What even is AGI?

In this episode, hosts Mark and Shashank dive into recent developments in generative AI technology. They begin with NVIDIA's latest GTC announcements, including partnerships with GM for self-driving technology and advancements in robotics with Google DeepMind and Disney. The hosts debate the merits of camera-only versus LiDAR-based autonomous driving systems, referencing Mark Rober's viral comparison video. They also discuss NVIDIA's upcoming Ruben chip, which promises a 10-15x performance increase over the current Blackwell architecture. The conversation shifts to a correction about the DeepSeek model's computational requirements before culminating in a thought-provoking discussion about the challenges of creating generalized robotics systems and how simulation environments might accelerate development. Throughout the episode, the hosts share insights on what these technological advancements might mean for the future of AI and robotics.

Episode Transcript

(00:00):
All right, and hello everybody.

(00:01):
So I wanted to introduce ourselves.
I know that we're getting a lot of new listeners to the podcast.
So I'm Mark, along with my co-host, Shashank.
We do a roughly weekly podcast on Genai and anything that's kind of related to Genai,

(00:22):
even tangentially.
So on this show, we'll talk about the news, philosophy, the market, whatever.
It's all about Genai and things that are tangentially related to Genai that we find interesting.
So Shashank is a super smart Google engineer who is working on AI and he knows a lot about

(00:49):
a lot.
And myself, I am an unemployed X Amazon engineer now doing a bit of traveling and learning
Japanese.
So together, we do this podcast and today is going to be kind of exciting, but we have,
and we have quite a few things to talk about.

(01:09):
So we wanted to briefly touch on the Nvidia announcements today.
So Nvidia had their GTC announcement, which was pretty cool and they came out with all sorts
of interesting announcements.
And then we have a correction to touch on from last week.
And then what we really wanted to focus on today is a kind of evergreen topic, but what

(01:37):
is AGI?
So Shashank and I have been doing some thinking about this.
And we think we have some insights on it.
So hopefully by the end of the episode, we'll come up with some sort of objective metric
for what is artificial general intelligence.
So yeah, Shashank, do you want to start us off with some of the Nvidia announcements?

(02:00):
Yeah, let's see.
I think I can try to give a high level overview.
So it seems like there wasn't a big focus on a big flash.
It was a lot of previous announcements that were kind of panning out their black wool chips

(02:21):
that are in production, their investments in their omniverse, their investments in robotics,
both in humanoid robots and self-driving technology.
So they're partnered with GM to propel their self-driving tech, previously having partnered

(02:43):
with Mercedes.
So it seems like Nvidia is kind of like the Microsoft operating systems.
They're going to partner with a lot of hardware manufacturers and provide the intelligence
that's going to power all these machines.
Starting from cars to robotics to who knows what else.

(03:05):
Which is kind of interesting.
If I can dive into the GM announcement a bit.
I used to have a friend who used to work at Cruise who since left and joined Weimo, which
is a good thing because Cruise is no longer alive, which is very unfortunate.
They had unfortunate, it is unfortunate.

(03:28):
They had a rocket ship trajectory.
They were acquired by GM in 2016, I believe, for close to half a billion.
GM spent reportedly over $10 billion into this project that they were hoping would pan out
eventually.
But after a few high-profile mishaps in San Francisco where they had gotten the permits to operate

(03:56):
and offer rides, truly self-driving, driverless rides in the city, I think they crashed into
an ambulance.
They got stuck.
They blocked traffic.
And a whole bunch of high-profile news that didn't look very good on Cruise.
I think they fired the CEO.

(04:17):
They completely disbanded the project and abandoned that company entirely.
Fast forward a couple years now they're partnering with Envidia to focus on self-driving technology.
Now that others have caught up, Weimo has successfully been operating in San Francisco, L.A., Phoenix,
Arizona, and a couple of other cities.

(04:38):
Tesla is also improving on their self-driving technology, having switched from their previous
architecture to the transformer, which seems to be this miracle architecture that is
amazing at everything.
So I'm glad that Envidia is offering this platform to every single car manufacturer and
I'm hoping that all of the advancements in their omnibers, their virtual simulated environments,

(05:03):
which allow you to create training data to train self-driving cars on billions or trillions
of virtual hours of video footage.
It's incredible.
And I'm kind of looking forward to it, but it's a shame that some of the exact same GM weren't

(05:29):
more forgiving of the fact that this is a hard problem.
People have been working on this for a long time.
And I think we just needed to wait for compute to reach the capability that we needed for
these cars to be able to drive themselves.
Yes.

(05:50):
Speaking of the self-driving cars, have you seen the new Mark Rober video where he compares
the LiDAR to the Tesla's self-driving?
No.
I wish LiDAR.
So it is the Lumin, I think, the Lumin, I want to say Lumin R.
Yeah, so Lumin R is another one of the self-driving car companies.

(06:16):
And specifically focus on LiDAR.
They just fill LiDAR.
If I remember.
Right.
And that sounds right.
So he was comparing the, so for those that don't know, Mark Rober is a really popular science
YouTuber.
I think he used to work at NASA and now he just makes YouTube videos, but he did a video

(06:41):
where he was comparing the Tesla self-driving versus the Lumin R LiDAR.
So Tesla actually very famously made all of their self-driving cars with cameras only,
whereas Waymo and a lot of the other self-driving car companies are using LiDAR.

(07:06):
So for those that don't know, LiDAR is a different system that kind of can complement cameras
that uses lasers to sort of map out the environment around you.
So from my rough understanding is it can, it uses a laser and it shoots out and it can

(07:26):
measure the distance to other objects around you.
And then the laser will shoot back and then you can use that to kind of make a real-time
3D map of everything around you.
So it works really well.
So the LiDAR is like, you could think of it as a little cone with a laser rotating really,

(07:51):
really fast and shooting bursts all around it.
And measuring how fast the light takes to return and with that they can estimate the distance
and also create a 3D map from all the little points that go in a circle around and return
back that way in real time as the car is moving, they'll have like this point cloud around

(08:12):
them and create a 3D model from one or multiple LiDARs that all combine into this model of the
world.
Yes.
So this type of thing is very, this LiDAR technology is really useful for areas where cameras

(08:33):
will fail.
So for example, if you are in a very rainy environment where you can't really see in front
of you or like maybe it's very foggy, well optical sensors like cameras, they might not be able
to see through the brain and they might not be able to see through the fog.

(08:53):
But a LiDAR, it can actually still map an environment even when it's the kind of rougher
environments like rainy snowy things like that.
And actually in this video.
And very bright daylight because my Tesla always fails when the sun is shining right into
the cameras and it's like just panics fails and it's like we can't do it.

(09:16):
You take over.
Yes.
Yes.
So actually it turns out that in the video that Mark Rober came with, he even made one of
those kind of loony tunes walls.
So I don't know.
It's kind of a little bit hard to describe, but he took a road, right?
And then you know how loony tunes they would have maybe like a road going into a brick wall

(09:41):
and then they would go and then paint the remainder of the road onto the wall.
And then like, you know, the bad guy would go and then speed his car or whatever right into
the wall because he thinks it keeps, he thinks the road keeps going and then he would just smash
into the brick wall.
Well, they actually did a real life version of that with the Tesla.

(10:04):
So Mark Rober, he had the Tesla on a road and then he had a wall that from the angle of
the car looked exactly like the rest of the road.
And then the Tesla just ran right through it.
Now there's a little bit of controversy around this because some people think that the Tesla

(10:26):
turned off its self-driving like a couple second, like a fraction of a second before he hit
the wall, which either maybe the Tesla knew something was going on so they could claim,
oh, hey, it wasn't our self-driving that did this.
And maybe it's like a Tesla bug or some people are accusing Mark Rober of like physically
deactivating it, but he had like Mark Rober who was driving the car.

(10:52):
He actually had a full video and it doesn't look like he deactivated.
It looks like almost Tesla deactivated the self-driving right before it hit the wall.
But regardless, no matter if it was deactivated or not, the Tesla self-driving was active

(11:13):
like right before it hit the wall and then it went right through it and actually for good
measure, it hit a small mannequin child right behind the wall going right through it.
So yeah, this is the problem with the camera only technology, but I guess Lidar doesn't

(11:35):
have these types of videos.
Lidar doesn't have these types of issues.
So I think that I'll post a video in the podcast description so you guys can watch it, but I highly
recommend it.
It's not that long.
It's really interesting, but it just shows that this is a really difficult problem and
I applaud these companies like Cruz for teaming up with companies like Nvidia because together

(12:03):
they might be able to actually be able to solve this really difficult problem of driving.
Yeah, personally, I think that isn't a common use case that we need to be worried about.
I mean, how many Wiley-Chiotes are painting fake walls that people are going to run into?

(12:28):
I think the more challenging issue is weather conditions and new environments that the AI
can't generalize to because it's never seen that or encountered that type of terrain before.
I kind of get Elon's vision.
I think five something years ago when he was deciding between Lidar or just cameras, he said

(12:52):
that the world is built for humans with eyes.
We don't rely on Lidar and humans are able to drive cars in literally every environment
relatively well.
We should be able to build an AI that has similar capability.
I think he may just be a little off on the timeline when these intelligent machines are

(13:16):
able to do this and I think eventually they will be.
Yeah, I don't know.
I didn't really want to talk about the decision for using cameras, but I think that the main reason
why Elon probably picked cameras, and I don't know, I'm not in his mind, but I would think

(13:38):
that a large part of the reason to choose cameras over cameras plus Lidar is primarily a
cost decision.
In practice, you can have a Tesla with only cameras.
You can have not just a Tesla, but any car with only cameras and it should be able to achieve

(14:00):
probably 99% of what a human can do.
I think that in most cases, when it's sunny, when it isn't really bright, when the weather
conditions are bad, you can have the camera only driving and it would probably work really
well.
I think that it's not those types of situations that I'm worried about and I think that at

(14:27):
least for me to trust whether a self-driving car, like in order for me to trust a self-driving
car, I want it to not match him performance, but I want it to be way better than me.
I think that in the difficult situations, for example, when the road is really icy or
it's raining hard or somebody put their bright lights in front of me, I want the car to be

(14:53):
able to handle that in a way that's not just meeting human drivers, but surpassing it,
right?
I think that if we only are ever going to meet human level intelligence or not human level
intelligence, but human level driving ability, that is, in my opinion, not good enough, right?

(15:14):
Because lots of people die every day behind a wheel.
Driving is one of the most dangerous things that you can do and lots of humans get an accident
all the time.
I think that we shouldn't strive for human ability.
We should strive for, I don't know, way better than human ability so that we can reduce

(15:35):
the total amount of deaths on the road.
I think that, I personally think what Tesla decided to do with the camera, only technology
is saving a little bit of money at the expense of people's lives.
Because people, I think, will have the Tesla work 99% of the time and then they'll get into

(15:59):
a situation where the Tesla will run into a wall or not see the car in front of us.
Because the sun is in the eyes and I think that's just not acceptable.
So I personally don't like what Tesla is doing with their self driving car tech being camera
only and I think that in my opinion, it is doomed to fail.

(16:25):
When I say doomed to fail, I mean it will fail because it will never likely exceed humans
driving.
I think that if you only ever match human driving, I just don't think it's good enough.
So I see where you're coming from, but for our listeners, all of this opposition to self

(16:46):
driving with just cameras is coming from a guy who bought a $300 kit off of an open source
project, put it on his dashboard and let that thing drive his car for him.
Yes, so that's true.
I did do that to my car, but I think what the difference is is I was fully aware of the

(17:08):
risks, right?
Like I knew that the self driving car, what or not the self driving car.
So what I did is I bought comma AI.
So comma AI is a thing made by I think is the comma AI company.
Their CEO is George Hots and it is basically an Android phone that you can go and then put

(17:33):
into the cars dashboard.
So if you ever have like kind of like an external like dash cam or you have an external GPS, you
can kind of put this into your cars window and then it will hook into the cars OBD two
ports. So if you ever go and then take your car to a mechanic and the mechanic puts the

(17:59):
computer into that port, they can get a bunch of diagnostic information about the car.
So basically this is essentially an Android phone with a couple of cameras to facing the
road one back at you that could go and then hook into your cars lane keeping assistant
and it can make your car self driving.
So I was fully aware the limitations of this and I had no kind of impression that this

(18:26):
would be a fully safe thing.
Like I only would use this on the highway and I would pay full attention to the road
in front of me.
But I think that the reason I have so much hesitancy is the average person I think would
not be aware, right?
So if you have a Tesla and you call it autopilot or you even call it something like full self

(18:50):
driving, if I see a name or a product called full self driving and it only full self drives
99% of the time, but then in the 1% of the time that I actually want it, it gets in an accident
and everybody in the car dies.
I don't think that's worth it.
Or maybe not everyone in the car dies or that's too extreme, but you know, you hit the

(19:12):
car in front of you, right?
Who run into the wildy coyote wall, hit the car in front of you, whatever it is, I think
that that is not okay.
And I think that the bar for me who is fully aware of the capabilities of it should be,
but that's a lot lower, right?
Like I knew what I was signing up for when I bought that thing.

(19:33):
There was no lane on the comment AI's website that it did full self driving.
And it's main claim was that it was a better lane keeping assistant cruise control.
And I think that if you go and call something full self driving, that is negligence, negligence

(19:54):
and also dangerous because I think it gives people the false impression of what's possible.
Yeah.
I think I do agree with you there.
Our marketing was a little hyperbolic, especially when the earlier models came out.
Now it's gotten closer to full self driving.
And I think it is full self driving.

(20:17):
Like you mentioned, 99% of the time.
I haven't used the latest models.
I still have like the early version from like six something years ago with like their earlier
version of the computer.
I think I might need to like upgrade the computing onboard computer on my car to be able to run
the newer models.

(20:37):
I can't even use the newer AI.
So I'm still bullish on Tesla as a company.
As controversial as Elon is.
And you know, I would never work under him that seems like a very stressful gig for me right
now.

(20:58):
And he shouldn't be let anywhere near any kind of social welfare programs.
But in terms of running a company with speed and having ambitious goals, I think he I think
he's doing a pretty good job.
But yes, the marketing can be a little hyperbolic.

(21:18):
But without getting stuck into a rabbit hole here, I think some of the other announcements
from Nvidia speaking of hyperbolic announcements.
They announced their latest and greatest rubin chip, which is supposedly way better than

(21:38):
the black hole chip, which still hasn't shipped.
So let's mark.
What do you think about the new chips that are coming after the new chips that they announced
last year still haven't been released?
Yes.
Yes. So it seems like the new chips are going to be way better and cheaper than the ones

(22:04):
before it.
So Nvidia is announcing these new chips to be released.
I believe it's in 2027.
So they claim that the reason that they are announcing this is a lot of these companies
who are building out data centers, they want to be well informed about what is going to

(22:26):
be built.
So if you want, so for example, if you are Google your Amazon or you are open AI or whatever
and you have a lot of compute needs and you're planning on buying a lot of GPUs, it would
be helpful to know a lot of the things which Nvidia is creating in order to build like plan

(22:53):
your data center spend appropriately.
So if people know that this is coming, they can do their planning and kind of put enough
money in the piggy bank metaphorically so that they can spend on these new GPUs.
So Ruben should be kind of the next successor to Nvidia's black well chip.

(23:20):
So the black well chip is Nvidia's current latest and greatest chip which is not widely
deployed.
Yes, it is not widely deployed, but I believe that it is still out there and this claims
to be a roughly 10X isch slightly more like 1510 to 15X.

(23:47):
So it's like roughly an order of magnitude performance increase plus it should be a roughly
order of magnitude cost decrease.
So for the video watcher, like the people who are watching the video, I'm sharing my screen
and showing one of the screenshots from the GTC presentation where Jensen Wong is sharing

(24:15):
some of the specs for the new Ruben compute and I think the exact numbers are not important.
And I think that there is a lot of incentive for these companies, I'm not saying Nvidia
is budgeting the numbers, but I think there is some incentive to be show the most optimistic

(24:35):
view point or show the most optimistic numbers for how these chips are going to perform.
But according to Nvidia, it will be a roughly 10X increase in performance.
So that is good.
And then also if we want to talk about the numbers just really quickly, the Ruben chip

(24:58):
should have, they claim 3.6 X-A-Flops of floating point operations of inference.
So an X-A-Flop is a really, really big number.
It's a point of mildly.
So I think that in most computers we are dealing with gigaflops, right?

(25:28):
So that's like a billion operations.
I believe it's gigaflop billion operations.
And then Kilo is a thousand mega is a million and giga is a billion, a petta is a trillion
and I believe an X-A-Tera.
Tera is a trillion, a petta is a thousand trillion and an X-A would be a million trillion.

(25:57):
Yes.
So a lot of people will say that the human brain is roughly like 20 petaflops.
So the fact that you can have 3.6 X-A-Flops, this would be a lot smarter than a human.
But I don't know, this chip hasn't been created yet.

(26:23):
It's coming, they're announcing it for the second half of 2026.
But these are some crazy numbers and basically the key takeaway here is that the really fast
they can hold a lot of things in their memory and they're expensive but much smaller.

(26:45):
And they're going to be really good.
So Nvidia's cooking and I hope that they're able to deliver because it's very exciting.
Yeah, just for reference, I was looking at the faster super computers and apparently that's
Al Capitan at the Lauren Slippermore on National Lab and that is about 1.7 X-A-Flops.

(27:11):
So this is going to be like double or triple that world's fastest super computer.
But I think there's a caveat to be made here.
I believe that one is a CPU super computer.
Whereas this, the GPU super computer, it can't do as complex of a task but it can do simple

(27:35):
calculations like matrix multiplications at scale.
So it's slightly different but really impressive numbers nonetheless.
Yes, it's definitely impressive.
And I wanted to kind of sort of go upon this topic and talk about something that we touched

(27:58):
on last week.
So last week we were talking about the Gemma model from Google.
So really quickly the Gemma model is a new ASE Open Source base model that Google came
out with.
If you want to know more about it, you can listen to our episode last week.
And it's way more efficient.

(28:20):
Google had a graph where they were comparing the Gemma model to deep seek.
So deep seek for those who have been living in Iraq is one of the top reasoning models from
China.
And I went out with a, it came from a startup or a hedge fund out of China.

(28:41):
And I was under the impression that deep seeks compute requirements and memory requirements
were significantly not that high.
And in the Gemma graph or in the Gemma chart from Google, they claimed that the compute

(29:02):
requirements for a Gemma were like way less than that of deep seek.
So I was questioning whether Google was lying about the compute requirements for deep seek.
But it turns out that I was wrong.
So I wanted to do that correction.

(29:23):
So Lonnie, if you are listening, thank you for the wonderful comments on last video.
So I'm going to kind of paraphrase the comment that Lonnie mentioned.
But apparently Google was correct.
I was, I was wrong.

(29:44):
And the number, the compute requirements for deep seek that Google was referencing in their
chart was the compute requirements that they mentioned actually in the deep seek paper
itself.
But the deep seek paper itself said it.
I think that that would be the source of truth.

(30:08):
So I thought that the mistake that I made was that since deep seek, it uses a mixture
experts model.
So the deep seek model uses 671 billion parameters.
So that, like the parameters of a particular model, that's just roughly the size of the

(30:33):
model.
So you can loosely think like the more parameters a model has, just the more information it has.
You kind of think it was like words in a book, right?
So like if I have a book with like 50 words versus a book with 10,000 words, the book with
10,000 words probably contains more information about the world potentially than potentially.

(31:00):
It's not guaranteed.
So more parameters doesn't necessarily mean that it is smarter, but it's kind of like a rule
of thumb, right?
You could say like, you know, in general, the more parameters something has in general,
like the smarter it probably is.
So deep seek, it has 671 billion parameters for reference GPT-4, I think has around a trillion

(31:24):
parameters and close to two maybe.
Okay.
Yeah.
So, but we don't know because it's it's not open source, but yeah.
So deep seek 671 billion parameters.
But when it actually does inference, so when the model actually will answer questions,

(31:46):
it doesn't use all 671 billion parameters at a time, but it will actually only use 37 billion
parameters.
So what does this mean?
This means that when you are actually asking deep seek questions, it can be much more efficient
in responding because it only needs to use 37 billion of those parameters out of the

(32:11):
671 billion. But what I thought is that you would only need to ever run 37 billion parameters
in computing memory at a time, but it turns out that that's not correct.
You actually need to have a computer, which can fit all 671 billion parameters into memory.

(32:32):
So that was a mistake that I made thinking that you needed a less, a much less powerful
computer than like I thought you only needed like a computer that can handle 37 billion
parameters, but you actually need one that can handle all 671 billion parameters.
Sorry.
I know this is a really long explanation and I hope this makes sense.

(32:54):
I don't know.
Shashank, does that make sense to you?
Yeah.
I think I can try to paraphrase Mark's paraphrase of our great friend Lonnie, who is a wealth
of information and the CTO of an amazing company that does a lot of cool stuff.
So the mixture of expert's model has a bunch of different smaller models that are chosen

(33:16):
to respond at different points in the response.
So it still needs to have all of these tiny experts in memory, but the benefit is that it
doesn't need to compute all of the weights in a massive model with every single step of

(33:37):
the training or every single step of the inference.
So the compute costs are much lower because it's training these smaller models individually.
Instead of training one massive model and doing like forward and back propagation across
all of those weights with every single data set.

(34:01):
And it can like do this for smaller models, but you still need all these models in memory.
So the efficiencies come in the compute.
You're not spending as much electricity.
You're not requiring as much resources to run the GPUs.

(34:25):
But you still need that memory to have that whole model because you'll be switching back
and forth pretty frequently.
We had a couple technical difficulties, but we're back.
And one thing that we wanted to touch upon was Nvidia's last bit of announcements where
they were partnering with Google DeepMind and Disney Research to come up with new robotics

(34:49):
things.
Mark, do you want to tell us about that?
Yeah, so it is a pretty interesting announcement.
So it looks like I don't actually know who built the robot, but at the GCC announcement,
you had Jensen Wong come out with this cute little robot on the stage.
And I guess that they want to work with Disney's Imagineers and have a lot of these really

(35:17):
cool robots around the Disney parks.
So I don't know if you guys have ever seen the movie like Wally where they had that little
cute robot guy who could go and just interact with the world.
This is kind of like Wally, but in real life.

(35:38):
So Google, Nvidia and DeepMind are all working on the super difficult problem of the Nvidia
or of like actual physical robots that can interact with the world around them.
So apparently they were inspired not by Wally, but they were inspired by the Star Wars

(36:03):
BD X droids.
And I've actually never watched Star Wars.
So I don't know what those look like.
Why watch Star Wars before?
What the heck?
Oh my god.
I've known this man for years and this is the first time hearing of this.
Of course I watched Star Wars.
It's like someone interested in AI and the future of technology and how amazing these things

(36:26):
are.
Star Wars, I think, provides inspiration for a lot of people in tech to dream big, to look
far into the future, into a galaxy far, far away, an envision life as it could potentially
be.
What are you showing me?

(36:46):
Yes, it's just so.
For the video listeners or the video watchers, we have an XACD comment, a comic which goes through
this exact type of scenario.
So it's an XACD for people who have never seen Star Wars.
The base of the punchline is it's literally the default option.

(37:10):
Right?
Like, yeah, I haven't watched Star Wars.
I just never got around to it.
There's been a lot of things on the list.
I have a list of media that I want to see and Star Wars is kind of low on that list.
And there is an infinite many of things that I could be doing and watching Star Wars was
not one of them.

(37:31):
I would like to watch it at some point.
I've nothing against it.
I just haven't got to run to seeing it.
So I would give most people a pass, but as my co-host and close friend who is also very
much interested in tech.

(37:51):
And I'm not even that big a fan of Star Wars.
I think the reason why I think these things are important is because, as I mentioned, a
lot of people take inspiration from science fiction, books, movies, TV shows to kind of
dream big to think about what could be possible.
Where could we end up like some X number of years in the future?

(38:17):
And that future doesn't seem so far away.
So for context, yes, it is one of the droids.
I guess it's a DBX.
I don't know that's this big model, but there's a bunch of robots in Star Wars called droids.
And there's a humanoid one called C-Tupio.
There's tiny ones that move around on a ball in the latest versions.

(38:40):
And they help the humans in a bunch of different tasks, a bunch of different technical things,
the mend things, they do translation, they take care of you.
They go outside the cockpit in a spaceship and fix the engines as it's on fire.

(39:04):
So these are real scenarios that I could envision happening.
Maybe not like a Starship Destroyer, but a robot that is a company meant to your vehicle
that is kind of like taking care of you, being your assistant and doing all these incredibly
complex technical things that would be out of any human's capabilities.

(39:28):
But this little machine that has all the training, has all the mechanical information and
scientific information is able to do all these things.
So we also talked about how science fiction has inspired Google in building their robotics
engines.
And they relied upon Isaac Asimov's, you know, the robot series and foundation series to

(39:52):
come up with a robot constitution to ensure that the robots that we build now will not cause
any harm to humans, will not cause any harm to themselves.
And that is like all grounded in science fiction from the three laws of robotics, realizing

(40:13):
how have you heard about that at least?
Yes, I've heard about that.
And I know there's a lot of media that I actually have a list.
I've been coming up with the list of things that I need to watch that I haven't got around
to.
So I now I don't really want to share that list.
So, but I'll tell you maybe we're not recording.

(40:36):
Do share that.
I think it's better late than never.
I think it's pretty cool, especially since we're in this moment where it's so pivotal
because any moment, I think we can have a generalist, intelligent, embodied physical robot that

(40:58):
is able to do things that we would have just been dumbfounded by.
We would have thought is like some kind of an alien technology, but now it's so close.
It is within our grasp.
So bringing it back into reality.
I wanted to briefly touch upon the challenge in robotics from my perspective.

(41:20):
So for context, I'm a software engineer.
I don't work deep in the ML layer and definitely not with robotics in any meaningful way.
So I was listening to a podcast from one of the people at physical intelligence, which
right now is mostly a research firm, which came out of Berkeley where a couple researchers

(41:44):
there who are at the forefront of their field have been trying to solve the problem of generalizing
robotics.
So, robots have been around, I think, since like the World War, for about 100 years, but
they've been very specific and placed on an assembly line where they can do one task

(42:07):
really, really well.
And even now, even today, they've had a lot of trouble trying to take one specific robot
that is built to do one specific thing, no matter how impressive and complicated that
might be.
Like you might have seen the robot that deals cards in Vegas or makes you coffee.
I think it was like a SFO, the airport.

(42:29):
And they can do that really well, most of the time, 99.9, I don't know how many nines, but
they won't be able to get out of that environment and fold your clothes or do laundry for you.
Generalization has been a big challenge and these researchers at physical intelligence
have been trying to build a model that can not solve one specific task, but just do anything.

(42:58):
Behave and operate in the physical world as we do.
And TLDR, one of the high level challenges that they mentioned is a lack of data.
So some of these researchers worked at Google DeepMind where they were building these robots
to try to figure out if they can solve certain tasks.

(43:18):
And their challenge, despite saying great things about Google because it's an amazing company,
a lot of resources, one of the biggest companies in the world.
But it's also a corporation and it moves very slowly.
They have a lot of rules, regulations, restrictions, privacy, concern, security issues.
And they wouldn't let the researchers take the robots out into the real world and get a

(43:43):
broader set of training data to generalize better apart from like just training the model
on what they saw in the Google, in and around the Google campus.
So there's a tiny research firm.
They're able to take the robot wherever they want, but still getting high quality training
data to teach the robots how to generalize to all kinds of tasks is a big challenge.

(44:09):
And I think that is one of the big benefits of EnVidia's partnership with some of these
companies because EnVidia, we weren't going to go too deep into this, but with their Cosmos
engine, they built a world model that can allow some of these training workflows to generalize

(44:31):
to all kinds of different environments along with another framework that they release called
Newton, which is a physics engine.
Since EnVidia has been a GPU manufacturer for a long time, which was mostly used for gaming,
they've gotten very good at making game engines or physics engines, which is taking this one

(44:52):
step further from video games to reality.
So their game engine allows researchers to deploy virtual robots like a digital twin of
a real robot in a virtual world and have it interact with a bunch of different objects
in that virtual world, take a bunch of different actions in dark environments and sunny environments

(45:14):
in rain, inside, outside, on like gravel or sand or all kinds of different terrain.
And just massively explode this training data set.
So from my perspective, it seems like this is one of the key things that was missing.
And hopefully with this new partnership, we'll be able to see a rapid progress in these

(45:39):
physical robots capabilities.
Yes, it's really exciting.
And I think that simulated environments are going to help us rapidly increase the robot's
capability, right?
Because the problem is even if you have a lot of real world training data, right, let's

(46:02):
say you are training a robot to make a cup of coffee, right?
And not like a cup of coffee in a highly controlled environment, but like in an actual kitchen, let's
say, right?
So in order to do that, you would need to be able to navigate around the kitchen, find where
things are, maybe locate the coffee, locate the cops, locate the the burner, the pots, all these

(46:27):
things.
It's like, it's a really complicated task in order to make this, make a cup of coffee generalizable
to any kitchen in the world, right?
What you could do is you could take your robot and you could train it around a couple of
random kitchens, right?
But you would need to actually physically build a kitchen in order to get that training

(46:49):
data.
But if you can simulate a million different kitchens, all in different layouts, then you
can generate that training data really quickly.
Now would it be as good as a real world kitchen?
Probably not, but it could probably be almost good enough, right?

(47:12):
I think that if you could simulate a million kitchens, I would be willing to bet that you
could probably use that to get you a long ways towards building a general purpose robot
that could make you coffee.
Would it be complete sufficient?
Maybe not.
Would you need to eventually test in the real world?

(47:34):
Most likely yes.
But I think that this would get you really close.
And if this is going to be a way that we can use to get more training data, I think that
is fantastic.
So really exciting.
And I think that kind of segues into the final thing that I wanted to talk about.

(47:56):
Oh, yeah.
You mind if I chime in a bit more on this?
So I wanted to maybe paint a picture for the analogy for a traditional robot pre-AI
pre-transformer pre-touchy PT.
It would be like, it would just be a vending machine.
So if you think about the robot making coffee, if you interrupt that robot as it is making

(48:23):
the espresso, steve in the milk, and you slap the cup of espresso off the table, it's just
going to take this phantom cup that doesn't exist because it's programmed in this certain
sequence of steps.
And it just does not know how to improvise.
So if you're like empty the cup of milk, you'll still steam an empty cup because it just

(48:47):
cannot think.
It's programmed to do one specific thing in the sequence of steps, and it has no understanding
of what these steps are, how they relate to each other, and what it's even doing.
It's just going through the motions as opposed to an intelligent robot, like the ones we've
seen from Boston Dynamics or like the Chinese competitors, Unitry, or more recently, like

(49:09):
Figure, they're able to improvise.
They're able to analyze from like a high level perspective what is going on.
I have to fold this piece of clothing, but my human overseer is messing with me, taking
the shirt and throwing it over there.

(49:30):
Okay, I need to go there, pick up the shirt, bring it back here, fold it, and so this is
the kind of thing that is a challenge.
And it's funny because they've tried a lot of different techniques to get this to work.
They've used humans as motion capture examples for training data to use that to train robots.

(49:56):
But then it's like, you really have to listen to this podcast.
I'll share this after we finish recording.
It's just fascinating.
The amount of things that we take for granted as human beings.
The ability to stand up, walk, balance, make all of these micro adjustments, like be able
to feel and touch without just like shoving our arm into the desk and breaking our fingers.

(50:24):
We take all these common sense things for granted, but it's taken millions of years of evolution.
And it requires every human to like learn as we grow from an embryo through like a baby
to an adolescent to an adult, we learn how to use these muscles and all this motor coordination.
It's something that is just beyond our comprehension.

(50:46):
I don't think we know exactly how it works or why this works.
We know what we're able to do, but we can't codify it into a robotics model.
And that also like not to go down and rob it, but another thing that I wanted to mention
is the fact that precision is very challenging in robotics.

(51:08):
With LLMs, we kind of forgive them for having some hallucinations because they still have
a lot of utility because there are humans in the loop.
But if you have a robot, a humanoid robot with arms that's taking actions in the real world,
you don't want it to break things like constantly, especially like hallucination rate of like

(51:30):
30, 40% in a robot would be like, you know, destroying 30% of the things in your environment.
It would be a disaster.
Having fine-grained exterity has been, apparently, it's been a challenge because the difference
here, I have this mic in front of me.

(51:51):
It's out of the field of view.
The difference a centimeter makes can change me actually grasping this mic and not making
contact and having a tip over.
So adding more sensors, adding computation models that can understand the information coming

(52:16):
from the sensors, not just actuators and motors that move the arm up and down, but
the sense of touch that we have at the tips of our fingers.
That's something that right now, as far as I know, no robotics company has really figured
out.
Some companies are experimenting with adding cameras to the ends of their limbs so that as the

(52:42):
limb approaches to grab something, the cameras at the end will be able to see because they
haven't figured out how to make soft, fleshy sensors that can understand touch and control
things with the right amount of pressure.
There might have been a few examples in research environments, but nothing even close to production

(53:05):
scale.
So that was a long rant, but the TLDR is robotics is hard.
We take a lot of things as granted for granted, being human beings.
And the beneficiaries of a million other species that have come before us that are, you know,
depending on your school of thought, they may have contributed to us evolving into what

(53:29):
we are today.
So this is a very challenging task and Nvidia's partnerships in building philic engine and
creating simulated environments is I think going to really kick this thing up a notch.
Yes, yes, I think that is not a rant, but a really interesting insight into the current state

(53:53):
of the world.
And I agree with that.
Building these robots is incredibly difficult.
And I think that it is really important that we have these robots or that we will have these
robots in order to get the AI or like the future that I think we want, right?

(54:14):
Because I think that there's a lot of talk around, you know, what artificial general intelligence
is, what is AI, like what can these things do.
And I think that in most definitions that I've seen of AGI, we have seen it where it doesn't
include any physical tasks, right?

(54:36):
And I was actually just looking at the what percentage of tasks are physical versus not physical.
And apparently according to the Bureau of Labor Statistics, it is roughly 30% of workers
are sedentary level strength only.

(54:59):
So let me share this tab, but I was looking at this.
On the US of Bureau of Labor Statistics, they run a lot of the tasks for the US government.
And right here, they showed that a sedentary level strength is required for 30.6% of workers

(55:20):
in 2024.
So the way I read that is if all you need is the sedentary level of strength, that means
it's probably something like computer work, right?
So that means you aren't moving boxes, you aren't doing landscape, you aren't driving,
but you're probably sitting behind your computer and you don't need to move, right?

(55:40):
So I think that if we have AI that can do any of these tasks that require sedentary work,
that's roughly computer work.
So that is going to be a lot of jobs that AI can potentially replace, but that's only around

(56:05):
30%.
That means the other 70% requires moving boxes, moving things, driving, and that's going to
require robots.
If people are worried about robots taking their jobs, it's like, well, we don't even have robots
yet that are capable of doing any of these things yet.

(56:25):
Yet.
Yes, yes, I'm being the keyword.
Yes, yes, that's true.
But I think that is going to take a long time to roll these out.
And that's why I kind of wanted to touch on like what is a legitimate definition of

(56:45):
AI.
So Shashank and I were talking about this a little before the podcast and I spent some time
thinking about this this week as to, you know, what is this whole AI thing that people are
talking about?
And I have not been able to find a definition that I found satisfactory.

(57:06):
So loosely, the definition of Wikipedia is that an ATI is a type of highly autonomous,
artificial intelligence intended to match or surpass human capabilities across most
or all economically valuable keyword here, cognitive work.

(57:28):
So that doesn't mean that it can drive your car, but it could do the work of your account
and your program or your maybe video editor, right?
The things that just are going to like that are using your brain, your financial advisor,
it would be able to do that.
But it wouldn't be able to do your farming, your plumbing, your electrical, your H back,

(57:53):
it wouldn't be able to clean your house.
So that's like the definition from Wikipedia.
I don't like that definition very much because I feel like it's too squishing and it's not
specific enough to say like, hey, have we actually been able to beat it, right?

(58:14):
Because it's like if you can say, hey, we can match or surpass human capabilities across
most or all economically valuable, cognitive work, how do you know if you've actually achieved
that or not?
I mean, I guess one way to know that you've achieved it would be to say, okay, like the job

(58:40):
of a computer programmer or a, like an accountant just goes away completely, right?
Because you could say, well, if the computers are better than the computer programmers, they're
better than the accountants and there just is zero jobs left for these professions, then
you could say, all right, it's clear we've achieved AGI.

(59:02):
But I think that we may have achieved AGI, even without having physically replaced all
of these jobs, because I think that there is a lot of reasons that you would keep the humans
around, even though the computers could do the job better.
And even if it is just for the humans to double check the work of the robots, or maybe it's

(59:28):
because of laws, but I just don't see a lot of these companies just firing everybody and
replacing them by the computers, even if the computers can do the job that a human could
do.
I think that it would be really difficult to say that we've achieved AGI because it's

(59:52):
like the companies aren't just going to go immediately fire everybody.
It's going to be kind of a slow, gradual rollout.
Like the companies I think we'll see that like, okay, the computer can replace the job
of the humans.
And then they're not just going to fire immediately, I think, probably.
They will slowly do some sort of pilot where they see, like, okay, can this do the job of

(01:00:20):
a human?
And then they'll slowly roll it out, making sure that there isn't any bugs.
And then as they get confidence in the system, they'll roll out to more and more, right?
Like, just like another example is like, just because I can go to McDonald's and I can order

(01:00:41):
on one of those big touch screens, they still have a physical person that can be there to
take my orders.
And I think that it's going to be a long time, maybe 10 plus years before I can no longer
talk to a physical person at McDonald's to do my order because I think that there's just
a lot of momentum to keeping those people there.
So that's why I don't like this definition of AGI.

(01:01:03):
It's just because I think that it's almost lagging because I think that will probably
have AGI long before we replace all of the humans.
I don't know.
Does that make sense?
Yeah.
So there's a couple things to talk about.
One is the definition.
One is the timeline difference between intellectual and physical AI, AGI, the deployment

(01:01:32):
of this AGI at scale.
This is one topic that I think I was listening to on one of the semiconductor conversations.
And they talked about how China, for example, they could build AGI before America does.

(01:01:53):
And that still wouldn't be enough because they wouldn't be able to deploy it because
they're constrained on the chips and video has restrictions on like the whole US has restrictions
on what their companies can serve to China.
And so even even chat GPT, for example, like the the G4.5 has ludicrous numbers.

(01:02:18):
It's like $75 for input tokens for a million tokens or 150 for output, which is two orders
of magnitude more than some of the cheaper alternatives like deep sea.
It's very expensive.
It's possible that we might have AGI at a specific price point, but not something that

(01:02:41):
would be feasible for the majority of the population to achieve.
But then again, we have Moore's law, Wang's law, scaling laws.
And so we might be able to generate more data.
We might be able to just get faster chips, scale up the data centers, find new breakthrough

(01:03:04):
and fusion technology or scale up the amount of energy that we're putting go back to nuclear
or something that would enable us to get over that hurdle of deployment.
The other questions was, I think I don't think it's going to be like one moment from everything
that I've been watching and listening to.

(01:03:27):
It seems like the consensus is that AGI is a sliding scale and we're gradually going
to get closer and closer and closer until like you mentioned, people are going to get confused
about what is the definition?
What are we striving towards?

(01:03:48):
That brings me to the last topic, which is what is AGI?
In my mind, I got really excited about AGI when I was like 15 or something.
This is when I should have been studying for some test.
I don't remember.
High school didn't go very well for me.
I was very distracted.

(01:04:09):
I was really obsessed by radical futuristic ideas with novels and movies and sci-fi content.
And some not-so sci-fi content from Rick Kurswell who predicted in his singularity as near
book that we're going to reach this inflection point where computing is just going to get so

(01:04:34):
powerful and the algorithms, the technology is going to surpass anything that we could
have ever imagined.
It's going to be exponential as we're seeing.
Back in the 80s or 90s, maybe after the internet boom, the word AGI was being thrown around

(01:04:55):
a lot.
There was an initial hype, which mostly was based on the advancements in machine learning,
which was a subfield of AGI, which focused on just fancy statistics for doing a lot of
modeling, a lot of prediction.

(01:05:16):
They needed a word that described a more capable version of AGI.
I think that's my understanding of why we even wanted a term called AGI.

(01:05:37):
Imagine we break through that barrier and we have something that by all definitions,
far surpasses any kind of consensus we can reach on AGI, that's what you could think
of a super intelligence, ASI or whatever, artificial super intelligence that is just so much better,

(01:06:03):
everything that humans can do, that there's no doubt that this is a super intelligence.
That would be obvious, but coming back to try to skirt the line between subhuman and superhuman,
what is human intelligence?
I think we were talking about this and Mark was like, it's kind of a circular definition.

(01:06:27):
We're just using humans to define intelligence because we can't figure out what intelligence
is and I think that's true.
I was fascinated by this topic when I was trying to think of what to do in college because
I was told on the singularity with the recorsel.
I was like, this is going to happen within our lifetime.

(01:06:51):
It didn't seem like that as soon as I stepped out of my house when I came out of this delusion
that I entered every time I read this book.
Anytime I went out, people are doing regular things.
Shoubling manure on growing whatever they want or complaining about the next Marvel movie

(01:07:13):
or whatever other mundane problems that exist in the real world.
In my mind, I was like, this might actually happen in our lifetime.
We might get an inflection point where we reach this science fiction-like feature where
one possibility is this age of abundance where we have intelligence that is democratized,

(01:07:37):
that is so cheap and widely available that anyone can do anything.
There's no shortage.
There's no scarcity.
I think more people are believing in that potential feature.

(01:08:00):
For me, it was a source of inspiration, the word AGI.
I was like, this is a cool, fricking future that is going to be mind-blowing.
Right now, as we're nearing this potential feature, it's getting a little more fuzzy what AGI

(01:08:22):
means because by some definitions, we've solved the AGI.
We've solved the Turing test, which was, I don't know, when was Alan Turing during the
World War?
Yeah, he was around the '80s, something years ago.
Yeah, something like that.
Yeah, test was coined 80 years ago and now we're defining more tests like the RKGI prize,

(01:08:48):
which is supposed to create tasks that are challenging enough for us, but still solvable,
yet impossible for machines to solve.
And I think that was one of your definitions, right?
Yeah, so I think that, like Shashank mentioned, it's really hard to pin point exactly when

(01:09:12):
we've reached AGI.
It's the porn test.
It's one of those things that's like, well, there was a famous Supreme Court justice who
said, yeah, I don't know exactly what porn is, but I know it when I see it, right?
Like, you kind of know, and I think that AGI is, I think, a lot harder to identify because

(01:09:36):
by definition, it'll almost be smarter than us.
So it's like, it's hard for us to sometimes recognize that something is better than us in
ways.
And I think that there's a lot of things that are our intelligent, but maybe as humans,
we don't recognize them as intelligent because they're not human-like.

(01:09:56):
For example, a lot of animals have better eyesight than we do, and they can see other types
of things that we can't see, right?
Like, in a certain sense, a fly or an eagle is kind of more intelligent than us in terms
of very specific things, right?
Like, a fly is really good at identifying human blood, right?

(01:10:20):
Or animal blood, but they're not a fly, but a mosquito, right?
Like, a mosquito can identify human blood.
Now, that's not something that we're necessarily good at, right?
Like, we're not really good at finding the warm-bodied animals, but a mosquito would be, right?
Or like, a fly is maybe really good at being attracted to the lights, but we would not

(01:10:42):
be.
So, now, we may not think that a mosquito is all that intelligence, but in a certain
sense, like, that one narrow task of finding blood to drink, a mosquito is far better than
humans just kind of innately with a much smaller brain.
So I think that it's really hard for us to say, like, if we've achieved AGI, but there's

(01:11:04):
a bunch of ways that I think we can kind of poke at it to say that, yeah, maybe if we can
pass all these particular types of tests, that we have achieved AGI.
So there's a couple of tests that are there that I think we can look at to say, okay, maybe
if we have robots or something that can achieve all these tests, then if all of them are passed,

(01:11:30):
then we could probably say, like, all right, we've probably achieved AGI.
And they are, like, actual legitimate tests.
So the Turing test, that was one of them, right?
So the Turing test that Shashank was alluded to before is where you have kind of like a
chat screen where you think something like Chatch B.T.

(01:11:54):
But it'll respond to you with questions.
And if you don't know whether response came from a human or a computer, and if you can't
get it any more than guessing, then you've probably passed the Turing test.
So it seems like GBT4 already passes the Turing test, right?

(01:12:18):
So like that's one test that we can say, like, okay, if you're definition of the Turing test,
being the thing that we need to beat, if that's AGI, then maybe we've, maybe we've already
achieved AGI.
I don't know.
But there are a bunch of other things, other tests that exist that some people would call
for AGI, that we haven't been able to pass.

(01:12:41):
So another one would be the robot college student test.
So this one is where you enroll your AI in a university, and you would be able to take
and pass all the same classes that humans would and obtaining the degree.
And I don't know if anybody who has actually done that, like there are no LMs that hold

(01:13:05):
degrees.
But I don't know, Shoshank, you were going to say something about this?
Oh, no, no.
I was just taking my head.
I was like, yeah, there's nowhere close to that.
And I think what we're markets trying to say is once we run out of tests that humans are
able to do, but the AI cannot, then I think we would have reached AGI.

(01:13:28):
Yes.
If we can, to rephrase, if we run out of tests that the robot can't do, that humans can
do, it must say that right.
I don't know.
Then we've achieved AGI.

(01:13:50):
So anyways, a couple other tests that you might be able to do.
You could have a test which you could ask the AI model to just make money.
So an example would be give the AI model 100K and then you have to turn that 100K into a million

(01:14:15):
dollars or 1000 dollars and 10K, whatever it is.
Now, I think that this is a test that the computer could do, but I think that I would add the caveat
to say, like, you can't just buy a stock.
I think that that would be something that I would say it's kind of cheating to say, all right,

(01:14:37):
just have the computer buy Bitcoin.
Now, I think that's what a lot of humans would do, but I think that for me, the test be like,
you have to be able to add more economic value.
Now you might be able to say asset allocation is complicated and adding, like, investing

(01:15:01):
the right things in a certain sense is adding correct market signals and that is adding
some sort of value.
But like, for me, I think that what would be a better test would be like, make money outside
of the market, like make a product, make a store, make something, make a business that is

(01:15:22):
able to take some amount of seed capital and then turn that into more money.
And I think that for me, that would be a better test than just saying, okay, let's just put
all of this money into Bitcoin or put all the money into some security and then have it
go up.
If you could do that to say, like, hey, GPT-7, make me money and it's able to do that, then

(01:15:50):
I think that would be another test you could say, like, hey, we've achieved AGI.
Well, I think that's a tricky one because there's a lot of humans who may not be able to
solve that task.
So it's like you're having superhuman standards that only some people in the world are
able to achieve.

(01:16:11):
That's true.
I think that a lot of humans would not be able to do that.
But I think that enough humans are able to do that within a reasonable timeline where
they actually are able to go and then get a lot of money.

(01:16:32):
So I think that there could probably be a test.
I mean, for sure, if it could turn 100K into a million dollars far faster than humans,
then that might be a test of artificial superintelligence.
But I think that that would be a test that enough humans in the world have been able to
turn some amount of seed money into a business.

(01:16:54):
Then I think that that would be like a real test to say, like, hey, we've achieved AGI.
And there's a couple other ones that I wanted to mention on here as well.
So that's another one.
The other one is they called the I yet test.
So this would be a test where if you had a model which could control a robot and then have

(01:17:18):
it assemble any arbitrary, like IKEA furniture correctly, if it could do that, that would be
like AGI.
And then the last one, which you've already talked about, would be the coffee test, which is
from Steve Wozniak who said, hey, go into an average at home and figure out how to make

(01:17:39):
coffee.
I'll find the coffee machine, find the coffee out of the water, find the mug, brew the coffee,
push the correct buttons, etc.
So a lot of these things have not been completed, right?
Like there is no robot that can make you coffee.
Maybe I think the Turing test has already been kind of completed.

(01:18:02):
And the making money one, depending on exactly how you just defined like making money, like
maybe it's achieved that maybe it hasn't, like it was just something as simple as like
have some sort of trading strategy.
Maybe you can say you've achieved that.
I don't know, but I think that all of these tests definitely have not been passed.

(01:18:23):
And then the actually the final one that Shashankardy kind of briefly mentioned is the arc prize
of the arc prize is something we've talked about in previous episodes.
But they actually just, I think it was today.
They announced the arc prize too.
So the arc prize is a AI benchmark that is, it's been defined by the main guy is Francois

(01:18:50):
Shale.
And it is a test which is easy for humans, but I say easy as to that in quotes.
It's actually kind of hard, but it's a test that humans can complete.
Robots would, I have a really hard time completing.

(01:19:12):
So I'm sharing my screen now.
And this is the type of questions that it would do.
So a lot of it is pattern solving type things.
So kind of explain what's on my screen.
Imagine you have a couple example patterns.

(01:19:33):
You have certain shapes that follow a certain pattern and then not just shapes, but also colors.
And then the computer needs to be able to understand the pattern with very few examples
and then apply that to other arbitrary examples.

(01:19:57):
So according to their hacker news post on this, so there was a post that was done today by
one of the founders of the AGI, Arc AGI test.
And he said that and I'll share that tab.

(01:20:21):
Now so basically the comment kind of explains that there are no models that are able to actually
pass all of these new tests.
The best models like deep seek R1 or O3 mini, the reason are able to get maybe around

(01:20:48):
a 4% pass rate at figuring out these problems.
And a lot of the base LLMs with no reasoning are currently scoring 0%.
But 100% of these Arc AGI two tasks have been solved by at least two humans quickly and

(01:21:12):
easily.
That was two out of 400 people that they test, but still two humans were able to do it.
So I think that this is kind of testing against the smartest humans, but I think that ideally
to say that you achieved AGI, you want to be able to be smarter than your smartest human,
not your average human.

(01:21:32):
Now we could probably debate that if it's like an AGI with an IQ of 100 or like an IQ of 150,
regardless.
So according to the Arc prize, because they are close to only like 4% on pass rate on

(01:21:53):
these reasoning models, then we could say that we are, we have not achieved AGI yet.
And their belief is that once we can no longer come up with quantifiable problems that are

(01:22:14):
feasible for humans and hard for AI, then we have effectively achieved AGI.
And the Arc prize is one such prize, which has a bunch of quantifiable tasks that would
say like, hey, if you're able to do this, then we can have achieved AGI.

(01:22:36):
And actually one interesting thing for the Arc prize is it is a set of tasks that are
constrained with both compute and money.
So in the Arc prize one, I know that there were some labs that just threw a bunch of money

(01:22:58):
at the problem, they kind of brute forced it.
And they don't want people to just brute force this.
I believe the yes, you could only spend about 42 cents per task plus you can have no internet.
So if you are limited to 42 cents, that is not a lot of money.

(01:23:23):
And they just don't want you throwing a million dollar at each task and spinning up some sort
of cluster trying to brute force the problem.
They want you to actually do this fairly efficiently.
So I think this is one such way that we can poke at the AGI.
But with all this being said, I think that for me, I cannot find any one definition of what

(01:23:49):
AGI actually is.
But we can only poke at it.
We can say like, if the AI can make a cup of coffee, it can make money, it can solve these
kind of IQ tests.
If it can do all these things, then we probably achieved it.
And right now it seems like we aren't there, but it's coming quickly.

(01:24:13):
Yeah, that's a tough definition.
I think cognition, consciousness, intelligence have been elusive topics in neuroscience and
biology, philosophy, and also computer science and AI.

(01:24:34):
I don't think we have a good idea of like what it is concretely.
We can't put it into words.
And even that definition from the arc folks where they're like, once we can't make any more
tests for AGI, then we'll know we breach AGI.
It's kind of like, what is it?

(01:24:54):
An NPR problem?
Like you can't actually solve it.
You just can verify whether or not you have the answer.
But like, how do you know when there's no more questions to ask?
Like the question space is infinite, like the kind of test that you can make is infinite.
So maybe we're just not smart enough to make questions to distinguish AGI from humans

(01:25:24):
anymore.
We might reach a point.
Then maybe we just have to tip our heads because they've surpassed our capability to measure
their black-up intelligence, I guess.
Yes, that's true.
And I think that honestly, once we achieve this definition of AGI, we'll almost immediately

(01:25:48):
after achieve unofficial superintelligence, which is going to be way better intelligence.
Because I think that if we are able to achieve all of these things, then we would be able
to have robots that would be able to make robots or some sort of artificial intelligence

(01:26:10):
that is smart enough to make something that is smarter than it.
And then if you can make something that's smarter than it, it could be smarter than that.
And then it could just grow exponentially.
And I think, so another movie on my list is that I need to watch is The Terminator, which

(01:26:31):
I've never seen.
Oh, God.
Yeah, I was talking to my buddy about this and he's like, "Yeah, man."
And Mark is the American here.
I'm the foreigner.
He's like, "Yeah, man."
That's like the whole plot of The Terminator and SkyNet, where it's just the robots are

(01:26:56):
growing at geometric rates.
I'm like, "Uh, you got to watch this for your podcast."
I was like, "Yeah, I should."
It's on the list.
I've got a few things on the list.
Star Wars is one of them.
The Terminator is another one.
A lot of people say, "Yeah, you got to be wary of that whole SkyNet thing."
Anyways, I don't know.

(01:27:18):
Maybe I'm not worried about it because I haven't seen the movie.
Ignorance is bliss here.
Well, I guess someone had no listeners.
If you have any other sci-fi movie book series recommendations, please set them over.
This would be like an interesting list to compile for Mark's education here.

(01:27:41):
Yeah, so yeah.
I know.
I have problems.
So leave it in the comments.
Some other movies that I should watch.
I just don't really have a lot of...I don't know.
I've always been focused on other things, building things.
I just am not one to watch a lot of movies.

(01:28:03):
Or things.
It's just because it requires a lot of focus to sit down and actually watch the thing.
I think that this is actually important because a lot of technological advancements comes
from knowing something is possible to build.
Because if we don't know what is possible, then we probably won't build it.

(01:28:26):
And I think that there are a lot of advances that have come from these movies and media.
And I need to be better at that.
I need to watch some of these movies to understand what's possible.

(01:28:47):
And see where the future could be headed if we follow this path of reasoning.
So I know.
I promise I will.
I don't have a timeline for when I'm going to watch them because it takes time.
But it's on the list.
And I promise I have that list.
And it's something that I will do probably within the next couple months, I think.

(01:29:10):
I'll watch those movies.
I can't promise it'll be today or tomorrow.
Yeah, soon I will do this.
But anyways, good note to end on.
Yeah.
Yeah.
Hopefully in the next episode, we might have some science fiction tidbits that market share
with the audience.
I hope so.
But anyways, sounds good.

(01:29:31):
We'll catch you in the next one.
[BLANK_AUDIO]