The AI Tools I'm Building to Replace My Spreadsheets with Anish Acharya

For years, planning any trip meant flipping between Google Flights and a couple of award search tools. putting all those prices and both points and miles into a spreadsheet and then doing the math to determine what to book and how to get the best deal. Then a few weeks ago I built a new tool that does all that in just a few minutes. It knows my rules, like never suggesting super early flights or having a layover unless the savings is really worth it.

And this is just one of about a dozen tools I've built recently that have collectively replaced all the spreadsheets I use to dial in pretty much every part of my life, whether that's travel, health, work, and a lot more. So today I'm comparing notes on this setup and the AI tools that power it with my good friend Anisha Charia. He's a tech investor at Andreessen Horowitz and a longtime friend who's been thinking about all this from a different perspective.

And we're gonna get into why I'm not even using OpenClaw anymore. We're gonna talk about the architecture mistakes that almost everyone is making, the case that personal AI is about to give every consumer the kind of optimized life we all dream about.

And what happens when the heavy subsidies on all these tools eventually go away? I'm Chris Hutchins. If you enjoy this episode, leave a comment or share it with a friend. And if you want to keep upgrading your money points in life, click follow or subscribe. Anish. Okay, we're doing it. This is a thrill man.

Forever. I know. I feel like we've been talking about this nonstop. Yes. And we haven't sat down to actually put our thoughts to paper. And so I guess now we're putting it to audio but and video. We should give e everyone a little bit of the backstory. We've known each other since Google. Yes. Since you were responsible for some of your original hacks. Your hacks before they were they were trademarked as hacks, which was helping the group of us go from Google Plus to Google Ventures.

And then later I don't embarrass you but being responsible for my marriage. Yeah, well you know I've only been responsible for one and you're are it. The most important one. Yes. Good. Well, dude, it's a pleasure to be here. You know, it's it's funny I the way I've been thinking about the world right now is it feels like there's two religions in Silicon Valley.

Claude Code and OpenClaw. Each of them have their sort of figures and I feel like you're the person who's been the most at the edge of the OpenClaw, but broadly AI enabled productivity ecosystem. So maybe a good place to start is like

What are you working on? I like the way you framed it because I've been having a, I don't know, internal crisis about this show. The show is not supposed to be the AI show, right? I don't want to have videos where every day I'm publishing stuff about these models. But the way I generally think about the world and the way this show evolved was like I really like to dial in my life.

And historically, dialing in my life has been making spreadsheets, doing tons of research. And now I'm just still dialing in the same aspects of all my life, my health, my points, my miles, my travel, my money. It just turns out that there's these amazing tools that make it so much more efficient. So one thing that I do a ton of, as you know, is I book flights with points.

And searching for flights with points has gotten so much better in that there are all these tools that you can sign up for, most with a subscription, but some for free, where you could say, oh, I want to go from point A to point B. You could maybe add some variability.

The challenge has been that when I want to book a flight, what I want really depends on what the prices look like and what the points look like. So if it's a really great deal on points, awesome. But if it's not, maybe I'll pay cash. If business class is a great deal, maybe I'll do that. So the way I've always searched for flights for a trip is I'll go to some award tool and I'll search June fifth, then June sixth.

Then June seventh, and I'll write all the prices and points. Then I'll go to Google Flights and I'll search the same set of dates in coach. Then I'll search the same set of dates in business. I'll put it all in a spreadsheet. And then I'll go in and like figure out the math to figure out what makes sense. Now I thought, well, I have a computer. I have an agent that runs on the computer. Can it just go do all this, aggregate all that same information, and present it exactly how I want?

And so in a couple prompts, I was like, I want you to use these tools, Google Flights, Award tool, Seats.air, whatever it is, and I want you to collect all the information for these dates. With the preferences you know I have. I don't wanna leave before 7 a.m. unless I'm gonna save more than a hundred dollars. Like I only want nonstops if available unless I'm gonna save like two hundred dollars. Like I have my rules. And then I get this response, maybe it takes ten minutes.

And it's like, oh, you wanna go to Cabo for five or six days and you can leave on these three days? Here are your options. We know how you value points. Here's what you need to make a decision. It's amazing.

Incredible. And okay, so take me back in time. Was this the beginning of your journey? What product were you using for this? This is actually very recent. It was part of an experiment to detach myself from one tool. So part of the experiment was: can I build this tool with Claude Code? And then can I use it with OpenAI's codec?

Okay. Because I wanted to s do things that were platform agnostic because I think I don't know, I felt like we were going down this path where it's like everyone's saying Chat GPT, we're all tied in. And I worry that in the future we're not gonna be tied in. And I want a platform that allows me to not be tied into one thing because who knows? Like people are pulling the plug. People are gonna raise their prices. They're gonna stop subsidizing. Yeah. And I want my skills to work.

wherever they are. This is later, right? I went deep on open claw for a while. We talked about it. Yes, we got I was like singing from the mountaintops about how cool it was. Yes. Now I'm not even using it. That's crazy. That's crazy. But how are you interacting with your cloud code and your codec then? At its core, Claude Code or Codex. I'm kind of interchangeable. This week I already maxed out my Claude Code Max plan, so I'm like, I can't use Cloud Clint. You're a token glutton.

I know. I've been uh you know, I feel like I've just been probably hitting it hard. I'm trying to automate a lot. Someone's leaving our team. So I've got a lot of new processes to automate. So I just kind of hop in. I'm almost exclusively using chat or code. Yeah. This kind of middle ground of cowork that Claude has. Yeah.

I feel like it's a great entry point until you realize that like you could do a lot of the things in cowork just with chat. And then you could also do a lot of the things in code and more. Yes. So I'm mostly doing claud code or codec sessions, no longer using the command line, mostly using the native apps for each of those.

platforms uh-huh and throwing it in a folder that I've now kind of treated as my agentic operating system, if you will. When I got after OpenClaw, it was fun. It was great. It did all this cool stuff. And I was telling everyone about it.

But one of the challenges with OpenClaw was everything was messy. There was no intentionality. And every two days they were launching a new version and it was upgrading itself, sometimes breaking. I remember the thing that I got me so excited about using Clawed Code was actually that it could fix OpenClaw. No more gateway down errors.

It was so bad. And I also realized that practically agents and humans are very, very different. And so you can't just have one threaded conversation forever with unlimited context. with OpenClaw. It just got so messy. You'd be like, hey, I want you to tell me what's on our meal plan tonight. And then next you're like, oh, okay, can you go build me a cool app? Oh, can you go search for these flights? It got so messy and I felt like

Tasks needed to be threaded. Yeah. Like they are when you log on to Chat GPT or Gemini. There's all these nice clean threads. That's right. That's what I was missing. And I feel like OpenClaw didn't have a good solution except building more agents. So I built more agents.

Talk about your sub agent architecture and your agent per channel,'cause I think that's very sophisticated. So I started with, ooh, I want a bunch of agents because I want one agent that can help me, you know, think of creative for the podcast, one that could do some research, one that could help family stuff. I had this idea that they all needed to be different agents. And now

I don't have different agents because the concept of different agents doesn't even make sense. Every agent, your agent, my agent, they're all exactly the same. They're just assuming we're using the same model. If our agents both work on top of Let's say Codex five five or ChatGPT five five or Opus four seven. Like they are the exact same thing and the only difference is what context we feed into it.

Yes. And so that context, originally I was thinking too much like a human and I was like, well, if someone's being creative, I want all their knowledge and their lived experience and their brain to be all about creativity. Right. But there is no lived experience. There's just what is in the session you have. And so for people who haven't gone too far down this path.

Every model has some number of tokens that live in its context. And every time you ask it a question, you're sending a bunch of information to it and it is giving you an answer. Anything that is not in that information, which by the way, that information could be Here's where to go find other things. Yeah, I use a tool called a search memory, et cetera.

But if you don't give it that information, it has no idea what's going on. Whereas humans, it's kind of interesting. So if you take an average model with, let's say 200,000 token context window, some are now up to a million. It's roughly a 500-page book. Okay, so that's about two hundred thousand tokens, five hundred page book. Human brains keep in context, we're not the same, it's that we can't measure it perfectly in tokens, but it's like four ish things in context at any given point in time.

Now, depending on our experience with those things. Yeah. And I was reading an example where it's like a chess grandmaster, one of those things might be the board. A newbie to chess, one of those things might be like, what does the rook do? So we have varying degrees of what each of those four things are. The difference is that we have like two and a half petabytes of long term memory in our brain. Yeah. That's right.

Which is like the equivalent of twelve million five hundred page books. Uh-huh. So like we have this immense ability to query all the stuff we know. Yeah. But when we make decisions and when we like actually run calculations in our brain, we're using much, much smaller amounts of information. So if you were to say, gosh, I wanna look at everything I've ever sent by email to this one person and I wanna understand our relationship. You could literally throw all of those emails into one contact.

And an LLM could process it really, really well. For me, I would have to read them and then I would have to like store experiences and then tap on them and it would take a really long time. And so I've just started thinking, okay, so if we're gonna structure how we work. We have to really think about it in terms of how these tools work, not how people work.

And the whole agent model was great, but it kind of personified these computers and it made me think, like, oh, I should set it up. Like I would set up my team. I have the CFO. I have this. I have this. Yeah. Maybe that's necessary. Okay. But what I started realizing was If you create workflows and LM seem to be really good at like repeating specific things, really bad at like

abstract thinking on the leveling chart of like hiring managers versus VPs versus directors, they're not directors. Okay. But at the lower end of the experience They're really good at these repetitive tasks, but you got to define them. And if you're gonna define them and you know that the context is whatever you give it,

then it doesn't really need to have another agent. You could just throw into the context, here's a thing you need to do, here's the style guide for this type of writing, here's the past examples. And so Yeah, I don't know. I didn't need a persistent thread with ten agents. I just needed a set of rules or skills for how to do things that any agent can hand off to another agent. This episode is brought to you by Bill.

You've heard me talk about Built as the loyalty program that lets you earn points on rent wherever you live, and they just leveled up even more. As of 2026, homeowners can also earn up to 1.25x points on their mortgage payments. This is all thanks to Bilt's three new credit cards, the Palladium card, obsidian card, and blue card.

All three turn your housing payments, rent, or mortgage into flexible rewards so you can choose the card that fits your lifestyle without missing out on points and exclusive benefits. Built points can be redeemed at top airlines and hotels, Amazon.com purchases, future rent payments, and more. Built points have also been ranked by top publications as the industry's most valuable point currency.

Your housing payment is already your biggest expense. Make it your most rewarding. Find the card that fits your lifestyle and apply today at joinbuilt.com slash all the hash. That's J O I N B I L T dot com slash all the hack. Make sure to use our URL so they know we sent you. Terms and limitations apply, subject to approval and eligibility. Built cards are issued by column NA, member FDIC, pursuant to license from MasterCard International Incorporated.

This episode is brought to you by Green Chef. I've tried keto, I've tried intermittent fasting, neither stuck. And honestly, no eating trend has made a real difference the way just having good food already in the house has. That's why I've prepared dozens of green chef meals over the years. The chicken with garlic spinach rice is a favorite, and we've reordered the Thai style coconut shrimp soup multiple times. What I love most is just not having to think.

No deciding what to make, no wondering if we picked up a veggie for the side. Every week you choose from over 40 recipes, Mediterranean, high protein, or the new longevity line built around brain and gut health. Meals come pre-portioned with organic produce and responsibly sourced proteins, plus unlimited one-on-one nutrition coaching.

Green Chef is the only certified clean meal kit tested by the Clean Label Project. If you can't tell, I'm a big fan, so you should definitely check it out. Head to greenchef.com slash 50 all the hack. and use the code 50ALTHHHHAX to get 50% off your first month, then 20% off for two months. That's code 50ALTHacks at greenschef.com slash all the hacks.

Okay, so just taking you on the journey then, you started with a single open claw instance that you message via iMessage or Telegram or whatever else, but then it started there's just too much in the context window as you were context switching, human context switching, it got confused. Then you move to subagents per sort of topic area, travel, family, finance, et cetera. Then you realize that was probably too skeworphic.

So now you're sort of back to a single agent which has different skills. Each of the skills sort of corresponds to pulling in different things to the context window in different topic areas. Is that right? I think that's pretty spot on. Okay. I would say It's not that I have one session every time now. I have multiple sessions because there might be things that you know I want to follow up on. Like if I have a task.

That task could either be something that just gets done once and then I don't ever need to come back to it, but it might be something where it's like, could you draft this thing and then I want to have a conversation. So it's important to spin up a task in a way that you could continue having that conversation, not that it dies at the end. Yes. But at the end of that task, I don't need that person to like exit

I might have a now take everything you learned from my revisions and save them to this like how Chris writes doc. But that how Chris Writes doc is actually relevant for all the tasks. It's not just relevant to one specific thing. And so when I look at my workspace, I started with the open claw folder and I was like, okay, let's start here. And I was like, okay, we need skills. Like skills are these repeatable tasks.

And so we have a flight search skill, we have a go reconcile invoices skill, we have a go transcribe something skill, and they could be pieced together. I've got connectors, which are just other things you might want to connect to. Here's one password, you know, here's Google Drive, whatever things that I don't need you to replace.

Go get connected with them. Projects, which are like, I want you to just know that there are multiple things happening. If there's multiple skills and multiple things, like sometimes I elevate a skill to a project where I'm like, it's not an app yet. Would like trip to Japan be a project? Mm. In my mind, project would be like Finance, where it's like there's just a bunch of stuff going on, and I want to like sub understand all the pieces. And the channels is interesting because.

I might want to engage with AI tools in different ways. I might have some things that I want to be triggered by, you know, sending an email, some things that I want to be triggered by sending an iMessage, something that I just want to go create a session in a native app.

And so every single one of these things needs to be separate. In OpenClaw, it's like there's one unit and it has a channel and it has skills. But here I might wanna kind of piece it together and say, well, I want a channel where if I message you on Slack. I can send a message on Slack and it will use these three skills. Yeah. And then there's delivery, which is how the output gets. That's right.

You might want to click a button on a dashboard you build to trigger something, but you might want the delivery to be an email. So for me. I might want to go do an award search and I can trigger that by using the Claude app. Like the thing we haven't mentioned is that most of my interaction with this is not with custom tools. It's with like download Claude or Codec.

which is like the desktop app for the two mainstream providers and use those tools on your computer to do all this stuff. Back when I first did this episode, I was like, here's how you can go use the command line and here's how you can go download these packages. And it's like I don't think most people need that anymore to have a really, really both sophisticated and simple version of how to use these tools to actually add value to their lives.

Incredible. Okay. So maybe talk to me. So that i I understand the architecture now. You mentioned the kind of award flight booking example. What are the other aspects of your life you've automated? And maybe give us just a window into what

you think the everyday consumer's life could look like. Because in a sense, you've always been automating your life. You've got a lot more leverage than you've ever had before. But I think one of the magic of these products and tools is that everyone will now have access to the kind of level of effort and sophistication that you've typically put into your own life for many years.

I will go there, but I know you've been thinking about this a lot. So I want to push you to kind of share, share your perspective on where have you been coming at this from? Where are you at? Where do you think we are? Yeah, so okay, in the world of the two religions, I think I've been more about codex clawed code than open claw.

I think open cloud is really, really interesting. And I think I you know, I mentioned this a few weeks ago. I think the ergonomics of it are actually fascinating because it meets you where you are, which is iMessage, Telegram, group chats with your family, which you told me about, WhatsApp messages. You can have this bots chatting with each other within these group threads. So there's just something about the way you actually consume.

the model and the skills and everything else that open class seems to have gotten right. With that said, I've just never been that productivity pilled. I probably run a relatively unproductive personal life, but I've so much more interested in kind of making new things. And that's what the coding agents did for me. With that said, it feels like the two big breakthroughs we've had, there's been many, but the ones that I'm paying the most attention to are

Cody agents in twenty five and personal agents, open claw in twenty twenty six. I've been thinking about this a lot lately, like what are the kind of implications for society and the world? And I think one of the big implications is that

you kind of have this zero marginal cost of doing work now or digital work, right? Just as software I think is very famous and important for a zero marginal cost of distribution. you now have zero marginal cost of doing work because of computer use plus personal agents. So all of these things where apathy or sort of being underinformed about a product area, like finances, is the most classic example.

Where if you sat down with someone and looked at their credit report, or if I did, I'm sure we could identify for the average person a dozen ways they can make improvements without changing their life at all. Like, I'm not saying don't drink the Starbucks. I'm not saying make different financial choices or change your retirement target age. I'm just saying with all of the existing constraints of your life.

Here are a couple of hacks that you can make to actually improve your score and decrease the cost for your credit, et cetera. But it requires motivation and information. And that all requires work. And now that personal agents can do the work on your behalf, something that I think could happen in the near future is that every consumer lives a fully hacked life. You know, it's sort of completely optimal as if you designed it yourself for them.

So I think the one thing I'll push back on is I've been playing the points of miles game for a long time. And I think when I probably got started, it wasn't as mainstream, right? Like not every person went and had a points card and was trying to think about these redemptions. And then they added award search tools, right? And so now the average person doesn't have to go figure out how to log into the Air France website to search there and do this.

So it's like, okay, cool, it's gotten better. However, if you have a United card, you might get a discount. And that's not going to service on an award search tool, at least not today. And you know, I could tell you, for example, one crazy thing is that for domestic united flights.

like just short haul flights. You could book them on United. It's often ten, fifteen thousand points. You could book it on Air India for thirty five hundred points. The award search tools don't search Air India. Air India is hard to even get points in. They're like not a transfer partner of a lot of programs. But they are for United? No no

Star Alliance partner of United. But to get Air India Miles is not easy. But like Rove Miles, it's a hotel booking portal. Like you could and a shopping portal. You could earn some rove miles. You don't have to earn a lot, thirty five hundred. And now all of a sudden you can book this three hundred dollar United flight. for what's like thirty dollars, like ninety percent off. And so

Everyone will always probably continue to be more and more optimized, just as we've gotten more and more productive and more and more healthy and life expectancy and all stuff. But there will still be, I think, as this evolves, knowledge that isn't shared publicly. I wonder if

Knowledge gets less democratized over time. Interesting. Because You know, I don't want to share everything in the world with an LLM. So does everyone start like turning off access to these LLMs and does do conversations about things get more private? 90% of the content that I consume about points and miles and stuff is not public.

in that it's not going to show up in a Google search result. It might be buried in a Reddit comment. It might be on a podcast. It might be in a Teleg or Discord group or a Slack group. And so I think yes, people will be living more optimal lives. But I also wonder if those of us who kind of like doing what I do, maybe not as much as you, like will still find ways to live a even more optimal life.

Well, I guess the question is there's secrets and then there's just things that are a pain to know. You know, if you had to go follow every telegram group chat and listen to every pod, like all of those things are highly doable by agents.

So this is my point about zero marginal cost of doing work. They will do all that work for you. They will discover every secret that is discoverable. Of course there will be secrets that are kind of whispered down shadowy hallways and the agents won't be able to find those because there's just no path to them. But for everything else that is knowable, I think the agents will know them, will operationalize them, and the system will have to change around us.

I'm excited to find out how this evolves. Like I don't know the answer. But actually what you said is something that r has been the kind of crux of a lot of these tools for me. Yeah. Is that There have always been kind of three things. And I'll use health as an example, but I don't want to go too deep because I've already put together an episode entirely on this. But

When I think about my health, there's kind of these three or let's call it four components. I've got all the knowledge. And there's probably two tiers of knowledge. There's like the public knowledge, which if you search health things on the internet, varying degrees of profitity and Yes, yes, yes. But then there's like the scientific journals, there's the medical research, the clinical trials.

Then there's what I'll call like, it's not private knowledge, but it's knowledge that doesn't today surface well in things. Maybe it's podcasts, maybe it's member sites, maybe it's private YouTube channels. I don't know. I'm sure you and I listen to various podcasts.

that aren't private, but the information in them doesn't necessarily show up on the internet when you search on an LLM. Maybe you pay for a course. There's another good example. Sure. So that I have this knowledge. So I have a bunch of places I like to get knowledge on every topic. Health, this, that, and the other. Then I have all of my personal data.

And that personal data for health is pretty obvious. It's my biomarkers, it's my PR nuvo scan, it's my doctor's appointment records and all that kind of stuff. And then separately I have where do I go to ask questions? And most of the time I'm going to Chat GPT or Claude or Gemini and I'm like, hey, what should I do? that has no knowledge of like, here's the people I trust to give me information and where I would get it.

And then here's everything about me. And so what I think we can start to build in every area we care about in life is tying these things together. And I don't know if they get tied together publicly and maybe they just get tied together by us, but that's where I think everything gets really fascinating. This is true, you know, I say health as an example because I've done a lot of health tests.

I've crawled a bunch of health data and built a little database of all this health information. And then I have all of my stuff. And then I can go ask questions and it hits those and public. And that's where it starts to get magical. And that's what's not

obvious today. A lot of people, my mom, like, oh, let me ask. I'm my stomach hurts. I ate this. What should I do? But it doesn't know your genome and it doesn't know that maybe you trust Andrew Huberman more than anyone in the world. And so you want his advice or whoever your medical expert is.

Chris, okay, so here's the real black pill in all of this. My counterpoint to actually the true sort of fully optimized consumer in the near future thanks to this technology. The black pill is if you look at what we've done for the last two years is we made it really easy for anyone to make a movie or a creative artifact.

and anyone to create a piece of code or software or anything they wanted to make. And Nikita said this and I think he's right, which is like funny enough, two years later, the exact same people who are doing all these things are still doing them except with the new tools and technology. So perhaps you're one of the most high-agency people I've ever met. You're probably top five in my world, and yet you continue to be high agency with these new tools and technologies.

does the mass sort of public actually have enough agency, interest, curiosity, motivation, even if the tools are right there to operationalize all of this? I don't know I think that Where this kind of takes a turning point is when the tools can start to both do things less expensively, we should talk about the cost, but actually prompt it. So right now, if you go to Chat GPT,

And you say, Hey, my stomach hurts. What might it be? It's like, oh, what did you eat last night? It doesn't say like, hey, here's a connector to connect to Apple Health so I can pull in that data. Who's your doctor? Can I email them and get your medical records?

It doesn't say, hey, you know, if we really want to do this, we should track what you eat. Can I build you an app to do that that will live on your phone? And by the way, that's actually probably an easier problem to solve than many other things. product problem, yes. It's literally like if these tools just and by the way, it's like, Oh, we're halfway there. Do you want to upgrade your plan so we can finish building like it it seems like a great thing for all these people to build.

But that doesn't happen yet. And so today someone's like, Hey Chat GPT, how do I use all my points? Yeah. And it's like, well, you should search for flights. Yeah. Not. Yeah. Let us evaluate everything you need to do let me connect to everything you have. It's great. And so I think once the connecting happens, it's really magical. So I'll give you one example. I built this app called Card Tool. It manages all your credit cards.

And the other day I was having a conversation about paying taxes with credit cards with Claude. Oh my god. Now Claude is tied into card tool. Okay. And so when I went in and I was like asking a question, it was like, hold on, let me just go look at what cards you have and what their available credit limits are. I didn't think Let me go open this app that I've built and collect this data. But because I'd connected Card Tool to Claude, it just knew

That's a tool available to me. Let me go check the tool and let me give you a better answer because I have the data. Now I've taken the time to connect all these things. That's like an onboarding thing. And it's a challenge that most of the things out in the world don't yet have those connectors.

But I have to imagine in I don't know, months to years, every single consumer product, service, data store, et cetera, is gonna have some way to connect to every single agent tool. And then it can start to do some awesome stuff. What is so beautiful about what you're describing, what's r got me thinking about is the fact that

Okay, maybe a subset of people in the world have really high agency. And actually Sam and Elon are maybe the two most high agency people in the world. And for them, funny enough, the chat box is actually the ideal interface because they just say what they want. But perhaps for much of the rest of the world, the agency has to be sort of externalized and the model has to be like, you know, a bit of a a trainer. Like, hey.

Do you want this really good thing to happen? I will just do it for you. Let me like push you to like connect all of your data to do this thing. Let me phone your doctor on your behalf. Like I don't even have to be high agency for that. I kind of just have to let it wash over me and for all these good things to happen in my life. It's a really interesting idea. This episode is brought to you by Netsuite. Every business is asking the same question how do we make AI work for us?

The possibilities are endless and guessing is too risky. But sitting on the sidelines is not an option because one thing is almost certain, your competitors are already making their move. No more waiting. With NetSweep by Oracle, you can put AI to work today.

I'm using AI across my business, and the hard part is knowing which tools hold up when everyone's shipping something new. That's what I like about NetSuite. It isn't another bolted-on tool. It's AI built into the system that runs your business. The number one AI Cloud ERP trusted by over 43,000 businesses, bringing your financials, inventory, commerce, HR, and CRM into a single source of truth.

From software and IT services to healthcare, equipment manufacturing, financial services, and many other great American industries, NetSuite delivers a customized solution for your business. If your revenues are at least in the seven figures, get our free business guide, demystifying AI, at netsuite.com slash hacks. The guide is free to you at netsuite.com slash hacks, netsuite.com slash hacks.

This episode is brought to you by Mercury, radically different banking now available for personal accounts. I've been using Mercury Personal for two years now, and I love it. It's fast, elegant, and designed to help you do everything you want with your money all in one place.

You can send wires with no fees, set up sub-accounts for your goals and projects, put your money to work with high-yield savings, and set automations to move money between accounts so everything is funded without you having to think about it. You can also spin up cards with custom limits for an assistant or a nanny. You can grant read-only access to your accountant, and manage shared finances with joint accounts built for clarity and control.

All of this for one all-inclusive subscription of$240 a year. Visit mercury.com slash personal to learn more. Again, that's Mercury.com slash personal. Mercury is a fintech company, not an FDIC insured bank. Banking services provided through Choice Financial Group and Column NA, members FDIC.

I don't think it's prescriptive enough and proactive enough. And that was like the biggest complaint when all these agents started. It's like I want you to do work for me. Well, the reason is like it doesn't think.

Like it's not a human. And so over time it will start to get better. Oh, let's spin this up and see what are the things this person is accomplishing and what can I do. If you said, I want to be healthier, it's like, here's a fitness plan. Why doesn't it text you the next day? And that comes back to

From a channeling standpoint, I think you have to do it yourself, but you can, right? So I stopped using OpenClaw, but I still use iMessage. And so you can both use out-of-the-box channels for Claude or you can build your own channels.

And one of the things that I think this goes back to like you have to want to do it. But I literally was like, Well, OpenClaw works with iMessage. I want Claude to work with iMessage. Claude has this channels feature, so you can work on channels. So I literally just took the open claw code. And I gave it to Claude Code. I was like, build a channel that does exactly what this other thing does. And in like 20 minutes, I'm I'm messaging an instance of Cloud Code. It just works.

But you have to have that thought. That's the thing I think a lot of people are missing. It's like, I don't want this thing to work. Can I fix it? Well, here's a great example. A lot of people know that I've done tons of research on random stuff. So people call me all the time and they ask me questions. Now they could go answer these questions somewhere else, right? Like, but they just know that I'm this person that's like probably answered it. Yeah. And you've sent me these text. HumanChatGPT

It's beautiful. Once people realize, oh, that's how it works, like their mind will be trained to be like, oh, let me just ask this question. Let me do this thing. And I've seen it happen. And I'm gonna give you an example because you're gonna actually tell me how it turned out. This woman asked me, can I have it reorganize all my photos? Can I like do all this stuff?

And I wasn't sure if it could. She wasn't someone who uses these tools like I do, but she had the idea when we were talking. And once more of those ideas happen and you see the output, then you'll just start to train yourself. Oh, I can share those ideas. I can get comfortable with that. And you did it.

I did. Well you should describe what the prompt was or I will if you like. Go for it. Yeah, so the prompt essentially was and you're just talking about your show with Kevin, which I thought was terrific. The prompt essentially was, Hey, can I have Quad or Codex go and look at all my iMessages, all my photos?

and put together a video montage that sort of interleaves and puts together, you know, sweet messages with my partner, photos of my family, and puts it all into this nice finished polished artifact.

And you were saying on the show, oh, I don't know, that might be a little beyond its capabilities. And then you tried it and it worked. So I too tried it yesterday after being inspired by your pod and it was incredible. It like selected A plus photos. It found all of these old iMessage. You know, I'm sure you've got

Thousands of messages with Amy. I've got thousands with Lindsay, and it went all the way to the bottom to find the very sweet and endearing ones. You know, one from Lindsay that said, Hey, it's true, I'm pregnant, things are so great right now. Just something that was such a nice digital artifact of our relationship that I never would have found otherwise. So it found messages and photos.

Yes. And it selected great photos. I wanted to put the music on myself, but I said, Hey, go and dig into my music library and give me ten ideas for tracks that might be cool and it found this remix of a Drake track, you know, sorry to all the Kendrick people.

that worked great and I threw it on there and I am so excited to show it to her. I actually showed it to the kids last night and they're they're super excited too. So it's just amazing what's possible. It's something that's richer and I think more touching than anything I would have made myself. You know, I think of my daughter, right? She's so young and she loves creating stuff. It's like that creativity that we probably don't have, we've kind of lost it as adults.

And I think that these tools let us do that in a way that we're probably even more comfortable with than like working with wood or working with paper. And I'm just excited for it to be retrained. I will give a kind of caution because when I think of the task you just described, it reminds me of why things often break in these things. It's like

If you describe everything you want all at once in one big blob, sometimes I found errors and problems. And so I'm gonna encourage people when they're trying to be creative. to do this like reverse prompt where it's like, well, I'm thinking about this, how should I structure it? It's like using the tool to ask how to use the tool. And I always find that when I do that, I get much better output.

And so if you want, you can send the exact prompt you used. I'll share it with people. Thinking carefully about the prompt is something that I think we probably won't have to do as much in the future, but right now it's like the difference between good and bad output.

Well it's also very interesting because I think for a lot of creative fields, people are unable to articulate what they want to see or hear or what they like. If I told you, hey, tell me what your favorite track sounds like, what does it really sound like? You'd be like, I don't know. It's cool, it's high energy, it's moody, like you'd have all these words that we're just grasping.

at what the song was or what the genre was or what the sound was and yet if you heard it you'd be like I know it when I hear it. So you should actually be able to put things like that into the context window, which you can to some extent, but In the future, I don't think you'll have to have the language as sort of like they're the keys to unlock all these incredible artifacts you want to see, and hopefully there'll be a lower bar for that.

I find sometimes now, if I want to do something in the style of something else, I like upload a photo and I'm like describe this photo. It's a great technique, yes. Then I say, Now create a photo that sounds like this. You know, and I find that when you break things into discrete tasks, the quality is so much better. And so I I did a whole session with Codex because I ran out of my tokens on why.

And you know, it doesn't have episodic memory. So it's not like I did this and then I did this and I it just kinda like does it all at once. Prioritization is tough. So they've solved these with like plan mode. But it's like so driven on software development that I think once the tools evolve to be better for human tasks. it's gonna get a lot better. But the thing that I learned building this kind of replacement for one of the processes the this person on my team who's no longer gonna be here.

is doing is if you want something to collect information and then rank it and then draft it. Oftentimes my instinct is like, here's 10 sources, go get this and try to figure out the best and do this and draft it. And then I was like, okay, go get the information.

then give me a summary of the information. Now run another task, take the summary of the information, and then go and rank it, and then take the ranked summary and draft it. And when you break it into discrete tasks, Maybe future models won't need this, but at present, it's wild how much better the response is when you're like not trying to have one session do everything.

So since we're on this topic about family relationships, kids, we should get to that as well. You know, my question for you is I understand what it looks like conceptually to have an optimized life, finance, health, professional perhaps.

Do you think that this technology allows us to be more or less connected to our partners, to our children, to the world, to each other? Are we more or less connected than we were three years ago? Like maybe give me the philosophical and the practical, because you've lived both. I'm really

torn on my answer to this question because I was trying to figure out how to better communicate with someone that I work with. We both took this personality test and I sent the transcript of everything we've done and our meetings and I was like, what could we do better?

And sometimes it just turns out that like certain managers have a style of work and certain employees have a style of work. We were the most incompatible possible. Right. Like, so that's not great. I was like, hmm. I wonder if I could write an agent where it's like, imagine this person's name is

Sarah. I have like fake Sarah and I DM fake Sarah on Slack. And then it translates everything I say into the version of it that I wish that I could say, but I don't have like the ability in the moment to say it in the way that it would be best heard. And then I was like, gosh. There could be a world where like I never even have a direct relationship with this person, but we kind of feel like we do.

If that world comes to fruition, then it's like our relationships are definitely not closer. So what I've been trying to do instead of that is every time I have a conversation with someone, I'm like, I send the transcript of the conversation. I'm like, tell me how I could be better at communicating. Tell me how I could improve this. And does it work?

I think it does. Like I think I've gotten to be a better manager. I'm gonna go deep on what I've learned doing this because I'm recording like everything that happens. And so I could have. all my personality, I could have all the conversations with my wife and all the texts we send and every month I could say, Hey, go look at everything we're both doing, talking, saying. What could we do to be better? So

That would probably make our time together better, which I think is great. Like I want to have a more exciting relationship with people. Okay, two things put together, and this is such a perfect microcosm of the entire debate around technology and AI, which is you have an abstract concern that perhaps if Virtual Sarah talks to virtual Chris, you never interact. That would be terrible. Meanwhile on planet Earth, you actually are communicating with Sarah

much more effectively thanks to this technology. You know, like there's this abstraction of a fear, which maybe it could exist, but like it's just not what's happening. It's not how you're experiencing the technology in the world. And honestly, I'm thinking about how does this affect our kids? Like what are how are they gonna communicate? What are they gonna do? There's a lot to unpack there. This is not gonna be that unpacking, but Tell us everything.

I think that I've been able to improve so many things and be more efficient. And ultimately, I think I'm probably a little bit too much of a spreadsheet productivity junkie to stop. But I think most people don't do what I do at all. At all. And so if they can free up some amount of their time each week, what can they do with that? I have a long list of things that I want to do with my free time that I don't have.

And I've been able to make meaningful improvements in personal things, not just work. I feel like we always talk about like, let's build a software to do this thing, but it's like spending time planning a trip. Spending time meal planning, spending time like all of those things. If you can outsource a lot of those tasks. Then

we have more time. Right? Like I could spend more time just hanging out with my kids. We could spend more time on the weekend just going to the park. I mean most most of the weekends off for AI, but can work in both directions, you know. So for example, maybe a future form of passive income. The same like way like I call Chris to get his quick advice on taxes, there's like the Chris tax skill.

which I can hire to look at my taxes and do a bunch of work and give me a bunch of feedback that I wouldn't otherwise get. And you've like written it, you maintain it. So perhaps there's a way to extend the one thing in the world that we're most opinionated at into this technology in a way that financially benefits us.

I would love to be able to have some version of my lens on certain things to give to other people. And then I would love to have your lens on certain things. Like imagine I'm like reading this article. I'm like, what would a niche?

So can we do that? In the bigger world, how do you get compensated for doing that? Because in a sense, that is your specialization. So you don't want to give it away for free. And I think a lot of people would pay to hire virtual Chris and it would cost one one thousandth of what it would cost to get you on the phone, perhaps, right? Right. That's a cool a cool world. Because I think we all have

different like random specialties. Yes. Like I might want my father in law to teach me this like strategy in a card game that he's like played for fifty years. Yeah. I could just go spend time with him. But like conceptually, like

You know, if someone else came to me and they're like, I'd love to learn how to do be a better cribbage player, I'm like, Oh, you should hang out with my father in law, but that's never gonna happen. But like he's got a little like, you know, you could basically virtually play with him and learn.

And it's this thing I've thought thought a lot about this, which is that there's intellectual knowledge, like the stuff that you like read in books and maybe read blog posts about, like knowledge that's quote unquote important with a capital I. Then there's tacit knowledge. Tacit knowledge is just the knack for getting things done. But the majority of the knowledge in the world is

When you call customer support at United versus Air India, maybe you sort of know what the transcript to talk to each is to get exactly what you want or what the tone or what the bedside. There's just so much information that is never explicitly expressed. that could now be captured and be used to benefit you and to benefit others. I think that's really beautiful.

You gave me this example or someone did about how could everyone just start to pay all these bills or fight all these bills? If you get medical bills, if you want to negotiate them, or if you want to submit appeals, you can do all these processes.

So stage one is gonna be everyone can just do that. You get a speeding ticket. Do you remember where it's like hire a person and they'll like write the defense? It's like stage one is I wanna fight my parking ticket, just go do this for me. Stage two though is like what happens next when

all of a sudden there's just every bill is getting fought. Like I don't know how that evolves because you're seeing this in software. Like the number of people contributing to software is crazy. And so I'm not a software developer, but I was using this one app as a teleprompter app. It's like a Mac app.

And I'd route all my audio through this roadcaster. It's like a mixer that we're using today. It ultimately means that the audio device on my computer has multiple channels and this app didn't work with multiple channels. So I was like, well let's download the source code for this thing, let's tweak it to do this, and then I was like, let me submit a fix.

that got accepted and now it's in the final app. And that's cool. But then I noticed I looked, I was like, there were like a hundred people submitting these things. And so it's gonna be overwhelming when the insurance company has eight million claims and appeals and fights.

I don't know how that gets solved, but in the short term, I feel like anything that you're like, Ooh, would this be worth it? We always have this question, what's my time worth? Is this worth my time to go fight this thing, to go do this thing? Yes. Well now You don't have to ask that question. That's exactly and I'll tell you a funny side story and then we should talk more about this, which is

Brazil is a country in which it's the easiest to sue anyone in the world by a long shot. So in Brazil, you can sue people digitally and people sue companies like they sue the Brazilian DoorDash because their order was late. So companies get sued thousands of times per day for otherwise trivial things.

And so in a sense, you already have this DDoS of sort of agents in the form of lawsuits, and they of course then have their own technology, but they've had to change the system on both sides to handle this sort of emergent behavior. And I think it would be beautiful if, for example, we had to change I don't know, the DMV.

and the way that we handle adjudication of parking tickets and all of these other systems which we assumed were so ossified. Like if you look at how society works, you know, I don't know too political, but you might say that the super rich get services and all the rest of us get products.

You know, where they actually have a team that manages things and makes calls and fights on their behalf and they never have to make the labor trade-off. But now with these technologies, nobody does. Everybody's got a team and it's incredible. So I'm gonna just pause. And we talked a lot about how I've set things up. Yes. And like tools, but practically, like if someone's listening to this and they're like, I haven't dove deep into this, you know. I said my advice is

to continue doing what you're doing, you know, use ChatGPT or something in the web. But I think it's really interesting when you start using one of these platforms and letting it talk to other things. My advice to start to play with this is because we didn't get into the like tactical, is Unfortunately, I feel like Gemini is a tool that I know a lot of people I've talked to are using. They have chosen to not integrate much beyond the Google E. Do they not have the M C P connectors?

So Google Gemini on the command line, which is like opening up the terminal on your computer does. But the consumer grade version of it, which is like on your browser, maybe on your phone, they've built like three or four connectors. But I see and there's no n marketplace or OpenAI has apps and Claude has connectors. And I went deep on both of them because I knew we were gonna talk about this. I'm disappointed with

OpenAI's version because they treat it like opt-in of the connectors. So I'm having a chat and I'm like, let's also include Notion in this chat. And the problem I have is that oftentimes like I don't know what I wanna do when I first start. And then in order to add custom ones, you have to switch into developer mode on Open AI, which anyone can do. But once you add this custom one, you switch into developer mode, there's no memory, there's nothing.

I have gotten so much value out of the memory in Claude and OpenAI in ChatGPT because I was like, oh, I'm thinking of doing an episode about Japan, which by this time this comes out will have come out. And I was like, what are some of the things I should include in this? And it was like I don't even know why I asked it. Cause in my mind I was like, well, it doesn't know what we did, but it does because every day I would pop in and be like,

If I'm looking for shoes, what are some great places to get shoes in Japan? It's like where should we go for X, Y, and Z? So it actually had all this context, and you lose all of that once you enable these custom tools. So if you can live in a world where you're using the default tools,

then both of them are pretty comparable. But if you want to use any custom tools, which might be something you've built or just something that exists, I really like the card pointers app. It basically aggregates all the Amex card offers, chase offers. They've got an MCP server. You can connect it. But they're not like in the marketplace. So you have to use it as a custom setup. That doesn't mean that it's like more risky. It just means they haven't like been approved to be in this marketplace.

So I think Claude gets the win there for just making the connectors available. So I would use one of these platforms. Most of the services you'd want to use are probably already in there until you start building your own. And if you don't, The thing that I haven't heard anyone talk about, and I'm sure I'm not the first person to talk about this because they big company is what Zapier has built. Yeah.

Tchau. So Zapier built this new Well, yeah. So they built this thing with a Zapier M C P server. And the reason why it's really cool is they are in the marketplace. So you don't have to do anything custom. But then you can go connect all these random custom things to it and they are in the business of that. So if you're like, oh, I use this one obscure thing.

There's a decent chance that Zapier's like 8,000 connectors work with it. And so instead of having to go and like wait for Company X to go build something specific for these tools, You could connect it to Zapier. And for now, you know, you're gonna pay 10, 20,$30 a month for hundreds of calls, which you probably won't need until it's maybe saving you so much time you're willing to pay it. But until then, it's like a great stopgap for I wanna do a thing.

And I can't quite do it and I don't want to go code it. I built this website for a friend that was like, how to get started and it was kind of like a fun, interactive way to get started. And he was like, All I wanna do is build a tool that can check my email every day and label everything that meets this criteria. And I was like, Oh, this is the easiest tool you could build. And now with Claude has routines and scheduled tasks.

And so a lot of what OpenClaw could do, you could just do natively in the app. You don't need to code anything. You don't need cron jobs. It doesn't get messy. And you can go in and say, like, oh, every day I want you to research this thing and send me a briefing. And I actually tested this. Like you can do a cloud

Claude code briefing every day. Now, what Claude can't do is send you an email. So you can go into Gmail and say send me an email, but it can't. The default Gmail connector drafts. It doesn't Only draft. Yes. But Zapier's does. Oh and so Zapier can plug Gmail in to send emails. And if you're gonna send an email every day, that's thirty, thirty-one emails a month. It's not gonna cost that much. You know, like so there are a lot of ways that you can extend without having to build. Now.

I wanna push everyone to go build because it's fun and you could do it, but it's just amazing that if you download these apps. They now can control your computer. They can go browse the web. You know, you want to go build a tool that searches for hotels? Great. Easy. This episode is brought to you by Fora.

Imagine if your love of planning trips could actually become a source of income. Your years of comparing hotels and putting together perfect itineraries for friends turning into a real business on your term. That's exactly what I've been doing. I am a travel advisor, and full disclosure, I'm also an investor in Fora, because Fora makes turning travel planning into income actually simple.

It's a modern travel agency that gives you the training, the booking tech, and a global community of advisors all in one subscription. With over 7,000 preferred travel partners, the trips you book come with upgrades, resort credits, and late checkouts. Every hotel, cruise, or activity you book earns a commission, and you can build it on your own terms, alongside your job, your family, or whatever else you've got going on. Become a fora advisor today at foratravel.com/slash all the hack.

That's f o-r-a-travel.com slash all the hacks. And make sure to tell them we sent you. foratravel.com slash all the hacks. Thank you for being here today. You can find all the links, promo codes, and discounts from all our partners at all the hacks.com slash deals. They're all brands I love and use, so please consider supporting those who support us.

Let me ask you a tactical question about this. How do you manage logins? Like today, I was downloading all my health data from Stanford. And you know, you've got to log in on behalf of Codex so that it can then operate the browser on behalf of you. And then every time you want to refresh it, you have to go log in again on behalf of it. So how do you sort of manage passwords logins, especially if you're a most sensitive data source is your health?

A few options. I was thinking about this a lot because there's a few different ways to get this health data, some of which are not available to consumers. But are available to medical companies and Apple Health. So we'll get to that. So one option is just one password. So I have one password set up.

and it's on the computer. I have one vault that the computer has access to. It's called like a service worker account. It's actually free. You don't have to add a user to your account. So I have a service worker account that has read access to a vault. And that vault anything I put in there, then my Claude Code session or my Claude Cowork session or my codec session, like anything can do it. I think the the beautiful thing that I didn't mention earlier.

The most valuable thing I recently did is every time I start working on anything that's going to create anything, skills, tools, I start it with build this. So that any agent coding platform can use it. And so when I built this award tool thing, I built it in Codex and then I immediately tested it in Claude Code and it worked. But it worked.

because I told it. Because if you don't, it's gonna say use the function X to go use a browser. That is a codec specific function. And so for health data specifically, you kind of have a few options. You can download all the data yourself. for the historical and maybe rely on other things for upcoming. I don't love that. A lot of these sites have logins. You can just put your login into one password and let it do it in the browser. Or Apple Health syncs to all of these platforms.

And so I I have something that's like it's so silly, but there is a shared Dropbox folder that is on my computer that has access to all my health stuff. And then in Apple Health you can export. And you could even there was a shortcut related thing you could do, but you could just go into Apple Health and export it once a month and it pulls from a lot of the records in all these other places. So you could use Apple Health as the hub.

You could, and this is like sounds daunting to most people, but it probably isn't, you could build an iOS app that literally all it does is runs on your phone and syncs with Apple Health. with the one permission that it just gets access to all your health data. So that's an option. For when I collected this, I I kind of did it a little bit more manually out the gate, but you could choose how much you want to give it access to.

Can it log into your Stanford My Health account? Like, is that that big of a deal? I don't know. You gotta decide. I mean I'd be fine with it. I just don't want to give it access to my primary one password vault. Yes. So I have a service where like a Yeah. Britt Vault. With a separate service worker account.

And so the computer can access that vault and it can only read it, can't write it, so it's not gonna override something. And that's changed a lot of the problems of like, oh, I need an API key for something, put it there. Everything can find it. Okay.

So now let me ask you a different question. Take me to the edge. What's not possible today that will be possible in three or six months, presumably? Like what do you wish you could do that you can't? Or do you feel like you're just, you know, now you're looking for areas in which to apply this stuff? What's not possible is is to not waste time because these albums make so many mistakes. Okay. Like like last night

One of the things we haven't talked at all about is how to manage this like on the go. If I gave people a lot of advice, install codex, install Claude. And then like what do you do if you have a question when you're on the road or something like that? Claude has a cool feature, remote control, that's like

better than it was when it launched, not perfect, but you can start a thing on your computer and like control it remotely. Then they build a thing called dispatch where you can actually just like tell your computer to start doing a thing remotely. At the end of the day it requires some persistent device on.

That's gonna be an offering of all these tools in some period of time where it's like if you want a thing that can run all the time, like you can have a Google Cloud Server, you can have a Cloud Cloud Server. You've got your codecs connected to remote terminal sessions as well. Right. So this is where I was like, I can't believe this happened. So I went to Chat GPT and I was like, I really love how if I'm using clawed code on my computer.

I can go to my laptop when I'm not home and control those sessions. Or I can go to my phone and control those sessions. How do I do that in codecs? Because I hit my quota. I'm switching to codex for the rest of the week. And it was like there's no good way to do it. And I was like, Here are three articles about other people doing it. It's like, well, there are some ways to do it.

But maybe they're not the way you wanted. Here's the limitations. Here's the problem. So what I had done was and this is probably a little bit farther than most people need to go, but once you have a like a Mac mini or any s computer setup or any desktop computer. That can run all day. If you're on another computer, I find it's better to just operate on that. And so why it's not enabled by default, I don't know. But there's this thing called remote connections on the code. In codecs.

That you have to add to this config file. And if you just Google like enable remote connections, it sounds like it's a beta thing that you can enable. And then it lets you do what's called SSH, for people who don't know, where you can connect to another computer and run on that computer. So I did that and I was like, I'm doing that. I would like to control that from another computer. And it's like you can't do it.

And I spent twenty minutes arguing like there's gotta be a way. Like just fundamentally, this should work. But I gave I gave up. I gave up and I was like, you know what? I'm just gonna create a new session from the other computer.

So I go to my laptop, go to the config, I enable remote connections, and then I put in all the details to connect to the computer, the Mac mini in my house, and then it's like, what working directory do you want? And I chose this directory that's like my agent directory, the former OpenClaw directory. I choose that directory. And then all the sessions just show up. Like they're just there. So one of the challenges, the things that it's tough is

this whole context thing. I even told it to research it and it couldn't figure it out. So I don't even know if that could have been solved. But a lot of times you'll ask a question like, Can you do this thing? It's like, well, I don't know. Like it doesn't know it can do the thing because it's not in its context or in its model. And so I think what I can't do now is operate in a world where models are like able to figure things out. And so you've got to give it so much instruction.

And so it's like if you think of when you've hired people, it's like when you first hire someone, you can't expect them to do the work. You have to show them how to do their job. These models can do things that employees I've hired have never been able to do. But sometimes at the fundamental level they can't. And so they need a lot of instruction.

And so when they're able to figure more out on their own through trial and error, and I think we're getting there. Like the browser use stuff where it's like, oh, I can't figure it out. Let me look. Let me take a screenshot. Let me see if it's there. Let me move it around. We're getting there. It's just getting better every week. You know, it's interesting. So Aaron Levy talked about this on X last week. Do you know what the Gelman amnesia effect is? Have you ever heard of this?

So I'll I'll describe it to you. So the gentleman amnesia effect essentially is when you read an article in the newspaper about a topic that you really understand and you're like, this isn't right at all. And then the next moment you read an article about a topic you don't understand at all and you assume it's a hundred percent accurate.

So I think there's a similar thing with the models where if you use the models for something that you're sophisticated at, you're like, uh, it's like 80% but not quite there. And then when you use the model for something you're not sophisticated at, you're like, oh my God, Profession X is totally cooked. So terrifying example was like in the health stuff. If it gets something wrong and it sends you down a rabbit hole of like you have this Gonna know.

You're not gonna know. And so that is, I think, the thing that I can't do is I can't rely on it. And so what I've realized though is If you change your framing from like, this is an agent that's going to go do this stuff to this is something that's going to take away the repetitive work I can do.

Then it becomes really, really interesting. If every day you want it to go check the snow report in three places and send you a message, every day you want it to go look if this thing's in stock, or every five minutes, go look if this thing you want to buy is in stock.

Those kinds of things that are really redundant and repetitive, great. For the more complicated stuff that matters, like there is still a component. Earlier I talked about this health thing. It's like you've got my data, you've got the knowledge that I like, you've got my questions, and then you've got a like

person that you can trust to help you until the models are actually at a place where they can make those decisions. It's like, I still have a doctor. What I want to do is combine all of that and then take it to someone who can sanity check it.

And the models are getting better at sanity checking, but you can't just rely like you need to ask them. How many times have you said, are you sure that's right? And gotten no, no, no. It's totally, I'm so sorry. It was totally wrong. Like that happens all the time. And so I find that it's better at just repeating. And this is why just the theoretical versus the practical, right? In practice, you still need your doctor.

Yes. And I think that doctor will hopefully not need to do as much. That's right. I was thinking about this when our daughter got sick. We went to the doctor and the only thing we needed to be in the doctor's office for. Was just for someone to put a stethoscope on and listen to her lungs. And so I said, I Google, I was like, there's got to be electric stethoscope because we have an

Otoscope, which is like goes in your ear, in your nose or something. Like we use the o ear one. We take a video, we send it to a doctor online so that we don't have to find out if she has an ear infection. I've now done this enough that I'm like, I now know what an ear infection looks like, you know? Like I don't even need to send it in. Not terrifying. But I was like, there has to be like a remote stethoscope. So like a lot of those things, great.

And then on one hand, looking at the full context of a person is something that like a doctor can't even do. Right. Like we talked about how many things you can keep in memory. Like a doctor can't keep your genome and all this stuff in memory, but a doctor also can do a lot that an AI can't.

and can sanity check it with lived experience because they've got that, what, you know, 2.5 petabytes of knowledge in their head about how medicine works. And until we have a what would it be, whatever 12 million times a million context window is? We're talking trillions of token condex window. Like it's just not going to have the ability to process on lived experience. Yeah, or a tiny context window with much better tool use.

Yes. One thing that I think's also worth talking about briefly is like the cost of this stuff. So right now there's basically the free version, the twenty dollar version, the hundred dollar version, and the twenty dollar version. And then there's the pay for the tokens version. And almost every consumer that I know is using one of these monthly plans. And that might be the free one. Like they might just go to Gemini, ask questions, and that's fine. I think it seems like

We're at a point where these companies are starting to realize that they can't just lose as much money as they are. And so for people that aren't aware, Like if you fully max out your$200 a month plan, my best guess is that you're using roughly 10 times as much usage as you're paying for. And I know this for a couple of reasons.

When I was doing this codex thing, I finally got it working and I was like, oh, this is awesome. But I didn't realize that when it connects to the remote machine, it uses whatever login is there, which didn't exist. So it used an API key. And for like 30 minutes. I was not using my$20 a month Chat GPT plan. I was using the API and I used like$100 in 20 minutes.

And I was like, wait a second, I've done tasks like this. It wouldn't have even used my twenty dollars. And instead I spent a hundred dollars. There's definitely subsidization that's happening. There just is. You know, I mean, there's gonna be a lot with token economics. It'll be really interesting. So, one, as you're pointing out, I think.

Some of the big labs are subsidizing, many of the big labs are subsidizing coding tokens in particular. Then I think the second thing that's gonna be interesting is just like what is the value of an incremental token? So a token that you use to click a button on a web page, probably not that high value, the token that you use to unlock some tax insight that saved you a hundred thousand dollars is hugely valuable.

And yet both of those tokens cost you the same amount. And then there's also, by the way, other things coming like local inference. So what happens? Maybe we're all gonna have a big GPU rack in our closet. And there'll be a subset of queries that get routed there and it costs us the same as electricity. So there's a lot that, you know, well let's see. Let's see. I only say this because I have two neighbors.

one side of me and across the street. And we're all like nerding out about this stuff. And we were like, should we just buy like two massive studio Macs? Yeah. And then we'll have like our neighborhood local computer and we can actually like hardwire to it. Yeah. And so we ran the numbers and like The break even is like

years away. Unless you are constantly using these things, it's years away. And with the subsidies, it's also years away. If the subsidies go away, it becomes interesting. And I think we have to realize something that I think is really funny. So right now, On most weeks, I never hit the max on my$200 a month plan.

And so every question I'm asking, I'm asking the most expensive best models. So right now it's like I'm asking Opus 47, I'm asking GPT-55. And You know, as much as I've argued that this is not humans, if we equate that back, it's like if I needed someone to wash my car

I'm not gonna call my neurosurgeon to come in and wash my car. But effectively, like I really struggle with like what tasks do I wanna use though? Like I don't wanna use the dumb model to like do anything yet, but we're gonna have to.

It's a funny human quirk. So I use a bunch of the browser plugins. So for example, when I'm using the Claude One, I never want to use Haiku. So I'm like, oh God, that's the dumb model. You know, even though haiku is far more sophisticated than three seven from a year ago, which was the most cutting edge expensive model. So cool.

But something about using the third best model just like irks me. I'm like, I'm not a third best model guy, I'm a first best model guy, you know? Agreed, that needs to get reconciled. But when doing what I'm doing goes from costing two hundred dollars a month to two thousand dollars a month, I've got to make trade offs. And I think as someone who's thought about this a lot,

The question I have not found a good answer to is what is a good practice for model routing? And I think someone's gonna have to solve this. It's gonna be like when I ask a question, it's like, we're gonna infer what level of effort you want to take. Well, Cursor and others try to do this already. Like I think there are many that are thinking about this routing problem. Also think But it's not a consumer problem yet. Because all the consumers are on some subscription.

They're getting subsidized and they don't really care until they start using it like an enterprise. Slight pushback I'll give you is that I think today we think of these things within our existing technology budgets, just like I think many enterprises are like, okay, the token budget goes in IT.

But the truth is that these tokens are going to start to eat into our spend in every other part of our life. It's going to eat into your healthcare budget, into your parenting budget, into your finance budget. So I do think the ceiling on how many sort of dollars of tokens will be consumed by consumers is going to go up a lot, but also on a pro token basis, consumers are going to need to get more efficient at some point.

It's crazy because I talk to my parents regularly about random things. And my mom, when it comes to how much money she's willing to spend on, I don't know, a meal or a flight or a hotel. It's a lot. And then I'm like, you should try this app. It's like$3. No. Why would I pay$3 for an app? That's crazy. That's right. And so I think we struggle as consumers. Businesses are really good at this.

Consumers are really bad at paying for time. It's like, oh, I want to hire a lawyer, I wanna hire a doctor, I'm gonna pay three hundred dollars to ask a question. And that's gonna get harder with these tools because you're like, well, I get the like a pretty good answer or do I wanna like the delta between the answer you get and the cost is is a lot.

I think we're gonna have to realize that as human tasks get replaced by tools, we should pay for them. And I had this moment the other day where I was like, gosh,$200 a month. So I ran out of tokens. For the week. Not even the week. I ran out tokens for the week. And so I upgraded my Chat GPT from the$20 to the$200. As you should. So I could keep going. And I was like, this is crazy. I'm spending$400 a month. And why am I spending$400 a month?

I am massively trying to automate all of these things this human did that cost thousands of dollars a month. And so I had this moment and it was easier because it was more of a business thing, right? Like I was replacing a cost. with something versus buying time. And I think buying time is harder for us.

It's not just buying time. Kevin and I were talking about this. Like I think entertainment budgets, you know, so how many of the things are you building because they're fun to work on versus they're important to work on? And sometimes it's hard to tell, but I've definitely built a bunch of things that are just fun to work on.

And if those things had to eat into my bottle of wine budget or my movie night budget, that would be fine because it's entertainment. So I just think that every dollar of sort of disposable income on the consumer in the consumer's budget is sort of up for grabs, at least on the digital side. Models might even eat into your restaurant budget at some point. We don't have a good mental model for how much value these things are going to create and how much we're going to pay for them in the future.

Yeah. I think that is something we're gonna have to figure out. Until then, I'm glad that they're being subsidized as consumers. Yes. Amex and Chat GPT. They've got this like three hundred dollar a year credit. Really? So if you have a business platinum or a business gold, they have announced it. Part of the reason I hadn't upgraded my stupid Chat GPT is because I was waiting to use it because it was like it's been announced. They said it's coming. It hasn't come out.

So you get three hundred bucks a year for every Amex platinum business platinum or business gold. It's a business thing. So I don't have to hit on this in another episode really quick. Unfortunately, it's only for Chat GPT business.

which only is effectively the twenty dollar a month plan. So like beyond the twenty dollar a month plan, you've got to pay. So for anyone out there who's like, I'd love the twenty, twenty five dollar a month Chat GPT plan, you've got three hundred bucks a month. You do need to pay for two accounts. So like find a buddy.

You guys sign up for a business. Oh no. You set up two accounts. One of you pays for the first three hundred, next person pays for the next three hundred, and you go back and forth paying monthly. And there's a way to get value out of that. Or maybe you have two. Oh my god. But unfortunately.

There is no and I get why there is no like two hundred dollars subsidized plan for businesses. So the Amex credit sounded really cool and I was thinking like, gosh, I've got five business platums and business golds. That's fifteen hundred dollars of open AI credit. It is fifteen hundred dollars of open AI credit that can only be used on a twenty five dollar a month plan. Just eat twenty five buttons. I need twenty five buddies. Let's make it happen. It's gonna be fun. I'm having a lot of fun.

Dude, I mean, the best ever. People are talking about, oh, token maxing, they're not sleeping, they're working harder than ever, like it's a bad thing. But everybody I know who's deep in it. They're giddy. They're having more fun than they've ever had. And so their digital lives or professional lives or personal lives, it's just a lot of fun. I would say my exercise has taken a dip.

But I feel like a kid when like the new Zelda or Mario game came out. The only thing I hate is that it's not a as social, but you know, you and I last night we hopped on a little call for a little bit just to share what we're working on. I think like the more we can make it a little bit more social, but like

I don't know. It's just so much fun. I feel like a kid again, just getting to build fun stuff and enjoy. And I'm watching in some of the groups I'm in with other people who aren't deep in it with tech. Like they try one thing and it's like, oh my gosh, I can't believe I could have done this thing. It's just simple tasks, hard tasks, tackle this, do that. It's never been easier to automate things. And the thing I'll close on is If you heard me talk about Opa Claw

couple months ago. You can do pretty much everything I described then in terms of function, maybe not in terms of flair, but in terms of function with first party tools, without any getting like building custom anything. And

That is really cool. And I haven't played with like Perplexity Computer, which is supposed to be similar, but I I don't think it works with any of the coding plans. So I'm like, you know, I uh like it's it's crazy that if you want to build tools that use AI models, yeah, if they're not first party tools, the tokens cost ten times as much. Yes.

It's great for consumers because we'll get all this choice of tools. Yes. But right now, you'd be crazy to go use non clawed open AI tools or tools that integrate with them because the tokens just cost ten times as much. Yeah. So this has been fun. What a blast, Chris. The well thanks for having me. Let's do it again soon. And uh yeah, and let's let's like check in in three months and see where we are on our journeys. That sounds good. All right.