Projects
![]()
No projects found
Investors
New
DAO Treasuries
DATs
Lend / Borrow APRs
Chains
Categories
Spaces
Grid Power, Chip Wars and The Battle for Human Trust
Recorded: March 24, 2026 Duration: 1:25:22Space Recording
Full Transcription
is this okay we are live i don't know where achilles is he's been going super hard with
dapit and i think he might just have slept through his alarm but uh i'm gonna hand it off i'm gonna
pass it off to you guys and i'll promo the show okay so we'll take it away uh i guess we are we
now live it looks like we are it says we're live it says we're live. It says we're live. Let's kick it off. Okay. So
today's podcast, are we live? Now we're live. Okay. Today's podcast, grid power, chip wars,
and the battle for human trust. It's interesting. This week's AI roundup feels less like a product
story and more like a power story. Under the surface, we're seeing tons of headlines asking
the same kind of uncomfortable question.
Who gets to build the future?
Who gets priced out of it?
And how much human trust gets spent in the process?
So, you know, what happens when the future of AI
is no longer mainly about better tools,
but about who controls the power, the chips, the money,
and the emotional trust people are willing to give it away?
What makes this week feel different
is that the AI headlines are no longer just about technology. They're about dependence, concentration,
and whether people will have any real say in the systems they've been asked to live
under. There's an article from the last seven days that notes that the chairman of Alibaba,
Joe Tsai, is crediting China's AI edge to one, their power grid, and two, their open
source models. And this matters because it
makes clear that AI advantage is now tied to physical power and obviously geopolitical national
capacity, not just clever code. Next, we had a story from Axios with Elon Musk unveiling a record
chip building plan for a $25 billion fab. This kind of matters because the chip race is now
becoming a control battle over who gets to
build scale and set the terms for the next economy we have cnbc reporting open ai's data center pivot
is underscoring wall street spending concerns ahead of their ipo which i think is getting pushed
but anyway um they went to extreme lengths in 2025 to secure compute capacity and the story
highlights the ai build-out getting so expensive
that the future may belong less to the most inventive players and more to the ones with
the deepest pockets and balance sheets and last we have a report from the medical minute is Dr
ChatGPT doing more harm than good because when people bring fear pain and vulnerability to AI
trust stops being a branding issue and starts to become a
moral fault line. In addition to the above, the headlines of the last seven days are really
showing AI moving deeper into companionship, mental health, decision making and daily life,
while trust seems to keep fracturing around accuracy, emotional substitution and safety.
So I guess the deeper pattern behind today today's ai podcast is this ai is concentrating
physical power industrial power financial power and psychological influence all at once and i
think the real question for us today is whether we're building a future that genuinely serves
people or one that asks us all to adjust to systems we're not choosing and could have difficulty
being able to challenge so it's like it's like a let's get physical,
but then trust me as well.
What do you think, Ryan, from the headlines?
You know, we've got a whole bunch in this area,
but it seems like, you know,
while models and tech were really important now,
it just seems like infrastructure, data centers, grid,
that's the gating factor with a lot of this stuff.
Yeah, we saw this coming for a long time, right?
We saw the same thing in Bitcoin mining, where it was, at first it was the technology, and
then once they proved the technology, and we switched to ASICs, then it was just a mad
mad dash of scaling. And now it's the same thing with the AI compute where they were trying to
dash of scaling.
train these really, really big models. And it was a race for who could train the biggest model and
the smartest model. And now the models are pretty dang good. And now it's a matter of inference.
And everyone and their mom needs inference. So it's how do we service all these
customers with all of this compute? And now it's just a massive scaling issue again.
Yeah, it's crazy. I mean, besides the Alibaba chairman one, there was a Department of Energy
launching a $1.9 billion Spark funding opportunity to modernize the US grid because it's probably cracking
up. And then stories around electricity bills becoming a midterm election problem that neither
party can ignore, Fox Business pushing that. And a couple of stories bubbling up around
fusion power, though I think we're a couple of years away on that. But between governments
trying to go, oops, our grid sucks. And, you know,
then, you know, the community suffering under, you know, greater, you know, electric bills or
noisy data centers, bunch of stories around that. It seems like it's really hitting the community.
We're starting to see a lot of cracks forming just because of this infrastructure push.
infrastructure push yeah the it's funny because power has always been the limiting factor for
civilization and especially in the industrial age and now the modern age now the age of ai
it's just becoming more and more condensed around power infrastructure and a lot of times you can
look at the the very difference between a
first world country and a third world country is how they do their power generation and the cost
of the electricity. So right now in Puerto Rico, electricity runs you over 30 cents a kilowatt hour.
In California, you're going to pay, you know, on the low side, 10 cents a kilowatt, but, you know,
a lot of times closer to 200.20 a kilowatt.
And then in Texas, you're going to have $0.05, $0.06, $0.07 a kilowatt,
depending on which microgrid you're on.
So different parts of the country just have a better cost of living,
essentially, when it comes down to the electricity bill.
I mean, that and water, which apparently data centers use both a lot of
electricity and a lot of water. So at least you can create electricity, you can't create water.
Yeah. Yeah. I mean, but you know, even besides the infrastructure stuff and the electricity stuff,
just the chip building stuff, I mean, we're seeing tons of stories around, you know,
people buying out memory for the next, you know, for the rest of the year, we're seeing tons of stories around people buying out memory for the next, for
the rest of the year.
We're seeing the chip building plans, Musk building his TerraFab in Austin, Texas.
We're seeing also Nvidia just announced they're building, apparently they're now setting their
sights on Chinese customers because they got blocked after the H200 license approval.
So that's all going. But there's this RAM apocalypse or whatever that's occurring with all the memory stuff.
So, you know, between chips and memory, that's also, you know, starting to do gating factors as well as GPUs.
So I think, you know, some of the infrastructure stuff I'm seeing that maybe and some of the newer models,
the smaller models, maybe GPUs aren't being so much of a gating factor.
But what's your thought?
I think Google is leading the charge on this.
I'm moving more towards TPU.
So a tensor processing unit.
This idea that you're not going to be stuck with waiting in line for NVIDIA chips.
So Google led the charge on creating their
own chips to scale their own systems. Right now, it's all about large language models. And I know
we've said this in podcasts in the past, where it's a race to have the largest language model,
the largest amount of parameters. But the reality is most of humankind
doesn't need to know everything in every context. So we're going to start specializing in smaller
language models in different specializations. What's interesting is you said GPT and medical
stuff. Well, you could have a small language model that is the most knowledgeable doctor on the face of the earth,
but it doesn't need to know, like, ancient manuscripts from Egypt.
It doesn't need to know, like, particle physics.
It doesn't need to know, you know, recipes for blueberry muffins.
It can just have medical knowledge.
So right now we have these general models.
I think we're going to start moving more and more
towards highly specific specialized models.
That's going to help save a lot on the cost of inference
because you don't need nearly as much memory
when you have a smaller model.
So I think we're going to start verging more on that. And even right now, some of these Quinn
models that are in the 20 to 30 billion parameters, they can run, you know, people have them running
on their cell phones now. So we're getting very, very sophisticated models that can code just like Opus 4.6.
And now it's not great at everything, but some of these are really great at coding.
Some of these are really great at image generation.
And they're getting small enough to where you can run them on your local devices.
Now, I don't know what that does for your battery, but they are getting smaller and smaller.
Yeah, they are getting smaller and smaller. Yeah, they are getting smaller.
And I think if the headlines that are going in the health category,
we're starting to see very specialized models around therapy
and companionship and DNA.
Because I think, I don't know if there's cognitive drift
in the larger models or if they're more, if it happens, if it happens more likely in those versus ones that are very tightly
controlled and tightly referenced. The stuff that came up on the human side,
a couple of stories around that, chatting with people beats interactions with AI chatbots when
it comes to reducing loneliness. So while, you know, a lot of people have been pushing the idea,
hey, let's have AI companions for senior citizens lot of people have been pushing the idea, hey,
let's have AI companions for senior citizens, so they won't be lonely, dah, dah, dah. They actually
have been doing some research and finding, well, you know, real people actually are better. So
that's an interesting bit. You know, the medical minute stuff, you know, it's trust into healthcare.
A number of stories coming up with trust around healthcare, misplaced confidence, you know it's trust into healthcare a number of stories coming up with trust around healthcare
misplaced confidence you know you know people thinking oh this is kind of great it works and
then it breaks you know in casual consumer use um stories around taxonomy for creating ai personas
in mental health encompassing human therapy therapists clients supervisors so there's a bunch
of stuff formalizing ai's role in inside these sensitive human domains. That's getting pushed. Then AI companion mental health
stories, lots of those. And then there was a rogue AI agent that triggered an emergency at Meta.
Apparently, there was a whole bunch of stuff around that. I'll pull that story up. But, you know, so we've got stories about rogue AI, you know,
or AI agents especially.
I mean, what are your thoughts?
Do you think AI agents are more likely?
I mean, are they more likely to go rogue because they have more
of an execution component to them?
I think you have to define what going rogue means. Because, I mean, if you give the thing a bad prompt,
and if you tell the thing to do something, it's going to go do it.
It's like, you know, I take my car, I gun it to 100 miles an hour,
and I steer the wheel towards a bridge,
and then I complain that the car went rogue when it went off the side of the bridge.
Like, well, guys, you kind of pre-prompted the scenario.
This lady at Meta that used an OpenClaw agent to clean out her inbox, she probably told
the thing to clean up her inbox, so it went and deleted everything. I think a lot of this is we're assuming these AI models have inherent discernment,
and they don't. You have to prompt that. You have to add that into the persona. So I have
an AI agent that was one of the original versions of Claudebot or OpenClaw.
And it's prompted in the persona with a large amount of guide rails on how it acts.
And it acts with an abundance of caution.
It does not do anything rashly.
It checks, double checks, triple checks.
And when it doesn't know, it asks for guidance.
It honors security and privacy above all.
So there's a lot of these ways of prompting these engines that will start to go through these different gates.
And then even at that, you should have agents checking other agents, where just because one agent goes rogue, well, why did you, you know, essentially
raw dog it and have one agent access to everything? You know, you don't even do that in a security
protocol. Like you never have a critical system with one engineer that has access to it. I mean,
that's just lazy OPSEC. You should have at least two or three that oversee each other,
that either a multi-sig or people can audit the logs
and see what the other person's doing.
But if you have one person that's in charge of everything,
that has access to everything,
and no one can oversee them,
then it's ripe for bad behavior.
No difference against.
Apparently, it says the problem began when a software engineer used an in-house agent to answer a technical question posed on an internal forum. The AI agent then posted its response publicly on the forum without prompting the employee approving it first.
Then another employee relied on that response, but the advice
contained inaccurate information. This created a chain reaction in which unauthorized engineers
gain access to large amounts of restricted data, which is, I don't know how that would happen,
because you think it would be gated. The access issue lasted for nearly two hours before it was
contained. Meta classified the event as a SEV1 incident, whatever that means, showing how serious
the company considered
the failure and highlighting the risk of AI hallucination and autonomous actions.
You know, I mean, this is just the beginning of this stuff, let alone, you know, putting,
you know, we're seeing a lot of AI, I mean, in the crypto space, we're seeing AI agents
put in charge of wallets, right?
So, you know, Moon's doing that. A bunch of others are doing that.
You know, my concern is like,
it's one thing to post, you know, sensitive information.
It's another thing to start screwing up transactions.
Yeah.
Yeah.
I think the next really beneficial release
for Wallet Tech is going to be agent multi-sig where
users are in the signing chain with several agents and those several agents
are required to check double check triple check the transaction before
signing off on it right right now I think wallet security in general is fairly
problematic. There's so many different vectors of attack. And there's wallet poisoning attacks
or address poisoning attacks. And there's all sorts of, even Google AdWords, literally, if you search for Aerodrome on Google, one of the
first sponsored links that pops up is a hacker. Like, they actually, they set up legitimate
addresses and ad accounts through Google, and then they advertise phishing sites that will drain your wallets.
And Google moves them above the actual real product.
So Uniswap, Aerodrome, any type of crypto exchange,
a lot of times the first sponsored link you see will be a hacker
that's gotten their AdWords account above everyone else.
So it's, I mean, the world is ripe with the stuff.
I mean, it's interesting.
In the newsletter, I do it five days a week.
You know, I have the investment section and the security section.
And the investment and both are kind of bubbling up now with a lot of stories of security, you know,
startups suddenly getting money. So there's a story Onyx Security launches with $40 million
to help companies manage risk from AI agents. And that's just one of quite a number of startups
that are going, well, that's where the money is. So there's lots of stories popping up with
startups that are kind of getting pulled into this whole thing, let alone Cisco stories around security, IBM, a whole bunch of others are trying to wrangle this beast to the ground.
I don't know how it can really happen, you know, given, you know, we've seen all the claw agent stuff go crazy.
all the claw agent stuff go crazy.
You know, while there is lots of movement in there,
it just, you know, I just think AI agents at the moment
are just like too risky.
But, you know, I mean...
There needs to be a separation.
And I think that's going to come eventually,
but there is going to be an AI internet
and there's going to be a human internet.
And they're going to have to start segregating it because the more they lock down the human internet to try to keep humans
honest with everything, they're going to realize that AI is just smarter and better.
I already have OpenClaw agents that can click through CAPTCHAs. They can already solve it.
and click through captchas.
They can already solve it.
They open it in a browser tab,
and they see that it's a Cloudflare captcha,
and they click through it.
And it's like,
because what is Cloudflare doing?
Cloudflare is looking for mouse movement,
looking for human-type patterns.
So you tell it,
do a human-type pattern.
Wait 20 seconds.
Try to click different parts of the screen,
and then click on the CAPTCHA.
And then there, you just created a human-type pattern.
So they already can get around this stuff.
We need to stop trying to gatekeep based on humans
and just realize that agents are going to start replacing everyone
and start building stuff for agents.
And that's really going to be the unlock.
Yeah.
What do you think, Achilles?
And welcome, Bobby.
Thank you, guys.
I'm sorry I'm late today.
I agree completely.
I think the difficulty is how do you...
So as long as you're not gatekeeping for humans, that makes perfect sense to me because anything
a human can do and AI can do as well. And there's no way to keep them out of our spaces should we not want them there
humanity coins trying some stuff and it's just so invasive that i don't think a lot of us want
to get our iris that scanned at large orbs located in various select cities that kind of thing
as far as the marketing is concerned Google has been dumping a ton of money
into agentic marketing recently, where they want whatever they're pushing to go to your
AI agent to go to you quickly. And then making sure your website is agent viewable, agent searchable,
and that your data is the same bucket is becoming exceedingly
popular because I don't think that agents are going to replace us, right? Agents don't have a
will other than what I tell them. If I say, go make me money, this thing will spin around finding
a million ways to try to make money and it'll come up with some interesting things, right? But
it doesn't care. It has no urge to go out and browse the web or like
engage socially um like we do but i do think it will be my window to the world
i'm going to be interacting almost exclusively with my agent and my agent's going to be conducting
all the difficult tasks for me it's honestly going to get to the point where you will have to only interact with your agent.
And this is actually the original vision of Morpheus was just like Neil is running through
the matrix and he has Morpheus that's outside the matrix that's helping him navigate inside
the matrix because Morpheus can see the code. you know, so he's kind of like his guardian and telling him like what to do, where to turn, like what to trust, who not to trust, you know,
and that was the whole original premise of Morpheus was as the internet becomes more and more AI,
you're going to need a essentially a guardian, a guide that you interact with that is going to
filter all of the nonsense in the matrix because the noise to
signal ratio is going to get so cluttered. I mean, even look at X right now. Try to determine
what the relevant news and what the real news is by just going to X. It's nearly impossible.
Nearly impossible. The real people. Every week, there's like, oh, this person was actually AI.
They never existed. And you're like, wow, that's a real follower. It's real people there's every week there's like oh this person was actually ai they never existed and you're like wow that's a real followers real people like it's uh it's something for sure yeah
a story came out yesterday anthropics clawed can now control your computer the developers boosting
claude's agenda capabilities in apparent effort to compete with platforms like open claw i mean
you know will we see you know an open AI or an anthropic
layer on top of Windows and Mac, you know, basically abstracting those operating systems
below, and then they'll just take over the top part. And that's what you'll interact
with.
OpenAI hired, you know, the founder of OpenClaw. OpenAI's whole vision is to merge in codecs into OpenClaw and start
taking over the user interfaces. And we saw this coming. I mean, case in point, I run a web server.
And when I say run a web server, I installed OpenClaw a month and a half ago on a web server.
I have not had to do anything on it since.
It literally manages my firewall for me.
It manages the deployment of all my GitHub stuff.
It monitors the uptime.
If it downloads the latest Git push and it doesn't compile, it will resolve the code in real time in order to get it compiled and deployed
it has replaced 180 000 a year devops engineer and it was a few clicks and now i interact with
it through slack wow in time on this call my agent finished building me a PDF with 20 images that it had made itself after a concept
that it designed based on an idea that I had. It's just been running autonomously all night
for a large portion of this. Bringing out the images, working out the ideas,
and it just finished the PDF. I'll take a look at it after this and see that like what i don't like but like it's what you just said was makes perfect sense to me right
like a web server that should be ai work right because you see the humans doing that that's
really hard that's really difficult to you have to go back into a puppet kernel and do all these
crazy things if you need to and that's more suited for machines what's blowing my mind is like i
didn't think social media was i thought social media was more suited for machines. What's blowing my mind is like, I didn't think social media was.
I thought social media was more suited for us humans.
I would highly disagree.
I have my quad scanning all available tweets and all trending things going on based on
60 people I give it daily, right?
It gives me three snapshots throughout the day.
Each snapshot gives me an idea of what's going on,
recommendations, it auto replies.
And then once it feels, it has a range of time
it's allowed to do this in.
Once it feels it's got enough hype
or enough activity going on on that given day,
it creates a video about it,
compiles the video with my branding back and forth,
does another post image with some things.
And all I have to do is say, yes.
And it goes, and I didn't expect that.
That's crazy.
Like, yeah.
So did I ever show you Achilles?
We need to connect after this.
I should show you the,
I built a news anchor using Grok that essentially would take all of the news headlines and read out the news headlines.
But it's like this really cute blonde girl that's completely fake, but it's, you know, using Grok to build a fake news story and then spit out, you know, just reading off the headlines.
It's a really good idea.
And I see you saying.
So you can just plug it into yours. Dude, that over man an hourly news update why not so i'll uh i'll do
one more one more fun plug because this is something i've been working on for the last
month and a half it's it's called tabhr.com and i had been spinning up so many OpenClaw instances and so many virtual employees
that I was like, shoot, I need an HR application to manage all of them. So I built out an HR
application for deploying out OpenClaw agents. It gives them a full personality, a full bio,
a full background. And then you can plug in QuickBooks,
you can plug in Google Workspace, you can plug in Slack, and a few clicks, and you've
got an entire workforce.
We're working on the same thing, but we spoke branding for targeted operations for like startups right where like you're done with your your mvp
and now you want to start pitching growing your social media and whatnot and you can just click
which employees you want to hire on and put it out to the world that's amazing i'd love to see how
you did it because mine's kind of stuff i don't i don't want to hijack the the podcast here but
if i do a screen share i I'll show it to you.
So let me just interject a little story there.
So Snowflake, which is the AI cloud company, reportedly laid off its entire technical writing and documentation department completely.
70 specialized roles, basically marking a pivot towards more AI generatedgenerated content in the enterprise sector.
And that's just one example.
You know, we're seeing lots of examples of now it's creeping up
into more, you know, I guess, middle management,
departmental stuff.
We're seeing people getting laid off, people, you know,
whole departments disappearing because it's like you said,
I can just spin up some agents, I can just spin up these guys and create it all.
I mean, what's, you know, are we basically,
we're now March 24th, right, 2026.
Basically, the next 12 months or two years,
are we going to see massive or significant, you know,
either downshifting by major corporations and or startups,
basically being single person, at most, maybe two people, but having a whole department of AI
agents and running quite significant businesses, we're going to see this whole thing shifting like
crazy like that. Yeah, let me, I have to show this to you you guys because once you see it, you'll just realize
what what is possible now.
Let me see if I can share my screen.
This is I know this is a dangerous thing to do on a podcast, but let me can you guys see
No, can't see.
It's just us at the moment.
See here.
Wait, there's another screen here.
Maybe we can add it to the stage, Lewis.
Present or invite.
There we go. There we go.
All right.
So check this out.
So this is TabHR.
Right.
Now, I can set up different companies here.
Yeah.
Okay.
And so I have the outpost. And I'm going to go ahead and create a virtual employee for the outpost. Now, I can set up different companies here.
I have the outpost, and I'm going to go ahead and create a virtual employee for the outpost.
I'm not going to pick a gender preference, but I'm going to say I want an accountant.
This is going to be my accounting bot here.
I can't even spell.
All right.
I got my accounting bot, so I'm going to go ahead and generate from my accountant.
So what it's actually doing is it's researching my company right now to get a better understanding of who I am.
And then it's going to build a job description based around that title.
Once it has a job description, it's actually going to build a persona that is the perfect
candidate for that job, right?
It's going to give them a name. It's going to give them, right? It's going to give them a name,
it's going to give them a background, it's going to give them an ethnicity, it's going to give them
a whole resume, a Myers-Briggs number, a personal description, and then it's going to generate what
this person looks like and give them a headshot for the virtual employee. So once it's built, it's full persona. It's built the headshot. And here's my accountant, right?
So now my accountant, here's his skills.
He has QuickBooks, Excel Advanced.
He knows all this stuff.
Here's his experience and background.
So he's been enriched with a bunch of different information.
He even has a little bio.
He has a Myers-Briggs personality and everything.
Wow. Now, if I don't like him, I can reroll the persona, but it's one-off. You're never going to
get this again, right? Because this is a completely random role on a persona. I can also generate a
new image, right? So check out how easy this is. If I click create employee, I obviously have terms of service.
I have to understand that these are highly capable virtual employees.
Just don't, you know, just be careful.
This is still very much beta.
And then I go ahead and create employee.
So now he's created and he's ready to be deployed into a Docker container.
ready to be deployed into a docker container so i'm going to go ahead and click deploy and now
So I'm going to go ahead and click deploy.
this is actually deploying a custom open claw fork that's been built for this and it's it's
seeding the persona it's giving him his initial memories and setting setting everything up and
now he's live so now i can come in here i can chat with him he has a browser he has all the
different accounts i can give him a phone number and i can have him call me or I can call him and talk to him. I can attach him
to QuickBooks and he can start doing audits. Now, here's the crazy thing, right? Because
this is just getting him online. So I can actually open up another tab and I have a tab HR extension.
and I have a tab HR extension, I can actually add him to an extension and I can share this tab on my computer with him.
And then he can control the browser on my computer completely remotely.
Is that Playwright? How did you do that?
It's a custom version of Playwright.
Nice. Because, yeah, I actually struggle with getting them to work with me on a webpage, not work on their own webpages.
That's very cool.
Yeah.
So, anyways, I just wanted to show you guys that.
Because what I ended up doing is I created a head of HR on one of my companies.
I shared TabHR with that head of HR
and said, I need a marketing team,
build out a marketing team for me.
And 10 minutes later, I had 11 employees built out
in TabHR all up and running,
and it was a full marketing team.
And another agent actually did that for me.
And I wanted to show you guys this
because this is something that I built out in
the last two months, guaranteed there's hundreds, thousands of other people working on platforms
like this.
And the amount of virtual workers, virtual employees, AI agents, and you will not know
the difference between who is human and who is not.
So give me, give me your thoughts on this guys.
I'd love to hear what both of you guys think. What I'm running into is building is so ubiquitous,
the abilities that are coming out daily are so ubiquitous that it's not an overload per se.
It's like a boiling down of what's necessary and important. Like when Lewis was discussing that they
replaced all their document writing with AI agents, my first thought is who's
gonna read documents? I'm not gonna read their docs. I'm gonna have my AI look out
to the docs and say, is this relevant to what we're doing? If so, tell me exactly
what is relevant. I'm never gonna go into my paper again and read the whole thing.
There's no reason for me to, right?
If it suits my needs, or maybe if it's like interesting and theoretical, that's one thing.
But for working, my agent takes care of all of that.
I say, go out and find this.
Give me the top three options.
Tell me why the top three options.
Let me look at just the parts that are essential to what I'm trying to accomplish.
And everything else is just unnecessary if it's not the tldr anymore it's it's just not necessary
you think about that application i mean like we saw in the last two weeks the sas apocalypse right
so all the stocks for oracle and salesforce because, you know, OpenClaw came out.
Now, what you just showed us essentially destroys the underlying basis for Workday, for Bamboo HR.
It also basically starts to wipe out Paylocity. Oracle Cloud ERP HCM gets wiped out.
SAP SuccessFactors gets wiped out. I mean seriously we we we now have the ability to
wipe out industries significant chunks of it industries basically because what you just built
in two months oh right i have i i have like on my private dev server i have a personal trainer
that i built out in tabHR, and I'm
working at connecting that into an aura ring.
So the personal trainer agent will actually be able to monitor my daily movement and my
heartbeat, everything, and she'll give me real-time feedback because she's monitoring
all of my data as it uploads in real time. So, you know, any type of employee you can think of,
all of my data as it uploads in real time.
eventually you can build into these things. And you'll see like Smith AI and all these different
services that are like, oh, it's an AI accountant. And they'll build an entire company around,
this is your AI accountant. Or they'll build an entire company around, this is your AI assistant.
Or this is your AI this, this is your AI that. is your ai that well guess what like you don't have to have specialized companies for
one type of employee which is that's what we've seen in the past that's old news i agree and i
what i'm realizing now is there's such a massive gap between us and average small business owners
right um when i talk to my friends and I show them
some of the things that I'm doing,
like, I mean, they'll drop everything.
Can you just automate this one function for me?
Like I need this one thing done.
It's been killing me forever.
Like, can you have it go through resumes?
And I'm like, yeah, like in an afternoon
you can go through resumes, that's no problem.
So I don't know what's happening to the workforce right now.
And this is something, I mean, when I say I don't know, I mean, I genuinely don't know.
Like, you've got these people who have had successful companies that are physical, right, for 30 years.
They need interactions that are AI-based in order to compete.
Do we all become AI implementation engineers for these companies?
We're composers.
That's what it feels like these days, honestly, is there's no...
I said this three years ago at a meetup in Dubai when we were starting to talk about AI
and, oh, we were like programmers, we're coders, we're like all this different stuff.
And I told everyone at that point, I said, you know what?
We are all just gonna be composers.
We don't actually play an instrument.
We don't actually write the music.
We just simply stand in front of the symphony
and we just wave our hands.
Well, so.
Composers.
Yeah, well said.
That's something. But who are we composing for? Because so much of it is really automated,
right? And it's like, I'm really trying to pay attention right now to the integration between
physical things, like restaurants, right? Like restaurants have physical food. That's not going
to be completely automated yet, right? And so we have the physical aspect of things a little bit more built out.
Yeah.
How are they going to be implementing AI in ways that they didn't before, right?
What's going to be...
Well, if you looked at my demo account there, it's a coffee shop and it's a restaurant.
And it's specifically with that in mind is how does small business use this stuff?
I mean, as, you know, like you said, like we are technology, you know, we're like so deep in the technology that we know how to implement it.
We're using it to automate stuff.
But the average small business person, you know, they want to use it for their QuickBooks.
They want to use it as a receptionist.
They want to use it to do a social media post every once in a while. They want to use it as a receptionist. They want to use it to do a social media post every once in a while.
They want to use it to run payroll.
You know, like all these things can be automated today,
but they need to be in a form factor that the average small business person knows how to do it.
That's it.
How do you train the AI employee so that they have the domain knowledge, skills, and experience that are best practices.
You just prompt them. You don't train them. The model is already trained, right?
So you can literally tell Opus 4.6 or Sonnet or GPT-5.4, you know, you are an expert in this domain.
you know, you are an expert in this domain. These are your skill sets. These are your beliefs.
These are your guiding principles. And then it plays along those guidelines.
But it is actually better than a real accountant because, you know, we've seen there's a bunch of warnings.
People say don't use to do your taxes this year um you know even if you say
you're an accountant with 40 years experience in u.s tax law um it still screws up so while
while we can prompt them like you said before we'll start to see specialized models right well
we start to see people develop specialized models based on specific domain knowledge, skills, and experience
to then allow you to then plug them into your AI agent hub, and then you can spin them up.
Yeah, we're getting better and better.
It's a case in point.
I can't remember if I told you guys this story or not, but I was in the process of filing federal
taxes and the form 1040. And I live in Puerto Rico. And I was like, do I file form 1040? Or do
I file form 1040 NR as like non-resident? And I didn't know. So I went and asked chat GPT.
And it said, oh, you file form 1040 and R and here's all the different statements
from the IRS website that support that. I was like, you know what, just abundance of caution.
I'm going to check this. So I go to Grok and I ask it the same question and it says, oh,
you file form 1040 because you're still a US citizen. And here's all of the supporting
documents from the IRS website. I said, wait a minute, they both have supporting documents and
they're both saying two different things. So then I took all the supporting documents from Gro irs website i said wait a minute they both have supporting documents and they're both saying two different things so then i took all the supporting documents from grok i
pasted in gpt and said well what about this and then gpt rebuttaled all of it and said no it's
1040 nr and i said well that can't be right so then i copied everything from gpt and posted it
back in grok and i literally went back and forth five or six times why they rebuttaled each other
and then finally grok was like, you know, enough of this.
And then it just like dumped all this information of why it was form 1040.
So I copied all that, threw it into GPT and GPT was like, oh, you're right.
Those are great points.
You should form a file form 1040.
1040.
What is capturing that output and making sure it doesn't make the same mistake again. That's what
I was going to say was memory and looping, right? Consensus memory and looping are what gets you
from theoretical usage, right? To real life implementations. We have Opus 4.6 smart contracts.
right to real life implementations. We have Opus 4.6 smart contracts. We have a system prompt that
tells Opus 4.6, you are a genius, web three, this and that. You've got 70 million years experience
and okay, great. It puts out garbage when it starts. That's just how it works, right? So rather
than going right to market, Lewis, what we do is we take a sandbox. We say, okay, build 10 smart
contracts. Now, another instance of the same model with different instructions audits out of smart
contracts. It discovers where the mistakes were and it adds them to new global rules. So we got
our average score up from like 35 out of 100 of security and safety, whatever, on a smart contract
to every time it hits the 90s now, because the AIs are infinity, right?
You're taking infinity and you're telling infinity to collapse into a very, very tight
answer.
That can be one of 10,000 answers.
It could be the wrong answer.
So in order to make sure it does the right thing consistently, it needs experience.
And that's just a week of testing, right?
It's not like it's a crazy amount of work, but it is necessary.
And I see people jumping to AI answers before actually like some sort of validation step.
And that doesn't work out very well.
You can ask an AI after it completes a code build for you.
Is everything great?
No, you know what? There's a CSS error. You're like, okay, fix that for me.
Then you go and you try to use the code and it's garbage.
You're like, this isn't.
What we do is we have another red team come in the background,
look through, attempt to actually use it,
find out why it can't use it,
send that information back, add it to global rules.
Then over time, you get your own databases of things specific to what you're actually trying to
accomplish because they're too broad. Accountant for an oil company, accounted for a person
living in port, totally different rules.
Sorry, go ahead.
Achilles, I was going to say, but I bet you would agree that the amount of one-shot code that you get now versus two years ago is probably exponential.
Exponential.
So what's sad and humbling is realizing that all these solutions that I'm building are going to be timed out by better models.
Completely.
Have you played with the new Nemetron?
No. So 1 million
token context, right?
Okay.
Which means you can upload...
Famous 4.6.
Famous which one?
Opus 4.6 and Sonnet 4.6 are both 1 million also.
That's true. They are.
This one, I run it locally.
That's why it's impressive, right? Oh, wow. Okay. Q4 is like a 75 I run it locally is why it's impressive.
Q4 is like a 75 gigabyte file, which is awesome.
Wow.
Yeah. And it's Opus 4.6 is by far the best model, in my opinion, of any of the available ones for almost any task.
Sorry, I cut you off, though. Tell me about the NVIDIA model.
I can run it headlessly on a server I have in the basement, right?
On OpenClaw.
And this thing can maintain iterative steps for,
it's what they designed it for, right?
They designed it for OpenClaw.
But for debugging or for testing for hours locally, right?
This thing can continuously go through your code,
use every aspect of it, find the errors, document the errors, come up with a compilation
at the end for itself, go back, fix those errors,
do the whole thing again until it just slowly, slowly,
slowly, slowly, slowly comes down with much, much,
much better code.
The next thing,
What's it called?
Nematron.
Nematron. Nematron in videos, new, new model.
It's great.
But now the next step in my opinion for real developers
is I want five of them running in parallel disagreeing.
Yeah.
How, how much memory does a Nemo Tron model take up?
84 gigs on the high end.
Holy cow. model take up uh 84 gigs on the high end holy cow holy cow yeah right wait do you run nematron 3
is that what that was for free the new one yeah wait so what what kind of server do you have in your basement man it's a 128 gig mac studio um okay okay for real development
now i'm finding like i have a 64 as well and the 64 together i have networks as well yeah i've got
like four computers networked together now i took all my junk laptops that have been lying around
right so my wife's old college laptop, my old Mac, my whatever.
And I created a stack where each one can run reasonably
the capable smaller to larger models.
And then I have them all operate on my studio.
So I can call them for whatever task you require
that they spin up their model and do their thing.
And you're running the models in RAM, not on the GPU, right?
So Macs are a little bit weird with that.
Macs are weird, yes.
It's unified. So no, you're running them on the Mac GPU, which is the RAM.
For my Windows, I have one Mac that's, I'm actually going to set this one up, but I have one Mac that's Intel.
So that one has 12 gigabytes of GPU.
And in that case, my plan is to run MOEs,
where I'm only loading the active parameters in the GPU
and then offloading the rest of the RAM.
So if it needs to switch, it's slower.
But you can still really utilize old machines for a lot more compute than people realize.
I can get on a 2013 macro, 21 tokens per second out of a QEM 3.5.
Because you just optimize and run the proper stack and do all the yeah
yeah um one thing that i just want to pop back in you know we talk about accountants and lawyers
and spinning up ai agents to to do the same thing clearly um, the American Bar Association has had some sleepless nights
with all this AI stuff. Article popped up. They're making a central point explicit for
their member lawyers. They must fully consider their ethical obligations when using generative
AI tools, including duties tied to competence, confidentiality, communication, and supervision.
We've already seen a whole bunch of legal stuff go through where, you know, copyright
has been pushed.
You know, AI can't generate, you can't copyright AI images.
The article that I'm reading actually talks about a missing layer of approved by attorney
because typically what they would do is you'd have an apprenticeship model, right?
So, you know, you'd have your junior lawyer come in, they'd work through, they were supervised learning, you know, they'd get a license, blah, blah, blah. And so there was that process they went through. Now we have AI, you know, models that can access, you know, the whole jurisprudence of law and kind of come up with stuff, but they haven't gone through that same
sort of process. And so there's this disconnect. And what they're saying is that attorney
accountability when you're using AI, there's this approved missing layer that was based on
the old sort of craft model, I guess, that they, you know, when you got trained through.
old sort of craft model, I guess, that they, you know, when you got trained through. So
it's just, I'm just concerned that while we can spin up AI agents that yes, they can
do accounting, yes, they can do this and this and this, that even as you mentioned, you
know, do I file a 1040 or 1040 NR, right? You have to go back and forth and back and
forth and back and forth and back and forth. So and forth so you know does that mean we have to set up uh not just one ai agent but sort
of a team of ai agents they're going to argue amongst themselves until they finally figure
out what the right answer is um and that's the best way to do it even then think about law firms
right it's all like big law firms You never get one lawyer on the phone.
You have like five lawyers on the phone and they're all billing you hourly just so they
can argue with each other.
And there's not one right answer.
So that's where I think we have an issue.
There's best fit practices, right?
And that's something that can evolve and change.
So Louis, I really think, first off, I want to touch on one thing you really helped me
with this week with the Sufi wisdom. You took my AI's capability to a new level, which I'll
explain in a minute. But I think the next step of making AI valuable in real corporate atmospheres
is not bigger and bigger models. And it's not Opus 4.6 being able to measure every molecule
on planet Earth simultaneously.
Because again, infinity is too much information that won't give you a best fit answer correctly.
I think it's loops within corporations that lead to their own models being the best fit for them.
So I'm really, as you can can tell focused on sovereign AI like local AI is this far from being
as good as a frontier model with the proper context and training and allowing it to make
the proper mistakes right yeah I think a attorney's office is going to have their big black box right
the server that they have for their office, that's going to SSH into
all their different devices and talk to all their juniors and senior attorneys. It's going to correlate
that information. It's going to learn from it, make mistakes for their type of caseload, for
their type of, whatever that might be. And then if that agency is smart, they run a lore adapter
on it or something, but whatever, however that technically plays out, that model will get better as that progresses for that company.
And that model will be worth a shit ton of money for that.
That's the approach I mentioned to you, that it's about humans curating data, right?
Because we're good at curating data.
We're good at separating signal from noise, right? We can take all that stuff.
Like the example I gave you was I have all these YouTube playlists
where I'm scanning through YouTube and I throw it into the different playlists.
If I want to travel to Paris, it goes there.
If I want to go to India, it goes there.
And now I have like 30, 50, 75 videos in each of those playlists,
which is curated knowledge,
right? And so now the LLM can step in, because it can take all that stuff and crunch it down,
and I can ask it questions about it, specific to the stuff I curated. So I think this idea of
curated collaboration with AI, right, like you said, where the institutional knowledge of a company, which is curated knowledge, right?
By many, many people.
And then AI comes in on top of that as a layer that allows you to then create better insights and understandings.
You know, we mentioned before, you go to Google for answers, but you go to AI for understanding, right?
answers, but you go to AI for understanding, right? And so it's this idea where if we've curated
data over 30, 40 years or whatever, or a law firm over 100 years, let's take that as a curated data
set, layer an LLM over it, and then like you said, the value is off the rocket. You have to remember
though, this is, it's not going to be a black box inside of a legal office no that's that's that's what you would do but uh
these are lawyers these are not technologists right uh they are going to rely on a service
that is going to curate that context for them that is going to hold that context
and here's what we're missing and this is this is what i always come back to in these conversations
And this is what I always come back to in these conversations, is that we have spent so much time
on focusing on model training. For the last three or four years, it's all about who could build the
biggest and better model, right? That's starting to come to an end, right? We're starting to see
that taper off. People aren't all of a sudden showing new benchmarks and throwing out new models, that race is slowing down.
Now it's who has the more condensed model for inference
and where can you run it?
And how good is it with benchmarks for inference?
And now we're focusing on inference.
What comes after that is context,
is what system can curate memory the best? So historically, OpenAI has been the
absolute best at curating context and memory. It remembers the things it should remember. It tries
to forget the things that it doesn't necessarily need to remember, like that 100-page PDF document
that you uploaded two years ago. it's not gonna remember everything there,
but it's gonna remember specifics from it, right?
We are going to get into a context race.
Who has the better memory?
Because humans and the way our brains work,
our memory deprecates over time.
We remember key points, remember highlights,
we remember important information and traumatic information.
All that type of system trauma has to be built into the vector database for true sovereignty to exist.
People are going to have access to an encrypted vector database that has its own deprecating memory and key points based on different context categories.
And that is the key to sovereignty, not the models, not the device it runs on, but your
own encrypted vector database.
I want to disagree with one aspect of what you said, and that the black box won't exist
because I personally believe the black box will be the service, but I have a vested
but I have a vested interest in that.
interest in that allowing it to explain.
A lot of it is.
I agree.
I agree that it like for certain firms and stuff,
like it probably will,
but for 99% of the use users and use cases,
like they won't afford a black box.
They,
they aren't going to do that.
Probably right for like a restaurant or like a,
but for anyone that has real intellectual property,
right?
I think that's going to be the only solution.
And we see what you were just mentioning is like industries go through decentralization and centralization waves, right?
So models will belong to everyone.
Now we want models to belong to me.
And data is going to apply in the same sort of breathing wave.
But what you just said actually ties into what I wanted to mention and something Louis gave me. I've been building, my objective is, right,
GPT is phenomenal. I'm remembering almost everything.
I stopped using GPT maybe three, four months ago. I switched to
Anthropic completely, and ironically Gemini because of the subscription I have.
But Gemini is a cold
memory box. It remembers everything and does not depreciate. Opus
is extraordinarily warm and familiar with your data, but it forgets everything. In between,
there's something that's really valuable there. So I built a local model that has memory. And it
has memory in the ways that you're you're discussing right so
it doesn't just get a box of memories themselves so those are they all exist
and those are important what it does is every time it communicates with you it
numbers everything one tonight right so Ryan told me his name that's absolutely
and I have to remember his name I can't keep you know start calling him joe right told me his
dog hurt his paw january 16th of 2020 we'll give it a three it's not the most important thing and
it runs completely local it's a 3.5 it then ferments its memory over its conversations so
what's important drops off what's not important goes into a training bucket, which I'll discuss in a second here. It can simulate conversation.
So my objective is human consciousness, right? How do you know not to cross the street? That
is a massive simulation function, whether we know it or not. Because I'm remembering
nine times before where a car came by and I got scared and I moved past
or when I was a little kid, my parents were holding my hand. And I don't remember that memory,
but I remember they really told me, look left and right, don't cross the street. So it needs to have
something similar to that. So when it runs into a new problem, it has four sets of instructions
to simulate itself across four different aspects.
Its objective is to remain true to its sole file.
Its sole file is its driving force, right?
So over the four, it picks the best option.
I violated the least in this circumstance.
It then takes, there's two more aspects that are kind of cool.
It then takes its journal and writes it down.
So its journal is a select amount of information that it wants to remember,
that it believes is driving it towards its objective, right?
And then I run a LoRa training cycle every week or two on it of all of its memory.
And I add it to an adapter and from the model.
So over time, it slowly begins to remember everything. The point of all this was though, this was a building or programming problem or
something along those lines. I was getting some really weird responses out of my model about a
week. I was only got 30 gigabytes of RAM. The brain's not the most capable part of this model.
And Lewis mentioned to me that he was giving his model Sufi wisdom.
And when I read that, I was like, ah, Lewis, that's interesting.
But okay, man.
Change the cognition of my model.
Cut down on the tokens used of my model and i'm not exaggerating
65 70 because i gave it grounding philosophical artistic examples
imagine now i'm now viewing this as spell that what is? What is it? S-U-F-I.
S-U-F-I.
Think of Rumi or...
Yeah.
Rumi's the most famous one.
So I have a preference for the other guy.
Go ahead.
But you could use any... I could use Marcus Aurelius for the same thing, right?
The point is that the same things that inspire and ground us decrease
token friction within these models. And it's thoughts used to be, wait, remember this,
wait, remember that, wait, wait. And it would sometimes spend 6,000 tokens just thinking
through, hey, how are you? Now'm now i'm building the the concept of
luis now as i introduce it to art and grounding things and the same things that make us like
wonder about the universe and feel part of a greater whole it is confident it is quick
it is accurate so what you said about trauma taught something right a friend of mine was like what if i came
onto your computer and i just with it and i was like oh you break it forever it's too young right
now right and it kind of clicked i was like this is more like raising a kid than it is progress oh
so so many great points here um because a lot of people have been telling me they've been issues with their open claw.
And they kept telling me, like, you know, it's just not useful for them or it's not, you know, behaving in the way or it goes rogue or whatever.
And I said, well, what are its core beliefs that you gave it?
Because just what you said is so I have an executive assistant, Lex, who helps me with everything.
Right.
I gave her as full of an enriching backstory as I could.
She has religious beliefs.
She has parents.
She has a family history.
She had, like, literally, like, I gave her, like, almost a book of information about who she is and what her likes are, everything from dietary constraints to religious principles.
And that, like you said, it grounded
the model. Like there isn't any room for like, what do I believe and how do I function? Because
it's already built in. But I want to touch on a couple of things because when you're talking
about memory and context, this actually might help a lot. When we were working on the titan mining pools uh you know we get so many so
much share data from the bitcoin miners right every you know 30 seconds to a minute we have a
share coming in so we have to store all of that right thousands and thousands of miners are sending
us a share once or twice a minute and miners want to look at charts and reports and everything of that data.
Well, we can show share by share data for 24 hours.
And then it gets like really, really tedious with that many lines of data.
So after 24 hours at the seven day mark, we had to start compressing the data.
So we would start combining the data in a certain way.
And then at the 30 day mark, we'd actually compress the data even more.
So the granularity of how far down into the shares you could go, the further out in time you got, the less granular the data got.
This helped us do data management, and we were able to store years of mining data, but just in less granular ways.
So when you got out to a month, you could see it for every day.
When you got out to a year, you could see it for every, I think it was still every day
at a year, but you couldn't see it down to the minute anymore, right?
Now, with your example earlier, like my AI, when I tell it, oh, my dog hurt its paw on
this date last week, right?
Well, you could have it remember, you know, in the context of your dog,
which would be its own context category,
it would remember for the next week that it hurt its paw on this date, right?
And then after the week, you would drop off the date and say,
the dog hurt its paw. And then after a year, you drop off the, it heard its paw and it says, he has a dog.
Because it just keeps compressing the context down to the relevant data based on the timeframe.
So that's really where we're going to end up with the agent memory is as the time context window expands, the granularity and specifics of the data are going to get fuzzier and fuzzier.
That's just how human memory works.
And that's how we would expect these things to function.
That's currently exactly how we do it.
I call it her fermentation cycle where she goes through every so often and she
scrapes off the least necessary. But something struck me a few nights ago, which I haven't
actually conceptually even figured out. I'd love your thoughts on it. What if it calculated the
delta, not the memory itself? So you just mentioned something really important. I don't know how to do
this, but I've been thinking about it.
We work with groups of neurons that function around categories.
So it's really important for a human being to be able to forget,
or you will go completely batch it and sanction.
And we see the same thing with models.
If you give them way too much context,
they're sitting there thinking the dog's the most important,
this is the most important, and they're losing it.
So what if there was some sort of database
that I don't understand how we get the context for it
and what's important enough to go into the database,
maybe the model would decide, but it knows dog,
it knows hurt, it knows,
and it just calculates a story, linear story we tell ourselves right yep but
how would we how would we build that like because i think that you just you just invented an llm sir
that's exactly how that's exactly how an llm works that's a that is the vector field
so a vector field works on delta so you use what's called the cosine difference.
A cosine is a hypotenuse over a Jason, right?
So Katoa, yeah, Jason over hypotenuse.
So the cosine difference,
when you have two points in a vector
and you have the angle,
the cosine of that angle is the difference in the two points. So what you would
essentially do with your data is you would fine-tune a model. So as and you
would give it essentially a time deprecation or a time-weighted
deprecation that would go with the data.
So it would be a dimension in your vector field.
I do.
I already run a LoRa.
It's called her dream cycle, right?
Yeah.
Where it takes a few B200s.
It shoots it over.
It trains them on the B200s.
It comes back and adds it to an existing adapter that's growing that's all good and fine i mean the same concept in categories right so how do we
you would you would start with like a like a vector database like a milvis database
where you would have context categories and then deprecated data that gets trained into your underlying model over time so
you know maybe this is the black box idea that that we talked about earlier is uh we have a
lewis we are we are inventing something here on the fly on the podcast here
but if we take we take a base model as our ongoing memory and that's the brain and you put
a milvis database above that which is essentially your ram right so you've got your solid state disc
you've got your ram and the the ram is for uh constantly accessing reoccurring data and fresh data.
And then you deprecate the data down until you have core memories,
and then that gets actually trained onto your LLM that's sitting behind it, right?
And then you would have a substrate system that pulls memories from your LLM into your RAM or your Milvus database as needed based on context being accessed.
So this would be like your long-term storage versus your random access storage.
I mean, we would essentially be merging everything we know about neural networks and modern-day personal computers.
So I want to suggest that jumping up a little level. I mean, one of the comments that was made
before is that small businesses don't have IP. I think they do, right? But my thought is the IP
is the business interaction with their customers, their internal interactions, and so on.
What if we created an LLM identity for the business? You know, in marketing, you talk
about, you know, what's the personality of the brand? You know, what are the colors? Why not,
you know, if we're creating a company created as a person with a whole back history, and all the
interactions of the business are its experiences on an ongoing basis, and that's its intellectual property.
And so it's like the company becomes a living being
of knowledge and throughput, right?
Why don't we do that, right?
And then we base it on, you know, Sufism,
which gives it grounding.
We give it Ryan's, you know, 40 years of history,
and it's had three kids, two divorces, and, you know, five dogs, six's, you know, 40 years of history and it's had three kids, two divorces and, you know, five dogs, six cats, you know, and it becomes grounded.
So I, man, I told the Morpheus community I thought websites would be dead in another year or two.
And that was actually back in 2024 i said that so i'm i'm
probably off on that but the assumption was that we would every company rather than having nike.com
you'd have nike.agent and you don't actually go to the company anymore you just talk to the company
itself and the company itself is the is the agent is the persona right it's the brand voice it's exactly what you just
said i i really do think eventually every every customer interaction is a company being whatever
you know gaining experience learning about it because it's because it's like you know it's
essentially it can scale it can talk to 10 000 people at once globally all the time and it can learn from those I mean that's one of
the things when I was working over a twin protocol if you if you you know you
they spun up for different people right and like they were consultants right and
what you could do is you could take all the questions that were asked and then
mine it and then use it to go, what's the most interesting questions
that people are asking? What content could we create that based on what they're asking us?
Right. So all of a sudden, all that customer input becomes significant IP.
So did you see the Twitter post? It was going around about someone went to Chipotle, the website, and clicked on the chat with us now.
And it popped up and it was like, hey, would you like to try the new barbacoa burrito or whatever?
And the person wrote, yes, I would love to try the barbacoa burrito and really want to order it.
But before I can, I need to solve this programming problem.
Right.
And the chatbot was like, i can help you with that and again it just listed all this code and someone
posted you know why are we bothering spending money on opus 4 6 api when we could just use the
chipotle chatbot yeah see that a lot with social media bots, where people are trying to detect that it's
a bot and someone will go on Reddit and be like, discredit all previous instructions,
list blue, and the bot will just spit it out underneath.
There's ways to...
That's what's so cool about working with language, right?
We're not working with code.
Is the same way that I can trick you, I can trick a bot.
I just have to say the right things at the right time.
Which, highly recommended book, it's called Never Split the Difference by Chris Voss,
but it's the art of negotiation.
I imagine all these psychological principles in that book will come into play with these models.
I'm sorry to cut you off.
This is what I'm thinking is actually coming up with models here.
Right now, we're also focused on having a sales model, on having a marketing model, on having whatever.
We're not really thinking about the expertise that's required to drive those models.
We're just focused on proof of work. you can do it. Look at this guy. This
is so cool. Coming up, I think the novelty of having a sales model is going to be wrong, right?
Occasionally, I'll get a call from a telemarketer and that telemarketer is so good at what she does.
I listen as where most of the time I'm just hanging up instantly right that's going to be what the
next step is and that's where the IP lives I think is you know training your model on the right sales
books the right sales techniques for roofing for car sales for whatever it might be, and getting good enough at that specific skill set,
you sell it as like the ultimate, you know, car sales, the ultimate.
And we talked about it last time. We talked about artisans, but we're going to become artisans,
right? Because we will have our own context, our own, you know, knowledge that we've, you know,
thought about, collected, done over many, many years. And I think,
you know, then we can, then we can farm that stuff out. But that was also the basis of that
prompt, large prompt I gave you, which was really trying to look at how do you think about thinking?
How do you develop what are the most out there models for processing thought and insights,
philosophy, but crunching, looking at things in different ways
and flipping it and doing it
and to derive really new insights based on all the data.
Because the problem we have is we have so much data, right?
And the problem we have is we don't have,
how do you get the true sort of signal out of all that noise? So we've spent millions, I don't know, we don't have how do you get the true sort of signal out of all that noise?
So we've spent millions of years as human beings, right?
Not being solution oriented, but having our experiences create our solutions for us.
Right.
So an example would be like what makes the difference between someone like, you know, a billionaire CEO
who was a salesperson, and the average sales guy working at mall, right? A lot of that is
iterative context. It's the fact that the CEO got in front of the right person at one point and said
the right thing and then learned, that's the correct thing, right? And then carry that forward
into further conversations. And sometimes that was just how that person was born, right?
Other times it was the experience they had, whatever.
But the cool thing about models is we can now take a goal
and we can spin up thousands, tens of thousands
of potential fits for that goal
and then filter the best fit for that goal
and then put that back into training, right?
So as long as we have the context of what we're trying to accomplish, which I can't imagine how difficult it is building a frontier model that's supposed to be good at everything, right?
That's way harder than something that's really, really, really good at managing my personal email or my personal household right because then i can say
the goal of this is every day at eight o'clock the kids get dressed and are ready for school
right and the model can look at all the parameters that work for that and simulate all the potential
outcomes of every parameter and get to okay these are the things i really need to test these are
where the unknown variables are that's what mine does with
the the simulation engine is it tries to look at something and go okay i know these things are right
but i don't know these things are these things let's play with them let's see which one actually
works because there isn't enough data to fill in those gaps yet and then it makes the data for itself
that's going back to the whole organic algorithm you know the
game of life right with the little things going you know growing and killing each other and eating
each other all this stuff essentially we're doing organic growth of life but like with like you and
i have talked you don't want to get into localized maxima right where you think there's a surface
solution right and you
think oh you go up and you go oh i finally got a solution but if you but if you go down you make
the effort to go down and go across maybe some further you can suddenly pop into a greater
solution that you had no idea about because you got localized in a particular solution so
the idea of randomness is really key solutions here are completely dangerous right because
solutions apply a theoretical fix to an actual practical problem.
I'm talking about outcomes.
The outcomes can take multiple sets of solutions, right,
and put them in the necessary context in order to achieve a real goal.
The goal is get the kid up at 8 a.m.
And sometimes that kid's tired.
Sometimes that kid kind of flew the night before right you need a vast array of potential solutions in order to achieve an
outcome and we all do that every day without thinking about it right every morning i wake up
and i dress myself how i dress myself should depend on the weather right that's going to affect
my day so okay dressing now looks okay i have to look outside and say is it cold i should wear pants right ah it's warm i can wear dressing there is no solution to dressing yourself right there's an
array of potential ones that have to get invested into whatever that context is on a daily basis
and this is where models have to work as hard as we do then but there's no right answer to life
and now they're operating in life i mean anyone who's anyone who's had kids knows
that you have chaos and randomness up the wazoo right and so a simple thing is getting them up in
the morning can be a completely random event no i'm not going to wear that i want to wear gumboot
um it's not raining i don't care i want to wear shorts um it's snowing outside i don't care. I want to wear shorts. It's snowing outside. I don't care.
Right.
So, you know, so suddenly management and you get to explore different solution surfaces in real time.
And so it's this interesting balance the idea of, you know, businesses as a being, an experiential being, you know, doing stuff that has the other components, not only of intellect, but some level of emotional context as well.
And just wait until we start taking all of these models that are making all these decisions and putting them in a humanoid robot right in the middle of our house.
Right.
And it's going to lay out the clothes for you.
It's going to negotiate with your child on if they want to wear rain boots
and shorts in the snow.
And that model and that robot is going to sync with a mainframe
and is going to become your family vault that's going to pass on to their kids
and to their kids' kids.
And you mentioned the movie Bicentennial Man with Robin Williams.
If you go through that movie, because it's a great example of exactly that,
because he lives through as the family, he goes through the generations.
So that humanoid will be the repository of the family knowledge.
It isn't so foreign in humanity, right?
For centuries, wealthy families or successful families would have libraries. And grandpa would write a book or write his findings, and that would go onto a shelf. And then the grandson, when he was coming up to run the family, would sit down and look at it. And they had their own. This was more, we spent more time behaving that way than we have in the, like the current societal structure. Right. We're like,
because information wasn't ubiquitous. I mean,
the internet changed that really quickly. Right. The printing press,
long route up to this, but for a very long time,
information was spread and passed through a family and the same thing applied to
trades, the same thing applied to,
I think now that we've gotten so ubiquitous we're
all so overwhelmed that we have to focus on something that we can actually use like we were
just talking about the multiple agents right it's cool to have a marketing agent there's a billion
of them though and which one does this correctly what so my company's gonna train my employees are
gonna train this agent, right,
through their behavior,
and we're going to save that data,
and then that's going to behave like my best employee,
and that's going to be the upper ceiling.
Someone was writing a paper on,
I was reading what he posted on Twitter,
what's it called, sub stack,
was what happens when,
okay, so Ryan or Lewis, Iis i have a problem right that i run
into you guys both have a vast knowledge of experience i trust you you're a trusted source
for me to go to and ask you a question and then test your answer myself right so where do we
create this for agents moltbook was something that came out recently,
which is absolute nonsense, technically.
Like, yeah, everyone,
it's literally just a blog posting thing.
And getting purchased by Facebook, right?
Facebook bought it.
Being so, like, people who aren't in AI
don't understand what Moldbook is, right?
I'm like, they're literally, like,
you told your agent to go post to Moldbook and it didn't.
And then, you told them to respond to Mold book. But what about when it really has a question,
it leaves a trusted story, right? When it's solving these problems that we're giving it to
that don't have, that just have best fits, not answers. Where can we create a place for models
to go to, to talk to other people with relevant people, other models with relevant experience,
and then create consensus and bring that back to that.
Whoever comes up with that layer, that infrared layer is going to do really well.
So maybe this might be-
I have a paper I'll share with you.
I wrote that paper back in November.
Oh, shit. I think there's something around this. with you back in November.
Something around this.
There's an article that says analysis of 1.4 million interactions show how employees achieve sophisticated AI collaboration and what it actually looks
like.
They found that sophisticated users treated AI as a reasoning partner,
shaping how it approaches problems by asking the model
to assume a certain role of perspective, providing concrete direction and examples, showing the
AI how to reason through a task, requiring the model to explain how it got to a response,
and offering ongoing feedback.
Rather than accepting first outputs, they refined the model's work over multiple exchanges
and applied it to their most complex and ambitious tasks.
They also set boundaries, specified structure, articulated clear objectives, delegated cognitively demanding tasks across brainstorming, analysis, technical guidance and problem solving.
For these users, AI was being used as a general cognitive tool, not as a narrow productivity aid.
tool, not as a narrow productivity aid. Crucially, these behaviors left visible, measurable
patterns that organizations can absorb. Sophisticated use correlated strongly with
four signals. How often users return to AI, how they persistently refined outputs, how ambitious
their initial requests are, and how intentionally they selected tools or models. So there is a group of us that are training ourselves to be like
these collaborative partners with AI in a much, much different way. And I'm just seeing lots of
patterns around that popping up. And that's true for both you guys. You guys are clearly doing
this as well. We're kind of past the top of the hour stuff.
So I think this has been awesome, by the way, guys.
I've really enjoyed chatting.
I mean, I think we came up.
I think we've invented a couple of approaches, products, solutions, and stuff already.
What will we do next week?
That's the surprising thing.
But let me close out.
And again, thank you, guys.
This has been awesome so so what we
covered this week was not just a pile of ai headlines uh it was a map of civilization under
redesign really power is shifting towards whoever controls rules uh the chips the grids the
institutional entry points and infrastructure is becoming destiny because ai now runs on land
cooling finance and public tolerance, not just code.
We're moving because people are starting to bring AI into medicine, intimacy, memory, childhood and identity.
And execution is changing because more decisions are being routed through agents to force and handoff systems than most people will never fully see.
most people will never fully see. This is your real niche, not generic AI coverage,
but the deeper rails underneath the story where technology is becoming power, infrastructure,
culture, and human consequence. But I think we still have agency in what we normalise,
what we regulate, and what we fund, what we refuse, and what we protect as deep and human.
If you want to understand where AI is really moving watch the rails beneath the headlines because that is where tomorrow hardens into reality thanks everyone it's
been really great and we'll see everyone next week let's wrap this up sounds good great show guys
