Intelligent Machines 862: Ménage à Claude

It's time for Intelligent Machines.

Paris has the week off, but Father Robert Ballacere joins Jeff Jarvis.

And just in time for a great guest, Roman Chaudoury is here.

She's the founder of Humane Intelligence.

She says we need to take back agency when it comes to AI.

You may remember her name.

She led the ethics team at Twitter until Elon Musk fired her and the entire team.

She's worked for the FTC, the UN, the US Senate.

She is a mover and shaker.

We'll talk to Roman Chaudry next on Intelligent Machines.

This episode is brought to you by OutSystems,

a leading AI development platform for the enterprise.

Organizations all over the world are creating custom apps and AI agents

on the OutSystems platform.

And with good reason, build, run, and govern apps and agents on one unified platform.

Innovate at the speed of AI without compromising quality or control.

OutSystems is trusted by thousands of enterprises worldwide for mission critical apps.

Teams of any size and technical depth can use OutSystems to build,

deploy, and manage AI apps and agents quickly and effectively

without compromising reliability and security.

Without systems, you can accelerate ideas from concept to completion.

It's the leading AI development platform that is unified, agile,

and enterprise-proven, allowing you to build your agentic future

with AI solutions deeply integrated into your architecture.

OutSystems, build your agentic future.

Learn more at OutSystems.com, Slashtwit.

That's OutSystems.com, Slashtwit.

Podcasts you love.

Front people you trust.

This is Twit.

This is Intelligent Machines with Jeff Jarvis and Paris Martin.

Episode 862, recorded Wednesday, March 18th, 2026.

Manage a Claude.

It's time for Intelligent Machines to show we cover the latest AI news,

robotics, and all those smart machines all around us.

These days are getting smarter and smarter.

Paris has the week off, but I'm very happy to say we've got Father Roberto.

I was going to call you Roberto.

I'm going to call you Roberto.

I'm very glad you're there, too.

Father Roberto is here.

From, of course, he's visiting us from the Vatican.

It's not a joke, folks, that's him.

Hi, Robert. Great day.

It's always wonderful to see you.

It's always a great day when I get to get to see you in the Twitter army.

I miss y'all.

Yeah.

Robert used to have a little place in the basement of the old Twit Studios.

That's not a joke.

That's actually not a joke.

It's true.

Also, of course, here, the Professor Emeritus of Journalistic Innovation

at the Craig Newmark Graduate School of Journalism.

The same works.

Craig Newmark.

York.

Jeff Jarvis, author of the Gutenberg Parenthesis magazine.

His new and hot type.

Now, delayed.

Did you can still pre-order it?

You can.

Yes.

Gives you no advantage.

It's now July, you said?

It's all because they were going to move it from production reasons from June to July.

And I said, no, that's death for books.

No, so they're moving it to the end of August.

So it's basically a fall book now.

A fall book.

Yes.

You've brought us, I think, one of our most interesting guests yet.

So would you introduce Ramancha Dury?

Well, I'm going to have the egotistical joy first of announcing something else that

we'll need to.

Oh.

Tell me about that.

So a big announcement.

I don't know.

I meant to make that big.

I mean, you know, you get some trumpets or drums or something like that.

So I'm proud and amazed to announce that Boomsbury Academic is a launch

in a new book series called Intelligence AI and Humanity, which is not a technical book

series, but it is a book series enabling writers from many disciplines to reflect on AI

and how AI reflects on humanity.

And Boomsbury, take the title again, Intelligence AI and Humanity.

Wow.

And I will be editing the book series.

Oh, man.

I can't believe it, but I will be editing the book series.

So I'm very proud to say that we have signed up our first three authors.

I'll mention the other two first.

One is Matthew Kirshenbaum, who's been on this show, who's writing a book about the

text apocalypse.

Another is Charlotte McElwain, who is at NYU, who's writing a book, a very hopeful book,

surprisingly, about race and AI and the opportunity to undo the oppression of technology on race.

And then we have with us a very happy, very proud to say the author that I was dying to

get to be the first author in this series, Dr. Ramon Chaudhary, who's writing a book

about asking the question is, what is Intelligence?

So oh, that's a great question.

Isn't it perfect?

That is the fundamental question, if you ask me.

So Ramon has a PhD in Political Science.

She is the founder of Humane Intelligence, which she'll explain to us, but as an effort

to hold AI companies accountable, I know her name from Twitter, where you were responsible

for ethics at that.

I was, I was the engineering director of machine learning ethics, transparency and accountability.

This is before Elon.

This is, oh, yes, I always say I worked at Twitter and not X, like shocker, I know you'll

be shocked to hear that my perspective and his don't align.

I know.

Who worked at the UN, the FTC at the US Senate.

Geeks might remember her from a death, DEF CON, where in 2023 she organized the largest

generative AI red teaming event in history, putting eight major AI models in the hands

of 4,000 people to probe for vulnerabilities.

I was one of them.

Yeah.

Yeah, Dave Robert was there.

Ramon, we're so thrilled to have you on intelligent machines.

What is intelligence?

Oh, okay.

So I've already written chapter one of the books, so let me just, let me preface this

with my, with my twisted point of view and you can say I'm crazy.

One of the things to me that's been most intriguing about what's been happening in AI, you know,

we've been trying for decades to duplicate how humans think with computing machines and

a lot of people say, well, you could never do it with a von Neumann architecture.

It's just, that's not how humans are massively parallel, blah, blah, blah.

But what's, I think to me, very interesting is that once we started using transformers

and started building these large language models with transformers, they have become,

they seem to have become more and more dare I say intelligent.

They seem more like humans, not, you know, a poor imitation.

But nevertheless, it has made me think lately a lot about, well, what are we then?

I mean, literally, all of us are just the sum of our, dare I say, training over this,

over our lifetimes.

Perhaps we're born like an LLM maybe with some instinct that it forms us to begin with.

But then as we grow up, we learn language.

And we learn all this through example, much like a machine does.

So I'm really thinking that one of the most interesting parts of AI is what it teaches

us about our own consciousness.

Well, absolutely.

So I want to tease apart many, many points you make that actually I've already started

exploring in the book.

And as I, I wasn't kidding when I said I've read it in chapter one.

This is an aggressive writing timeline because Jeff, what I think they want to launch the

first book, Q1 or Q2 of next year, which means I have to be done writing it by on again.

And we want your book to be in the first one out there.

The first book in the series, right?

So back to the fundamental question, what is the intelligence question?

And then there's the, how do we measure intelligence question?

And then there's the intelligence versus sentience question, right?

So cognition does not necessarily mean sentience or consciousness, because you said the word

like consciousness, right?

So one is every measurement of intelligence that we have today is fundamentally rooted in

economic value.

So the first part of the book really goes to intelligence is a social, economic and political

construct, right?

So why do we care?

So the basic question I ask is, what is it that is really like striking us all existentially?

And it's not just that these machines are performing the way we perform.

It is at our sense of self-worth and value is driven by this notion of intelligence.

But if you go back to how intelligence has been measured, it was constructed in the

first industrial revolution.

It was constructed around, so this is Alfred Benet, who was asked by the French government

to find a way to classify kids in classrooms to determine who would be a good factory worker,

who might be a good manager, who would be organized, who would be, so it was always rooted

around productivity.

So today, when Sam Altman says artificial general intelligence is the automation of all tests

of economic value, and we're like, what?

And it hits us hard in our core.

It's because the fundamental basis of what we call intelligence has always been about

workforce productivity.

But is that what intelligence really is?

And then we get into like the social and political magnifications, right?

So politically and socially, why do we care if we are intelligent or not intelligent?

Well, one aspect of it is that rights are given and denied based on it, right?

So justification of why it was okay, quote-unquote, to enslave black people was in large part rooted

in concepts or intentional misconceptions about intelligence.

I.e., you can treat these people like animals because they are no smarter than animals.

Women, why are women not allowed in higher education?

So, oh, because your little brains could not handle it, your intelligence is not there.

So we make these presumptions.

We design these tests to prove the points we want to make.

So to your point on the, you know, AI is a mirror, I would even say our construct of intelligence

is more about the fears of the economic ruling class and their attempts to categorize

us and put us, quote-unquote, in our place, then it is an objective measurement.

Anything.

So the problem is when this goes into computer science and we have the Dartmouth conference,

these men, they're all computer science and mathematicians, sit down with actually a very

simplistic understanding of intelligence.

So they presume intelligence has been mapped.

We know how to measure intelligence and people, that's their starting presumption.

So the second presumption they make which is incorrect is that, okay, well, we can break

down this thing called intelligence into its aggregate parts and you could just sum it

back up and it'll be intelligence and break it back down.

So if you know like basic systems theory, there is no system in which you could just sum

up the parts and then you get the system.

The system itself has some residual impact.

So there are like a lot of things.

One last thing.

So the other thing that interested me is in science, right?

How have we explored measuring intelligence in not humans?

Because one assumption about computer intelligence is for some reason, because we are very species,

centric, you know, animal, we have just presumed that human intelligence is the thing to model,

right?

But then what if we look at other ways of looking at intelligence, animal intelligence,

mycelial intelligence?

There's a whole field called extraterrestrial intelligence.

If we go to Mars and there's a moving slime, how do we know that slime is intelligent?

And like whether or not we should, again, why does this matter?

Well, because it can lead to ecological ramifications, it could lead to so many other things, right?

So there aren't fields of study.

And by the way, like newsflash, in 0% of these fields do they base intelligent measurement

on human capabilities?

In fact, that is almost the first thing you are told not to do because animals in mushrooms

etc have different ways of perceiving the world that are actually better than ours in some

way worse than ours.

But what you don't do is give a monkey a set of physics questions and say, well, obviously

we're smarter than you because you don't know what physics is.

So again, you flip the script and say, well, then why have we decided that these machines

need to be modeled after?

It seems like a pretty self-fulfilling prophecy then because these CEOs sat down and they're

like, oh, we need to do this model of the human brain and automate all the economically valuable

things as human brain can do.

So what we feel is really not an attack on our intelligence, but it's more visceral.

It's more visceral.

They just want to get rid of us as intermediary economic bodies.

I saw this TikTok with this woman said something like, companies seem irritated that they need

to go through us to get to our wallets.

And that is how AI feels.

Let's just go ahead and take the money directly, it is so much easier.

But I also think that there is an existential dread that comes from the thought that maybe

we're not special, that maybe what we have is a kind of intelligence.

When you say slime mold might be intelligence, that's threatening, too, right?

We want to think that we are somehow special.

Well, absolutely, and to your point, it goes back to how we construct intelligence, right?

So if it is constructed on economic productivity, and then we make an economic productivity machine,

then we're like, wow, we're not that special.

So then the last part is really, I've been playing with the idea of calling the book something

like the new intelligence or something like that.

It's like, wait, let's go back and then let's say, given that we have created a machine

that can surpass us in the way we have defined intelligence, a current measurement of intelligence,

right?

So let's actually create a method of understanding intelligence that maybe is divorced from

workforce part, because there are, by the way, many methods of intelligence.

So gardeners, multiple intelligences, right, there's kinesthetic intelligence, like

spatial intelligence, like dancers, for example, have this, they've built an intelligence

where they understand proprioception, their body and space in a way that like you and I

could not, right?

Because we are not trained in that intelligence.

Empathy is a form of intelligence, resilience is a form of intelligence, right?

There's all sorts of things that are not measured in SAT tests that we therefore do not value

that maybe we should.

I mean, you're going to hear a broad race view on this, yeah.

I absolutely love this idea of linking our understanding of intelligence back to the

Industrial Revolution, because yes, that was such an upheaval in society, that it makes

sense that that's when we were trying to quantify the definition of intelligence that

we use today.

In my tradition, there's a little bit different of an angle on it, and that is to separate

this idea of knowledge and understanding from intelligence.

Those two things are treated separately, because knowledge could be broke memory.

It could be the knowledge to be able to do a task, the knowledge to be able to complete

a process.

However, intelligence requires agency, and agency is that intentional desire to act upon

knowledge in order to affect the environment in which we live, and not just to affect

the environment, but to take accountability for the intentional actions that we take.

For us, for my tradition, intelligence looks like knowledge, but it has that additional

step of agency, which we still don't think that LLM said current AI has, because it

cannot act as an agent.

It can only act as a source of knowledge.

I am absolutely tickled.

I love this idea of using the Industrial Revolution, because you may know that Pope Leo is big

on the document Ray Romanovarum, which is what the Catholic Church released during the Industrial

Revolution, to introduce this idea of agency and bring this idea that there is something

innate and special about humanity, which is what Leo is talking about.

So how do you have a working definition of intelligence?

For us, intelligence would be the ability to take knowledgeable understanding of the world

and act in an intentional way to influence the environment based on values, goals, and

beliefs.

So for us, that's human agency.

That's the step in intelligence that we don't think AI currently has.

Roman, is this part of your book defining intelligence?

In a way, I frame it, because I'm a social scientist, I frame it more like sociotechnically.

What is it?

It's not enough to just define it, like I'm not a philosopher.

What I want to do is understand it in the context of the world, right?

So what are the ways in which we have defined intelligence, maybe even just sort of judgment

diagnostic and say, what has that meant in how things have been executed?

Because again, the fundamental question to me was always like, why is this idea?

Why are we so scared of this thing?

Why are we so scared of it?

What is it forcing us to look at or question about ourselves and what do we feel threatened

about?

Again, that's how I got to where I am.

But Robert, I love what you're saying about this idea of intent in agency.

And this is where we shift from whether it's intelligence to sentience or conscience.

And people conflate the two a lot all the time.

And again, if you talk to the average person on the street and you ask them what they

think artificial general intelligence is, they think of something like the Terminator,

her, the Scarlett Johansson's robot and her, the AI.

And those things had intent.

They acted with desire.

And there's nothing about these machines.

And also, by the way, this narrative is being pushed by tech companies.

It's very, very intentional.

Why?

I coined a phrase back in what 2017 or 2018 moral outsourcing, where essentially companies

anthropomorphize these models on purpose.

So that when something goes wrong and something is bad, they can say, the AI did it.

Right?

The AI did it thing.

Oh, and you see them doing it today.

Right.

And you see it starting with all of the tech layoffs.

Jack Dorsey saying, AI is taking jobs because AI is making it easier.

Like, sir, you invested in a bunch of crypto that tanked and you overhired like that.

Not our fault.

Yeah.

Exactly.

But now there's this very, very convenient intelligence shaped thing that you can put

the blame on when bad things happen.

Such a great phrase.

I'd like to hear you.

Go ahead.

If I can introduce one more uncomfortable truth about our system, our way of thinking of

intelligence is if you look at intelligence as that combination of agency and knowledge,

there is this fear.

And it's a very rare fear, real fear that there are humans who do not meet that definition

of intelligence who do not reach that level of agency.

So that is, that should also be on the board.

So I love that.

And I especially like it because one of the things I'm very, very focused on right now

is the future of education and the future of work.

And these are like institutional flaws that predate AI.

AI did not make, you know, the educational systems fail our kids.

AI did not make it difficult for a child, for a recent college graduate to translate their

degree into a job.

Like, that existed before.

How many of us work in the field?

Well, maybe some people in this work work in the field that they studied when they

are around it.

And most people don't, right?

Most people studied some thing and they ended up somewhere totally different.

And we've just sort of accepted that.

Most people will say what I learned in college is nothing to do with what I did even at

my first job, right?

I certainly am not in.

And that's fine.

There's nothing wrong with that.

But then we need to re-examine our institutions of pedagogy and say, well, how have we been teaching

and what have we been teaching?

And I have very strong thoughts about like decisions that have been made in the educational

system, but fundamentally the purpose of education.

So just to get that very specific, because again, AI and education is something I'm looking

a lot at lately, the pedagogy of AI is very, very problematic because we teach AI in

general as a tool of productivity, not a tool of mastery, right?

So if the purpose, and we've done the same in education and smart, quote unquote, smart

kids know how to game the system.

They're good test takers.

They know how to do all the SATs.

They know exactly what to say to the teacher and what they should write in their essays.

Some of them happen to love learning, not all of them do, right?

So we have taught education as an institution of productivity, produce X, Y, and Z, and

then you'll get into Harvard or MIT or Stanford.

And then we make, again, we make this tool that we're teaching as a tool of productivity.

But there is research, by the way, which is excellent, into AI as a tool of mastery.

But none of it's being taught to kids that way.

So like it is not, I guess my fundamental point is like it actually has nothing to do

with the technology specifically itself, but how we are framing our usage of it.

And that's also what's driving a lot of the fear.

We're talking to Raman Chaudhary, sorry, we're talking to Raman Chaudhary, the founder of

Humane Intelligence.

There's a non-profit, and there is a public benefit corporation.

Tell us about Humane Intelligence.

What's your goal here?

Yeah, so the non-profit was founded to build the independent community of algorithmic

evaluators, which is very, very needed.

So right now, essentially tech companies right their own homework, grade their own tests,

and pat themselves on the back about how smart they are.

And then when anybody listening to this podcast or sitting in this room tries to use AI,

however impressive it may be, it's like a smarty-pants technology.

If you try to use it for something very fundamental and real, you'll see it falls apart very quickly.

And as all these memes, I can't spell strawberry and all this up, but then there are the bigger

issues.

There's a lot of embedded bias in it, like that CEOs of these companies, and especially

thinking about GROC, have dictated how they want these bottles to answer certain questions.

So there's biases baked into it, and also the average person, if our lives are meant

to be impacted by AI, we should have a right to say how this tool is being used.

So the non-profit started as an organization that would try to cultivate and get people

excited about evaluating AI models.

The for-profit is specifically looking at how to build the infrastructure to do this.

So things like algorithmic transparency, technical methods of evaluation.

One thing I want to say, I like to get a little bit, like a little in the weeds on it.

So machine learning and AI, like narrow AI, like pre-generative AI stuff.

Those are like largely statistical models.

And as a statistician by background, like we know how to math those things.

Like we have over 100 years of, you know, mathy-mathing to figure out things, right?

The generative AI, you have probabilistic outcomes, and the way I describe it to people

is 2 plus 2, sometimes equals 3.9, sometimes equals 4.2, usually equals 4, but sometimes

equals 98, right?

So you don't have this consistent answer.

So in trying to evaluate this, it is hard to make a test that is scientifically

sound, something that's reproducible, something that's generalizable.

And these are all things we need to know if this model is going to work or fall apart,

and we don't have that yet.

So the for-profit, it's a structure of the public benefit corporation for many reasons,

is dedicated to creating the infrastructure.

So the nonprofit creating the community, and then the for-profit creating the environment

that they can do these tests on.

I'm looking at the Humane Intelligence Nonprofit webpage, and you talk about AI-RED teaming,

which is so important, but instead of having it be done by the companies that make the

models, having it, I presume, done by a community of people.

AI contextual evaluations, what's that?

Condextual evaluation is actually a phrase coined by my colleagues, Riva Schwartz and Gabriel

of Water.

They were both actually previously at NIST, and now run their own consultancy called Sivitas.

Contextual evaluations really mean, how do we give a test of a model that understands

the context in which it will be used?

I don't mean to use the word in the definition.

So for example, if I am a car company and I want to understand, I want to build a voice-activated

AI system in the car to help people whatever get directions or find the nearest gas station,

how do I do an evaluation of that that's not just some sort of a generic evaluation?

So things you might want to think about in that situation, how does the AI give an answer

that will be correct and not lead somebody to an unsafe place or distract somebody when

driving?

These are very specific things that today, the very generic and superficial testing tools

put out in Silicon Valley really don't answer.

So they don't answer their questions.

I do a lot of work with companies, and these are all tech companies.

These are companies trying to use AI, banks, insurance companies, et cetera.

And zero of them have told me that they have found the tools being built in Silicon Valley

to be useful for them, and they just do all of their evaluations in house.

They try their best to do with themselves, which is not a formula for success.

So they feel there's risk.

I mean, you have AI-red teaming, AI-contextual evaluations, a bias bounty, which I presume

challenges to find bias in these AI models.

That's right.

Yeah, so all of this really would be under the rubric of AI safety, yes?

Yeah, so that's a tricky term.

So yeah, well, because there's a lot of like, you know, like in the family fighting of

responsibility, safety, governance, and you know, sometimes the word, I don't mind the

word safety, I think it's fine, but for some people it's coded as existential risk, which

means there's a community of people that say, you know, AI has a 25% chance of killing

us.

And again, it's like, it very much anthropomorphizes the AI, it uses language like manipulation.

It talks about things like bomb threats and scenarios, and frankly, from my perspective,

I think sometimes that narrative is somewhat intentional, somewhat naive and privileged,

and distracts from the real harms we are seeing today, because we are busy speculating

on future harms that are not possible to happen, right?

So today, what do we have?

We have algorithms that deny people jobs that, you know, unfairly, you know, accuse them

of crimes that, you know, are used for surveillance.

Like we know those are actual harms that happen.

And instead, an overly significant part of this community funding, brain power policy

is spent spinning on terminator stories of AI has gone rogue.

Right.

Yeah, we've said this many times.

This is one of Jeff's favorite drums to beat.

It actually is the flip side of the coin moral outsourcing.

So on the one hand, you say, well, it wasn't us, it was the AI, on the other hand, you

said, but the AI could kill us.

It's all kind of the same.

Right.

Well, then you say, then I think it's so important, is that when you blame the AI that way,

you take away our agency, which is right, on the browser said, right?

And it acts as if we're powerless, that the AI is just going to take over everything

and there's nothing we can do about it.

And these companies want it to be that way.

There's a payoff.

There's a hubris to it.

There's an extreme hubris.

I'd love to hear you riff on the notions, the hebristic notions that they add of general

intelligence, super intelligence.

I'm not enough to say that you're as good as super human intelligence super, super human

who prevention intelligence.

I don't know.

Big go.

Exactly where it goes.

I said, I happen to say to you, and I also said, Leo, a paper this last week, that

John Locot is one of the co-authors arguing against this notion of generality that humans

aren't general, that we are good at some stuff and crappy and other stuff.

But this idea that these people are so smart, they can build the machine that is smarter

than all of us.

Is that a new plateau in this notion of intelligence as privilege and power?

So both Jan and Fei-Fei has raised money for her startup and Jan is also doing something

similar.

And I think this is what Sarah Hooker is doing as well.

Sarah just raised 50 million for her startup called adaptation, which I cannot claim to

know anything about.

But it sounds like what Fei-Fei and Jan have been talking, which is building world models.

So their argument and Jan is making this, I love Jan's hilarious, I think everything

he says is always correct, and he is not afraid to offend people.

What I say offend people like, you know, who fight the powers that be, not like us.

But he's always very, very correct on what he says is that, you know, that there is a

belief in the general public populace that these models are just linearly improving over

time.

And actually, they're not, right?

So the newer versions of chat GPT are better in some ways and worse than other ways than

previous models that came out.

So it is not true that models are simply linearly, exponentially improving.

And all you got to do is give them more data and, you know, more energy and that they'll

solve all of our problems.

So he argues and Fei-Fei argues that you need world models, which are AI models that do

more than just absorb a specific language knowledge it needs to understand the world

around.

This could be vision, it could be voice, it could be like lots of different things.

So I don't know, we don't have a world model yet, but this is what they're betting their

careers on.

And given that, you know, they are a quote godmother and godfather of AI, you know, I figure

they know what they're talking about.

So I find that very interesting.

And I was on this, like, debate show last two weeks ago, discussing, you know, whether

AI will take our jobs.

And that was the point that I made that actually these models are not simply linearly improving.

And we may have actually reached pretty close to saturation at the capabilities.

And right now, really, what everybody is doing, the bait and switch have happened in the

last year is actually that the models have not been improving.

What's been happening is they have focused from building foundation models to building

applications.

So you may see, if you have Google, right, all sorts of new AI stuff dropping every week.

So yes, they're building Gemini, but now they're actually saying, let's take Gemini that

exists today and like build these little tools, which is not necessarily a bad thing.

But again, this is not this world of super ubermensch intelligence that's going to, you

know, sit at your desk and drink your coffee and take your job.

That is a very, very different world we're talking about here.

That's kind of what, yes and while I was talking about on Monday, right, Jeff was that

we've moved into the age of inference that we've moved away from the age of building models.

And now it's about what the models can do.

It's, you've said, Ramon, that part of the problem is that the people in charge of all

of this are, well, the companies are making it.

So they obviously have a dog in the sun.

They have an axe to grind government, which you say doesn't really have the tech and you've

worked in government.

So you know, it doesn't have the technical capability to understand what it's doing and

what it's regulating this.

And then you also point out that we in the public really don't have any way to measure any

of this.

You know, it's a, it's a little bit of a black box for us.

What's the solution to that?

It sounds like nobody knows what's going on or nobody's expected to do anything about

it.

Maybe that's better.

Yeah.

Well, okay.

So I can, I can talk all day about policy.

It's funny because I was just talking to, I talked to a lot of policy makers and I'm very

heartened to see a lot of young, this would be like older Gen Z, you know, people interested

in running for office, specifically on a tech platform.

I think, you know, Mamdani has really emboldened a lot of people who want to see positive change.

So there are a lot of junior, but, you know, they will be the next generation of people

and all of their heads are in the right place.

So I'm very heartened to see they may not have the wisdom or maturity yet.

They'll get there, right?

If they are smart and they sound somewhat right people, they, they will get there.

So I, I think we are going to see in the next, like, I would say five to 10 years, which,

you know, maybe too slow, a sea change happening in DC that I think will, will in some ways

be quite positive.

Um, but, you know, one of the things that I, my kind of pie in the sky, like what is

this ambitious shoot for the star is things I, I gave a, I gave my Ted talk on this idea

of a right to repair and right to repair, especially for the reason I chose that phrasing is,

it really appeals to kind of like the old heads, right?

This idea that if you own a piece of technology or piece of technology influences your life,

you have a right to tinker with it and do stuff to it.

And, you know, the right to repair actually is more about physical devices like iPhones

and McDonald's off-serve machines, um, but I do give the example of AI tractors, John

Deer versus farmers, who actually learn to work with hackers and hack into their, into

their tractors, uh, because John Deer required that you work with a licensed, you know,

part technician from them, and again, this is a community of people who are used to just

tinkering with their own stuff, but, you know, it, they can't wait three weeks for someone

to show up, crops grow when crops grow, right?

Um, so this was a fundamental problem for them.

And I think we all, like, we all need to think about, like, what are our rights as people?

And it was, it was sort of meant as a thought exercise, and we've never had technology

framed that way to us before.

You know, when I was at Twitter, we did this, you know, we did this exercise, or we wanted

to understand what would it look like to give people more ownership of their timeline?

And we, I worked with, um, Dr. Sarah Roberts, who's the author of Behind the Screen, which

was the first book that exposed content moderators and, and, and all of the, the horrific things

that they have to see and do, just to make sure we get a sanitized internet.

And we worked with her to really understand how people feel about agency and ownership.

And, and like, the TLDR is that, like, everybody said they wanted agency, but nobody understood

what that looked like.

Nobody could help.

Nobody could articulate what that meant.

And to be fair to them, we have never been given that.

We have never been given ownership in agency.

So what does it look like to have a right to repair?

I'm not sure if I know, but I think a starting point is something like public red teaming,

right?

We're regular people.

So bring back to the red teaming.

And we, we, we purposely do these exercises with teachers, students, policymakers, like

the point is not AI experts in the room.

And it's, it's to break down that initial barrier people have when they say, oh, I've

done the AI expert great, but you're an expert in being you.

You're an expert in being a teacher or being a multilingual sociologist or a cultural

expert.

That's what we need more than more tech people in the room.

So that's a starting point.

A bug bounty for, uh, social harms kind of.

Exactly.

And that's, that's kind of what we did with NIST.

We did a project with NIST called ARIA.

And what we did was ask literally anybody in America to go on to our platform and evaluate

Gen AI models.

And that information went to NIST to inform their, their, their standards development.

And when I say to, I got literally gave it to the guy who manages my gym.

And by the way, he was super interested in it because he's like, hey, I have like a sidehouse

of where I make websites and I really worry that AI is going to come take my job.

I really want to do this.

So when I say everybody, like we, and the thing is like, that's the dirty secret.

Everybody can interact with it.

But this, like this mythos around it.

This like we're too smart for you and the technologies to, like, it's all on purpose

to make us not feel like we deserve ownership.

Yeah.

Pay no attention to the man behind the curtain, Robert, you wanted to say something.

Yeah.

I was wondering.

So, uh, back in the 2023 at Defcon in the AI village, um, two things that really stuck

me from the final analysis that came out of the event was the, uh, first, the recognition

that you had that sometimes closed models are required for security and intellectual

property, but that the creators needed to, to provide transparency on capabilities.

Uh, and I'm wondering how much of that you're seeing?

How do you actually see the, the creation, the creators of the foundational models explaining

what it is that they're, they want their model to be able to control what they wanted

to be able to do?

The main part was, and if I, I'm sorry if I'm not remembering this correctly, you were

talking about the democratization of desirable behavior that, that was absolutely something

that needed to come out of the red teaming.

We needed to be able to, to get together and in making policy decide how are we going

to regulate the, the reward behaviors of these models?

How much progress have we made since 2023?

And does the anthropic sole document do the job?

Yeah.

No.

So I'll work backwards.

No.

I had a feeling you might say that, but I, did you ask?

You know, I, I'm very cynical, if you've not gathered, uh, at the intentions of the people

who simultaneously are going to be billionaires in building the technology and yet, you know,

proclaim to also be the public philosophers who will cure it all and save humanity.

I'm like, I'm okay.

Uh, so you're going to point out the problems, but not do anything about them.

Um, but, but I want to talk about your, your question.

So first is, you know, this, this idea of, uh, closed versus open, and this is one of

the reasons why we need an independent committee of evaluators, right?

So think of literally any other industry, like that, that is impactful finance, education,

airline, safety, they to protect intellectual property, right?

But if you are a licensed evaluator, let's say a financial auditor, right?

If you have this license, you get, you have professional standards, you are allowed access

to things that a regular person off the street would not have access to.

You have guidelines in which you can do this testing.

I mean, we do this in healthcare as well, right?

Uh, so this is not completely on, like tech likes to think, you know, we're, the first

time anyone thought about, blah, it's not, right?

So we have done, we've created institutions, professions and systems in place to protect

IP, while also enabling independent evaluation.

This is why that independent committee is needed.

I think the, the public right teaming is a great tool for awareness raising, for people

to get demystified to learn how the technology is using, but working, but if we want to

talk about improving these models, writing good regulation, really understanding performance

and harms, that is like a different animal.

And this is again, like, with the for profit, why I want to build this infrastructure, we

need people who are skilled in doing this.

We need a way of, you know, understanding their expertise and, you know, getting, giving

them access to, and this could be legal protections, legal access to it.

It could be, you know, professional certifications, but this is why you need the, the profession.

And then the second of, you know, democratizing, uh, cywoops of race again, you said,

a desirable behavior.

Yeah.

Democatizing desire behavior.

Democatizing desire behavior.

Yes.

So one of the, uh, you can tell none of these people care about philosophy or social sciences

or anything like that.

Right.

There's this, like, very arrogant notion that they can arrive at, like, this universal

good.

Um, and I, I always find it really funny when people are trying to make these models, and

they claim it will, you know, have this constitution or, or values, the universal values.

Like, I have actually heard people say, like, oh, obviously, we all believe X and we

actually don't all believe X is actually very, very hard, it's not impossible to come

and even if you think about the most fundamental universal value, one might argue, which is

that, okay, well, human beings say we shouldn't kill other human beings.

Don't we though?

Don't we have the death penalty United States?

We do have state sanctioned killing of human beings.

We've actually said it's lawful and okay.

So we don't universally think that it's wrong to kill other people.

And one would argue that would be the most fundamental thing, right?

The most fundamental thing that we could theoretically say the universal agreement.

And yet we don't.

So yeah, there's this arrogance of this idea that we can come to universal, uh, you know,

this universal list of values, you know, one of the things I, I love to laugh about is

one of these benchmarks is called humanity's last exam.

Yes.

Uh, what a dramatic, but you look at it and so much of, like, physics and math questions

and I'm like, like, opt to be out of humanity's last exam, like, it's real thing, bro, isn't

it?

Yeah.

Yeah.

Like, like, really is that?

And by the way, there is this section, there was this counter paper by a bunch of people

that, you know, was trying to make a benchmark for, quote, universal values.

And I want to, like, I remember, like, the first thing I ran, I went to was what, what

did they consider to be global historical knowledge?

And it was Europe, America, Asia and other cool, cool, cool, cool, cool, yeah, yeah,

other.

You're living out there.

You might not agree.

Like the literal, the literal cradle of civilization, which is the Middle East and Africa

is.

But Europe gets its own America, which is the youngest of all of the nations gets its

own.

Right.

But, uh, yeah, it's like, this is, this is what these people come up with.

We're really glad we could spend some time with you.

I wish we had more time, uh, Ruman, uh, Chaudhury.

I look forward to your book, but I know you have 50,000 words to write by August.

So I don't want to keep overstay our, uh, can I ask one more time?

They call it painful.

Yes, please.

So I'm, uh, Leo and I watched, uh, Jensen Wands, uh, keynote, uh, I'm a connoisseur

of the showmanship of it, uh, every time he does it and we're going to talk about it later

in the show too.

Right.

Yeah.

And so at the end, he went Gaga over open clock and, um, I'm curious to hear about that,

but I was thinking about this in my world in media.

We went from a world where you couldn't make media unless you had the tools of, uh,

production and distribution, unless you had the capital, unless you had the equity to

do that.

And what the internet has obviously done is, it means that, that we can all entertain

ourselves.

We can all make media, uh, the culture makes itself fashion determines itself, and I celebrate

that immensely.

For all its, for all arms, I think it's better.

Um, the internet ended up being, uh, top down in a lot of ways, corporate, right, just

just just happened to media, uh, along comes AI.

And your co-author in the series, uh, Charlottes McElwain, uh, surprised me with a, with

a surprisingly optimistic view, having written the book, Black software about the oppression

the technology caused, uh, in Black America, he sees an opportunity to break out of that.

Now, so I'm finally getting to the point of open claw.

It does this mean that, um, we can all make technology, just as we can all make media

on our own.

We can all make creativity on our own now.

Um, is it possibly to not be over optimistic that this opens the door for us all to make

our technology now?

Does it, does that, is that a step to give us all more agency, even though the models

have to be made by the big boys and their old boys, I accept Feifei, um, um, but is there,

is there an opening here that the technology gives us the chance to take it over?

Well, and this is where right to repair comes in.

I fully, I fully agree with you.

Uh, I think that that is not, that has to be intentional and it needs to be, because

the tech companies will not frame things that way, right?

So they will, because they don't want that, we need to do that for ourselves.

And just, just as an example, my partner has been messing with us, uh, making like an

IOT system for our house, but one that's done in a way where we're locally hosting all

of our own data so that we're not sharing information.

We don't have like a ring doorbell, but, you know, like right now, we have, let's say,

like, simply safe, right?

So instead of that, and the thing is all of the tools now exist to do that.

And my partner, who by the way is an architect, not a programmer, but who has always

had a, like a passive interest in like IOT and automation and, you know, out of necessity

because we move, we're, we move around a lot.

We have it.

We are actually able to do that today in a way that we were not able to a few years ago.

So it just says one example, but, but again, like no one's going to sell you that, right?

So either we have to raise awareness among people that you can go do this or, or create

a counter movement to provide that service and, and give people that world.

Cause like I said, like AI and this new wave of technology was not given to us the way

the internet, the internet was given to us as a tool of free use and democratization algorithms

and this new wave that they get Sabier and Sabier every time they consolidate more and

more power and wealth and they're not going to give that up randomly.

We can make a counter movement that maybe is designed around things like right to repair

where we can just do stuff like this.

I was telling her that like she should build a side hustle.

I think it would go over well in places like New York, right?

Where you just do this for people, somebody pays you a bunch of money and you're like,

I'll buy you a server and a bunch of raspberry pies and set up a dashboard and like,

there you go, you can monitor your whole house and not, not a one bit of that data is going

to go to open AI or Amazon or anybody else.

Leo, that's your new business.

Maybe I love it.

Basically that you're, it sounds like your general argument is for human agency in all of

this, not to let the companies that are creating this stuff take that from us.

But in fact, and not to assume that it's a black box that we cannot have any understanding

or agency of, but in fact, to take that back just as we have with right over or trying

to was right over pair.

Is that fair?

Yeah, absolutely.

I think paramount to all of this that I think is important is the ability for people

to choose their path in life.

Like I may not agree with the way somebody, maybe somebody does want to give their data

to Amazon.

I don't know.

I don't care.

But like we don't have a market with choice right now and our choices are getting fewer

and fewer.

I want to create a market where we actually have choices that we can act on our values

because that's what a lot of people are expressing.

They have particular values about their personal information data, even passive data,

like your ring doorbell and how that's being used in ways that they are not okay with.

Right.

I look forward to your book.

You better get writing.

Well, I really look forward to it.

Your editor is sitting right here and he seems nervous.

So no, this is very exciting.

Tell us again, the, the working title, you can change it.

We're not requiring it.

Oh, I don't know.

I really don't know.

Maybe it's something like the new intelligence, critical thinking and cognition in the AI era.

The other one I've been working with is measuring minds.

That is the title of the first chapter because that's what, that's what all of this attempts

to do and do poorly.

Our publisher, Horace Nakhvi is very good at titles.

So he'll have a, he'll have a good, I will, I'm actually very, very bad at naming things.

When I, I built the first enterprise bias detection mitigation platform and I just called

it the fairness tool because I'm like, it makes things fair.

So, well, thank you for the work you did at Twitter.

I'm sorry that Elon didn't think it was important.

Rush news.

But hey, you know, it's worked out well for you, right?

You know, you're probably better off, to be honest.

Thank you so much for being here.

Thank you so much for a month, Chudori.

Thank you for having me.

Really look forward to the book.

We'll have you back when the book comes out.

That's what will happen.

Yeah.

Yeah, if not sooner.

Take your time.

You have one definitive future reader.

Yes.

And really, I really support what you're saying, which is we, we, we need to fight for our own

agency and all of this.

That's clearly not the, we can't let the frontier labs and the hyperscalers dominate this.

Just because we don't, I don't understand it is not, it's not going to work.

It's not enough.

And it's our reality too.

Government need isn't the solution either.

Unfortunately, maybe it will be in the future with a younger crew, but not now.

Thank you, Ramon.

I'm glad more of intelligent machines in just a bit.

Yay.

Wow.

Lovely.

Is this what you've been doing on the show?

This is, this is excellent.

Oh, you haven't listed me.

I heard one episode like a year ago.

Do you want to hear like an interesting story that my friend, Sarah Fina told me?

She's the, before you say it, I want to let you know, you're still on the air.

It's not part of the podcast that we stream live.

It's fun.

No, no, no, it's, no, it's totally, it's actually, it's a, it's an interesting story.

And it's going to be a, at some place in my book, it may actually even be in the introduction.

So, you know, we worry a lot about young people and overvaliance on technology and

come in, you know, critical thinking, et cetera.

Do you know that Socrates and Fadres was very, very concerned with the

advent of writing because to him, memorization was what was the mark of intelligence.

And he was concerned that all of his students would become stupider because now we have

this thing called writing and we have like freely available paper.

And they're no longer going to memorize.

So it's interesting because again, like as I work on things like AI and education,

people like spout their fears about critical thinking and overvaliance.

And, you know, as somebody who measures things, what I think about is,

well, what is the, that means the existence of overvaliance,

presumes the existence of the appropriate amount of alliance, which means there's undervaliance.

But nobody can tell me what like appropriate alliance is because they benchmarked on themselves.

But like our parents all told us we watched too much TV or sat on our computers too long,

right? Like, you know, like any of us who have kids probably told tell our kids

they're on their phones too much.

You know, like that is just, that is just, you know, parent to child,

like how, how that goes and we all worry the next generation is getting dumber.

And maybe they are and maybe they are and or maybe intelligence is shifting, right?

Because that was, this is socrates here.

This guy, nobody was talking about, right?

I think socrates is right.

You should start memorizing all that stuff right away.

Stop writing it down, stop, stop writing.

No computers, no phones, no more writing.

Well, and it's, it's funny because how many of us memorize phone numbers?

I could tell you my childhood phone number, but I couldn't tell you exactly how many people know

their own phone number because you don't need it, right?

Or I don't, I have a phone number, yeah, exactly.

What is a phone number?

We number phone?

Oh, the scientific American was a very good piece today, by the way, just in the side.

On the, it's the kids today arguing against that and saying the kids today are in fact,

in good shape and brings data to it.

Good.

So I hope that's true.

Love that.

Future's good.

I hope that's true.

Take care, Ramon.

By the way, love the hat.

It is not the IEEE, IHSTO, it's actually Portuguese.

It is, it's a, it's a, it's a, it's a sustainably sourced B-Corp in Portugal and they make amazing

organic cottons and linens, et cetera.

So I, I want to support a local sustainable business and by I will quit tech and do something

else job would be to open a textile shop in Lisbon because I, I, I have found joy in

tangibles.

The more I work on intangible things, you know, people are out here saying they want to

open a bakery like, hey, I am not waking up before I am to make croissants and be I'm not

dealing with the 9 a.m. coffee rush.

Absolutely not.

What I'm going to do is open up a shop where we sell beautiful linens and cottons and

fab and ceramics and only the people who I want to come in will come and still.

But I just want to let you know if people ask, you could say it stands for the industry

standards and technology organization that they have here.

I can, or I can get people to buy organic organic sustainable cotton.

Thank you, Ramon, take take care of my, thank you, I looked up ISTO and that's what I

found.

Wrong ISTO.

Wrong ISTO, exactly, exactly.

We're going to do an ad.

I've been installing Nemo Claw during the interview and I'm ready to load the claw.

All right guys, I'll see you later.

See you later.

Bye.

Thank you very much.

Thanks for coming.

Pleasure to meet you.

Very interesting stuff.

Yeah, I have more at least intelligent machines and our special guest, Father Robert

Balace, they're filling in for us, we know, and just a little bit.

Our show today brought to you by my domain registrar, spaceship.com, spaceship.com slash

twit.

When the member Paris wanted to do a website, secretly, British, and we registered a domain,

secretly, Britto.sh, well, I did it at spaceship because it was so easy, plus we had searched

around and it was also the best price.

If you've heard us talk about spaceship before, there's a reason it keeps coming back.

Spaceship is now really that one of the fastest growing domain registrar in history, it's

because spaceship is rethinking how people register and manage domains.

Its fresh approach has now led to six and a half million domains under management and

record time.

We just started talking about it a few months ago.

That kind of growth comes from, well, I guess giving people what they actually want at

a fair price, spaceship offers transparent low pricing on domain registrations.

By the way, if you're somewhere else, move your domains over, their transfer pricing

is fantastic.

Their renewal pricing is fantastic.

This means there's more clarity over what you're paying for over time.

So often the case that you, it's a dollar for the first year and it's a thousand dollars

for the second year, not at spaceship.com.

Alongside great value, the platform is especially built for flexibility.

You can instantly connect your spaceship, register domains to spaceship products.

We clicked a button and secretly British had an email site, you get web hosting.

If you want, we haven't, she hasn't set up her domain yet.

So I pressed a button that connected it to her existing domain.

But when we have a website for it, it'll be very easy to do hosting it on spaceship.

That professional email is first rate, even virtual machines.

So a great place to host your open claw, for instance.

And you can build and test before committing because almost every spaceship product comes

with a 30 day trial.

But if you prefer third party tools, don't worry.

No problem, just point your domain to what you need by updating your DNS records.

It's easy to do, we're name servers.

And actually they have a nice little AI called Alph that can do that for you.

So now you have the freedom to build your stack exactly as you want because they know this

is what we geeks want.

It's basically the best of every world.

Visit spaceship.com slash twit to learn more that's spaceship.com slash twit.

We thank them so much for their support.

Open claw.

Let's see.

I installed it.

I needed docker.

During the year, we were watching the keynote on Monday, you, me, and Michael Sargent,

Jeff Jarvis.

The keynote.

What did I say?

Just the keynote.

Just the keynote.

The keynote.

Yeah, right.

The keynote.

You know what?

And I will stand by this.

As I'm watching Jensen Wong masterfully spend two hours and some minutes describing all

their products.

I said this.

There is no CEO in technology today that can kiss the hem of his robes as the mastery

of his topic.

Yeah.

And boy, is that company doing all the right things?

So one of the things he talked about is the fact that open claw is the fastest growing

open source history, a project in history, more stars than Linux in just a few months.

And he said, and so we're going to support this with a enterprise focused, safe, open

claw using something called open shell.

It's installed a bunch.

You can see my screen.

It's installed a bunch of stuff here, open shell, CLI.

It's apparently, I don't have, it says NIM requires an NVIDIA GPU, oh, of course, but

we can use cloud inference.

Well, you can buy that now there's a new Dell that has the small cost.

And now I have clicked the link.

It does say security risk, because it's HD to be it, whoops, this is it.

What am I seeing?

What is this?

What is this?

That's not, I mean, let me go back to local host, hold on a second, that's weird.

Is that what they wanted to show me?

He says, warning.

Oh, I suppressed go back.

No, no, no.

I want to go forward, accept the risk and continue.

And now, ladies and gentlemen, nothing.

Excellent.

I was going to introduce you to my new NIMO claw.

Well, go to advanced or go to view certificate, right?

That's true.

No, no, this is it.

Accept the risk and continue.

I can view the certificate, the open shell server.

I think it's running in my doctor, doctor.

So anyway.

So what makes it safely?

I'll explain that.

Well, it's running in Docker, which most people recommend you do with open claw anyway,

because if you're running in a virtual machine, you're a little bit safer, although Docker

can be misconfigured easily to be not safe right, Robert.

Oh, yeah.

Oh, yeah.

We just did that last week.

So yeah.

Oh, yeah, absolutely.

So that was one of the things they call it open claw with guard rails.

I think this is, you know, we've said this before open claw show that this is the era of

the year of the, of the agentic AI.

In fact, I thought that was the thing that was very interesting.

And I mentioned this when we were talking to Ramon and we certainly noted it during the

keynote.

Jensen Wong saying it's the era of inference now that it's about what you do with these

things.

Right?

And it was a good thing that he bought into Groc with a queue.

Oh, yeah.

Those chips now, those chips, very important to what they're doing.

You kind of rolled your eyes, Robert.

I mean, yeah, we're still building models, right?

They're not still building models.

They want to get us into this post foundational era of AI, but we're not there yet.

Yeah.

Yeah.

I understand why they want to though, because all the new money making applications seem

to be in inference.

Oh.

That was CES.

CES is AI booth and the part of the West Hall was all about inference, using inference

in driving, using inference in home appliances, using inference in security.

So yes, I get the push because they see that as an untapped market.

But if you look at the sustainability of the current inference model, it's not there.

I don't think it's nearly as profitable as they think it's going to be.

Jensen, for take a little bit of a victory lap, here's the picture.

I said this would be the picture, this is the picture of him with his WWF belt or something.

He said, and this spiked the stock briefly until people thought about it, that in video

it was poised to sell a trillion dollars worth of black, well, and Vera Rubin chips next

year, trillion up from 500 billion.

Is inference just another word for application?

Does this mean that this is an effort to get the industry going into retail channels?

So when we talk about the foundational model, that's the traditional, we're going to shove

a bunch of data into this training and then we're going to get something out that we can

use.

The inference model is sort of continuous training.

So we now deploy a foundational model, but it starts to learn through its interactions

with the real world environment and it goes back into the training.

To be fair, Claude code, which is one of the hottest things in AI right now, is basically

an inference machine, right?

Yes.

Actually, one of the things I've been really thinking about lately, my goal with Claude

code is to basically replace myself.

I know.

Coat do that.

No Leo, no.

Well, actually my job, the reason I started to it in the first place, the whole point of

this was to, for me, just to do the stuff I like and have, initially, how other people

do the stuff I didn't want to do, like edit the shows and produce the shows and technical

directors.

I just want to walk.

My goal was, from day one, 20 years ago, just to walk in, sit down on the microphone, turn

it on, do a show, get up and leave, being done with it.

But instead, I've created what Claude Dr. O'Call's a reverse centaur, which is basically

AI is making it more work for me, not less.

So he talks about a centaur, like a computer is like a centaur, which is a human machine beast

instead of a human and horse beast, is a human and machine beast, where the horse does

the carrying and the human gets to, you know, beyond top and look around.

And so the work is being done by the bottom part of the centaur, a reverse centaur, the

humans doing all the labor, doing all the work, while the AI is sitting there looking

around.

And I kind of, in a way, created that with my workflow, because now I have to spend hours

every day going through stories.

Admittedly, once I have the stories, it generates the rundown, and it does a lot of the, you

know, the busy work.

But I realize the piece that's missing is I want it to, and this is the hard part, somehow

encapsulate my editorial judgment.

Now in the past, you would train a model, maybe for that, but I'm thinking I can create

a small language model based on a bigger model.

The bigger model has all the language capabilities, and in the small model, train it to have my editorial

judgment.

You think that's crazy, Robert?

It's not crazy.

I see an exceptional expansion of requirements for power and other resources, because if

you are using a small model and then training it with inference, you are constantly going

back and you have to retrain, retokenize your data, otherwise it's not really truly learning

from the inference data that you're giving.

Oh, yeah, you're right.

Well, one of the things that we were looking at doing is using some sort of, maybe, Bayesian

system or something to train it using articles I didn't choose and articles I did choose.

I haven't.

Pretty big database of articles I looked at and didn't use and articles I looked at and

bookmarked.

That could train it, actually, Darren Oki suggested something, kind of an exotic technique

that he says is working really well for him, something called, what was it, SLM, do you

remember SLV?

I think it was.

SLV.

I'm sort of, he said he tried SVM, he says SVM, that's it.

He had tried Bayesian and other statistical tools, linear SVM, same idea, I guess.

Anyway, I'm going to try, I'm going to play with that, but the point, my point being,

that's kind of inference.

That's like, I'm not going to train a big model.

That's already done.

I'm going to, it's not exactly trained, but you know, is it way, it's really well, it's

never ending.

But that's fine, because it's always going to get better.

I mean, it would end eventually, I guess, if it's somehow said, oh, I get it, Leo,

you just like this kind of stuff and don't like this kind of stuff.

Every time you tell it what you want, you are training it to know what you want.

Right.

And at some point, even it's possible to conceive of a time when it's done, but maybe not.

One thing I do like about the inference model is it lends itself to local models.

Exactly.

To individual models, because cheaper local models.

Right.

Exactly.

I mean, the reason why it was so hard pushed in the automotive section is because they were

saying, look, we want to create a model for full self driving, but it learns your style

of driving.

Exactly.

It learns how you drive, not just how everyone drives.

And we don't want your driving to affect the model for other drivers.

We don't want, because right now the Tesla drives like Elon does.

Yeah.

Exactly.

And he rolls through stops.

That's not good.

Actually, you know, BMW has announced this new class model and their new i3, which they

just announced yesterday.

They say the whole point of the self driving is we're going to learn your style.

So they're on top of this.

What if you are a bad driver?

Well, they're a bad driver.

Shouldn't it correct for you?

Well, I think it will keep you from running stop signs.

And it'll stop at stop lights, even if you maybe wouldn't.

But perhaps you're more aggressive about lane changes or less aggressive.

My Tesla was always more aggressive than I would be.

If it would learn, no, don't change if there's somebody 100 feet behind my wife.

Yeah.

Did Jensen Wong's announcements about yet more auto companies he's working with on self

driving?

Does that torpedo Tesla and Musk, to a great extent?

At this point, you can't really torpedo Tesla and Musk, because they are in such a

bubble.

They're in such a bubble that they can torpedo themselves.

That's about it.

They can torpedo themselves.

That's right.

Yeah.

Yeah.

Grock, which was a multi-billion dollar acqua hire, really.

Grock with a Q.

Since we talked about the Q.

Since we talked about the Q.

The Q for Nvidia is a server chip.

They licensed the technology that didn't actually buy the company designed to make AI servers

more cost efficient, but things like AI coding for inference in effect.

And the Grock system will begin shipping in the third quarter of this year, according

to Huang.

And it's going to be made by Samsung, which was kind of a surprise.

I thought that Nvidia was a big TSMC client.

I think they still are.

But they still are.

They're just backsed about.

Yeah.

Samsung is going to be making these.

It's not a GPU.

Grock integrates memory onto the chip.

It's really built to do this kind of thing.

To speed up this kind of communication with it.

Yeah.

The other thing they announced, DLSS-5, which really, I don't think is important.

No.

No.

No.

But it really made people upset, certain people who make certain things.

Gamers really didn't like it.

The idea was it takes existing assets in a game.

And locally, you know, pretty's them up.

Somebody else.

Remember when we did this to TVs?

How much people loved it?

Oh, you think it's like, it frame interpolation?

Basically, yeah, that's the same thing.

It's creating something out of limited information.

So maybe you get a couple of frames that look great, but most of the time, you're going

to be going, huh, man, hey, this has been even good.

My problem with this is that it changes the art direction of the game.

I think here's an example that I think is quite good myself.

This is me.

Actually, I think this somebody generated this on this.

It looks like you as Tom Cruise.

Yeah.

It's the same place.

It even made my eyes blue.

Yeah.

You see, it's exactly how I look.

Can you do a dissolve instead of a jump cut?

I think it'll be a little more authentic morphing.

The morphing into, well, so, and Jensen Wong was a little, actually, a little pissed, shall

I say, at all of this.

His reaction was, well, he just, he just don't get it.

You don't get it.

You don't get it.

You can control this.

I'm not going to control this.

You can control this.

Tom's hardware asked, and Paul Alcorn from Tom's hardware asked about the criticism.

He says, well, first of all, they're completely wrong.

The reason for that is because as I have explained very carefully, now I don't, I haven't heard

the recording, but I can see him saying this, as I have explained very carefully, DLSS

5, fuse his controllability of the geometry and textures and everything about the game with

Generative AI.

Oh, well, in that case, no problem.

That's right.

You know what?

I think Anthony Nielsen, our own, Anthony Nielsen got it right when he said, it wouldn't

have been so upsetting to people.

Had he shown it with the backgrounds instead of the four grounds?

Yeah.

But really what bugged people was it, he showed it with people.

Well, so Benito, is it your fear that it takes away artistic agency that it changes the

graphics?

It changes the graphics in a way that don't screw with what I made with what the people

have made.

Yeah, it changes it.

Well, the people who would use this, I presume, are the game companies themselves, right?

That's what I was thinking.

No, that's happy with Nielsen.

This is something that will be a technology in your computer that you could eternal.

Yeah, this is so that the companies don't have to do this themselves, so that they're

right.

DLSS has it traditionally been something you would turn on as it's like ray tracing and

something you turn on.

Yeah.

If they're able to, I don't know if the company show, show my screen because this is some

example.

It's not always, by the way, beautification here is from Hogwarts legacy.

It's turning the older woman into a really older looking woman.

It's adding lighting and shading.

It is changing the look a little bit.

I don't know.

It doesn't bother me as much.

Gamers historically have really been negative about AI.

Their personality is a bunch, yeah.

Yeah.

It doesn't bother me as a game-rated bothers me as an artist.

Yeah.

But if a game company can use this to make their stuff look good.

Yeah, that's pretty good.

If you're an artist, you can use this to make it more realistic.

What's wrong with that?

Yeah.

If it's in your control.

Anyway, this is one of those demos where we'd have to see it.

Anyway, this is just a video that Nvidia created.

But it probably got more attention than anything else, Jensen Wong-Match.

Well, in certain circles.

Yeah.

Well, I mean, I'm willing to test it out in about five years when I can actually afford

to buy one of their new ones.

Well, well, that's another problem, so yeah.

Nobody can afford this technology.

Well, Audrey Carpati doesn't have to afford it.

He got the first Dell machine given to him by Jessica Wong.

That was the DGX Spark.

They announced that a number of third-party OEMs are going to be able to make these spark

base.

With black well chips in the machines.

Our own Daron Okie has purchased one for $5,000 Australian.

So do you have this end of NB and coveted this?

Yes.

You want to do that?

Yes.

I have that's in in spades, my friend.

I mean, no.

I got one downstairs.

Haha.

Oh, I forgot.

I forgot.

Robert has one.

From CES.

Yeah.

I forgot.

You brought one back.

Yeah.

Yeah.

It was swag.

It was boot swag.

Yeah.

Was it on your seat when you went on the way home?

You get a spark and you get a spark.

Everyone in the spark.

I honestly don't want to have to need a $5,000 piece of hardware to do this stuff.

And that's why I bought the framework, which was expensive, $3,000, but it had 128 gigs

of RAM as a strict sale.

I'm interested in models.

I can run on that locally.

That's where the retail level excitement comes.

Well, do you know what a DGX spark is going to cost from Dell and they didn't put the

price up?

Well, it's around $3,000 to $5,000.

It's $3,000 to $5,000.

Yeah.

Um, and they'll be, um, I think DARREN's was a sus if I remember correctly.

Yes.

I saw several OEMs.

They'll make it.

Super micro headwind that was the ugliest thing ever.

It looked like a tower.

Yeah.

You don't need to be.

It doesn't need to be a tower.

Yeah.

Yeah.

It is, it is very, very nice to be able to do a model locally and have the firepower to

basically go cold hog on it.

So what does that let you do that you wouldn't have done before?

Uh, so, I mean, or first of all, out of the box, anything artistic, anything you want

to do with video or photo, that's, that's a no brainer.

But what we've been doing it, we're using it for is for translation models because we

do with a lot of languages here and at the same time to do summation of the conversations

that are happening happening in different languages.

It's extremely effective at that and I will, I will not lie if you don't need a frontier

model to do those kinds of things.

We don't.

We don't.

But it is, but we do need the privacy because the conversations that we have here are,

are closed for.

So we cannot in any way, shape or form use cloud-based infrastructure.

Right.

Like sense.

Yeah.

Yeah.

I think also a lot of us will do a hybrid thing, for instance, with cloud, you know, we'll

use Opus 4.6 for the really high-end stuff, but we can use Quen or Kimi or something else

for stuff that doesn't need so much power.

What model do you use locally, Robert?

Uh, I don't know.

I handed it over to, um, to our IT guy.

So he's running all of our models for us.

This sounds like Chinese models.

This sounds like running a Sun Microsystems computer circa 1995, you know?

Yeah.

Yeah.

Yeah.

Yeah.

Yeah.

In a couple of years, it'll be, you know, a lot less expensive.

It will be a lot more affordable.

I mean, I honestly, I'm not sure I agree with the Blackwell ones.

I'm not sure I agreed with Raman, and I wasn't going to challenge her because she's way

smarter than I am, but I would, I'm not sure I'd agree with her that, uh, we're flat lining

with LLM.

And pull lots of discussion.

We have all the time.

I don't think that's at all in evidence.

It's the argument that, that, that Yon and Fei-Fei make is that maybe we're not flat

lining, but it will only take us so far.

I would point out that they are just as self-serving as Sam Altman.

I mean, how much money do they just raise?

1.03 billion.

Yeah.

Yon.

So, yeah, of course.

Oh, arm.

Our way of doing it's much better than what the other guys are doing.

Yeah, but they had an argument that a lot of people bought.

I think that, um, and you even have, I mean, this is a big deal and not much was made

of it.

Yeah, because I went to a debate between Adam Brown at DeepMind and Yon LaKoon.

It wasn't meant to be a debate.

It turned into that overall just this and DeepMind was still scale, scale, scale models

against there.

And Devices Abbas has switched recently and has been talking about the need for world

models and that LLM's alone won't get us there.

Of course, defined there as the other issue.

Right.

Uh, and certainly I wouldn't argue against that.

I mean, the more kinds of data, the better, but I've been thinking about this lately,

because their argument is, well, you can't describe the world in text, except isn't

that how, uh, we work as humans, essentially, that cats don't, is there argument?

Uh, okay.

But humans, this is a slime argument.

My brain, there's other, most of the stuff that I, when I'm thinking about something,

I'm thinking about it in words, right?

How about words perhaps informed by my knowledge of the physical universe, but ultimately

words?

What if you weren't limited by words?

I have to think about that.

I have to translate everything into words.

Because that's the way I operate, but a dolphin doesn't.

Yeah.

Thought to, thought to text is lossy.

Sure.

Yes.

Okay.

Yes.

So now you're saying we can make something that's smarter than humans.

Ha!

Trapped me.

Did you?

So, see those words were pretty good.

The limits of tokenization.

This is the limits of tokenization, and I see the technical plateau, because when you're

dealing with, with tokenization and the need to address so much storage at any given

time, for any given answer, it's, we're at that level where, until we get to quantum

computing, we can't advance that much further.

Do you have a computer downstairs?

I have.

Traffic just gave us a million token context window, which is absolutely fine, but being

able to run that model as fast as we would need to, what we're, what we're doing instead

is we're creating models that are specifically good at a thing versus the human brain, which

can be very good at many, many things.

We can switch gears very easily.

Models cannot.

Oh, wait, not against the idea of having physics models, and as many models as you can.

I'm just quibbling with the sole argument, oh, we've tapped out LLM's.

I don't think that's true.

Well, but the other thing that, the paper that I sent you to, that was a co-author of,

his argument was against this notion of general intelligence, saying that every human

being is good at some stuff, and crap, you know, the things.

There's no such thing.

And same as machines.

And it has interesting outcroppings as well, because what LeCoin argues is,

that if you think about specialized models, you can also limit the model to what it does.

Right.

Safe rates are safer.

No, and that's what I was just saying, which is, we've got these general-purpose LLMs,

but the future lies with special purpose, smaller language models, you know, specially trained

models, special, I mean, absolutely.

We're not going to throw out the LLMs, we're going to still use those as the base.

But I really absolutely think that we are going to specialize.

I was watching a video this morning, Australian fellow, who was a video about small language

models, because it's very interesting in this notion, who he's an Australian, he said,

one of the problems we have in Australia is a lot of sun and a lot of skin cancer, but

we don't have a national skin cancer screening program.

So he created an iOS app, this is part for a cackle competition, an iOS app that is really

interesting.

And tell, it's not diagnostic.

You take a picture of something a mole or whatever, you take as many pictures as you want,

and it saves that.

It does describe it, and then next year you take another picture, it remembers the things

you took pictures of, and then you can look at the change from year to year.

So it is like a self-exam that you can then send to your doctor, and that's based on

a very language model that can run an iPhone in just about three or four gigs of RAM.

And it's just categorizing, not diagnosing.

And I thought that was very interesting.

That's a perfect example of a specialization that is safer and useful.

The coin at all, making this paper, is that if you have two models, one is trained to

just do protein folding, the other one is trained to fold proteins and you're longer.

The first is obviously going to end up better faster at protein folding, because it's not

an essence distracted by other tasks.

And I think that that makes perfect sense, and it doesn't distract at all from the power

of the model.

In fact, it's a way to get more powerful models.

Yeah.

And pool says in our discord intelligence is what you are capable of.

Inferences, what you do with what you know, these models already know so much, that's

why the focus is moving to inference.

I would agree.

I would agree.

The specialized models, that is the inference model.

That's inference.

Yes, exactly.

Exactly.

All right, let's take one more.

Not one more.

No, no, no.

Let's take another rate.

Oh, no, no.

Let's take another rate.

And so that's GTC.

I was, I enjoyed it.

I'm really glad we covered it.

Jensen Wong is an amazing fellow.

And Nvidia clearly is fine-lank firing at all cylinders, and they have many cylinders

to fire.

They are very, very hot right now.

And you know what?

I wouldn't put it past them to have a trillion dollars in revenue in the coming years.

She claimed he would have.

In a year, I think.

Yeah.

He said, 2027.

Mind blowing.

It hits nice to have Father Robert Balacea, Paris will be back next week.

But it's great to get you on.

You've never been on this show.

I think this is a good show for you.

I haven't.

Right.

Well, I actually worked you on once when I wasn't here.

No, I was going to take it once when Leo was going on vacation, and then he didn't

go on vacation.

Right.

Right.

That's right.

So we haven't been together on this show.

Well, I'm definitely going on because you know what I got, actually, I'm, I got, so one

of the things I was very excited about when I first heard about Starlink way back in

the day before Elon, when Elon was still someone even.

Was the notion that I would be finally able to travel and do the shows from anywhere.

And I'm going to order a, because I'm going to Hawaii in May, and I want to do the

shows from there.

Oh.

And I, so I ordered a Starlink, the shirts, I can't wait for the shirts, a Starlink mini.

You could put it on your balcony.

I should be able to do the shows anywhere I can get a clear view of the sky.

It has plenty of bandwidth to do the shows.

In fact, we often fall back to Starlink in the studio when Comcast dies on us.

So I'm setting up a portable studio.

How much does it cost?

It's not much at all.

If you do have the consumer version, right now we have a business account, which means

I have to go to Costco or Best Buy and buy it as a, I have to wear a hat and a mustache

and buy it and stick it under my raincoat and say, well, it's, yeah, I'm a consumer.

Did you know they don't let you take those on cruise ships?

For good reason.

Yeah.

And they don't want to buy new YC.

Yeah, you don't want to use it at sea either, right?

It's not as fast at sea because there are fewer downlink stations if you're in the

ocean.

Oh, yeah.

And because cruise ships use Starlink.

Exactly.

Yeah.

They all want to know this.

Are you taking Claude along on vacation?

Claude's coming.

I've already set that all up.

Well, I really do want an agent.

I really do want an agent.

I can't wait.

Do you start playing with that?

Yeah.

You're going to go crazy.

Well, so my personal opinion on all this, I have tried all of these, everybody, the

latest is the president of a white combinator, his name, Gary Tan, who just made, who's

just put out his own, these are basically skills for Claude.

Everybody's done it.

There's GSD getting stuff done.

There's superpowers.

I've been playing with something called PAI, personal AI assistant, open clause, just

a variant on all of these.

The idea is you load it up with skills and API keys and loops so they can run continuously.

That's just, really, it's just Claude.

The whole thing is, but people love about this is how good Claude is.

And then they're just putting plastering layers on top of it.

And sometimes I think it really is better just to use vanilla Claude.

So I think what I need to do is kind of strip it all out, take all that crap out, all these

skills and stuff.

And I'll keep your improved skill.

That's a good one.

Darren's skills are very good.

But I strip out most of that stuff, maybe right a few of my own skills are a combination

of prompts.

And then you can't put code in there.

It's one of the reasons coders still have an advantage.

You can put bash commands.

You can put code in.

It's a combination of all of those.

For instance, a good skill, a skill I want to write is a tweet API skill, which would

be everything that Claude would need to access our API.

It would be a first step toward, I don't know, replacing all the humans.

And then anyway, no, I'm kidding.

I think I am.

No, I don't want to replace the humans.

I think the humans are the most important part of our whole workflow.

What I want to do is replace the busy work.

Well, I had to be in day with my colleagues at Lotter State and also at the New Jersey

Hills Media Group, which is a small newspaper company who's bored.

I just joined.

And the AI genius from, who we ought to have on the show at some point, who watches the

show.

Hi, Joe.

Amditas was taking them through things they could do.

And at some point, there are some writers who aren't good at copy editing who always make

the same mistakes, blah, blah, blah, blah.

And the one hand, everybody could use Claude.

On the other hand, you could just email the article to a project on Claude with your

instructions already there.

And it could do its magic and send it back.

These things can be that simple that you don't have to have everybody be able to do that

now.

That's right.

That's what I'm saying.

The triviality is the power of it.

Right.

So really, I think that's what a lot of this is a gentics stuff really.

Yeah, it's just that kind of thing.

Other ways to interface with the brain that is somebody saying, at least there's going

to be jealous, that ship is sailed.

Honey, I'll be back in a bit.

I just want to go up into the attic and visit Claude.

My little friend Claude.

Does Lisa play with Claude?

She does.

I've been working on her bit by bit.

In fact, now she says, how many subscriptions should we get for the team?

Because she's blown away by the kinds of things she can do.

It's a binagia Claude.

My bag.

My bag Claude.

Show title.

Yeah, I think so.

Thank you.

Thank you, Jeff.

I'll show it.

We'll have more in just a bit.

I'll show it today.

Brought to you by OutSystems.

Oh, I love OutSystems for this.

They're the number one AI development platform.

OutSystems helps businesses bridge the enterprise gap to this agentic future we've been talking

about where the constraints at the past give way to unlimited capacity and scale.

And the thing I love about OutSystems, they've been doing this for decades, that they're

not new to the game.

OutSystems enables businesses to build AI agents that can actually do work, take actions,

make decisions, integrate with data, much more than just to answer questions.

OutSystems provides the only AI development platform that is unified, agile, and enterprise

proofing because they have been doing this for a long time.

They started with low code.

And now with the addition of AI, they have the most powerful tool I've ever seen.

You can build, run, and govern apps and agents on a single unified platform.

It's agile.

You can innovate at the speed of AI without, and this is important, compromising quality

or control.

It's really important in enterprise that your AI is doing the right thing, not the wrong

thing.

And this is enterprise proofing.

OutSystems is trusted by enterprises for mission critical AI applications and durable

innovation.

OutSystems is the secret weapon behind the world's most successful companies.

And by the way, not just for small one-off apps, OutSystems works with the massive complex

systems that today, right now, are running banks, insurance companies, and government services.

OutSystems even helps companies with aging IT environments bridge the gap to the AI future

without a rip and replace nightmare.

OutSystems provides the safest, fastest way for an enterprise to go from, yikes, we need

an AI strategy to, we have a functioning AI application, and it does it safely.

Stop wondering how AI will change your business and start building the agents that will lead

it.

Visit outSystems.com slash twit to see how the world's most innovative enterprises use

out systems to build, deploy, and manage AI apps and agents quickly and cost effectively

without compromising reliability and security that's OUTSYSTMS.com slash TWIT book a demo

you will be impressed outSystems.com slash twit, we thank them so much for their support.

Of this week in intelligent machines, let's see, so much news, so much I'm going to skip

through the Google.

Oh, this is interesting, Meta, taking a little left turn, it's kind of the, maybe more

of a U-turn.

It's a bit of a drunkard's walk, shall we say, remember when they spent billions of dollars

to acquire Manus, and I'm sorry, a scale AI, sorry about that, a scale AI, they're doing

a reorg according to the times of India, maybe this is suspect, I don't know.

They are reorganizing their guide they got from scale AI, Meta's chief AI officer, Alexander

Wang, is still there, but he announced the company is going to cut 600 people from the

super intelligence labs division, Wang wrote by reducing the size of teams, fewer conversations

are needed to make decisions, and everyone will carry greater responsibility with broader

scope and impact, and we'll save a lot of money.

The teams include Wang's research lab, the applied AI engineering organization will also

receive big cuts, this is Sabah, Amar Sabah's team, another acquisition, Iraqwa higher,

and so it's a complete reorganization, only two people left when they're equity vested

in November from Wang's team, so that's good, but maybe we're just going to move some

people around.

It seems like Meta, remember they're Avocado model, which is going to be their new big replacement

for Lambda, was pulled back, it's not good enough, it's not good enough.

Is Meta the new altivista?

Yeah, they're struggling, but you know what, it's interesting to watch all these companies

except anthropic and open AI, and I guess Google, kind of journal today had a story, maybe

the times, Google's in the capitol seat, I don't know if that's true, I don't know if

it is either, but I don't know, I don't know who's in the capitol seat right now.

Google's really doing what our guest was talking about where they're looking more at applications

than they are at big models, they did release Gemini 3 deep think, right, but they're

also like, in fact, I skipped through the Google section, but they're adding maps, stuff,

they did scrap the health tips because they were getting those from Reddit, turned out,

not a great source for health information?

No, no, well, it might be better than RFK Jr., but not much.

They are going to do an agent builder for the Pentagon, but it's only unclassified work.

Nonclassified, I thought.

Yeah, that's what I said, unclassified, yes.

And non-same thing.

So it was not classified, although you saw that now open AI is kind of jumping in the

fray.

They have not up to now been approved for classified work, but the Pentagon says, okay,

we don't like these anthropic guys, so maybe we'll let open AI into the behind the

iron curtain.

Man, the OG tech companies have more running room here.

I mean, you've got Google that's burning billions of dollars for AI, but they have income

as with that income.

Right.

I know if the AI deal doesn't go through, they die.

And actually, you could even extend that to Oracle.

Oracle has bet so much on AI.

They're heavily leveraged.

If it fails, they lose not just Oracle, but the Ellison media empire crumbles.

I do.

I mean, I agree with you that these companies that you throw Apple in there too, Apple,

Google, and Meta have other revenue streams, so they don't have to make money on AI

right away.

But we're not seeing the results.

Meanwhile, anthropic and open AI who are running on a razor's edge are our big leaders

right now.

Maybe that won't be sustainable, you know, that's the possible.

That's probably what these companies think is what we can sit back, certainly Apple's

thinking that we can see that.

I mean, they're just leaders because they're investing in each other, but like Meta through

how many billions away on the Metaverse?

I mean, yeah, it's embarrassing, but it didn't kill them.

They're killing, by the way, they're killing MetaQuest's horizon world.

It's going away.

It's over.

Wow.

Well, similarly, open AI, MetaLike is saying, okay, we're going to concentrate now.

We're going to concentrate on B2B, which is Hello, Anthropic.

Well, they go with it.

They go with it.

What is others doing?

Yeah.

Well, Anthropics doing it in a prize.

They said, yeah, maybe all this diversification, the chat and all that stuff, maybe we should

do the same thing.

Maybe this device thing.

How much should we spend to get Johnny Ive here?

Hi.

Okay.

Here's my thought on this.

If agentic is the thing, and I think it certainly looks like it may be a thing, you need

an interface.

What OpenClaw and a lot of others do now is you use Telegram or a Discord or Apple

message or something to talk with it.

What you really want is a much more convenient way of talking to it.

I was thinking I really would like to write some sort of tool that I can use with one of

my pins or maybe my Apple watch that I could just say, hey, Claude, I got an idea or

remind me later to do this.

That's the way it should be.

I think that's what they're going to end up doing.

It's part of the agentic.

It's the interface to agentic.

Do you need a device to do that?

You need a unique device to do that or do that or do that or do it or do it or do it.

I don't know.

I don't know.

I don't think you want to take your phone out of your pocket.

I think you want something ambient, whether it's glasses, earbuds, watch, ring, pendant,

you want ambient intelligence.

The same way you really would, what I would really like to do is just shout into the void.

Well, the ambient intelligence belongs to Amazon and their deployed base of ambient

devices is second to none.

Here's another example of this.

It's a company that has great revenue streams and cannot seem to make a decent AI.

Alexa Plus is horrible.

Oh, yeah, it is.

Even the people inside the company don't want to use it.

So maybe the urgency of we are going to run out of money any minute now is pushing

anthropic and open AI faster and they're doing better because of it.

Yeah, they have to spray.

Maybe these other companies.

They have to.

They don't have to.

They don't have to.

They don't have to.

No.

But who won that race?

The tortoise or the hare?

Oh, the tortoise did.

Okay.

Meta didn't buy the malt book for bots.

Says tech crunch.

It bought into the agentic web again.

They bought agents.

Malt book for the hype.

That's what they bought.

Well, malt book is a social network for AI.

So a meta is social, right?

I think you're right.

They bought the hype.

I'm trying to get that far on.

Meta knows how to panic.

They're a panic here.

Meta's the one who knows how to mine that data.

That's all about the data.

There's no data.

That's why I was sad when Meta bought the limitless pin.

Well, that's what I bought.

Remember about the B computer?

Remember I bought the B computer and then Amazon bottom.

Then I bought the limitless pin and then Meta bottom.

I, you know, if Apple does an ambient, I think ambient intelligence, that's the phrase

I'm thinking.

Well, it's like the Leo is just walking on the street screaming, yeah, I want a milkshake.

Exactly.

I drink you up.

I mean, the problem with ambient is there's so many of these purchases that feel like panic

purchases.

Yes, exactly.

It's going back to the early dot com where you had to do something with your money, especially

with Meta, right?

Meta is the king of, I don't know what we're doing, but write a check.

Well, I imagine somebody's running into Mark's office saying, oh, we'll keep us.

We can buy this one.

And if somebody doesn't come to him before that, he's going to get mad.

He's feeling no about it.

I know the feeling where you feel like, I got a lot of money, I'm going to buy that

stupid computer.

There's also the privacy issue when it comes to ambient, that ambient stuff is that it's

listening all the time.

So there's always a privacy concern there, right?

Maybe you, I don't mind.

Maybe most other people.

We don't do ambient computing here.

You can't.

Right.

It's no.

But that's at this advantage.

Yeah, I mean, that's one, I mean, honestly, the, the way you do it is you have it.

You tap something or you, you, well, you need to trust the third party, like

prayer is ambient.

That's true.

I didn't even think of that.

You're asking the ultimate intelligence for help, exactly.

Somebody once told me there are only two prayers in the world.

Thank you.

Thank you.

Thank you.

And help me.

Help me.

Help me.

Is that fair, Robert?

I would add one.

Oh, my God.

And that can be taken so many different ways.

That can be help me.

Help me.

Yeah.

And there is one more, which is help them help help help help help help help help.

But you asking for help, or you're giving, you're giving thanks.

Manus, the AI agent startup, the meta acquired the Chinese company, the meta acquired last

late last year, has, as of the 16th launched a new desktop application called my computer.

It's on my head now, Leo.

I appreciate that.

Thank you.

Oh, you are a lucky one.

Bringing Manus's agent directly into your personal device through my computer, the agent

can read, analyze, and edit local files, launch, and control applications, execute multi-step

tasks, including coding tasks, without the user having to upload anything to a server.

It's local.

It's going to compete with perplexities, computer, branding is not, well, not great.

And the Chinese government is a little actually concerned.

Manus is a Singapore company, but it runs out of mainland China.

And the Chinese government says, hey, I would give you, I'm on a little bit of a word with

you.

So, how are they doing that?

Is it an act like a virtual machine or a container on my local?

This is from the next web.

The key architectural difference between Manus and OpenClaws, the model layer beneath the

agent OpenClaws open source can be run with any model, right?

It's quality depends on which model you choose.

Manus runs on Metas-owned proprietary model stack, which the company says is more consistent

and capable at the cost of a subscription fee, but is it local?

I don't see how that could be local, it has to call out to the server.

Yeah.

So, you know, analyzing what's on your desktop, it's sending it somewhere.

Your computer doesn't have the power for that.

Sending it to Meta.

Exactly.

And Anthropic has this, of course, Cloud Code work, OpenAI created their version of that

as well.

Everybody's trying to do that.

Basically taking the coding platforms, Cloud Code and CodeX, and making it so that non-coders

can use it, but I don't know.

I still need to know how much, it's some of it goes out.

It sounds like they're making a good faith effort to keep everything local, but local

when it can.

The intelligence doesn't work like that, local when it can't.

OpenAI released two new models, today chat, GBT-5, for Mini and Nano.

Oh, you're complete Mini and Nano.

How big is Mini?

How big is Nano?

Let's see.

Nano is the smallest cheapest version of 5.4 for task, or speed and cost matter most.

It's a significant upgrade over 5, Nano.

There's a new Buick.

Yeah, there's the benchmarks, which I don't pay too much attention to.

Let's see the numbers.

Show us the numbers.

Yeah, come on, size should be relatively right, the first thing, right?

All these benchmarks.

In the API GPT-5, for Mini supports, text and image inputs, tool use function, calling

web search, file search, appears, 400k context window.

That's good.

That's bigger, twice as big as Cloud Code's context window, it's only about 75 cents per

million input, 450 per million output.

Mini uses only 30% of the GBT-5 for quota, letting developers quickly handle simpler coding

tasks and codecs for a third the cost, and you can use Mini subagents, which I do that

with Claude.

I use Haiko and Sonnet for subagent work, with it aren't too demanding.

Nano, let's see, Mini is available to free and go users via the thinking feature in

the plus minute menu.

For other users, Mini is available as a rate limit fallback for GPT-5 for thinking.

Nano is only available in the API, and Nano is 20 cents per million input.

Significantly cheaper, about 25 per million output.

That might actually be a good foundational model for like an inference build, because

you're already living in the scope.

So yeah, that's all the information I have.

This is from OpenAI.

OpenAI has signed a deal with AWS to sell its AI services to government agencies for

classified as well as unclassified work.

This is their opportunity to get in the door.

Microsoft has now threatening to sue them.

I say, no, you're ours, OpenAI, we can't do a deal with AWS, you're ours.

Additionally, Anthropic is owned AWS, right, and that was a big advantage for Anthropic.

But OpenAI has really jumped in the breach, but Microsoft says that's a breach of our

contract, and they are threatening to sue.

So trouble?

That's a trouble and paradise thing.

They were friends.

Well, I mean, come on, that's been going on for more than a decade, right?

That kind of a break.

Microsoft and AWS.

Oh, yeah.

But I was talking about OpenAI and Microsoft, right?

Microsoft gave them $10 billion.

Microsoft also gave Apple the money that saved them.

So 150 million.

Yeah.

They're very good at that.

Yeah.

Maxwell's F writing and wired inside OpenAI's race to catch up to Claude code.

This is what you were talking about, Jeff, kind of a repositioning.

So they still want to do the adult chat.

There is no controversy within the company.

The safety people there are saying, this is really a bad idea.