Look at that. Not half bad. Okay, notice 0:12:41.710 --> 0:12:43.190 that it's kind of sepia. Got a little bit of 0:12:43.190 --> 0:12:46.110 purple here. Got the dark charcoal. Oh, look at this one. 0:12:46.110 --> 0:12:49.110 Really cool. Do you notice that the concept of the 0:12:49.110 --> 0:12:53.110 art matches the content? All right. So I just went 0:12:53.110 --> 0:12:57.150 and edited this. I actually used this device and which 0:12:57.150 --> 0:12:59.590 is recording. I can just go for a walk and 0:12:59.590 --> 0:13:01.430 talk to this device. Or I could call Kai on 0:13:01.430 --> 0:13:03.630 the phone and say, hey, here's an idea. And I 0:13:03.630 --> 0:13:05.950 just ramble for the whole one hour walk. I'm just 0:13:05.950 --> 0:13:09.510 rambling about text and I then say, go write a 0:13:09.510 --> 0:13:12.510 blog post. Well, I didn't make any of these prompts. 0:13:13.230 --> 0:13:16.960 Another crazy experience with AI. My AI found what was 0:13:16.960 --> 0:13:19.360 killing my Mac and helped me replace two apps in 0:13:19.400 --> 0:13:23.319 about an hour. Okay, that is talking about specifically this 0:13:23.360 --> 0:13:25.680 app right here, which we made together. Well, really, I 0:13:25.720 --> 0:13:28.280 made it for me, but I gave the requirements and 0:13:28.280 --> 0:13:30.280 guess what? It built an Isa. There is an Isa 0:13:30.280 --> 0:13:34.840 for actually the UL bar app. Anyway. Diversion there. This 0:13:34.840 --> 0:13:38.080 piece of art was made from the content that I did. 0:13:38.520 --> 0:13:42.000 All of that is in my blogging skill. My blogging 0:13:42.000 --> 0:13:44.280 skill only has it in one place, which is in 0:13:44.280 --> 0:13:48.800 the blogging skill. Every single thing that I love that I. Okay. 0:13:48.840 --> 0:13:51.480 My typography. You have no idea how many hours I've 0:13:51.480 --> 0:13:54.800 been blogging since 99. This system is perfectly optimized to 0:13:54.800 --> 0:13:57.720 be exactly the perfect blogging system that I like the 0:13:57.720 --> 0:14:00.120 way this subtitle is done, the way the title is done, 0:14:00.120 --> 0:14:04.400 this little marker right here. The entire thing is perfectly 0:14:04.400 --> 0:14:07.480 in my ideal state, which is articulated perfectly in the 0:14:07.480 --> 0:14:11.680 blogging skill. So this is where I'm disconnecting from. Theo. Okay, 0:14:11.800 --> 0:14:15.700 he's saying memory systems are bad. I'm saying that the 0:14:15.700 --> 0:14:18.460 entire harness. This is all text files, right? What's the 0:14:18.460 --> 0:14:22.900 difference between a system prompt and a cloud.md file and 0:14:22.900 --> 0:14:25.700 your skill files, which is also a markdown file, the 0:14:25.700 --> 0:14:29.500 main skill.md and your prompts. Okay. The code is quite 0:14:29.500 --> 0:14:32.780 different because that's deterministic in code, but ultimately the entire 0:14:32.780 --> 0:14:34.860 harness is a whole bunch of guidance. It's a whole 0:14:34.860 --> 0:14:39.820 bunch of stewarding and shaping and adjusting to get things 0:14:39.820 --> 0:14:41.860 the way that you want them. Right? So all I'm 0:14:41.860 --> 0:14:47.500 doing is I'm making that extremely explicit. So this entire thing, 0:14:47.660 --> 0:14:49.940 I literally just talked now and I say, go make 0:14:49.940 --> 0:14:51.740 a blog about this and it goes and makes a 0:14:51.740 --> 0:14:55.540 blog about it. And this technical diagram right here is 0:14:55.540 --> 0:14:59.820 created by my technical diagram workflow inside my art skill. 0:15:00.700 --> 0:15:04.020 So look at this. I didn't link this chi link 0:15:04.020 --> 0:15:06.900 this for me by following the skill which perfectly articulates 0:15:06.900 --> 0:15:10.100 my ideal state for a blog post. So all of 0:15:10.100 --> 0:15:13.700 this is all the same idea. Okay? It is. How 0:15:13.700 --> 0:15:15.830 do I get to the perfect blog? How do I 0:15:15.830 --> 0:15:19.150 get to the perfect blog post? The perfect design? Okay, 0:15:19.190 --> 0:15:21.910 same if I'm building an app. Same if I'm building 0:15:21.910 --> 0:15:25.070 a new platform, a new security tool, I have ways 0:15:25.070 --> 0:15:27.710 that I want it to be done perfectly. And that 0:15:27.710 --> 0:15:31.990 is what I articulate inside of the skill system. So 0:15:32.270 --> 0:15:34.110 I wouldn't put that in the system prompt. I wouldn't 0:15:34.110 --> 0:15:36.030 put it in the cloud.md file. I put it in 0:15:36.030 --> 0:15:39.590 the skill system because it is a particular task. It's 0:15:39.630 --> 0:15:42.550 a specific task. We are blogging. I've got other ones 0:15:42.550 --> 0:15:45.230 for making outbound voice calls, so I can call and 0:15:45.230 --> 0:15:48.030 schedule a restaurant for me and put it on my calendar. 0:15:48.030 --> 0:15:49.830 So it's got a calendar thing. It's got an email thing, 0:15:49.830 --> 0:15:51.950 it's got an outbound call, it's got blogging, and I've 0:15:51.950 --> 0:15:56.470 got about 150 skills, small specific skills that articulate how 0:15:56.510 --> 0:15:58.870 to do this perfectly according to what I consider to 0:15:58.910 --> 0:16:02.150 be perfect. So this is why I disagree with what 0:16:02.190 --> 0:16:04.870 Theo is saying about memory systems. He's thinking of it 0:16:04.870 --> 0:16:09.110 as a whole bunch of like little tiny mini rules. No, no, 0:16:09.150 --> 0:16:14.230 the memory system is the cohesive system that merges your 0:16:14.330 --> 0:16:17.530 brain with your harness. That's what it is. Okay. In 0:16:17.530 --> 0:16:20.770 my opinion. And the whole centerpiece of the entire thing 0:16:21.130 --> 0:16:24.250 is this ideal state artifact. It is the fact that 0:16:24.250 --> 0:16:27.090 for anything we are building, we can always come back. 0:16:27.530 --> 0:16:29.810 Let me give you another example. Let's say that we 0:16:29.850 --> 0:16:32.130 are building as I am with a buddy of mine 0:16:32.130 --> 0:16:36.010 named Andrew. Building kind of like a gaming system, a 0:16:36.130 --> 0:16:40.490 system of building systems, actually of building other gaming systems. Well, 0:16:40.770 --> 0:16:44.170 when I go to do that, I can't remember everything 0:16:44.170 --> 0:16:46.330 all at once. When I go on my first walk, 0:16:46.330 --> 0:16:48.810 I'm talking about this other area. And I actually have 0:16:48.810 --> 0:16:51.370 tons of detail about this other area, about the integrity 0:16:51.370 --> 0:16:53.770 of the combat system and the simulation system to be 0:16:53.770 --> 0:16:57.010 able to simulate auto battlers or tank tower defense like 0:16:57.050 --> 0:16:59.770 type of games or whatever. Right. Well, I can't remember 0:16:59.770 --> 0:17:02.250 all of that at any given time. In fact, someone 0:17:02.290 --> 0:17:04.929 asked me, oh, what makes your gaming system better? In fact, 0:17:04.930 --> 0:17:06.850 people ask me what makes life OS better? I can't 0:17:06.850 --> 0:17:09.689 articulate all of this, which is why I'm doing this video. 0:17:09.690 --> 0:17:13.570 I'm trying to articulate all at once cohesively, right? The 0:17:13.570 --> 0:17:16.420 point is, at any given time you might not remember. 0:17:16.420 --> 0:17:18.740 You yourself cannot remember because the idea is so big 0:17:18.740 --> 0:17:21.460 and it has so many components and there's so much detail. Well, 0:17:21.460 --> 0:17:22.859 where are you going to put that? Where are you 0:17:22.859 --> 0:17:24.059 going to put that stuff? Are you going to go 0:17:24.060 --> 0:17:26.180 and gather prompts and talk to your AI? And it's 0:17:26.179 --> 0:17:29.500 going to put it in some session file. No, because 0:17:29.500 --> 0:17:32.059 that will be lost and it will not be unified. Right. 0:17:32.100 --> 0:17:34.500 So the purpose of this discussed thing is to keep 0:17:34.540 --> 0:17:37.699 adding pieces, as you remember, as something breaks, as you 0:17:37.699 --> 0:17:39.660 have a new idea, as you realize one of your 0:17:39.659 --> 0:17:42.700 previous ideas is total garbage. You're like, hey, remember, we're 0:17:42.699 --> 0:17:44.859 going to do that combat integrity system. We're going to 0:17:44.899 --> 0:17:47.179 use this method. Yeah, I watched a bunch of videos. 0:17:47.179 --> 0:17:50.059 It's total garbage. Let's do this instead. Boom goes and 0:17:50.060 --> 0:17:53.260 cleans it up. The ideal state is now improved, not 0:17:53.260 --> 0:17:57.100 in some random location in the actual ideal state artifact, right. 0:17:57.139 --> 0:18:00.580 The authoritative artifact. All right, so that's the skill system. 0:18:00.580 --> 0:18:03.379 That's the ideal state artifact system. This is basically the 0:18:03.379 --> 0:18:05.899 Kanban board that shows the work that's actually happening. And 0:18:05.899 --> 0:18:08.980 the algorithm is really just this hill climbing system. It's 0:18:08.980 --> 0:18:12.419 also got a model router that switches between like fable 0:18:12.459 --> 0:18:16.760 GPT five six, soul GB. Five six Cyber or Daybreak 0:18:16.879 --> 0:18:19.919 for all my offensive security stuff. It actually uses a 0:18:19.919 --> 0:18:24.240 whole separate agent. Actually, I could show you that real quick. Yeah. 0:18:24.240 --> 0:18:28.320 So down here, if you look at this haiku, sonnet, opus, soul, Gemini, grok, 0:18:28.320 --> 0:18:31.519 cyber and Max and Forge and Helios. So basically Helios 0:18:31.560 --> 0:18:34.680 is his own agent that does only cyber stuff anyway. 0:18:34.719 --> 0:18:38.159 Not important. Not sure why I showed you that. All right. 0:18:38.159 --> 0:18:40.159 So going back to this. Oh yeah. Evals. That's the 0:18:40.159 --> 0:18:42.079 other really cool part of skills. But we can talk 0:18:42.080 --> 0:18:45.359 about that later. This basically shows all the different content 0:18:45.760 --> 0:18:47.840 that's being executed. I think I showed you that earlier 0:18:47.840 --> 0:18:49.959 or it's in another video. You should watch that as well. 0:18:50.280 --> 0:18:53.600 But the algorithm is just how you're actually moving through 0:18:53.600 --> 0:18:57.119 the ideal state. Which models are you using, which intelligence 0:18:57.119 --> 0:18:59.519 level are you using? And it actually pivots up and 0:18:59.520 --> 0:19:02.239 down based on the work task. So if you shouldn't 0:19:02.240 --> 0:19:04.559 be using fable for this, switch out to soul or 0:19:04.560 --> 0:19:07.399 switch out to haiku in some cases, although not very often, 0:19:07.399 --> 0:19:11.600 or sonnet or opus or whatever. So ultimately the system 0:19:11.600 --> 0:19:15.650 is euphoric, surprise and ideal state. The system prompt is 0:19:15.649 --> 0:19:18.730 just like super minimalist. This is just a context routing system, 0:19:19.409 --> 0:19:21.969 and the skill system is quite unique and I think 0:19:22.010 --> 0:19:25.129 really helps with the determinism and basically making it more 0:19:25.129 --> 0:19:27.769 dependable to actually pursue your ideal state as things are 0:19:27.770 --> 0:19:30.570 being executed. And the algorithm is kind of just like 0:19:30.609 --> 0:19:34.090 managing the whole thing, managing the general hill, climbing from 0:19:34.090 --> 0:19:36.969 current state to ideal state for anything you're trying to accomplish. 0:19:37.090 --> 0:19:39.770 So this is what I want to say, just responding 0:19:39.810 --> 0:19:42.889 to Theo here again. I think me and Theo and 0:19:42.889 --> 0:19:45.490 Matt also are going to agree about this, I think, 0:19:45.490 --> 0:19:47.449 in the future. And I could be wrong about certain 0:19:47.449 --> 0:19:50.689 pieces about this as well, right? I've incorporated tons of stuff, 0:19:50.689 --> 0:19:53.650 for example from Wayfinder or from Grill me or some 0:19:53.649 --> 0:19:56.290 of his other skills. I've incorporated tons of stuff into 0:19:56.290 --> 0:19:58.929 life OS from Theo, from all the videos I've watched. 0:19:59.330 --> 0:20:01.490 Mad respect for both of them. I just think that 0:20:01.490 --> 0:20:06.050 this concept is so fundamental and first principles based, that 0:20:06.050 --> 0:20:07.809 I think everyone's going to be moving in this direction. 0:20:07.810 --> 0:20:09.490 And if you like it, if you try it, if 0:20:09.490 --> 0:20:11.730 you hate it, if you absolutely love it, either way, 0:20:11.730 --> 0:20:13.609 just comment and we'll see you in the next one.