22:03Jake Van Clief · Talking Head
Folders Over Agents: The AI Layer Nobody Teaches
A 22-minute argument that context architecture — not agents, not prompts, not frameworks — is the only AI layer that survives the next model update.
May 29thThe argument in one line.
Your markdown context files are the durable licensable asset in an AI workflow, and machine-learning optimization of those files can produce capability gains that normally require a full model upgrade.
Read if. Skip if.
READ IF YOU ARE…
- You are building AI workflows for clients or your own business and want a methodology for organizing context files that survives model updates.
- You are a freelancer or consultant curious about how structured context files create defensible value that can be licensed to companies.
- You work inside an organization where AI adoption feels stuck — outputs are hollow, teams are overbuilding, or adoption is inconsistent across departments.
- You want a first-person report from London Tech Week 2026 on where venture capital is moving: infrastructure, data structure, talent platforms, and agent-to-agent commerce.
SKIP IF…
- You want a step-by-step beginner guide to using AI tools — this session assumes you are already practicing structured AI workflows.
- You are looking for prompt templates or model comparisons; this is about methodology and market positioning.
The full version, fast.
The durable unit of value in AI work is the structured context file, not the prompt or the model. ICM formalizes this: markdown files in a routed folder tree let any AI navigate your knowledge base and amplify your specific opinion without re-prompting. London Tech Week confirmed money is chasing infrastructure and a new layer where AI agents are the buyers. Chicago executive workshops revealed most teams produce hollow output because they give AI no context or opinion, and the 60/30/10 rule (90% traditional structure, 10% AI) beats custom-agent overbuilding. Microsoft's SkillOpt paper shows ML-optimizing markdown files yields 30-57% capability gains without model changes. Two community products launch this week: the Ledger talent platform and an ICM deployment layer on Azure.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 02:07
01 · Welcome: the Ledger and ICM deployment
Two community launches announced: the Ledger talent platform and an ICM deployment layer on Azure. Both free for VIP members.
02:07 – 04:07
02 · Member intros and what people are building
~12 global community members introduce themselves: IT managers, agency owners, real estate developers, software developers across five countries.
04:07 – 08:15
03 · Engelbart, 1962, and software as collaboration
Engelbart's 1962 paper frames ICM: software was always about collaborating with information, not controlling machines. The mouse inventor's deeper work on augmenting human intellect.
08:15 – 15:18
04 · Fable pulled, and why output beats features
Anthropic's Fable model pulled overnight on government pressure. ICM workflows got better, not worse — proof that output is the goal, not model features.
15:18 – 19:47
05 · Getting unstuck on ICM
Live Q&A with an Austrian real estate developer stuck trying to implement ICM perfectly. The three questions framework: delegation, complexity, outcome.
19:47 – 33:08
06 · The three questions and a live ICM routing demo
Live demo of routing folders in Claude for a content client. The top markdown file routes the AI to the right subfolder without burning tokens reading everything.
33:08 – 38:40
07 · AI as your runtime, humans in the compute layer
Core ICM thesis: AI becomes your runtime; humans remain in the compute layer. ICM organizes processes to be automated so collaboration happens at a higher abstraction.
38:40 – 41:40
08 · Productionize your opinion
The fuzzy, opinionated output creates durable value. Every edit to your context files is productionizing your opinion — what the AI amplifies is the way you think.
41:40 – 46:39
09 · Ingest agents and distilling your brain into files
Building an ingest agent to process transcripts, extract patterns, and distill voice and decision-making into structured markdown. Referenced community member building a neuroscience-based approach.
46:39 – 49:23
10 · You are not behind
Being surrounded by experts creates a false sense of lag. If you are in the room, you are already far ahead of the general population. AI skill-building is not overnight.
49:23 – 52:25
11 · London Tech Week: the rooms and the money
CEO of AMD, AWS researchers, NVIDIA, mayor of London all present. Silicon, government, and capital in one room. Medicine and defense attracted the most attention and money.
52:25 – 54:08
12 · The buyer is changing: selling to agents
VC insight: companies are building payment and data layers where AI agents, not humans, are the buyers. More internet traffic is now bot/AI-generated than human.
54:08 – 1:01:13
13 · Everyone is overbuilding, and the talent layer opening
Most startups at London Tech Week are over-engineering. Simple folder systems get the same output as expensive custom agents. The real opening is the talent layer.
1:01:13 – 1:06:45
14 · Placement fees, freelancing, and the college problem
Ledger platform walkthrough. Placement fees from companies; free for community members. Vision: license your ICM brain to companies without working there full-time.
1:06:45 – 1:10:43
15 · No best, and a bet on humans
No such thing as best intelligence. Infinite human desire means there is always more work. Net positive thesis on humans in a world of increasingly capable AI.
1:10:43 – 1:13:56
16 · Chicago: hollow output and the 60/30/10 rule
Executive workshop findings: most teams produce hollow AI output. The 60/30/10 rule diagnosed. Engineers naturally get it right; most other teams invert the ratio.
1:13:56 – 1:18:22
17 · When to hire a human instead of automating
Sometimes hiring someone is cheaper than building an API automation. Humans are compute-efficient. AI-native means the value of human roles shifts, not that they disappear.
1:18:22 – 1:30:29
18 · SkillOpt: training your skill files
Microsoft paper: vectorize markdown files, run ML optimization loops using LLM-as-judge, auto-improve skill files. Optimal files end up under 500-800 tokens. 30-57% capability gains. Speaker building open-source hybrid.
1:30:29 – 1:31:43
19 · Launching this week and close
ICM deployment layer launching Wednesday on Azure. Ledger platform. Speaker stops recording but continues the call.
Lines worth screenshotting.
- A routing markdown file at the top of a folder tree lets the AI navigate directly to the right context without reading every file, saving tokens and improving accuracy.
- When a model gets pulled overnight, ICM-structured workflows get better on the replacement, not worse — proof that output is the goal, not model features.
- Microsoft SkillOpt paper shows ML-optimizing your markdown skill files produces 30-57% capability gains that normally require a full model generation upgrade.
- The most effective markdown skill files optimized by SkillOpt ended up under 500-800 tokens — smaller and more separated, not longer and more comprehensive.
- More internet traffic is now generated by bots and AI than by humans, and startups are already building payment layers where AI agents, not people, are the buyers.
- Hollow output is not a model problem; it is a context problem. Teams that give AI no opinion get output that sounds polished but is built on nothing.
- The 60/30/10 rule: 90% of your AI solution should be traditional code and process structure; the 10% AI layer on top is what makes it feel like magic.
- Humans are compute-efficient. Before automating a task with expensive API calls, calculate whether hiring someone for $40k a year is actually cheaper.
- AI-native does not mean zero human roles — it means the value of each human role shifts to where their specific opinion cannot yet be automated.
- Productionizing your opinion means encoding your judgment, naming conventions, exceptions, and aesthetic preferences into markdown files — that is what compounds across model updates.
- The back end is moving farther from the front end. Most people will eventually build their own interface layer while sharing common infrastructure.
- An ingest agent that processes your own transcripts and decisions into structured markdown is how you distill your brain into a licensable knowledge asset.
- At London Tech Week, seven different companies were pitching themselves as agentic builders of the future — a clear signal the space is overcrowded and underdifferentiated.
- If you can route a folder structure so the AI knows what is in each subfolder without reading them, you have built most of what an agent does without the engineering overhead.
Context files are the durable asset in an AI workflow.
WHAT TO LEARN
When models update or get pulled overnight, the thing that keeps working is the structured opinion you have already written into your files — and that file quality is now measurable and automatically improvable.
01Welcome: the Ledger and ICM deployment
- Talent platforms built around verified AI workflows — not just credentials — solve the signal problem for companies hiring in an AI-saturated market.
03Engelbart, 1962, and software as collaboration
- Software was designed from the beginning to amplify collaborative thought, not automate individual tasks — ICM reconnects modern AI tools to that original design intent.
04Fable pulled, and why output beats features
- Output-focused workflows survive model churn. When a model disappears overnight, an ICM-structured setup gets better on the replacement — it does not break.
05Getting unstuck on ICM
- Getting stuck trying to implement a methodology perfectly is a business decision problem, not a technical one — the three questions (delegation, complexity, outcome) diagnose which kind of stuck you are.
06The three questions and a live ICM routing demo
- A routing markdown file at the top of a folder tree lets any AI navigate your knowledge base without reading every file — cutting token cost and dramatically improving task accuracy from a single prompt.
07AI as your runtime, humans in the compute layer
- When AI becomes the runtime over your structured files, collaboration moves up a level — you stop prompting for tasks and start directing a system that already knows your workflow.
08Productionize your opinion
- Productionizing your opinion means encoding your judgment, naming conventions, exceptions, and aesthetic preferences into files that compound across model updates — that is the asset that cannot be commoditized.
09Ingest agents and distilling your brain into files
- An ingest agent that processes your own transcripts and decisions into structured markdown is how you begin distilling your brain into a licensable, deployable knowledge asset.
10You are not behind
- Feeling behind is a function of being surrounded by experts, not a function of actual lag — most organizations have not yet run a single structured AI workflow.
11London Tech Week: the rooms and the money
- Tech industry conferences are more useful as signals of what not to build than what to build — when seven booths say the same thing, the category is already commoditized.
12The buyer is changing: selling to agents
- The buyer is changing at the infrastructure level: more internet traffic is now AI than human, and startups are building payment layers for agent-to-agent commerce.
13Everyone is overbuilding, and the talent layer opening
- The real gap at London Tech Week was not in software but in humans who can build and operate structured AI systems — that scarcity is the business opportunity.
14Placement fees, freelancing, and the college problem
- The path from using AI to deploying AI runs through the talent layer: companies that need structured workflows are already willing to pay placement fees for practitioners who have done this work.
15No best, and a bet on humans
- There is no such thing as best intelligence, just as there is no best programming language — the differentiation is always in the opinion encoded on top of the general capability.
16Chicago: hollow output and the 60/30/10 rule
- Hollow output is a context problem, not a model problem. Teams that give AI no opinion, no rules, and no structure get work that sounds polished but has nothing underneath.
- The 60/30/10 rule: 90% of a robust AI solution is traditional code and process structure. The 10% AI layer on top is what makes it feel like magic. Inverting the ratio creates brittleness.
17When to hire a human instead of automating
- Before building an automation, calculate the true cost: sometimes hiring someone for $40k a year is cheaper than the API calls, the engineering time, and the maintenance. Humans are compute-efficient.
18SkillOpt: training your skill files
- Microsoft SkillOpt shows that ML-optimizing your markdown files — not fine-tuning the model — can produce 30-57% capability gains. The optimal files end up under 500-800 tokens, smaller and more separated.
Terms worth knowing.
- ICM (Interpretable Context Methodology)
- A folder-and-markdown-file system for organizing AI context. A routing file at the top describes each subfolder so the AI navigates directly to relevant content without reading everything. Based on file-tree design principles from Unix and 1960s-70s computing research.
- Routing file
- The top-level markdown file in an ICM folder structure. It describes what is in each subfolder so the AI can navigate to the right context from a single prompt, without spending tokens reading every file.
- SkillOpt
- A May 2026 Microsoft research paper proposing a machine-learning method to automatically optimize markdown skill files. It vectorizes the file, runs hundreds of LLM-judged trials, and edits the file until capability scores peak — typically yielding 30-57% gains over the original baseline.
- Fable
- An Anthropic model described as a smaller, controlled version of Mythos that scored highly on benchmarks but was pulled from public access due to US government national security concerns. Used as an example of why output-focused workflows beat feature-dependent ones.
- The Ledger
- A talent platform built by the speaker's community. It allows members to upload ICM portfolios and either hire or be hired by companies needing AI workflow expertise. Placement fees paid by companies; free access for VIP members.
- 60/30/10 Rule
- A heuristic from the speaker's consulting practice: 90% of a successful AI solution should be traditional code, process, and structure; 10% should be the AI layer that makes it feel like magic. Most teams invert this ratio and create brittle, over-engineered systems.
- Hollow output
- AI-generated work that looks polished but has no substance because the prompt gave the model no context, no rules, and no human opinion to amplify. Identified as the most common failure mode across the Chicago executive workshop.
- Selling to agents
- An emerging commercial model where services and data are sold to AI agents rather than human buyers. Startups at London Tech Week were building payment layers and data APIs designed for automated agent-to-agent commerce.
Things they pointed at.
Lines you could clip.
08:15
“The output is the goal, not necessarily another feature.”
33:50
“It's almost like the AI becomes your runtime.”
38:40
“You get substance by giving opinion. Substance comes from something that you cannot just read or copy.”
46:39
“There's no such thing as behind when you're building your own future.”
54:08
“Why would I wanna build a software to do something when I can build a system that gets that outcome?”
1:06:45
“Human desire is infinite. It is why the markets exist.”
1:25:00
“I now have mathematical evidence of why my ICM works well.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphor
00:00Hosted. Super fun. Uh, and I pretty much do SEO and small to medium business size automations, or that's what I'm trying to do.
00:10Uh, and I've been just learning so much from you, and I've been really, like, integrating my own workflow. And I'm looking actually to hire people now to help, um, use kind of AI to do all the work that we do for websites, SEO. And, yeah, and it's been great.
00:32Oh, you're muted, Jake.
00:36Well, I love that. Thank you so much. Um, when it comes to the hiring thing, uh, we're actually launching it, uh, probably this week.
00:43Our ledger. We've been working on it, and it's a place where people, uh, like, you know, our competitions, you can actually upload them as kind of portfolios, and then you can put yourself out there as freelance to work, uh, full employee to work, and it'll show what your kind of work you've already done kind of with all of that.
01:01So keep an eye out for that. We're doing two portals, and we're gonna make it free to everyone who's VIP and premium. Of course, there's not gonna be anything y'all are gonna have to pay.
01:10You're already paying, so I don't see a reason to to make you pay more. And so, hopefully, it'll be a platform where you're either able to get some of the top talent or vice versa, some of that top talent can, you know, find their own ways.
01:22And then eventually, we're launching the ICM kind of deployment layer. So, basically, we're we've created a a kind of dockerized deployment environment.
01:31So you can just upload a ZIP. We deploy it, and then, you know, let's say you do hire someone. Instead of them having to install a whole bunch of stuff on a new computer or get it all in, You just log in, and you can use that workflow, those files and folders you're building, or you can deploy it to other clients.
01:45And, of course, we're gonna make sure you all have access to that for free. Um, and then, actually, the goal for that one is to have it where you can charge other people to kind of, like, have a a access to your workflows.
01:55You're licensing a version of yourself is my idea, which we've chatted about before. But, um, I'll talk on that later. I don't wanna take up too much time.
02:02Uh, but, Marcelo, I'd love to hear from you next.
02:06Hi. Hi, Jake. Hi, everyone.
02:08I'm Marcelo from Brazil. I am a trained psychotherapist first, but always loved technology.
02:17And, you know, like, end of last year, I started to hear about OpenClaw and agents, and I jumped back into, you know, the tech space.
02:31And it I just it just blew my my mind. And and I'm I'm working with that since since then. I've I'm creating products to event management companies and small businesses.
02:48And I was just working on a supplier management dashboard with a chatbot attached to it.
02:58And I'm using a lot of ICM, and I've created something called the workspace creator.
03:07So I created an an ICM style folder structure that creates workspaces.
03:15So I go through all the things and and it speeds a workspace for me.
03:22So it's it's been fun and and amazing.
03:26Thanks so so happy. No. Thank you, Marcel.
03:29And that's kind of, like, the idea behind why I've been pushing this out there is I knew right? I I saw the structures. I've spent years diving into old papers from the sixties to the seventies and how we were building file trees and why we built them.
03:41And the whole goal wasn't to help us control the computer more. It was to allow us to collaborate with information among other people. We want an environment in which we can move ideas, move intent, move contents.
03:53That's the entire goal of software is to take our intent, amplify it with the machine, and collaborate and share that with other people, share that value. And I built ICM based on the idea of how do we take that into the next step, the next stage knowing that AI exists as a platform, as a process. And it's at the core of it, it's organization.
04:13Right? It's how do we organize thoughts, concepts, movements, value in a way in which these tools and myself or others can access it?
04:22And the idea of workspaces, I'll chat a little bit about it today, but I'm really glad you're diving into that. In fact, I've been reading a paper.
04:29Um, it's from 1962. It's called augmenting human intelligence, um, and it's from the guy who invented the mouse.
04:36So the mouse that you're using was invented by this guy. And I'm gonna be making a YouTube video, so don't worry. You all won't have to to go through and read it.
04:44Um, but it is absolutely amazing. Right?
04:47Uh, it's oops. Hold on. This is the newer one.
04:49That's the that's the let me see. 1962. Let me grab the actual one from it.
04:55There we go. Or sorry. Augmenting human intellect, not intelligence.
04:59And it is just it's by Engelbart. He's done a lot more than just inventing the mouse.
05:04He invented a whole bunch of things. Um, but it is just an absolute amazing paper about looking about how do we look at these processes. And it's very funny because if you'll notice a lot of the things I talk about, right, about breaking them down into these artifacts, these processes.
05:19Language is a control space. This idea of needing to train the human to get to the understanding of how to use that control space.
05:27It's why I made my community, and I didn't read this paper before starting my community. So I've just been having an absolute blast going through his work. So expect a video on that soon, and I think it'll help you a lot with what you're building around your workspace.
05:41And I highly urge you, actually, while I was at London, this was a common theme I saw people missing, is this collaborative effort. They keep automating single people tasks or automating processes rather than creating environment in which those processes can be collaborated between each other, between yourself.
05:58So, um, very excited to have you here, Marcelo. So I like that. Thank you.
06:02Um, and, hopefully, I'll teach you some stuff in the call today alone, let alone everything else. And a lot of other people in here. Again, there's amazingly brilliant people in here.
06:09Um, let's see. Think, uh, Joaquin, you also raised your hand and said you were new. Yeah.
06:14Uh, you can hear me? Yeah. Yes.
06:15Okay. My name is Joaquin. I'm from Uruguay in South America, and I'm a software developer and I own a software dev agency implementation of AI.
06:26And this course was, like, I have so many issues, problems, how to build stuff that is not gonna be built later for the for cloud or for OpenAI. So this course has been, like, a opener in a way. And so many ideas are being, like, popping in my mind the last week that, yeah, I want to write new stuff.
06:45I want to discuss new stuff. And, yeah, thank you very much for everything.
06:50I'm so glad. No. I I just I kept getting so angry, um, because, right, I'm working with large, like, Fortune 500 clients, and I have been for a while.
06:58And, um, I kept seeing these people, like, just pitching stuff online and sharing certain aspects. I'm like, we're never gonna it wasn't even a, oh, let me do it out of spite. Part of it was there, but it was if we can get everyone on the same page of building and creating, we can move past all this BS of, like, patching features.
07:15Because my whole goal is, right, when how often raise your hand if I'm incorrect. But how often do you all freak out and look at news articles when, um, iOS updates or when Excel updates or when QuickBook updates. You don't, really.
07:31You're focused on other things. Maybe if you're in compliance, there's a different area. Um, I think I see, yeah, Tiffany's in here.
07:36She she'll probably scour those because there's always data changes there, and that's important. But most of the time, you don't really care when a software that you use every day is integral to your life updates because your workflow, your process, your employees already know how it's supposed to work. The updates just make your workflows better.
07:54We eventually wanna get to the point where uh-oh. Did I did you guys lose me there? I just saw, uh, okay.
08:00Cool. We eventually wanna I saw my screen kinda paused for a second. We wanna get to the point where when Claude updates, when OpenAI updates, it's just another update.
08:10It's just, oh, cool. That's lovely. I enjoy that.
08:13And keep on working the exact same I'm working. If you're not at that point, we haven't fully understood how the technology works. Like, when Fable came out, I was now it's gone again.
08:22We've all lost Fable thanks to the US government. Um, but is I don't know if anyone's familiar with that news article that just came out. Is anyone not familiar with what happened there?
08:31For those of you who aren't, um, Claude and dropped a model called Fable, uh, which is a smaller or more controlled version of Mythos. You know, it scored a lot higher on a lot of benchmarks, but then the government was like, no. We don't want foreign nationals using it.
08:46We think it's a security threat. And, unfortunately, Anthropic has foreign nationals inside of their company, so they decided just to shut it down access to everyone. I think it could still be part of marketing personally, but I also think it's, you know, some political stuff.
08:58But we don't have to dive into that right now. Um, but long story short, that dropped. I tested all my workflows.
09:04My ICM methodology just got better. Right? That's that's all I care.
09:07I was like, alright. Cool. It's better.
09:09I still could use less tokens because it works well. The output is the goal, not necessarily another feature. So, yeah, absolutely with you on that, Joaquin.
09:16Super happy to have you here. Um, who else is new? Alan, uh, you're you're not new.
09:20Right? I I think you've been in the the past ones. Yeah.
09:22Yeah. I was gonna say you've you've brought up a bit. I just you're I don't know if you've changed your background or something.
09:26It looks different. Uh, We haven't chatted much before with you, unfortunately. Um, Oh, boss.
09:33Yes. Boss and Roy. I think both of you are new.
09:35Uh, would you guys, uh, be willing to, um, introduce yourself to everyone here? And then we'll dive into some some of the structure. I just think it's important we all know each other and what we're doing because, again, I'm valuable here, but everyone in here, this is a value point.
09:48This is a place of potential business, of a potential collaboration. We want to know everyone in here is at the higher levels of understanding AI. They're where you are.
09:57You don't have to sift through all the social media and be like, does this person really understand what I'm working on? And I was like, no. Everyone in here should have a pretty solid, uh, uh, idea for it.
10:06Um, but yeah. So, boss, would you, uh, would you be willing then to say hello? Uh, if you don't want to share your screen, that's also perfectly fine.
10:13Um, but be happy to to hear from you, possibly, if you're allowed to. I do not have permissions.
10:20Let's see. There we go. It should should let you share your screen and audio now, hopefully, allegedly.
10:29Oh, and Sandra. Yes. Hello.
10:31Go for it. Let me know when you guys can hear me. Can you guys hear me?
10:34Can you hear me? Yep. Absolutely.
10:36Fan fantastic. Fantastic. Fantastic.
10:38I woke up about, uh, forty five minutes ago, so I'm not gonna get on camera. I am a very, very busy person.
10:45I'm an IT manager for a very, very large company, 160,000 employees. I have seven direct reports, and I I manage seven different buildings.
10:54One of those has 2,100 people in them. So I work around sixty hours a week.
10:59I commute twenty hours a week as well. That stands for eighty hours a week that I am not using AI. Yeah.
11:06I know that's crazy, which means that everything you see me doing is in the in between. Everything I build, every every competition I enter is in the in between time of that two hours. Should tell you what AI is capable of, not what I'm capable of.
11:23So, yeah, I am I'm really grateful to be in this group.
11:28I think that we have a lot of talented people, but I think there's something more than talent here. I think there's real community, people that actually care. That is what it's about.
11:37And I think that that is what keeps me coming back. So I absolutely love ICM.
11:43I think that it's I'm an autodidact. I love to teach myself things.
11:48I love to learn new things. I love to pull things apart and see the structure behind them and see how they're running, and I think that that's at the core of ICM, knowing how things are running at the core. Mhmm.
11:57Once you understand something fully, you can break it down, and you can really hammer it out. Iteration, also, absolutely key.
12:06It's it goes along with the whole thing with, like, being afraid of failure. If your builds are failing, that's a beautiful thing. It gives you opportunity to find out where they're failing, why they're failing, and to iterate, fix it, and move on from it.
12:17It is absolutely crucial that we fail as much as possible to be able to 100%. A 100%. But Well, maybe Yeah.
12:24That's oh, No. I was just gonna say that's just me, and I will talk forever. So I will be quiet.
12:28I am I am I am I am autistic. So, yeah, sometimes you'll have to be like, hey, boss. Be quiet.
12:35Well and this is an important note. I always remind everyone too if, um, I always let these calls go even after I get off for today, so feel free to kind of chat amongst each other. And then, of course, the Discord.
12:44Um, I set up the Discord for us to be a place to just make it easier because school looks good, but it's not so great when it comes to, you know, kind of collaborative, being able to talk on the fly, things like that. So make sure, uh, you're in the the kinda high t VIP side of Discord so you can connect with everyone there.
12:59But, no, amazing to have you here, and I I'm so glad. Right? The whole ICM methodology was built off this idea of, like, output and creating and looking at things.
13:06Like, if you go into my research paper, one of my core parts is looking at how myself and, like, 50 other people well, technically, it was 32, but then I added another 10. We're editing stuff, editing AI outputs. The rest of the paper was built around the patterns that I saw and then just rereading stuff from the seventies.
13:24Right? Unix methodology, stuff like that. It's the fundamentals.
13:27Once you learn the past, you can really predict the future, not in a predicting of, hey. This is when events are gonna happen, but, like, this is probably how people are gonna work because this is how we've always worked, that there's patterns in it all. So yeah.
13:38Absolutely. And, also, thank you for giving your time on a weekend, and that includes everyone else here. Um, I really try to make these as valuable as possible.
13:46Again, for those of you who this is new, I take the transcript. Um, I pretty much create all sorts of different extra artifacts for every single IT that you can either give to your AI, throw into your knowledge base. If you have a second brain, that's I kinda build them for that there.
14:01Um, and then it's just nice to be able to review. So so thank you for that. Thank you for your time because, yeah, managing an IT, especially in today's world, is a nightmare.
14:09So thank you for your time. Um, okay. Let's do, uh, Sandra.
14:14Do you want to introduce yourself real quick?
14:18Is that okay?
14:20Hello. The mic work?
14:22Yes.
14:23Hello. I'm from Austria. Hello to everyone.
14:27I didn't know that talking a lot is an autistic trait. So now I know why.
14:35So please stop me because I was going once I started. I'm a real estate developer in Austria. I build buy basically grounds, develop apartment houses, and sell them again.
14:50I have three running project that project at the moment with 50 apartments and houses. So not a totally big thing, but also not really small.
15:01And I have a marketing agency longer than I have that, than I have the real estate development.
15:09In the marketing agency, I was really lucky to got a really professional and good foundation for for graphic design and for programming.
15:19I actually programmed more than I did graphics.
15:24So I write HTML and CSS like a letter. I'm I I always say so that kinda got me into to connecting me more with Clodo, maybe.
15:35Lower the barrier maybe a bit to the to get in touch with it.
15:40And then I was really lucky. I got I have one tenant in the house, which is a full stack programmer and IT guy, and he introduced me in February to Claude.
15:50So I started programming something, an app immediately, and very simple thing, a desktop app and an app for my bookkeeping.
15:59My bookkeeping basically is digital purely since 2018.
16:06So I'm really strict I have a really strict naming convention for my for my for my files since 2016, something like that.
16:18So I I basically have this structure already. So I'm I'm very clear in my structure. Also, my the the the files for the graphic agency were always very structured and always the same in every in every client's folder.
16:32So I had that kind of and and then I I I started immediately, when I started working with I actually programmed I I got programmed an app for my bookkeeping already in 2024, I think, or something or '22.
16:51But today, it's like with Claude, it's done, like, in a week or so. So it costs, like, a €100,000 back then.
17:00So and now it's I just did it all over again, basically. Or I'm doing it all over, basically. So I the way I got to know you so I got to know you right in the moment where I started thinking, wait a minute.
17:16I need to make a memory file. I don't wanna say things over and over again. I need to put a file into the root where is that's my identity file.
17:24I called it maybe different than you did, but, basically, I did the same. And then I got to know you and thought, oh, wow. There is somebody doing that actually, like, way more advanced.
17:32So let's get into it. But now I'm a little stuck because I tried to implement your method, like, perfectly, and, yeah, that's not happening.
17:45Instead of just going along
17:47This is a good question then, and and we'll move on here in a second. But what what is the main issue that you're having? What is, like, may if you were to name, like, okay.
17:55This was the goal that I was gonna have, and this is what's not happening even though I'm trying to use this methodology. What what is if if you were just to do a high level kind of description?
18:08I'm asking myself exactly that. So I think one part is in the moment in the moment where I where I saw somebody's doing that professional or more advanced, I got stuck because I also wanted to do it Ah.
18:27Perfect. That's maybe one thing. The second thing is that I'm really busy and that my my day is running.
18:33So for instance, next step next very simple step would be that that I improve the naming convention also for because my son is actually using it a lot and also my assistant, so it needs to be improved.
18:48But I'm I'm just not doing it because I'm working on implementing this ICM. So so I have it now in different places.
18:56And, yeah, and then I I I open Cload and and I do a thousand other things, and I'm gonna well, what I wanted to do.
19:06So I'm I'm stuck kind of in the in the setting up or yeah. I'm just sorting out again new or so.
19:13Like, I I I don't know how to sort it out new. And today, I thought then Versus Code, yeah, I wanna implement Versus Code. What again is it for?
19:21Yeah. Okay. Do I really need it?
19:23And then I'm asking five times, do I really need it? Because one thing is maybe that that I don't want to use any more tools. Because in the graphic agency, you know, it was like every year, another project management every year and yeah.
19:38As you say, the updates and everything. So I'm really tired of that. So I don't wanna immediately start to use another tool.
19:45So I'm very reluctant whatever comes.
19:49Me let me say two may help and then give you some questions to ask yourself. Right?
19:54Because right now, there's three main things you wanna ask yourself. It's a delegation question, it's a complexity question, and it's an outcome question.
20:02And what I mean by that is right? So my ICM, right, it's it's a methodology for organizing thoughts for yourself as well as for any sort of AI so that you can start getting an outcome, start getting it to do something well.
20:16And one of my favorite ones, this is one of my things I use most of the time because I just love creating, editing, writing, scripting videos. I have some for clients. Again, picture what I'm saying and apply it to what you're doing.
20:28Don't worry about my paper. Don't worry about any of the details right now. Think about why you want to implement it.
20:35I wanted to be able to use Claude, and, again, maybe you wanna use it in here.
20:41Right? Let's let's start even simpler. Let's get rid of that.
20:44Let's say I'm doing something for a client. Right? And you have maybe you're doing your real estate.
20:49You have a whole bunch of name conventions and data. Right? You have it maybe you have a whole bunch of folders that are, uh, you know, they have everything you need, whatever the explanation is.
20:58Right? But let's say you don't wanna look at that. You're not looking at Versus Code.
21:01It's just, you know, some folder on your desktop or wherever you're putting them. Maybe it's on Google Drive or something like that. Right?
21:09We want to reference that. Right? Every time I'm using Claude, my goal is to be able to do something with it.
21:15And for you, maybe it's like, hey. I am working on property a, and I need this information, information b, to do task c.
21:30Right? And what I mean by that is, right, when I say this to an AI if I was just to put this in right now, Claude would have no idea what's going on.
21:39Be like, oh, what are you doing? What are you I don't know. What what's the property?
21:42What does any of that mean? But if I'm if Claude is sitting inside of a large kind of work base, let's say you're just in Cowork, right, which I don't know if you've used Cowork, you can use that over oh my gosh.
21:55Hello. What's going on there? There we are.
21:58Let's say I'm sitting inside of my project here, and let's say that that project has a couple folders about those properties. It has a markdown file that describes, hey.
22:11You know, this folder is for property a, and here's all of the different information on that property and why I might need them. Right? Um, and in that case, when I say, hey.
22:22I'm working on property a, Claude immediately ignores everything else in that folder. It ignores all these other files and just goes to the file where property a is.
22:32Let me actually try to do this right now in front of you. It might be good.
22:36Let's do But you mean you mean because of the because you you tell it in the prompt, not not because it's rooted somewhere else or it's already set up like that. So do you do you need to tell it if I say I'm working on property a, then ignore the rest?
22:53Do you need to say that somewhere, or is it by default like that?
22:57So you wanna say it once, but you would wanna say, like, from a description. So the whole idea is, let's say you have that top markdown file. Right?
23:05And I'm gonna do it in here because it's already let's say let's ignore all of this because this is where it is. But let's say I'm sitting there, and you have folder for property.
23:16And, again, this concept works for everyone else. Right? This is the most basic version of ICM.
23:21You have prop you have another folder inside of it, property a oh gosh. I can't spell.
23:29Property. And you don't necessarily have to write this. You can have AI kind of structure some of this too.
23:35Property c. Alright?
23:39And then you'll have route file, or you can call it your Claude dot m d.
23:45Right? I most people call it a Claude m d, but the whole goal of the top markdown file is to basically tell Claude right off the bat, anytime you open up a fresh version of it.
23:57Right? If I open it up in here, if I open it up with a different AI, OpenAI, whatever AI reads it, it's just giving it a description of what's in there. And for your case, right, you might you wanna think back to why you're using it.
24:10What are you doing? I was assuming you're gonna want to be able to look at a property. Right?
24:15Someone emails you about a property or some sort of issue is there, and you need to be able to pull information from the information you've been already labeling. Right? And I wanna be able to take that information and write an email or give a report or send a payment or do something for someone.
24:32Right? And that would be that task. Each of these is a workflow or a description or a process.
24:38If you were to describe to Claude a prompt of how to do that.
24:43Right? So if I sat here and spent twenty minutes, right, just writing a whole bunch on how to do that, how I would do it, what your naming conventions are, what you actually can do is copy that. And instead of having to type that every time, you're putting it into this routing folder or into a subfolder.
25:00So let's say, you know you know, property, uh, like, I would be listing this routing file would list what the folders look like.
25:10Right? So here, I have source materials, pillars, ideas, brand voice, source materials, and everything that's that's underneath it.
25:18Right? I have pillars. I have, uh, ideas, brand voice, and it's just what's underneath each one.
25:25And then inside of those documents, I'm just simply saying and you don't, again, you don't necessarily have to type or write all these out. You wanna describe that folder structure to Claude, and it can write out a solid routing file like this, um, especially if you gave it my paper, and it'll do it well.
25:42And then I'm just saying, hey. This folder is for this. Right?
25:46In your case, you would say, hey. Property a.
25:52Here's all the things in it, tax documents, tenants. Right?
25:59Can you give an example of, like, one of your labeled conventions or something like that? Yeah. I just thought of it, but I only know it in German.
26:06Like, the the specs of the the specs of the how do you say that? How do you say that?
26:13Yeah. Like, the what the tiles are, the the flooring, and things like that. The materials used.
26:20Okay. Perfect. Right?
26:22And so you would wanna separate those based on how you would kind of think about it or read it. Right? So it's easy for you to read as much as it is anyone else.
26:30And, basically, this immediately gives Claude.
26:34Just by reading this, it knows what's in there without having to waste tokens to read what's in there. Right? And when you say, hey.
26:43Right? If you set that up, I'm working on property a, it immediately will read this and go, oh, well, property a is in this folder, and I know that it has these things in it.
26:54Right? And I go, okay. And I need this information.
26:57Perhaps it's I need to know which tiles were used and when they were installed.
27:05Right? And in order in order to give an email again, you might have a much more complex task, but this is just to kind of get the idea there.
27:17Now if I was just to send this to Claude or to any AI, it would kinda be kinda crappy and or people would say, oh, I can build you a real estate agent, and it's gonna have all of your properties in it, which can be useful. I've seen people build good systems that way. But the amount of effort it takes or to even have someone else describe this, I think, is minimal once you once you understand the simplicity behind it.
27:37Well, it's maybe not that simple, but in that aspect. So in order to give me an email, when I send this to Claude, even if it's a fresh version, not a random agent, or I send it to OpenAI, and as long as it has this routing folder, it's gonna know, oh, okay.
27:51Then I'm gonna go read property a. I'm gonna go pull up. I'm gonna ignore tax documents.
27:55I'm gonna ignore tenants. I'm just gonna go to the spec of the property folder, and maybe you describe kind of where each thing is. Right?
28:02Maybe you have different documents. You're just organizing them based on when you need it.
28:07Then where ICM gets kind of more special or more focused is when you add another MD file, right, that then describes all of these.
28:18So, like, maybe you have a naming convention. Maybe there's organization of how those materials, maybe where those contractors are.
28:26Your goal is to make it so that the AI has to think as little as possible to get to the information it needs. And so this top markdown file can be huge. Right?
28:37This is a huge one. But you can break it down. You can break down all of this context into smaller ones.
28:44So when I go into and I say this is from my side. Instead of property, I wanna make a video. Instead of, you know, the materials, I have different styles of video production.
28:55So when I say right. Let me get rid of all of this. And I won't even do it in here.
29:00I'll do it in this one. When I say, I want to make a video for NLP logics.
29:11Now, usually, that would require an agent, all you know, a whole bunch of prompting, things like that. This is a fresh co work.
29:18And, again, I could do this with OpenAI. I could do it with Hermes. I could do it with other things.
29:23And what it's gonna do is immediately oh, well, I didn't have it actually read the clot MD, so it's gonna take a second to dive in. That's okay. I hope it'll find it there.
29:31It's gonna immediately actually read that and go, okay. Cool. So my clotMD is right here.
29:36Here's the what the folder looks like. Then it's gonna look at the actual project. Oh, wait.
29:42You have an NLP Logix project. It's immediately becoming the agent, right, that you would have created for each client or each problem and immediately catching up.
29:53From one single prompt, it's immediately knows, oh, hey. Do you wanna make a new video?
29:57Because you've I see you've made five more before this. Do you wanna change videos? Do you wanna rework an existing video?
30:03Right? So I'm just gonna hit brand new. It immediately is then understanding just by reading and navigating the folder what I want, how I talk, what my things are.
30:15It's going through and saying, oh, okay. You said NLP Logix. Right?
30:20You said you wanna make a video. Well, I know if you wanna make a video, you go over here.
30:26Right? Record and cut pipelines. If you wanna do a project, then you have projects for different clients, and NLP Logix is inside of that.
30:35If you want to do something with that video, inside of these projects and inside of here, I have a description of what I do and how those videos are. So it's very quickly, without having to go through and read every single file, read every single thing to assume what I'm doing, it immediately knows, oh, that's what that workspace is for.
30:55The the way you organize the information and describe it for yourself becomes what the AI then understands.
31:04Now I've talked a bunch, and I'm sharing a bunch. And, again, this can all be done. You never have to touch Versus Code technically.
31:10You could just upload my paper into Claude and say, hey. I'd like to restructure this folder with my information to have a couple of these markdown files to route everything, and you could build something like that.
31:24Or you could have your IT person do something similar for you. And that's also on top of all of this, which we will talk about here in a second, and I'll shift off of this. I felt this was an important to answer.
31:36We are building kind of the layers. So you could just upload your stuff, and it kind of gets organized that way, um, to make it a lot easier.
31:46And then you don't technically even need another tool, um, where you'll actually have those kind of workspaces, uh, organized through ICM and whatnot.
31:55Um, and I'll let this load up and kind of come in. But, like, you'd be able to come in, just upload this folder that I'm showing you here.
32:02Right? And then it's easily just kind of just there, and you can just connect your AI to it. It's all kind of in there, and you can just start a chat with it and call it a day.
32:11Um, but going back, does this making more sense kind of how I described it there or no? Did I confuse you?
32:19It it's a it's a bit like it acts a bit like the PHP between front end and back end. Beautiful.
32:26May it doesn't maybe make more sense if I say, I have for instance, in my agency, I have these files with the specs in it, for instance. I produce myself in the agency.
32:36So I have an agency folder for the for all the graphic stuff with these files in it, the specs, for instance, or the brochure or things like that.
32:45And I have for the real estate for the pure real estate area, I have another an an another folder, another big folder with projects in it, the same projects in it, just the one on the graphical.
33:01So I could say, pull pull the the brochure from from that and pull the contract from this folder and things like that.
33:11So it acts a bit like because you said you could build, of course for a real estate developer, you could build a real estate agent's page or whatever, and you have to feed all that content constantly into it, and it needs to sync and everything.
33:27But that's what the PHP is is is doing all the back end and front end. So I keep only my back end? Essentially, yes.
33:35And pull with and and pull on demand my
33:41content. Exactly. Exactly.
33:43100%. Fly. My my formalization, my structure is there based on the patterns I'm seeing everyone else doing and what I think is the average of all of them.
33:53You don't have to take it as concrete. Hey. It has to be this way.
33:58My goal of that paper and my process was described, hey. Based on every single workflow and every single industry, this format seems to be a good methodology in which to implement it.
34:08But at the core of it, the reason I formalized all of it is that, is I noticed when someone like you, as long as you can route where it is, you know more about your information than anyone else will, even myself. And I also believe that the human should be in the compute layer.
34:24I think the argument that, um, agents are replacing people forgets the fact that humans are infinitely abstracting creatures. Uh, not everyone.
34:33Uh, this is not to say that jobs are not being lost, all of this jazz. But I think we are able to take technology that has collapsed extremely complicated processes and do more complicated processes on top of it.
34:46Um, so I yes. Essentially, that's what you explained, kind of this idea of PFE. If the back end is farther away and you're just kind of saying, hey.
34:54It's almost like . It's almost like the AI becomes your runtime. Oops.
35:00Someone Got to hear myself twice, though. That's good.
35:04But, yes, I think that's a really great way. But does that so that helps you? That kind of gets you in a in a better mind state for that?
35:09Obviously, there's still a little bit of extra work that needs to be done there, and that goes into the first question, which is, is your time more valuable spent on something else in your business and you offload this to someone else?
35:23Or is this valuable enough for you to do yourself that then it will return that value? Because that's a question I I tell a lot of people to ask. Yes.
35:32All of us can use AI and can build with AI and should be using it to a certain extent. But should you do you build?
35:39Right? You said you have an IT guy. Why why would you it?
35:44No.
35:45Not anymore. Unfortunately, not anymore.
35:48I would I would need one. I thought I'd need one, but that's actually not the problem. So if I would go if I would continue with the app but I I actually didn't see so much need anymore to to do things with with apps.
36:04What what I have now is is good and helped a lot. If I would continue that, I could continue I could easily continue that also with with a code editor. I'm used to working with that.
36:15I would probably get somebody I would still if in the community, somebody would like to do that to to get to get the security audit and Uh-huh.
36:23So, you know, somebody taking if it's if it's secure and everything. That I that I couldn't do. But I but but I totally appreciate the fact that me as the person that needs something to be to to just to get it done, you know, I I know what we need, so I can easily program that.
36:40But a a full stack programmer or IT guy probably would have to make the security audits. So that that's not that's not the problem. It's it's more like it's more like these little things that, for instance, now the naming convention could be done better.
36:59Okay. Worries. It's it's it's just not too too I I I don't even know.
37:04It takes longer to explain it to someone than to do it myself, but there are so many little things that I could do myself that I'm overwhelmed with it.
37:12Well, and and like I said, that I think that ends up becoming a business question, not a technical question. Right? Is Yeah.
37:18Like, is the time of being overwhelmed stopping you for because is your goal more money? Is your goal to scale? Are you enjoying learning these processes?
37:27And, like, I'm sure Natalie has some questions because she's chatting about this, and this is the important part. Being an expert in AI in your industry does not mean you are the one using it, right, or you're the one building it. If you're you've you've been in construction or you've been in real estate or you've been in some industry that is not software your whole life, it's okay if you wanna transition.
37:48That's what you wanna make your passion, but think about opportunity cost. It might be better to bring someone else on board and direct them your ideas. Right?
37:56Some of the greatest companies I know were great at delegation, and it's okay. Yes.
38:00I'm really
38:02yeah. Sorry. I really tried, but we really don't get people, though.
38:06If if I would get someone doing that, really, it would be it would have been lovely if but he he still he himself is is so busy. Yeah. Somebody wants to do that then.
38:16Sounds like sounds like I actually hired now a CEO. So I hired a CEO to replace me so I can focus on Oh, in that case, you are doing the delegation on that. But sounds like Bob is down to collab as well, though.
38:31Now I do wanna shift off this. So thank you. I was answering Thank you.
38:35Thank you. Because I felt like it resonated to the rest of the room.
38:40Before I kind of respond to more stuff, did anyone learn anything? Or, Ruben, you had mentioned that, um, that was kind of important there. But is there anything that anyone wanted to kinda hit on that you thought was really important about that conversation?
38:51We always wanna look back at dialogue because there's always hidden things that we we learned during it. Does anyone have yeah. Go for it, Krill.
38:59I think one of the actual things that you've taught me in the, like, the courses and the content that I've is also the fact that so you have your way of doing things, whatever it might be. If you're doing video editing or if you're doing SEO or websites or things like that, on the fly, as you're doing something, you could take a course from another person, and then you're like, oh, I like this part.
39:21So now every single one of my client processes, I just wanna slightly alter it. And with the file structure that you teach, it's so amazing that I'm literally just taking every single SEO course that I can.
39:33And the point is that they will add something new, I just tailor it. And I was like, oh my god, that's money well spent.
39:40Because now I can just use that, and then I just build my workflow, and it breaks it down into every single one. So the way that I, like, edit clients, for example, portfolios or the way that I edit their services, and it makes a huge difference.
39:55Like, these are, like, multiples. So I feel like they're working, like, at a 100 or a thousand times better just because I'm able to take outside information that people have been teaching and really apply in my business almost instantaneously.
40:10Yes. And that's the important part, right, is I'm trying to create systems where it allows you all to productionize your opinion.
40:18Because Anthropic, OpenAI, they're gonna make way better integrations. They're gonna make their models stronger.
40:23But the things that are fuzzy, the metrics where you think it should be done differently than the industry does, or you know in your niche, it needs to be done this way because of your opinion, how you've been it, that's hard to automate. That's hard to train an AI. I can't create for those of you who might be in machine learning, I can't create a reward function for your opinion that will then be generalized to the rest of the world easily.
40:48You can in certain stages, but it's very challenging. And in that challenge, you find value. You find something that lasts past model updates.
40:57And that's important. It's every little update you're adding to your files, right, every little word in here, every little process is your productionizing your opinion because now your AI is amplifying the way you have put data together, the way you think business should be run.
41:15And that's where I, again, think this is important. It's not that ICM is automating processes. It's that it's organizing processes to be automated so that you can then collaborate with information better at a higher level, a higher abstraction.
41:30So, yeah, 100% keep focusing on that because then that when when a folder becomes super valuable, it it it it becomes a great production layer. And, again, that's why I'm trying to create the thing that allows you to deploy that.
41:42Because in my head, imagine you or the six other people in this room could deploy that onto a platform, and I'm able to protect. And so I'm creating all the security on Azure, and we're using c c plus plus and c sharp, uh, kind of back ends to, like, build our own infrastructure because we don't trust anyone else.
41:58Um, and we wanna kind of make it so that certain parts of your folders can be editable by a client or a company that licenses your workflow and attaches it to their AI, but other parts of it can't be. And if you get paid for for licensing that workflow out to people.
42:14And so we're trying to create the infrastructure that allows that easier to where it's like, you can either hire someone and then also license their ideas out to the rest of the company, or b, you hire the virtual version of that person or their their workflows. Like, they say, hey, Krill.
42:30I'll pay you this much a month, or I'll pay these tokens to access your folder setup to do that workflow. And I believe that solves a employment problem while also creating a lot of money for people. So Jake, can I add something real quick?
42:44Yeah. Please go for it, Chris. Um, hello, Chris.
42:46So,
42:47um, I'll make this quick because I know you wanna move on. It's okay. We're here.
42:51I'll I'll make sure. Definitely, leaders in organizations, CEOs, whoever, who aren't in front of their computers every day, if you can, and it's legal wherever you are, get something like Plaud or be taking your thoughts and recording them on your phone, whatever, talking out loud, and have your first agent or one of your first agents in your ICM should be an ingest agent, where you take all of those transcripts, drop them into an inbox, and it pulls in your conversations, your thoughts, how you make decisions.
43:27You know, all of those different type types of things will form how your sort of scaffolding, your folder structure, all of your context will get built out.
43:39So I don't know. That's you don't have to make like you're saying, you don't have to sort of build this all yourself. You can take raw material, then use the ICM research paper, gives Claude a starting point, and then use your materials to make it tailored specific to you.
43:59Yes. Yes. Yes.
44:00Yes. Yes. Yes.
44:01And I the I'm just pulling up one thing that I'm doing right now, and I'll I'll I'll throw it up on GitHub here soon when I'm done. I'm I'm testing a distillation process. So, right, you don't wanna just simply grab transcripts and things like that.
44:14I mean, you do you should right now. Like, if that's what you can do and that's what you can move. But I Ari's worked on an amazing one.
44:20And hi, Ari. I got to see Ari in in London, which is amazing. She's building it based off of brain structure and doing a whole other level, which I'm I'm excited to have her share here soon.
44:28But I've been doing a very simple version because I I wanna start with a simple one that kind of is there. And it's like taking transcripts, taking the my papers, things that I've written, and then I'm trying to essentially go through and be like, okay.
44:42What information would I want everyone to know that I wrote and why I wrote those things? What types of things do I wanna keep private? Right?
44:50But then on top of that, how do I want the signals from all of those things to be broken down, to be, like, actually looked at? I wanna be able to ingest that input. I wanna be able then to separate it into these process, into these units.
45:03And then I actually wanna extract not just the physical words, but the patterns between those words. Why do I talk a certain way?
45:11When did I say it? How did I say it? Then I wanna distill it into its processes much like we do with the brain.
45:16And this is kind of how we train models, but it's in a much more structured way. And I like to do a lot of versions of this.
45:23And in a lot of ways, ICMs kind of already naturally doing that. Right? As I'm building this, I'm like, oh, I don't like how the AI did that.
45:31So I'm gonna change, you know, my the voice constraints here. I'm gonna add another one.
45:36Right? I don't like performed questions or, hey. I don't like, hey, guys, energy type stuff.
45:40Right? I'm adding these in, which is then kind of creating a version of my brain there. In a very simple way, I can then hand this to someone else in my company.
45:51My employees, they won't be me, but their work can be influenced by the way I think at a lot higher level, and it creates a kind of, I don't know, speed to production.
46:02Again, it's not perfect. It's never perfect, uh, but I think it's something interesting there. So, yeah, thank you, Curtis.
46:07I I agree 100%. Within, you know, rules, obviously, there's limitations with depending on the companies. But, yeah, thank you so much for that.
46:16I really appreciate that. Um, Oh, yeah. And so I just saw some text in there.
46:20I just put it in there. But for those of you who are not in the Discord, uh, if you go into, uh, the classroom and you come down to, uh, the drawing room VIP Discord setup, this specific link will give you VIP role and whatnot.
46:35And then you should be sitting inside the Discord. And this is just a great place to be able to link up with everyone, especially if you're not level five yet.
46:44And I apologize that we had to shift that, but there was just so much bot and spam stuff. And someone's even botting my Instagram right now, trying to send a bunch of comments onto me, and I've seen, like, six fake accounts for me, which whatever. That's the price of growing, I guess.
46:57But, um, definitely hop in here. There's a lot of talks going on, a lot of, um, kind of people diving into what they are and if you wanna connect. And then, of course, there's talk rooms as well if you wanna get into it.
47:07So highly highly recommend diving into that if if you are in VIP. I think it's extremely, extremely useful. Because, again, I'm here to answer lots of questions.
47:17We're here having these talks, doing these kind of masterminds. But and I wanna give more value than any of the other AI groups out there in these. Like, my goal is to just make these, you know, almost Harvard level lectures if I can, in in certain ways kind of back and forth.
47:31But the other value is everyone else in here. Right? Ari is insanely brilliant.
47:36Like, when we sat down in London, we got to talk about a whole bunch of, like, really cool weird stuff, uh, that she's building for distilling information using, like, neuroscience for the folders.
47:46It's such such cool stuff. But there's a lot of other people in here who are I mean, every single person who's introduced themselves today has just been really cool stuff.
47:55Like, all of you are doing really cool stuff. And that's the other thing I want to remind you of. And I say this all the time, but if you ever feel like you're behind, of course, you do.
48:04You're surrounded by experts. If you look at the rest of the world right now, most people if you go walk down to a coffee shop, most of them have barely touched Claude or ChatGPT. A lot of people I've worked with companies today who none of their employees have even used it.
48:19Right? And it's like, you have to remember how far you are ahead of get not only are you using it, not only have you seen the downsides of it, you've discovered a structured way in which to organize it and use them into such a level that it feels like cheating almost. But we're surrounded by brilliant people, so we're like, oh, man.
48:38There's still so much more me to learn. It's a process. Slow down.
48:42You're you're here, so you're not left. You're already not behind. I can promise you that right now.
48:47Like, you you are so far from not behind, it's not even funny. And it all takes time. It's not a overnight process.
48:55If it's an overnight process, it's marketing. That's them saying, hey. It's that's gonna if they're telling you, you know, this AI thing is gonna change your whole game, it is if you haven't used any AI for the last three years.
49:08That would be like if I someone hasn't been using AI for three years and I gave them a Claude account, it's a pretty big difference. But it's not like you eventually then need to learn the skills. You need to learn the process.
49:18Right? Excel is a hugely powerful technology. But guess what?
49:22It's still not overnight. If someone who's been using Excel for a year versus someone who's been using Excel for twenty years is quite literally a different software.
49:30Like, the guy who's used Excel for twenty years is a monster, um, and even then, right, in those areas. So, yeah, do not feel like you are behind, please. It is it is it is a learning process.
49:41I'm still studying and reading stuff every day. I I I was at London Tech Week sitting down with with PhDs and weirdos and all sorts of fun people, and I still felt like and even conversations with Ari, I'm like, damn. I didn't even think of that.
49:53That's such a good idea. There's no such thing as behind when you're building your own future. It's the world is hugely complex and infinite.
50:00So, yeah, don't stress it. Just motivation thing. Relax.
50:03Enjoy it. That's what we're here for. Um, but, of course, stay competitive.
50:06We gotta dive in. And speaking of competitive, um, I know we're coming up on the hour.
50:11For those of you who are used to my VIP calls, I always go over. So I apologize if you only slotted in an hour. Don't worry.
50:17I'm recording. You're more than welcome to leave. I'm not finished, and I love talking like this.
50:21So we're gonna be here until I finish, um, all of this. So, um, yeah, we'll make sure we're we're getting our time out of it. But, again, it's all recorded.
50:27So if you only slotted an hour, that's perfectly fine. So what did I hear in London? I got to see all sorts of crazy individuals.
50:34I mean, we had the CEO of AMD was there. We had people and researchers from AWS, NVIDIA, Microsoft, HubSpot, eleven Labs CEO was there.
50:43We had the mayor of London. There were some amazing people there that were doing some crazy talks. Uh, a lot of it was kind of government capital and silicon focused, but there was a lot of really cool startups from Turkey and The Philippines and just all around the world, and it was really cool.
50:58But the reason I I kinda wanna chat about it is to these startup and these tech events are a really good place to see where the industry is going. Right? This is a really good place to find out what are people looking at and also to see what not to invest in.
51:12Because if you see 27 startups doing the exact same thing, which I did, then you know, oh, okay. I probably shouldn't waste my time there because someone has already solved that problem or is solving in a good way. I can just pay them later.
51:25Right? Because when you're running a business, it's not always about being innovative because you built it. It's about knowing who else is being innovative.
51:32And when you buy someone else's software, you have to remember, if you wanna waste the time and build it yourself not waste the time, but build it yourself. That's good. But sometimes it's like buying a development team because they're handling software updates.
51:44They're arguing about the same questions of AI. And so you wanna think about, is this software that I'm building, is this AI solution that I'm building, am I the best person to build this, or someone else gonna do a better job?
51:56And I need to focus on some other value acquisition in my company. And it is highly dependent on your company where you are, but I really wanna make that important there. Also, there was a lot of VC and private equity investors.
52:08I sat down with the guy who invested in Perplexity and OpenAI five years ago or four years ago for both of them. And so he was really one of the investors. There's obviously a lot.
52:18And he was really cool. He wanted to to chat a lot more. Kind of a it's so weird.
52:22Once you get into the tech world, there's so many weirdos and stuff there. But I will say out of all of the investors I talked to, all of the money that seemed to be going, a lot of it is around structure and infrastructure. So I'm seeing a huge amount of effort going into this kind of like, hey.
52:37You know, the models are important, but what they're seeing, how they're seeing it, right, of, obviously, what we're working on in here. The kind of top level data is going into physical infrastructure as well as data infrastructure.
52:49So, like, hey. You know, how do I get it so that my data is being seen by the AI better? Or, um, hey.
52:57We need more chips so that the AI can run locally on my computers. AMD was sharing a new chip that they had, um, and it was a big service center that you could buy for your home so that you could run a trillion parameter models at home. Um, and NVIDIA obviously dropped their version of it too, which is is really interesting.
53:15Also, I find that there is, uh, and this isn't guaranteed. This is this is kind of a the buyer is changing is a weird thing here. But I noticed that there's groups of startups that are not selling to people.
53:27They're selling to the agents. Right? So HubSpot, Salesforce, um, a couple small startups were creating payment layers and data layers where the AI can actually buy directly from you, and it facilitates that process, be that with crypto or financial stuff.
53:44And their thesis, and I thought it was important, is right now, I think it was as of the last week, actually, more Internet traffic is being caused by bots and AI than real humans. So if anyone's familiar with the dead Internet theory, we are now one step closer to that.
53:59Um, but, uh, hopefully, uh, we'll never realize it fully. Um, I don't think we ever will.
54:05What is, you know, dead on the Internet anyway? But it's it's this interesting concept of if you have a service that's sold and it's blocked by payment and there's information and value behind it, someone else's AI might want to pay to get that information.
54:20Case in point, imagine I could let an AI sign up for this course to be in these calls to get these transcripts. If someone thought there would be value in in their AI agent going to these calls and paying for it, then they might approve that because that data has created enough influence and value much like how Krill said.
54:40Right? Krill's mentioning this. He's being the agent in this case, is he's the one deciding where the value should be.
54:45I'm gonna pull from these courses, these processes. Then why wouldn't an AI might be wanna sit in on the call and kind of pay for it, right, if you can kind of build those layers in? Um, so it's become an interesting area.
54:56I think there's certain areas where that will work. I think there's a lot where it won't work, um, but I did see that there. The one thing I saw, and I just this is opinion.
55:06You can disagree with me. I just think so many people are overbuilding huge like, I there were six booths, and they were spending so much money on huge booths on, oh, we're the agentic builders of the future.
55:19I must have seen that at, like, seven different companies. Like, don't get me wrong. If you're, like, niching down and, like, you're a consultancy, you can make a lot of money in that right now.
55:26And probably for the next couple years, you will. But if you're at the top level and you're trying to innovate into the next stage, just overbuilding these crazy custom solutions when you could just, like I said, make a folder and hire two people to get the same output.
55:43It's it's wild to me. And I I felt a lot of companies and startups were trying to just build these hugely overengineered things when you might not even need a software in that situation.
55:52You could just have a structured system, have the AI do something, and you would get the outcome of what that software was anyway. Right?
56:00Like, why would I wanna build a software to do something when I can build a system that gets that outcome, and I can skip most of the software BS anyway? Minus, like, UIs.
56:10UIs and and front end. Right? I think we're gonna see a shift over the next decade where the back end is even farther away from the front end, and there's kind of, like most people are building back end, only build back end.
56:21They don't even touch front end stuff, and everyone else is kinda just building their own front end and connecting to it. I mean, you all have your own, like, kind of layer. At least that's what I think.
56:30That's why my I don't know if you ever noticed my company's tagline. It's called the faces of interface, and it's specifically around that concept.
56:40I think people are gonna just make their own interface. Every human is gonna have their own fingerprint on it, but just personal opinions. You know?
56:49Let's see. Oh, yeah.
56:53And then, um, they they the opposite side is that they create platforms that, like or systems or fee that are just gonna be a feature in the next kind of, like, update, not because they're not production on theirs their opinion, but because they're trying to solve for, like, an integration problem. Like, oh, you know, none of the AI integrate with these three tools, and we made it so much easier to integrate with them.
57:14Well, okay. That could easily be a feature in the next update. Now if it's the way in which you process data is something you're building, that's fine.
57:23Again, the opinionated stuff. But I think solving the problems of what the big models can't do now in terms of, like, integration and, like, very structured data stuff. I again, it's okay if you give me pushback here.
57:36Not everyone is is this is true, but I think most people are not gonna solve it any better than Anthropic or Gemini or even some of these open source teams. Um, I did see a couple people like myself building in the talent layer. Like, hey.
57:50How do we right? I think I've heard almost every single person who talked today said, hey. I wanna hire someone or I need someone to help me with this, right, this process.
57:59Um, obviously, if there is a need and you are all business owners and you're saying, hey. I need more talent that can do this stuff, that means there's a market.
58:08And I think there's a huge market there. So I do believe, Alright? And I I urge you if you all wanna compete with me, like, I'm I am making a kind of talent layer process thing.
58:17Um, please build. Right? That's what we're we're gonna figure out.
58:19I I don't think because someone is making a version of your idea as a start up, you shouldn't make it. Right? There's Lyft.
58:26There's Uber. There's a a 100 different rideshare platforms. They're all making millions and billions of dollars.
58:32Right? Like, just because someone else has the idea, you the thing that makes it valuable is your opinion on that idea. Like, Lyft has certain opinions that are different than Uber that is why they think they should be successful, and so they compete.
58:45And so I think there we're gonna see over the next five years a lot of really big talent layer, like software as a service almost, but not quite, um, talent, uh, things. I think Handshake is one of the ones that is pushing already to dive into it.
58:59I think they're doing it the wrong way, um, but they used to just be like a hiring, um, kind of platform. I don't wanna sign up. I just wanna look at your front page.
59:08And but now they're, like, almost fine. Alright. Now they're almost entirely oh, yeah.
59:12You get to look at my 27 emails. I I think they're kinda diving heavy into that AI kind of layer space. Oh my gosh.
59:21Of course. This is this is what I mean. Just tell me what you do as a company, dude.
59:25Let's see. Here we go. Create the oh, Jesus.
59:28Of course. Alright. Funnels.
59:29Funnels. I'm not talking about handshake then. That was gonna be a good plug for them, but now it's not.
59:33But and that's my point. Right? It's like, if I was Handshake, I would not do that.
59:38But my my platform, what we're building looks more it's just very simple, and I'll I'll give a link to this once we're ready to kind of launch it. But the idea is right? Let me sign out real quick.
59:50Is you can come in, and you either say I'm hiring or I'm a candidate. And you you let's say you're a candidate. You come in.
59:58You know, we're we're doing
1:00:00Where you have a signing code and whatnot. Let me, uh, actually get that real quick, um, so that I can show you what this looks like. Um, and, basically, the kind of idea is if I was gonna sit there, I might try to get hired for a certain lead, a certain process.
1:00:17But the difference is everyone inside of this platform is going to be trained on ICM or that method. That's gonna be my opinion, is that every person who's in here is going to be someone who can actually understand this, actually understands ICM.
1:00:32They're able to upload their workflow so you can look through it. And if anyone knows portal or has played portal, it'd be really cool if we could name a company Aperture.
1:00:40But or maybe not. That could be a problem. And the idea is that you can go through and actually, like, send your intros, push them there, as well as, like, literally share your work and processes.
1:00:52And so this is just my solution to try to, like, get people hired who know these problems and processes. Um, and, yes, I'm thinking it will be a monthly charge to companies.
1:01:03I don't think I'm gonna charge candidates unless they're freelancing. Anyone who's VIP in here is gonna have it for free, and that's, like, gonna be no matter what.
1:01:12Yeah. Don't stress about that, especially if you're all the first, and I'll make sure that you all get contracts that if we get bigger, it stays that way.
1:01:20If you're in here, Gangal, yeah, I'm I'm not gonna charge any of you for access to it. In the future, to monetize it, uh, or at least anyone who's not in the community, um, I think what I really wanna do in the really big moneymaker would be, um, placement fees. So right now in the recruiting industry, if you wanna hire someone, uh, a lot of the times, especially at the executive level, you'll go to a recruiting firm.
1:01:43They'll find you. They'll do the her interview process. They'll put them in front of you, and they'll charge a percentage of that salary that you're hiring them.
1:01:50So if the person's a 100 k, they'll get 20 k for getting them for you and doing the whole process. Um, and so I think, uh, being able to build a platform that automates the recruiting industry in a certain way, but not in, a annoying AI way, but, like, just makes it better specifically around the problem of AI talent and ICM talent.
1:02:08Uh, it would just be great. And then the other side of it is if that works well, then I'm gonna connect it to the other platform, which is the one that allows you to upload your ICMs, and then the freelancers can just immediately license themselves to that company without even working there.
1:02:24Right? So the companies can go log in and give you, like, hey. You have the skills and the opinions I want, and you have some workflows.
1:02:30Can you just do this work, or can we connect to your ICM? And then you get, like, a small extra payment. Um, and I see that being, like, a really cool industry, and I see it solving the college problem.
1:02:41Because imagine you go through, you get your degree, but instead, when you leave, you have a productionized brain that you leave with college or with school that you can then license to companies right out of the gate.
1:02:54Um, your whole goal of getting a degree is to build a proper workflow process and understanding of your industry niche, which is what degrees are supposed to be, but most of the schools don't do that. Um, but now there's a software layer, and the guy idea of my whole fucking excuse my language. My whole thing is basically, hey.
1:03:13What if we were to give IP to the person for that? Rather than me taking all that data and then training a model, I think there's more value in growth where you have your own models that you're then licensing.
1:03:25That's my kind of, uh, niche thesis. People disagree, but I'll try it and work it out and see what happens. Let's see.
1:03:33I have been ignoring some of the, uh, things here. Uh, Ruby, I'm just wondering if I can build a service that is just a chat interface for house cleaner specifically. Connects to do a calendar.
1:03:42If they have a job that day, it text. One thing that I did for my buddy, um, this isn't house cleaners. Let me see if I can find it.
1:03:50Kanban. Where is it? I made a Kanban board that connects to his calendar, and he can chat with it, and it updates it.
1:03:56And he's, um, he's a construction worker, but he, like, he runs a bunch of construction jobs. And, um, let me see if I can find it. And the idea was, like, the Kanban board becomes oh my gosh.
1:04:07Where is the GitHub repo? I know I have it here somewhere. Hold on.
1:04:10I've gotta find this. Alright. Let's see if we can log in without breaking everything.
1:04:15Oh my gosh. Of course, I have to, uh, no one look at this. Hold on.
1:04:18I don't think oh oh, I put the wrong one in. There we go. But the idea was, like, hey.
1:04:24You have a Kanban board or or some sort of way that 2,000 repos.
1:04:30There we go. And it's basically, like, a database that updates on what he wants and how he does it, and it's like his little command center.
1:04:38And he can just kind of wander it over. This is his active one, so I'm not gonna screw with it right now. But actually, no.
1:04:45This is the fake one. But and he can just say, hey. This is what we're doing.
1:04:48This is what we have this week. Here's the backlog. Here's what I can filter it by customers or by, like, style or what's going on.
1:04:54Right? Okay. We have materials that we have to get, and then it's directly linked to his Claude and his Google Calendar by proxy.
1:05:03So he just tells Claude, hey. Update the Kanban calendar. Um, I'm doing a, b, c, d.
1:05:08Um, so I think that might be something really cool, um, that they would like, uh, in terms of scaling that.
1:05:15So, um, yeah, I would check that out. Let's see. Uh, I wish I knew enough to freelance.
1:05:21Okay. Uh, start freelancing.
1:05:23You know enough. You're in this community. I promise you.
1:05:25There's someone out there who doesn't even know how to prompt an AI who'd pay for your time. So I started freelancing, like, six years ago, five years ago, in writing and kind of scripting because I just loved writing and talking and speaking. And I didn't really I mean, I was in the Marine Corps for a while.
1:05:41I had a huge technical background, and I had been writing myself. And so what I ended up doing is just making a website where I just wrote a whole bunch on it, and it's still up. I don't really use it anymore.
1:05:51Um, but I just kept writing, and then I would tell people, hey. I'll write an article for you where I'll do something. Right?
1:05:57And even if it's, like, $50, right, I as long as you can get that first payment, it'll tell you why or why not they won't hire you again, what they wanna do. Um, you you always know enough to start freelancing because you know enough to have a job. If your jobless in here, well, then there there's a whole other conversation, I guess, we have to have.
1:06:14But if someone is paying you for something, it means they're getting more value out of you. That's how business works. And if they're not, it's a bad business.
1:06:21But if you are getting paid, you then can freelance 100% or start a company. Like, I just I just zeroed out in my mind that that's the case.
1:06:30Oh, this was a fun one I wrote back in 2023. Conscious is not binary. So I'll share some of that later for you all.
1:06:36But, yeah, just wanted to say that, Don, a 100%, especially if I've seen your posts, you could totally help companies in freelance, um, or team lens. Yeah. Work with others.
1:06:46Let's see. What do you think of stays protected defensible versus ubiquitous, uh, because of increasing intelligent models?
1:06:53I think once we get to a certain level in every single industry, there is no such thing as best. Right?
1:07:01So what's the best politician? What's the best writer? What's the best artist?
1:07:05Even programmers, we argue about what the best programming language is. Is Rust better? Is Python better?
1:07:11Is c sharp better? Right? And you could build the same exact software in three different languages that gets the same outcome, and the opinion is what ends up being the label.
1:07:21And I think it's the same thing for intelligence. We've been arguing about general intelligence way before robots. Right?
1:07:27What IQ tests are built on the idea of, uh, value g, which is this concept of a general intelligence. But we all know that that's general is so relative. Right?
1:07:36So I think what's gonna end up happening is we are gonna replace and automate a lot of stuff that we didn't think we're going to. But, again, there's so many problems in the world that aren't solved because we have not enough time, not enough work.
1:07:49We can't agree on how it's done. We're just gonna do that at a higher scale. So I I really I think as long as you're learning, you're breathing, you're creating, I just don't see a replacement for humans because and this is a weird opinion.
1:08:06It's never enough. Human desire is infinite. It's why the markets exist.
1:08:11I mean, if you go back two hundred years, most of the jobs that exist now are unfathomable.
1:08:18And I don't even mean we're gonna make new jobs. I just mean, like, the reason we would even need the job wouldn't even exist in our head. So I think there's there's a very I'm a very strong bet on humans, which some people disagree.
1:08:30Obviously, I was in the marine corps. I saw war. I saw hatred.
1:08:33I saw anger. So there's definitely some bad sides to humans, but I I I think there's a small percentage, but it's a net positive over the bad percentage in my opinion. Um, and if anything okay.
1:08:45Let's say it reaches superhuman intelligence. Okay.
1:08:48Well, then there's just more of us. Right? More humans means more work.
1:08:51More intelligent creatures means more work we're gonna do anyway. I I think it's it's just gonna be at a higher scale. Could it hit a point where problems happen?
1:08:59Absolutely. Um, I mean, humans do that all the time. We're at war constantly.
1:09:03So I think it's just gonna be the same world we've always lived in. Some is gonna be amazing. Some is gonna be disgusting.
1:09:09Uh, there's a quote I remember. It's I forget who it's by, and this is a beautiful quote. And it goes, no matter how many bombs we drop, no matter how many politicians become corrupt, no matter how many people spew hate, the trees still grow, the birds still chirp, and art is still made.
1:09:30And, yes, uh, you know, there is a lot of negative points to put there, but I I really do have faith in a in a small percent of net positive. And that's why I do what I do is there's a lot of bad ways AI is being used. And I but I think if we actively create the good ways, the good things, that is an act of defiance.
1:09:48That is a fight. By creating the good, you negate the bad often. Um, but if you constantly focus on crushing the bad, you often don't create enough good, if if that counts.
1:09:59We're getting really philosophical there, so I won't dive too deep. Um, go back to the business stuff, but I hope you all kind of get what that's going on there. Um, Yeah.
1:10:07Great great questions, by the way, everyone. Thank you for all of this. Let's see.
1:10:13Oh, yeah. So after London, I flew to Chicago. My flight got delayed four times.
1:10:19I got in at 3AM, and then I had to give a professional executive work talk to the CEOs, presidents, CTOs, CIOs of a very large data company at 7AM. So to say I was a little tired was a little rough, but it was a really cool workshop.
1:10:33We did a lot of work. I had dove into their teams. I talked to every single one of their teams, and we brought up, like, an entire kind of process of all of their workflows.
1:10:42And this is something that I do with a lot of companies. And so I went through and looked at their product team, their marketing team, their engineering team, and I asked how each of them and I gave them a workbook and said, hey. You know, what tasks do you do?
1:10:55What tool are you using? How often do you use it? What problems are not working?
1:10:59And I did it across which ones they think are priorities to get better at. Right? All of these things.
1:11:04This is a standard thing I do with every company I work with. And, basically, we went through, and I'll share a version of this template for all of you after this call. I just realized this might be useful for all of you to do with yourselves.
1:11:17And but, basically, I did that for each of their teams, and then I kind of created a workshop where I took all of their leadership. So every manager or vice president or whoever's leading each of these teams, we all got into a room and kind of broke down the process. And one thing that I noticed that was consistent, not just across this company, but a lot of them, so if any of you are running, hollow output.
1:11:39Right? A lot of people think what the AI is giving them is super solid and polished. Um, but at the end of the day, the steps to create the polished work actually made it almost have zero substance.
1:11:51It sounded really great, but it it just was built on nothing. Right? Um, and so I found that being constantly a problem.
1:11:58And that's, again, that's just because they're not giving it context. Right? They're not giving it rules, processes.
1:12:04More importantly, they're not giving it their opinion. You get substance by giving opinion. Substance comes from something that you can't just read or copy.
1:12:13Right? It's it's that weird kind of fuzzy area. Again, everyone was overbuilding.
1:12:20Oh my gosh. There was people building AI tools that could have been, like, a Zapier, uh, solution.
1:12:25Like, you could have just done let's just do a Zap. And, again, don't get me wrong. Like, saving money can be important if you're rebuilding a tool because Zapier is super expensive if you, you know, get a lot of things.
1:12:34Like like, I have Zaps for my school community, and it gets pricey very quick. But the amount of time effort to get the same output might not be worth it versus the cash that you can put down. So it's like, dude, just pay for the tool.
1:12:47Like, Like, it's really not that expensive unless you have a huge industry.
1:12:54Like, you have a thousand employees, then it becomes worth it to build your own because, you know, the search seats per head. But they were just overbuilding, which is wild.
1:13:03I'm sure some of you have heard about it, but the sixty thirty ten rule, right, my concept that, like, 90% of your structures, your processes should be, like, just traditional processes code, and 10% should be the AI on top of it all, which makes it all magic, so to speak.
1:13:20A lot of them were struggling with that except for the engineering team. The engineering team really was doing well on that, but that's natural. Right?
1:13:27Engineers naturally do that kind of rule process. Also, there were so many things they were trying to get an AI to do, and I'm like, how much you know, like, we calculated how much it would cost an API costs versus having a human for $40,000 a year copy and paste.
1:13:44And, um, I was like, dude, just hire an intern. Like like, just because you can automate it doesn't mean you should.
1:13:50Humans are compute efficient. We don't need a whole bunch of energy and stuff to do some pretty complex tasks. I just need a Red Bull and sandwich most of the time to get most of the work I do.
1:13:59And, like, that's very that's that needs to be in the calculation, especially I don't know if there's anyone who's in a big technical role or a data role in the in the call right now, but your efficiency of humans should be in that calculation.
1:14:14Right? Like, if you can give a AI tool to a person and they can get a lot done, that is more efficient and cost effective than trying to build an automation that does that thing kind of well, right, um, within reason. Again, larger companies, they get this, um, economies of scale, so it makes more sense to automate more stuff.
1:14:34But it's smaller bespoke teams. Like, AI native, in my opinion, doesn't mean you don't have any human roles using AI that you have gotten so efficient you don't need to hire certain roles or the roles themselves change what their value is giving.
1:14:53So, yeah, that's kind of a big thing. Red Red Bull and sour candy. Oh my gosh.
1:14:57I have a problem with sour skittles, and I keep getting the sour boss, um, boss. Was it sour boss? I forget what it's called.
1:15:04Final boss. Final boss sour. I keep getting ads for them, and I really wanna buy them.
1:15:08If they look so good. Um, Yeah. Mira.
1:15:10Oh, Mira. Wait. You're new to this community.
1:15:12I think I meant I remember you saying earlier in the chat. Uh, you didn't get a chance to introduce yourself. I'm so sorry.
1:15:17Would you like to introduce yourself?
1:15:21Uh, if you can unmute. Yeah. There we go.
1:15:25Hopefully, it lets you.
1:15:29Maybe. Possibly. If not, you if it's it becomes trouble, we can but I just I just realized that.
1:15:35I I recognize I was like, wait a second. You didn't get to introduce yourself. Oh, maybe?
1:15:39Nope. Still muted. Oh, yeah.
1:15:41Browser permissions. It's always fun. Yeah.
1:15:43Darn school and their their movements. I'm talking with the CEO of school to tell him to he told me to send him a Loom video of everything I wanted him to change. So I'm I'm I'm doing that down.
1:15:52Like, these are all the things that your software should update. If you're able to get it, Mira, just can go ahead and interrupt me while I'm talking.
1:16:00But if if not, just introduce yourself in the chat. That'd be great. Good night, Sandra.
1:16:06Thank you for staying on. Like I said, I'm gonna be probably another ten, fifteen minutes, or I'll try to close it up, um, but the recording will be there. Thank you for taking the time to be in here.
1:16:14And, again, make sure to connect in the Discord and and rewatch some of the older recordings if you all have time, or just give them to your Claude. That's another actually important note.
1:16:23If anyone does have to hop off, all of the prior recordings and sessions are in there. Um, and at the end of each one, if you go to the bottom, you you can download a ZIP file that has, um, kind of all and I describe what it is.
1:16:36Like, I have actual slide deck that I make from all of the transcript here. I have, um, actual, uh, kind of, like, markdown files, like, skill files that basically talk about what we were doing or how to do it, and I'm gonna do the same thing. Right?
1:16:49They're opinionated packages, not, um, you know, kind of theirs, and it's it allows you to just add that little bit extra to your own, um, kind of files.
1:16:58So it was my way of being like, hey. Instead of you having to go filter through all of this, I'll do it for you, then you add it to your second brains. It's instead of saying, hey.
1:17:06You know, I'm gonna put all this behind a gate, and then you're gonna try to take the transcript and use it to AI. I'd rather just give it to you. I know you're gonna use AI with it.
1:17:13I'd rather just help you out and make it valuable, becomes part of the product rather than something I'm trying to hide. But, anyway, yeah, have a good night, Sandra. That'd be great.
1:17:23Also, I was gonna spend a lot of time on this, but, uh, seeing as how long we were here. Um, but there is a paper that came out called SkillOpt.
1:17:36And, essentially, it is it's interesting.
1:17:40Some people are talking about it. I think I saw a couple individuals chat about it. It's from Microsoft.
1:17:46It's a lot of their teams. And it's this idea for being able to have your skills, your markdown files, essentially be updated and upgraded to a very specific set based on machine learning.
1:17:58So in the same way we train an LLM, we can train to make your MD files, your skills better. And it's a really long paper.
1:18:06They use a lot of really cool stuff and methodologies. They're doing kind of a mixture of neural network and using LLMs as a judge in a a kind of validation set. Long story short, to kind of get past so you guys all can kind of go into it.
1:18:20The idea is alright. Uh, actually, let me use Paint.
1:18:23This will be much easier, um, to describe here. Um, and Microsoft said, hey.
1:18:30You know what'd be really cool? What if you have an MD file, and it's got a bunch of text in it? What if we convert all of those words into a vector space?
1:18:41Right? So there's a bunch of numbers. And that's how we make language models now.
1:18:45We just take a word and we assign it a number. And we say, cool. So that is what this this markdown file is.
1:18:53Right? So, like, this markdown file oops. Nope.
1:18:55Not this one. I'll do
1:18:58this one's a good one. While you're pausing, I managed to find the browser setting, which was annoying. Yes.
1:19:05Hi. I'm Mara. I'm in New Zealand.
1:19:08So hello from Sunday morning because I'm usually in the future compared to most people. I I'm about five months into a new role.
1:19:20FMCG company, we're probably one of the biggest food 100% privately owned food companies in New Zealand.
1:19:27What got me excited about the job is the owner who is not the CEO, but he has a data background and a BA background from back in the day.
1:19:38So he's basically created this role, and he's put data at the big table reporting into the CEO.
1:19:44So all data will take and massive growth, grown by about six times in the last six years in terms of top line revenue.
1:19:55And now at the scale, we have joint business plans with all of the major supermarket chains in New Zealand. So, yeah, very, very cool.
1:20:02Very exciting time.
1:20:03Wow. Thank you, Mira. That's amazing.
1:20:05I love New Zealand. I've been wanting to to kind of wander out there. Actually, almost almost did my master's there instead of Scotland, uh, just because I knew one of the only ways to to spend more time in in New Zealand would probably be through academics for me.
1:20:18Um, but it is that's really interesting, and that's a crazy growth process. The food industry as well is is a really hard place to be in right now because you have to deal with compliance. You have to deal with opinion.
1:20:29You have to deal with people. You have to build with logistics. Then the logistics are getting upset by politics, uh, and then you have to deal with AI.
1:20:35Is AI the right way to be a man? So it's a hard place to be in So, yeah, thank you, Mira. I'm I'm glad to have your expertise here and your opinion.
1:20:41What's fun is the logistics side.
1:20:44Tell you what, what wasn't fun was, you know, literally hundreds of thousands of dollars of FF per week with with some of the things happening in the world. Just like turning up and each week, it was like it was like rolling dice.
1:20:54So what's gonna hit our books this week? Because they were literally adjusting their percentages on a daily basis.
1:21:01So that was that was interesting. Seems to have plateaued at the moment at about double what we had budgeted Oh.
1:21:07Based on last year. So that's, again, interesting place to be. Um, but we we bring or make a lot of Asian food into the country.
1:21:17So what what's done that growth path is more and more of our products are now ranged at supermarket rather than at the Asian supermarkets.
1:21:25Interesting. I'd love to I I we don't have time for it now, but I'd love to look at your data and and kind of what you all are looking at because, um, I think there's a huge market for for Asian foods in in the the South Of America or South South not South America, South US.
1:21:40Yep. Right now because a lot of veterans are you know, spent a lot of time in Asia and things like that. We have a kind of huge veteran population paired with kind of, like, this desire of of Korea and and Japan, and I think The US could be like if if you got a Japanese seven Eleven in Florida right now, you would make so much money.
1:21:58I don't know if anyone's been to Japan, but the seven Elevens there are top tier, top notch. So
1:22:04we we are about fifth the the thing that is nice in these economic times, about 50% of the instant noodles in New Zealand, we they're our products that are sold.
1:22:15So when times are tight, people go to ramen and instant noodles, and that makes us money. So that's quite a nice place to be when things go wrong.
1:22:24That's wow. That's yeah. No.
1:22:26That is hard. But wow. Cool.
1:22:28Well, thank you, Mira. I'm really excited to have you here, and I hope, uh, hope to see you next week as well at the next one. Maybe, um, you can bring some questions.
1:22:34And I apologize that there wasn't a questionnaire for this one with London Tech Week and everything else. I just did not I knew I wouldn't have time to really sit down with them. I'd rather just do it live in the in the call.
1:22:45But this coming coming high tier, I'll definitely have the questionnaire out. Anyway, so, basically, they created this process for basically automatically upgrading or testing scale skills.
1:22:57Right? And so let's say the goal of a skill is to route the AI somewhere, or the goal of a skill is to create a skill. Basically, you can have it go through and have an AI use this skill hundreds of times and every time judge the output.
1:23:14Right? Did it do it the right way? Did it create a good skill?
1:23:17What didn't it do? And it does this over and over again, and that will make edits to this your markdown file until you have a good one.
1:23:26Now there is a version of this that already existed by our lovely man, Andrew Kaparthi. He gave this one, uh, I think, about three months ago, which is called auto research, and it's a much simpler one where, basically, the AI goes in, says, hey. Is this the right skill, the right process?
1:23:41No. Let's fix it, and it keeps doing it and redoing it until finally the agent or the skill is actually really solid.
1:23:49Um, these people at Microsoft took it to a whole other level and actually brought in machine learning algorithms to keep track of every single token and word, and they actually vectorized the markdown as if you would as an AI and did that at a huge scale. Um, I'm telling you all of this because I'm working on creating a open source version that's really easy for everyone to use inside of ICM.
1:24:12So my goal is to create either a version or a hybrid between Kaparthi's and this so that you could just throw it at your database or whatever you're doing and say, hey. Um, like, case in point, let's, like, go into, uh, like, you have a brand voice document.
1:24:27Maybe your brand voice needs to be better. Right? It actually looks at your chat history, and every single time you've said, no.
1:24:35You're off. No. You wanna do this.
1:24:36It's doing that on auto overnight, and then you wake up in the morning, and you end up having a condensed skill. And what's cool is they did this already, and you can give those skills to any agent, and it is increasing them by a disgusting amount in some cases.
1:24:52So, um, there's current processes that already do this. Right?
1:24:57Like, all of these are processes that you have to do. So, like, searching, um, through a spreadsheet or using spreadsheet, there's skills for that.
1:25:05And there's actually, um, these are all optimization techniques, text optimizations that already exist. Human skill is literally someone going in and reediting the skill file.
1:25:15But what they found is in a lot of situations, the original baseline will go up by over 30 to 40% by using the So, like, it's using the scale.
1:25:25It tests and, like, reads through spreadsheets, and it's scoring, like, 40%. You use SkillOpt on your skill or on your your markdown file, your process, and it it can get up to twenty, thirty, 40% increase in its ability, which in the AI market usually requires fine tuning the entire model.
1:25:45Like, to get these type of numbers, um, like, every single one of these, 19%, 18%, 9%, 57% for the codex harness and the and that's the other thing on the Claude and codex harnesses.
1:25:58It did the best. To get a jump from 27% capability to 85%, that's, like, two or three model generations.
1:26:07But all you're doing is upgrading your skill and context, which I think is just really crazy. Again, the reason I'm sharing this, if you all wanna dive into this, let me actually drop this here or if you want your AI to dive into it.
1:26:18There's no good YouTube videos on it right now, so I'm gonna make one as well. But I think there's some stuff that were they're, again, overengineering a little bit, um, and I think there's a hybrid that could be done. I'm gonna try to make something.
1:26:30It'll be open source, and I'll drop it for all of you. And the goal is to be able to come in and say, okay. Let's go ahead and aim it at these couple markdown files, or let me aim it at my ICM process for animation.
1:26:42And I'll start small, or let me even aim it at my ClaudeMD. Is my ClaudeMD routing to the folders as good as it should be? And it basically gets rid of, like, even individual words that are adding too much noise to the model, and it does it through a series of processes.
1:26:57Um, and so I think this is super sick paper. Uh, it just came out this May, um, and very interesting, and so I'll be diving into that. If you all wanna dive into it, again, by all means, uh, they also have an open source GitHub for it.
1:27:09Um, but I'll be I'll I just wanted to let you know I'm doing the work so you don't have to, um, and I'll I'll put that all out for you, uh, here shortly. And I do think this is gonna be a huge thing or people you're gonna see companies and people talking about this soon. Um, I will say Anthropic already kind of does a version of this.
1:27:27So if you actually go to where Claude is installed and you go to the, um, skill creator, they actually have, um, scripts that force it to revalidate the markdown file that's being created.
1:27:41Um, and their skill creator is kind of doing a mini, hey. Is this a good skill? Or it but it's mostly doing it with this.
1:27:48Right? So it's mostly, hey. Is the description here good?
1:27:52Is what these the description about how a skill should be built? Um, but it's not doing it to, like, a high level. And right now, a lot of the other automated processes, what ends up happening is after, like, a million tokens, it starts really having bad context rock, uh, which is what the paper was trying to solve by using, like, traditional code, which, again, just kind of supports my theory.
1:28:12Not saying it. Just saying a little bit. They're using traditional code to solve AI problems.
1:28:16So, know, you just putting it out there. It's it's you know, you still need the traditional stuff. Just because AI is doing the coding doesn't mean the fundamentals go away.
1:28:24But, yeah, I just wanted to share on that. I thought it would be cool. I I I saw, I think, David and Roy or someone someone else posted about it in in the community, and I just wanted to chat about it.
1:28:37Like, hey. This is what I get out of it. It is pretty interesting.
1:28:40But long story short, it reads everything you did during a process. It edits your markdown files, makes them shorter.
1:28:48And that was the other cool thing about the paper. Almost all of the markdown files ended up less than 500 to 800 tokens. So it found that the most efficient markdown files were really small, and it was better to break them apart, which all of you know why.
1:29:04Because construction and separating thought is what allows the model to work better. So now I have mathematical evidence of why my ICM works well.
1:29:12Um, but, uh, I told you about the deployment layer for the ICM, and I'm gonna make a video and, um, hopefully get that all to you. We did some questions in show and tell, um, so that was kind of built in there. But other than that, I think let's see.
1:29:26We're coming on an hour and a half here. So we're thirty minutes over our usual hour, but, uh, you know, I always love talking to you all, and I can sit for a while. Um, I guess then for the biggest things coming up, I will try to get all of you access to the kind of employment layer freelance, um, uh, app that we built, uh, and then I will try to get you all access to the ICM this week.
1:29:46Uh, that that's my goal is to launch it Wednesday ish. Hopefully, um, right now, I've been spending, like, a disgusting amount of time on security, um, because that's really all I care about right now.
1:29:56Um, and, luckily, we're using Microsoft Azure as a main deployment layer, so, um, there's there's a lot of good security there just naturally. Um, and we're trying to create kind of more enterprise solution grade technology, but for everyday people as well.
1:30:10Because the idea is everyday people will be able to be inside of this. We can rely on Microsoft Azure for a lot of compliance, and then you would be able to deploy these kind of folders.
1:30:21And that kind of creates the you you as a freelancer or as a small company building these these things, it's really hard to deploy at enterprise level, but this hopefully helps out a lot.
1:30:32Allegedly, that's my goal. Uh, we'll find out with testing, and it's probably gonna be terrible the first month. So I'll I'll tell you all that.
1:30:38Uh, but that's how software goes. Um, so we'll roll with it.
1:30:42Other than that, this has been an amazing day. I hope everyone here got some awesome work out of this. I appreciate everyone sitting for an hour and a half on a Saturday.
1:30:50Uh, I mean, it just shows you're committed, and I really appreciate that. Um, I hope that you all kind of see a better future. You're coming out of this more clear every time we go to one of these.
1:31:00And more importantly, I hope you all find people that you wanna work with. Again, my whole goal is to teach you all and and be here and give you my time and just give you updates on all the crazy stuff that I'm seeing, but I also really, really, really want this to be a space where you're all like, cool. I know if I go to talk to Mira or I sit down with Curtis, we don't have to spend an hour making sure we're we're we're caught up to this stuff.
1:31:21Well, I don't have to go through this filter of, do you really know what AI is? No. It's like, okay.
1:31:26They're in this group. We can get right to the hey. This is what I'm trying to build.
1:31:29You already are caught up to it. And that just it just cuts out so much BS that you have to deal with in the AI world now. So I I hope that's what all you do.
1:31:38Um, knowing that, I am not gonna end the call. Um, I am gonna stop the recording.
The Hook
The bait, then the rug-pull.
What you are watching is a session that was never meant to be public — a 91-minute VIP community call shared once, as a single exception, before being locked away for good. What's inside is the kind of briefing that usually costs a consulting retainer: a live debrief from London Tech Week, a real-time ICM routing demo, a frank readout from a Chicago executive workshop, and the first walkthrough of Microsoft SkillOpt — a paper that may make your markdown files more powerful than your next model upgrade.
Named ideas worth stealing.
04:07model
ICM — Interpretable Context Methodology
- Routing markdown at folder root
- Subfolder descriptions without reading files
- Opinion-encoded context per task type
- AI as runtime over structured human knowledge base
A folder-and-markdown-file methodology for organizing AI context so any model on any update can navigate your knowledge base from a single prompt.
Steal forAny client workflow where the same AI context needs to be reused across projects or team members
15:18list
The Three Questions
- Delegation question: can someone else do this?
- Complexity question: is this more complicated than the return?
- Outcome question: what are you actually trying to get done?
Framework for deciding whether to implement something yourself, hire it out, or skip it entirely.
Steal forClient discovery calls, onboarding sessions, evaluating build vs. buy decisions
1:10:43model
60/30/10 Rule
- 90% traditional code and process structure
- 10% AI layer on top
The ratio that produces durable AI solutions. Inverting it creates brittleness, hollow output, and high maintenance cost.
Steal forAuditing client AI implementations; scoping new AI projects
1:18:22model
SkillOpt Loop
- Vectorize the markdown file
- Run AI task against it hundreds of times
- LLM-as-judge scores each output
- ML edits the file to maximize score
- Repeat until capability plateaus
Microsoft's automated process for improving markdown skill files using machine learning. Produces 30-57% capability gains without model changes.
Steal forAutomatically improving any CLAUDE.md, skill file, or routing document in an ICM setup
How they asked for the click.
VERBAL ASK
1:30:29product
“I'm gonna launch it Wednesday ish. Right now I've been spending a disgusting amount of time on security.”
Soft mention in context of community launches. No hard sell. Members are already inside the community.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
