Are current AI models smart enough to rule the world — or just house cats with fancy vocabulary?
This week, a tectonic shift is happening in AI: Meta's chief scientist Jan LeCun quits to chase world models, Fei-Fei Li launches Marble, a spatial intelligence engine, and DeepMind drops CMA-2, a self-taught gamer bot that might be the blueprint for AGI.
Meanwhile, OpenAI releases GPT-5.1 — and China’s Kimi K2 and Ernie 5.0 roll out shockingly powerful, ultra-low-cost models. The AI race isn’t just about intelligence anymore — it’s about who can afford to scale.
If you lead a business, this episode explains why spatial intelligence, not language, may soon be your competitive edge.
The next wave of AI isn’t just about better answers, it’s about deeper understanding, real-world interaction, and models that scale affordably. If you’re not watching spatial intelligence, you’re already behind.
About Leveraging AI
- The Ultimate AI Course for Business People: https://multiplai.ai/ai-course/
- YouTube Full Episodes: https://www.youtube.com/@Multiplai_AI/
- Connect with Isar Meitis: https://www.linkedin.com/in/isarmeitis/
- Join our Live Sessions, AI Hangouts and newsletter: https://services.multiplai.ai/events
If you’ve enjoyed or benefited from some of the insights of this episode, leave us a five-star review on your favorite podcast platform, and let us know what you learned, found helpful, or liked most about this show!
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Isar Meitis (00:00):
Hello and welcome to a Weekend News episode of the
Leveraging AI Podcast, thepodcast that shares practical,
ethical ways to leverage AI toimprove efficiency, grow your
business, and advance yourcareer.
This is.
Isar Meitis, your host, andtoday is gonna be a short news
episode because I just came backfrom a trip from Chicago doing a
keynote in a workshop, and I'mleaving today to Europe to do
(00:20):
two different AI workshops.
So I have a limited amount oftime, and I actually considered
not doing an episode today, butthere are two really big things
to talk about.
So we will do a shorter episodein which we're going to focus on
these two things.
One is going to be world models,and the other is gonna be models
who are ruling the world.
And you'll understand in aminute as we get started, there
are a lot more news from thisweek.
A lot of really interestingnews, and you can find all of
(00:40):
those in our newsletter, but wewill focus on these two topics.
So let's get this started.
So world models is somethingthat we mentioned several times
previously on the show, but itwas never a big deal and
definitely never the main topicof the episode.
But it is today because we gotthree different things in the
(01:01):
world models that happened thisweek.
And the first one we're going totalk about is the departure of
Jan Koon from meta.
So we talked about young Launmany times before in the show.
He's the chief scientist atMeta.
He's a touring award winner.
He was an NYU professor and he'sone of the pioneers of the
current AI era.
He has been the chief scientistin.
(01:22):
Back then Facebook since 2013.
So he started AI research waybefore the whole ChatGPT craze,
and he built the Facebook AIresearch lab, also known as
Fair.
But before that, he hasdeveloped Alex net, which he
developed together with astudent of his, that was a
really advanced visual analysisnet, which was able to beat the
legendary image net that existedbefore that by 10%.
(01:44):
For the first time.
Any other model was beatingthat.
So he was in the world ofresearching and developing
visual.
Aspects of AI way before thecurrent era, he won the Turing
Award together with JeffreyHinton and Yoshua Bengio, uh,
also known as the Godfathers ofai.
So he is been around the AIspace longer than most
scientists that are in it today.
And he just announced that he'sliving meta and he's living meta
(02:06):
to develop his own company thatwill focus on world models.
So a little bit of background.
Laun has been very loud sincethe launch of ChatGPT stating
that large language models willnot lead to a GI.
He always suggested that theonly way to get to a GI will be
to develop models whenunderstand the world as the
world is and learn like babiesas we learn from the environment
(02:26):
and not through just language,which is a very narrow way to
view the world.
he's known for a tweet thatsays, it seems to me that before
urgently figuring out how tocontrol AI systems much smarter
than us, we need to have thebeginning of a hint of the
design for a system smarter thanhouse cat.
Basically comparing largelanguage models to a house cat
in their level of intelligence.
Now there are multiple reasonsof why Jan would leave meta.
(02:49):
The biggest one is obviously theturmoil that's happening in meta
right now.
So he was the head of the AIshow.
And after the not so successfulrelease of LAMA four, he was,
and I'll be very gentle,sidelined by Mark Zuckerberg who
spend$14.3 billion into scale AIin order to poach Alexander
Wang, their CEO to become thenew head of Meta's Super
Intelligence Lab, which has alsopoached multiple scientists from
(03:13):
the top leading competitorsspending hundreds of millions
and sometimes billions in orderto get this kind of talent.
And since then, there's been acomplete chaos over there.
But it was very obvious that YLaun is not in the center of
this.
And there's other people whowill be.
And so from a politicalperspective, it was a good time
for him to leave, but I ampretty sure it is also going to
make him extremely wealthybecause I'm pretty certain he'll
(03:33):
be able to raise a few billionsof dollars with evaluations of
even more than that to build hisAI world's vision because.
There are more and more voicesthat are saying that this is the
real way to move to an A GIfuture.
Again, he's been definitely oneof the key people that was
driving this forward.
Now, is that tied to his sourcesof visual aspects of ai?
Potentially because there'sothers in a similar path that
(03:55):
think like him.
So since there's not much toreport right now on exactly what
Jan Koon will be doing, uh, I'llswitch to another person with a
similar background who is makingsimilar waves, and that is Fei.
FEI Lee.
So who is Faye?
Faye Lee.
She is another one of the earlyscientists in the modern era of
AI and one of the leadingvoices, and she was a Stanford
professor.
(04:15):
Since 2009, and she's the onethat in Stanford developed
ImageNet, which we mentionedearlier as a early AI tool that
could read text, and evenhandwriting, which was then
surpassed later by alexNet thatwas developed by Koon, so these,
how all these people are tiedtogether and been working on
similar things for many, manyyears.
She developed ImageNet back in2009, but she also co-founded AI
(04:38):
For All, which is a nonprofittwo Boost AI education, and she
did that in 2017 where most ofus couldn't spell ai.
She won multiple award and she'sconsidered the godmother, if you
want, of AI in the modern era.
And in 2014 she founded WorldLabs, which is a company that is
building AI driven worlds.
Now she just published asubstack called From Words to
(04:59):
Worlds Special Intelligence isAI Next Frontier.
A quick summary of what is inthat letter that will help you
potentially understand why thisis so important, so the idea is
that ai, in order to make thenext progress, need to move
beyond language and maybe imagesto understanding and interacting
with a 3D physical and orvirtual environments.
She says that current ai, likeChatGPT, are wordsmiths in the
(05:22):
dark, that they're eloquent, butnot grounded in physical reality
or with a real deepunderstanding.
Of how our world operates.
So she's asking the question ofwhat is spatial intelligence?
And she's saying it's theability to perceive reason about
and interact with the physicalworld in a 3D space.
And she's also saying that thisis the foundation of our
intelligence of human cognitionand how we navigate, how we
(05:45):
create, how we imagine, how wesolve problems.
It's all built on how we learnas babies before we even have a
language.
And she gives examples.
We know how to park a car.
We know how to catch keys whensomebody throws it at us.
Firefighters knows how tonavigate buildings when they
can't see anything, and it's allbecause we have general
understanding of the environmentaround us in a very deep level.
And to prove why that is aproblem.
(06:06):
She's saying that current AI ishighly limited in understanding
the actual environment.
She's saying that LLM strugglewith very basic spatial tasks,
like estimating distances orrotating objects in a space, or
placing them correctly ornavigating mazes or.
Predicting physics, all thesethings, ai, current ai, large
language model based AI is notvery good at even AI generated
(06:27):
videos that seem to understandthe world physics, which by
itself is amazing, are limitedto X number of seconds, and now
maybe mere minutes, but that'sit.
You can't run a persistent videowith clear physics and
continuity.
That will run for 30 minutes,not to mention 30 hours right
now.
And this is why there arecompletely new sets of models to
run robotics who need to operatein the 3D world.
(06:48):
So the solution that sees she'ssuggesting is what she's calling
world models.
And world models have three mainessential capabilities.
One, it's generative.
They can create geometricallyand physically consistent
virtual and or.
Real worlds what you're saying.
Real world is obviously arepresentation of a real world
versus a completely made upworld.
The other thing is they'recompletely multimodality.
They.
Process, diverse inputs.
(07:09):
They understand image, video,text, gestures, and physics to
produce complete world states.
And the third one is thatthey're interactive.
They can predict the next statesbased on actions and goals that
are driven inside that worldthat they have created or that
they operate in.
So from a timeline inapplication perspective, she
sets three different times.
One is the near term, basicallyright now, which is creative
(07:30):
tools for filmmakers, gamedesigners, architects, that
allows you to build 3Denvironments without the need
and the overhead of traditionalsoftware development.
Just by giving it a prompt andit will generate everything you
want with accuracy andconsistency, which is the key.
The mid to long term, three tofive years, the idea is to
develop whole environments forrobotics and what's called
embodied ai.
(07:51):
That we'll be able to drive awide variety of different kinds
of robots that will be able tooperate effectively and safely
in a wide variety ofenvironments.
And in the long term, she'stalking about scientific
discovery, healthcare,diagnostics, and immersive
education.
Basically being able to buildworlds to teach people anything
that you want to learn, beinginside that environment, most
likely with some kind of virtualreality headset.
Now, she's not just talking orwriting blog posts, she's also
(08:14):
making it happen.
But before I tell you what she'smaking happen, I want to quote
one line out of this manifestothat she shared, because I found
it very, very powerful.
She said, and I'm quotingextreme narratives of techno.
Utopia and apocalypse areabundant these days, but I
continue to hold a morepragmatic view.
AI is developed by people, usedby people, and governed by
people.
(08:34):
It must always respect theagency and dignity of people.
Its magic lies in extending ourcapabilities, making us more
creative, connected, productiveand fulfilled.
Spatial intelligence representsthis vision.
I think this is beautiful andwonderful, and I really, really
wish that all the people who aredriving AI today would feel the
same way and act based on theseprinciples.
But now to the big news fromWorld Labs, again, FEI Fey's
(08:56):
Company.
So after raising$230 million,world Labs is finally launching
its first commercial productjust a year after emerging out
of stealth, and it even has afreemium access.
You can use it for free withlimited generations, but what it
knows how to do is it knows howto generate either realistic or
completely made up worlds fromprops.
(09:17):
So, as I mentioned, you can useit for free for up to four
generations.
Uh, the pro tier that is$35 amonth give you 25 such
generations.
So she quoted on this launch,the new generation of world
models will enable machines toachieve spatial intelligence on
an entirely new level.
So this model is called Marbleand in addition to the ability
to generate worlds from aprompt, which by itself is
really cool, but there's othercompanies who do this.
(09:38):
Google has a similar tool aswell, but Marble comes with a
few really cool tool.
One is called Chisel.
Which allows to decouple thestructures such as walls and 3D
blocks from the visual styles,so you can control both the
building blocks of the world aswell as the style of the world.
Separately.
It allows you to move objects inthe world and place them in
different places to have bettercontrol on how the world is
(09:59):
built or like their co-founderJustin Johnson said, I can just
go there and grab the 3D blockthat represents the couch and
move it somewhere else.
But then you can also define howthe couch is going to look like.
So this allows more control tocreatives to create the worlds
they want.
Exactly.
They what they want them.
You can even expand the worldsby providing them a growing
edges, and you can even use whatthey call a composer to merge
(10:22):
scenes, to create vast, biggerspaces, whether photorealistic
or game like outputs.
That can be exported in severaldifferent format, either GA and
splats, which is theseenvironments that you can look
through or meshes, or videosthat can be ideal for either
gaming or just to navigate withVR headsets.
Now, in addition, they releaseda tool that allows marble to
create assets Existing graphicengines such as unity and
(10:44):
unreal.
So instead of just creatingentire new worlds you can create
accurate 3D objects for existingengines, which dramatically will
increase the pipeline ofgeneration of 3D objects into
the existing gaming universe.
So to summarize this componentabout Lee's ambitions, she said,
our dreams of truly intelligencemachines will not be complete
without spatial intelligence.
And this aligns perfectly withwhat Jan Koon is doing.
(11:07):
And this is, and this alignsperfectly with what Jan Laun is
saying and now going to bedoing.
As well.
And there, as we mentioned,coming from a very similar
aspect of the AI universe, so itmakes sense that they think the
same way, but this is not thelast piece of news about
understanding and operating in3D environments.
DeepMind just released CMA two,which is a.
AI that can reason in 3D worldsand play extremely well in every
(11:31):
computer game, including reallysophisticated computer games.
How does it work?
Well, they basically let it playany game, including really
advanced, sophisticatedmultiplayer games in complex
environments.
And it is learning on its ownhow to play the game.
Well, I'm not a gamer, but inthe release notes, they shared
several really sophisticatedgames like Mind Dojo or asca,
(11:52):
that it was very successful inplaying, learning over 600
skills and contextual reasoningin these games on its own.
So how does it work?
It's powered by Gemini modelsbehind the scenes and it.
Understands the environment ofthe game and it understands how
the game works.
And it can even communicate thatsuch as I'm going to the village
center, or I'm finding acampfire, or things like that
because it understands how thegame actually operates.
(12:14):
It understands how to do reallysophisticated aspects like
mining different things in orderto achieve other goals in the
game.
And it comes with zero humanassistance.
It basically.
Plays the game on its ownGemini's reasoning, trying to
understand what the roles andthe goals of the games are, and
it learns how to operate inthese worlds by itself.
In the release notes of CMA twoDeepMind highlights that gaming
is a, quoting incubator forgeneral intelligence, and it is
(12:37):
currently available only for agated preview, but they are
planning to release at leastreports on everything that it
does and can do later on.
Now, why are we talking about anAI that can play games, and why
does DeepMind think that it's anincubator for general
intelligence.
Well, here's a little story foryou.
I've been traveling literallyevery single week in the past, I
don't know, eight to 12 weeks,something like that.
(12:58):
Seems forever delivering eitherkeynotes or AI workshops to
companies and organizations.
And I just came back fromChicago and I had the
opportunity to finally watch theThinking Game.
The Thinking Game is a movieabout Demis AEs.
And his quest for AI forHumanity.
If you want, those of you whodon't know Demi, he's the
founder and still the chief ofGoogle DeepMind.
Well, when he founded, it wasjust DeepMind before Google
(13:19):
bought them.
I highly, highly recommendwatching this movie.
It is really, truly inspiring.
But it connects well to what wejust talked about because it
shows how AI learning from gamescan deliver in the end, very
significant results.
So before I tell you what thatis, two words about the movie
and about Demis.
First of all, the movie made mefeel completely worthless.
I mean, I really liked Demi'sapproach before the movie.
(13:41):
It always felt a lot moregenuine, realistic, and really
grounded in.
True drive to make the world abetter place with AI compared
with some of the other leadersof the other labs.
And this movie made me feelsignificantly more like this
about Demi.
He's absolutely incredible.
He's really driven by making theworld a better place with the
usage of AI and watching hisjourney and thinking how he
(14:03):
feels and how he drives otherpeople around him.
Made me feel completelyworthless compared to what he is
doing.
It is really inspiring, but toconnect the dots to this segment
of this podcast.
As you probably know, Demis wonthe Nobel Prize for alpha fold.
Alpha fold is a model that canaccurately predict protein
folding.
Now why the hell does thatmatter?
Because proteins are, if youwant the machinery of life.
(14:24):
Like everything we know as aliving, anything is built on
proteins that are foldedtogether, but trying to
understand how they're folded,basically, how they create a 3D
structure that then becomes lifeis a crucial component of A
understanding life.
And B, being able to understanddisease and designing new drugs
to solve every illness on theplanet.
Well, previously, before alphafold, it was extremely difficult
(14:47):
and very inaccurate tounderstand the structure of a
protein.
It would take years in a lab tofigure out one protein.
Well, alpha fold is a predictionmodel can predict protein
structures in minutes.
And get it highly accurate.
And they've used it to predictstructures of over 200 million
proteins and then open sourcedit so anybody can use it for
(15:08):
research.
So how does that connect togaming?
Well, in the path to getting toalpha fold, the previous
iterations that gave DeepMindthe capability to even go down
that path was gaming.
They have developed, many of youprobably know of Alpha Go, which
has beat Lee Seel and afterwardsthe Chinese World Champions in
the Game of Go.
so these attempts to have amodel learn through a gaming
process with clear goals andlearning on its own how to play
(15:31):
a game, were the building blocksthat allow them to later on
develop alpha fold.
And most of the bigbreakthroughs that DeepMind have
created, and there's a long listof them, came from teaching AI
to play games.
So going from being able to playa single, really sophisticated
game like go to, being able toplay any computer game on its
own is a very big step on thepath to a GI.
(15:53):
And this is what DeepMindreleased right now.
So that ends this topic.
Now what you need to do is goand watch the movie and we'll
switch to the next topic, whichis the arrival of GPT five one
and the other models around it,and maybe the imminent release
of Gemini three.
So OpenAI just released GPT 5.1,which doesn't sound like a big
spread from five.
It just adds a 0.1 at the end,but it's actually a dramatically
(16:13):
different model that has a lotof big, significant
improvements.
So on the high level, theyreleased two different models,
GPT five, one instant, which isfaster and more conversational,
and it has a warmer tone foreveryday tasks and conversing
and GPT 5.1 thinking, which isslower, but has much better
complex reasoning, math, logic,multi-step planning, and all
these kind of things thatthinking models are good at now.
(16:34):
What are the biggestdifferences?
Well, first of all, it is.
it is a good combination ofbeing a smarter and B, more
natural and human-like.
So it is better at instructionfollowing an accuracy.
It is more conversational andless robotic.
Tom, which was one of the bigpushbacks against GT five, and
it has fewer hallucinations.
The GPT five, all of these areawesome.
It also has adaptive thinking,so they dramatically improved
the amount of time that it needsto think for specific tasks.
(16:56):
So for simple things, it willanswer faster than G PT five,
and for more complex things, itwill think longer than G PT
five.
In both cases, providing youbetter answers, aligned with the
relevant amount of time thatneeds to be invested in actually
getting there.
It also comes with severaldifferent personalities that you
can select from professional,candid, quirky, friendly, and
efficient.
And the basic model is finetuned for being warmer and more
conversational.
(17:17):
and it even proactively offersto adjust the tone during
conversations depending on whatis actually happening.
It is supposed to be better atcreative writing, better at
coding with more tools that itcan use, and it has enhanced
planning and multi-step task, exexecution.
So on paper, a significantlybetter model than GPT five.
Now I didn't get a chance toplay with it enough again.
(17:37):
I just came back from oneworkshop and I'm shortly leaving
to the airport for another, but.
I can say a few things.
On a high level, better, promptcoherence is always a good
thing, especially as we'removing from basic, simple things
to more advanced and complextasks that we want AI to
achieve.
And building agents andmulti-step agents and
multi-level agents.
So following instructionsaccurately is a very big deal,
but also being able to optimizethe amount of time that the
(18:00):
model thinks is a big deal.
I've been recently using Claudea lot more than I've been using
ChatGPT and a lot more than I'vebeen using Claude previously.
So right now Claude is my numberone go-to tool.
I feel that it's better thanChachi pity in basically.
Everything.
I have only two annoying thingswith Claude, and one of them is
that I feel that it thinks waytoo long on some things that are
sometimes really basic.
Like when I ask a really complexthings, I understand, think as
(18:20):
long as you want and get back tome.
It actually dings you when it'sdone, so it's actually really
cool.
You don't have to actually sitthere and wait.
So what I've been doing recentlyis a lot of AI multitasking.
I have multiple tabs open, somewith Cha g pt, some with Claude,
some with gr, some with vibecoding tools and they're all
running in parallel and I'mjumping between them, uh, and as
they're thinking and doing theirthing, I'm checking the status
of another task, giving it myinput and moving forward.
And I find this to be extremelyeffective in being productive
(18:43):
and generating a lot more in agiven amount of time.
And yet Claude, in some cases,things way too long on stuff
that is very basic.
And if GPT five one can fix thatand maybe write as good as
Claude, that will be fantasticand it will give more points to
chat.
GPT.
By the way, the best tool rightnow from my perspective as far
as balancing speed andefficiency is grok.
(19:03):
Grok will spit out answersimmediately like the old models
did when it has the answer andwill think longer for more
complex tasks.
And I, again, I think from abalance perspective, they got it
right.
So if GPT 5.1 moves in thatdirection, that would be great.
From a personality perspective,I actually mentioned something
that I think about it on ourcommunity Friday, AI Hangouts.
Our AI Friday Hangouts is acommunity of people that care
(19:25):
about ai, learning about ai, andwe meet every Friday at 1:00 PM
Eastern.
More on that in a couple ofminutes, but yesterday in this
meeting, I shared that I thinkthat AI personality should adapt
to the task versus to theperson.
Because for brainstorming, Ineed a specific kind of
personality.
When I'm assessing strategicdecision in my different
businesses, I want a differentpersonality.
And when I'm asking fortactical, mundane tasks, I don't
(19:46):
need any personality.
Just do the thing that I'masking you to do.
And so if the model will learnto adapt to what I am doing and
understand which personality I'mlooking for at that time, that
would be the most impactful forme.
So when I share that, one of thecommunity members actually
shared with me in real timeduring the meeting, a link to
Ethan Malik's post on LinkedIn,so those of you who don't know
Ethan, Ethan Molik, he is aprofessor and he's been sharing
(20:06):
amazing insights both onLinkedIn and on x and on his
blog about ai and its actualreal day-to-day impact and
different use cases, ordefinitely a person you want to
follow.
And he basically said similarthings.
He said that OpenAI has aninteresting task and that 5.1 is
really trying to balancebetween, and I'm quoting people
who want to chat with an AIseeking a quirky old buddy
(20:27):
against pros that are craving.
Every ounce of smarts around thestuff that they need to do for
their business.
And he shared something verysimilar to what I thought, that
he thinks that AI should takedifferent roles in specific
instances versus having apersonality that is fixed for
specific individuals.
He argued, and I'm quoting whowants to talk to a cynic all the
time, but then he said that youactually would want somebody
(20:48):
cynic when you're trying to getreal hard feedback on something
that you need the feedback for.
The bottom line, it adds only a0.1 to GPT five, and yet it
seems to be significantly betteracross multiple different things
and while maybe the technology'snot a big jump forward, the
access to the technology becomessignificantly more effective,
which as I mentioned multipletimes in the recent few months,
the way we interact with thesemodels, the tooling around these
(21:09):
models will play a much biggerrole than the technological
advancements in these modelsbecause it just makes them more
helpful, which is what youactually need.
So whether the underlying modelis the same, but you can get
more out of it right now, thatis a big step forward.
Now, the release of GPT 5.1without any lead time to it, and
not too long after releasing GPT five.
is hinting together with a veryserious rumor meal on X that
(21:30):
Gemini three is imminent andit's potentially might be
released in the next few days.
However, other stuff happenedthis week that puts a question
mark to that.
So several different, extremelypowerful Chinese models were
released in the last 10 days.
Moonshot just released Kmi K twothinking, which is not just
another model.
It is a model that plays a verysignificant role in the crazy AI
(21:53):
race between the US and China.
So Kmi K two scores 44.9 onhumanity's last exam, which we
talked about several times onthe show.
It's 2,500 questions of thehardest questions that the
people who put this benchmarktogether could harvest from
experts across multipledisciplinaries.
44.9 puts it at the number onespot in the world on this
(22:13):
benchmark outpacing G PT fiveand anthropic CLO sonnet 4.5 and
it's currently number one andit's outpacing GPT five and
anthropic clo sonnet 4.5, whichare the most advanced models.
In the Western Hemisphere now,this new reasoning focused
variant of Chemic K two alsoexcels in coding and agentic
tasks and logical problemsolving, and it is showing very,
(22:35):
very clearly that China is notbehind in the AI race.
Did das, who is their partner atMenlo Ventures said, and I'm
quoting, today, is a turningpoint in ai, a Chinese open
source model is number one.
So this is toward, so we arevery close to the end of the
year and this is another deepseek moment, which is how we
started the year.
So if you remember in thebeginning of 2025.
(22:56):
The world was shocked by deepsix's.
First release a Chinese model,open source model that was at
par or very close to that withthe leading models from the West
at a significantly lower cost ofdevelopment and significantly
lower cost to use the model.
Well, this release of Chemic Ktwo is that on steroids.
It's actually better than theWestern models on several
(23:17):
different benchmarks.
Now, to put things inperspective.
On the LM Arena, on the overallchart, the one that combines all
the different aspects,including, uh, hard prompts and
coding and math, and creativewriting and instructions
following and all of that.
It is currently sharing theeighth point where it's number
one in math.
Number two in creative writing,and number three in coding.
And again, number eight overall,but it is not the only model
(23:39):
that was just released fromChina.
Two different models from Baiduwere released this week.
One is a thinking version ofErnie 4.5.
Which is using only 3 billionactive parameters out of 28
billion total parameters throughmixture of experts architecture,
which many of the new models aredoing.
But the fact that it can runwith only 3 billion parameters
enables it to run on a single 80gigabytes GPU, which is
(24:00):
something you can have on thecomputer in your house.
One of the things that excels atis visual reasoning, and it can,
as they're saying, zoom in andout on specific aspects of the
image, finding specific details.
It is very good at trackingevents across videos.
And even solving STEMphoto-based problems outpacing
GT five High and Gemini 2.5 Proin document and chart
(24:21):
benchmarks, despite using afraction of the resources that
they need to use.
But they also released Ernie5.0, which is the next version
with some additional tools thatcomes with it, and it's acing.
Several different benchmark likethe OCR, bench doc VQA and the
chart QA benchmarks.
Again, all in visualunderstanding, recognition and
comprehension of visual cues anddata analysis from both visual
(24:44):
and numerical information, whichenables it to do stuff that is
really, really important for anybusiness process.
So the ability to analyze data,visual data, and numerical data
is the key of running successfulbusinesses, and it can do that
better than the top models ofthe west.
Robin Lee, the CEO of Baidusaid.
When you internalize ai, itbecomes a native capability and
transforms intelligence from acost into a source of
(25:06):
productivity.
And I agree with him a hundredpercent if you understand how to
use AI effectively.
And now that they're developingthese extremely efficient models
that are practically free.
You can completely transformbusinesses as they run today
with the ability to analyze hugeamounts of data and make better
decisions on how you're gonnarun your business, grow your
business, operate your business,and so on.
Now, when I say they'resignificantly cheaper, here is
(25:28):
the cost of this model comparedto the models from the West.
So if you don't know all thesemodels, when you run them
through the API are measured bythe cost per million tokens.
GPT 5.1 price is one point$25for every million tokens for for
every million input tokens.
Basically what you type into themodel, the prompts that you're
giving, and it's$10 per everymillion tokens on the output.
(25:50):
The answer you're getting fromChad GPT, so a dollar 25 and$10
Ernie five, cost 0.000.
Eight$5 per input token and0.0034 for an output token.
That is almost 1500 timescheaper for input tokens and
almost three times cheaper onthe.
Output tokens.
(26:11):
But they're not the last modelthat we're going to talk about
that came out of China.
That is extremely powerfulZippo, which is a new company
that just launched a model inSeptember called GLM 4.6, just
claiming really high scores onthe leaderboard as well.
So as an example on the webdevelopment leaderboard on the
LM arena, GLM 4.6 is now numberfive.
(26:33):
Ahead of it it's just ClaudeOpus 4.1, Claude Sonnet, 4.5 GPT
five Medium, and another versionof Claude Sonnet 4.5.
Behind it there are Coin three,minimax, which is a Chinese
model.
N Gemini 2.5 Pro and Grok Codeand other tools.
So on numbers five, six, andseven on the web development, we
now have AI tools that aresignificantly cheaper than the
(26:55):
equivalence from the west.
As you remember, I told you lastweek that Jensen Huang said that
China is winning the AI race.
Now while he has a clear agenda,when he's saying that it is
currently unclear who isactually winning the race and
definitely not who is going towin the race eventually, but
there's one thing that is asclear as daylight and that is
when it comes to building highlycapable and highly efficient
models.
China is currently kicking ass.
(27:17):
It is developing models that aresignificantly more efficient
than the Western models, both inthe cost of creating them and
definitely in the cost ofrunning them.
And why does that matter?
It matters because while peopleare right now willing to pay a
premium for the most advancedmodel in the world.
This will stop sometime very,very soon because when the
models become good enough to doa task, then you start looking
(27:37):
at cost, right?
So maybe the other model, maybeClaude 4.5 sonnet is the best
coding tool out there right nowthat is obvious across multiple
benchmarks and on the EllaMarina, but it is 15 x and in
some cases a hundred x moreexpensive.
Is it worth it when you're goingto generate huge amounts of code
with it?
The answer is, it depends on theuse case, but it's definitely
not the correct answer everysingle time, which means more
(27:59):
and more companies are going tomove to Chinese models and use
them as the backbone of whatthey are going to develop next,
just because of cost.
Which leads me to the bigcritical question that is the
outcome of that, which is Geminithree really far ahead with a
big enough gap from theseChinese models?
Will it be able to compete withthem on quality and on cost?
(28:21):
Because if not, Google may delaythe release of Gemini three.
So if you are open ai, goingfrom G PT five to GPT 5.1 makes
a lot of sense.
It's just a 0.1, it's not a bigdeal and yet we got a better
model out there.
But if you are releasing a majormodel that is supposed to be the
backbone of your AI deliverablesin the next six months.
It cannot be behind open source,nearly free models from China
(28:42):
across more or less anyimportant benchmark.
And so I'm certain that a lot ofpeople inside of Google are now
testing their models and tryingto optimize their model.
And if it's not dramaticallybetter or at least close in
price to the Chinese models, Idon't know if they're going to
release it.
They may release a Gemini 2.7 orsomething like that in between,
and I obviously don't knowwhat's the level of readiness of
(29:02):
Gemini three.
I obviously don't know how goodit is.
I only, all I know is the rumorsthat I'm seeing on X, which are
rumors, but it will be veryinteresting to see how this
evolves.
And I will obviously report assoon as I hear what's happening.
Now there are dozens of moreimportant pieces of news this
week, but sadly, I need to go tothe airport and you can still
learn about all these news justby signing up to our newsletter.
So there's a link to that in theshow notes.
(29:24):
You can click on the link andsign up and get access to all
the news every single week.
By the way, it's not just thisweek.
Every single week there's newsthat do not make it into the
recorded version, and they allexist in the newsletter.
But I promised you moreinformation about our AI Friday
Hangouts.
So there is an incredible groupof AI enthusiasts that meets
every single week.
Friday, 1:00 PM Eastern, forover a year now.
And we are talking about ai,we're talking about big picture,
(29:45):
where the world is going, howit's gonna impact our society,
but we also talk very tactical,reviewing tools and use cases.
That everybody from thecommunity is sharing and showing
other people new tools, thatthey found new use, cases that
they implemented, discussdifficulties and other people
are trying to help solve them.
It's an incredible community,and I really want you to learn
more about this because I wouldlove other people to join us as
well.
And so I decided to give you aglimpse into what's happening in
(30:07):
the AI Friday Hangouts, and thatglimpse is going to come this
Tuesday as a mix of severaldifferent discussions that we
had in the Friday Hangouts, andit is going to be magical, I
think for some of you, becauseyou will see a great mix of
different kinds of discussions.
Some of them are highlytactical, explaining very
specific things and how theywork, and even sharing specific
use cases with prompts andeverything else.
And some are very big pictureand.
(30:28):
In both cases today and onTuesday, there's a link in the
show notes for you to come andjoin those Friday Hangouts.
It doesn't cost anything it'snot mandatory.
You just join the community andyou can join us whenever you can
on Fridays.
That's it for this week.
I hope you all keep onexperimenting with ai, sharing
what you learned.
If you are enjoying thispodcast, Please subscribe on
your podcast and platform, andif you're on either Apple
(30:49):
Podcast or Spotify, please giveus a review that helps us get to
more people and share it with afew people that you know can
benefit from this.
Just click the share button onyour podcast platform and send
it to a few people that you knowcan benefit from it.
I'm sure you know some of thesepeople.
That's it for this week.
We'll be back on again, a uniqueepisode on Tuesday.
And until then, have an amazingrest of your weekend.
Hello and welcome to a Weekend News episode of the
Leveraging AI Podcast, thepodcast that shares practical,
ethical ways to leverage AI toimprove efficiency, grow your
business, and advance yourcareer.
This is.
Isar Meitis, your host, andtoday is gonna be a short news
episode because I just came backfrom a trip from Chicago doing a
keynote in a workshop, and I'mleaving today to Europe to do
(00:20):
two different AI workshops.
So I have a limited amount oftime, and I actually considered
not doing an episode today, butthere are two really big things
to talk about.
So we will do a shorter episodein which we're going to focus on
these two things.
One is going to be world models,and the other is gonna be models
who are ruling the world.
And you'll understand in aminute as we get started, there
are a lot more news from thisweek.
A lot of really interestingnews, and you can find all of
(00:40):
those in our newsletter, but wewill focus on these two topics.
So let's get this started.
So world models is somethingthat we mentioned several times
previously on the show, but itwas never a big deal and
definitely never the main topicof the episode.
But it is today because we gotthree different things in the
(01:01):
world models that happened thisweek.
And the first one we're going totalk about is the departure of
Jan Koon from meta.
So we talked about young Launmany times before in the show.
He's the chief scientist atMeta.
He's a touring award winner.
He was an NYU professor and he'sone of the pioneers of the
current AI era.
He has been the chief scientistin.
(01:22):
Back then Facebook since 2013.
So he started AI research waybefore the whole ChatGPT craze,
and he built the Facebook AIresearch lab, also known as
Fair.
But before that, he hasdeveloped Alex net, which he
developed together with astudent of his, that was a
really advanced visual analysisnet, which was able to beat the
legendary image net that existedbefore that by 10%.
(01:44):
For the first time.
Any other model was beatingthat.
So he was in the world ofresearching and developing
visual.
Aspects of AI way before thecurrent era, he won the Turing
Award together with JeffreyHinton and Yoshua Bengio, uh,
also known as the Godfathers ofai.
So he is been around the AIspace longer than most
scientists that are in it today.
And he just announced that he'sliving meta and he's living meta
(02:06):
to develop his own company thatwill focus on world models.
So a little bit of background.
Laun has been very loud sincethe launch of ChatGPT stating
that large language models willnot lead to a GI.
He always suggested that theonly way to get to a GI will be
to develop models whenunderstand the world as the
world is and learn like babiesas we learn from the environment
(02:26):
and not through just language,which is a very narrow way to
view the world.
he's known for a tweet thatsays, it seems to me that before
urgently figuring out how tocontrol AI systems much smarter
than us, we need to have thebeginning of a hint of the
design for a system smarter thanhouse cat.
Basically comparing largelanguage models to a house cat
in their level of intelligence.
Now there are multiple reasonsof why Jan would leave meta.
(02:49):
The biggest one is obviously theturmoil that's happening in meta
right now.
So he was the head of the AIshow.
And after the not so successfulrelease of LAMA four, he was,
and I'll be very gentle,sidelined by Mark Zuckerberg who
spend$14.3 billion into scale AIin order to poach Alexander
Wang, their CEO to become thenew head of Meta's Super
Intelligence Lab, which has alsopoached multiple scientists from
(03:13):
the top leading competitorsspending hundreds of millions
and sometimes billions in orderto get this kind of talent.
And since then, there's been acomplete chaos over there.
But it was very obvious that YLaun is not in the center of
this.
And there's other people whowill be.
And so from a politicalperspective, it was a good time
for him to leave, but I ampretty sure it is also going to
make him extremely wealthybecause I'm pretty certain he'll
(03:33):
be able to raise a few billionsof dollars with evaluations of
even more than that to build hisAI world's vision because.
There are more and more voicesthat are saying that this is the
real way to move to an A GIfuture.
Again, he's been definitely oneof the key people that was
driving this forward.
Now, is that tied to his sourcesof visual aspects of ai?
Potentially because there'sothers in a similar path that
(03:55):
think like him.
So since there's not much toreport right now on exactly what
Jan Koon will be doing, uh, I'llswitch to another person with a
similar background who is makingsimilar waves, and that is Fei.
FEI Lee.
So who is Faye?
Faye Lee.
She is another one of the earlyscientists in the modern era of
AI and one of the leadingvoices, and she was a Stanford
professor.
(04:15):
Since 2009, and she's the onethat in Stanford developed
ImageNet, which we mentionedearlier as a early AI tool that
could read text, and evenhandwriting, which was then
surpassed later by alexNet thatwas developed by Koon, so these,
how all these people are tiedtogether and been working on
similar things for many, manyyears.
She developed ImageNet back in2009, but she also co-founded AI
(04:38):
For All, which is a nonprofittwo Boost AI education, and she
did that in 2017 where most ofus couldn't spell ai.
She won multiple award and she'sconsidered the godmother, if you
want, of AI in the modern era.
And in 2014 she founded WorldLabs, which is a company that is
building AI driven worlds.
Now she just published asubstack called From Words to
(04:59):
Worlds Special Intelligence isAI Next Frontier.
A quick summary of what is inthat letter that will help you
potentially understand why thisis so important, so the idea is
that ai, in order to make thenext progress, need to move
beyond language and maybe imagesto understanding and interacting
with a 3D physical and orvirtual environments.
She says that current ai, likeChatGPT, are wordsmiths in the
(05:22):
dark, that they're eloquent, butnot grounded in physical reality
or with a real deepunderstanding.
Of how our world operates.
So she's asking the question ofwhat is spatial intelligence?
And she's saying it's theability to perceive reason about
and interact with the physicalworld in a 3D space.
And she's also saying that thisis the foundation of our
intelligence of human cognitionand how we navigate, how we
(05:45):
create, how we imagine, how wesolve problems.
It's all built on how we learnas babies before we even have a
language.
And she gives examples.
We know how to park a car.
We know how to catch keys whensomebody throws it at us.
Firefighters knows how tonavigate buildings when they
can't see anything, and it's allbecause we have general
understanding of the environmentaround us in a very deep level.
And to prove why that is aproblem.
(06:06):
She's saying that current AI ishighly limited in understanding
the actual environment.
She's saying that LLM strugglewith very basic spatial tasks,
like estimating distances orrotating objects in a space, or
placing them correctly ornavigating mazes or.
Predicting physics, all thesethings, ai, current ai, large
language model based AI is notvery good at even AI generated
(06:27):
videos that seem to understandthe world physics, which by
itself is amazing, are limitedto X number of seconds, and now
maybe mere minutes, but that'sit.
You can't run a persistent videowith clear physics and
continuity.
That will run for 30 minutes,not to mention 30 hours right
now.
And this is why there arecompletely new sets of models to
run robotics who need to operatein the 3D world.
(06:48):
So the solution that sees she'ssuggesting is what she's calling
world models.
And world models have three mainessential capabilities.
One, it's generative.
They can create geometricallyand physically consistent
virtual and or.
Real worlds what you're saying.
Real world is obviously arepresentation of a real world
versus a completely made upworld.
The other thing is they'recompletely multimodality.
They.
Process, diverse inputs.
(07:09):
They understand image, video,text, gestures, and physics to
produce complete world states.
And the third one is thatthey're interactive.
They can predict the next statesbased on actions and goals that
are driven inside that worldthat they have created or that
they operate in.
So from a timeline inapplication perspective, she
sets three different times.
One is the near term, basicallyright now, which is creative
(07:30):
tools for filmmakers, gamedesigners, architects, that
allows you to build 3Denvironments without the need
and the overhead of traditionalsoftware development.
Just by giving it a prompt andit will generate everything you
want with accuracy andconsistency, which is the key.
The mid to long term, three tofive years, the idea is to
develop whole environments forrobotics and what's called
embodied ai.
(07:51):
That we'll be able to drive awide variety of different kinds
of robots that will be able tooperate effectively and safely
in a wide variety ofenvironments.
And in the long term, she'stalking about scientific
discovery, healthcare,diagnostics, and immersive
education.
Basically being able to buildworlds to teach people anything
that you want to learn, beinginside that environment, most
likely with some kind of virtualreality headset.
Now, she's not just talking orwriting blog posts, she's also
(08:14):
making it happen.
But before I tell you what she'smaking happen, I want to quote
one line out of this manifestothat she shared, because I found
it very, very powerful.
She said, and I'm quotingextreme narratives of techno.
Utopia and apocalypse areabundant these days, but I
continue to hold a morepragmatic view.
AI is developed by people, usedby people, and governed by
people.
(08:34):
It must always respect theagency and dignity of people.
Its magic lies in extending ourcapabilities, making us more
creative, connected, productiveand fulfilled.
Spatial intelligence representsthis vision.
I think this is beautiful andwonderful, and I really, really
wish that all the people who aredriving AI today would feel the
same way and act based on theseprinciples.
But now to the big news fromWorld Labs, again, FEI Fey's
(08:56):
Company.
So after raising$230 million,world Labs is finally launching
its first commercial productjust a year after emerging out
of stealth, and it even has afreemium access.
You can use it for free withlimited generations, but what it
knows how to do is it knows howto generate either realistic or
completely made up worlds fromprops.
(09:17):
So, as I mentioned, you can useit for free for up to four
generations.
Uh, the pro tier that is$35 amonth give you 25 such
generations.
So she quoted on this launch,the new generation of world
models will enable machines toachieve spatial intelligence on
an entirely new level.
So this model is called Marbleand in addition to the ability
to generate worlds from aprompt, which by itself is
really cool, but there's othercompanies who do this.
(09:38):
Google has a similar tool aswell, but Marble comes with a
few really cool tool.
One is called Chisel.
Which allows to decouple thestructures such as walls and 3D
blocks from the visual styles,so you can control both the
building blocks of the world aswell as the style of the world.
Separately.
It allows you to move objects inthe world and place them in
different places to have bettercontrol on how the world is
(09:59):
built or like their co-founderJustin Johnson said, I can just
go there and grab the 3D blockthat represents the couch and
move it somewhere else.
But then you can also define howthe couch is going to look like.
So this allows more control tocreatives to create the worlds
they want.
Exactly.
They what they want them.
You can even expand the worldsby providing them a growing
edges, and you can even use whatthey call a composer to merge
(10:22):
scenes, to create vast, biggerspaces, whether photorealistic
or game like outputs.
That can be exported in severaldifferent format, either GA and
splats, which is theseenvironments that you can look
through or meshes, or videosthat can be ideal for either
gaming or just to navigate withVR headsets.
Now, in addition, they releaseda tool that allows marble to
create assets Existing graphicengines such as unity and
(10:44):
unreal.
So instead of just creatingentire new worlds you can create
accurate 3D objects for existingengines, which dramatically will
increase the pipeline ofgeneration of 3D objects into
the existing gaming universe.
So to summarize this componentabout Lee's ambitions, she said,
our dreams of truly intelligencemachines will not be complete
without spatial intelligence.
And this aligns perfectly withwhat Jan Koon is doing.
(11:07):
And this is, and this alignsperfectly with what Jan Laun is
saying and now going to bedoing.
As well.
And there, as we mentioned,coming from a very similar
aspect of the AI universe, so itmakes sense that they think the
same way, but this is not thelast piece of news about
understanding and operating in3D environments.
DeepMind just released CMA two,which is a.
AI that can reason in 3D worldsand play extremely well in every
(11:31):
computer game, including reallysophisticated computer games.
How does it work?
Well, they basically let it playany game, including really
advanced, sophisticatedmultiplayer games in complex
environments.
And it is learning on its ownhow to play the game.
Well, I'm not a gamer, but inthe release notes, they shared
several really sophisticatedgames like Mind Dojo or asca,
(11:52):
that it was very successful inplaying, learning over 600
skills and contextual reasoningin these games on its own.
So how does it work?
It's powered by Gemini modelsbehind the scenes and it.
Understands the environment ofthe game and it understands how
the game works.
And it can even communicate thatsuch as I'm going to the village
center, or I'm finding acampfire, or things like that
because it understands how thegame actually operates.
(12:14):
It understands how to do reallysophisticated aspects like
mining different things in orderto achieve other goals in the
game.
And it comes with zero humanassistance.
It basically.
Plays the game on its ownGemini's reasoning, trying to
understand what the roles andthe goals of the games are, and
it learns how to operate inthese worlds by itself.
In the release notes of CMA twoDeepMind highlights that gaming
is a, quoting incubator forgeneral intelligence, and it is
(12:37):
currently available only for agated preview, but they are
planning to release at leastreports on everything that it
does and can do later on.
Now, why are we talking about anAI that can play games, and why
does DeepMind think that it's anincubator for general
intelligence.
Well, here's a little story foryou.
I've been traveling literallyevery single week in the past, I
don't know, eight to 12 weeks,something like that.
(12:58):
Seems forever delivering eitherkeynotes or AI workshops to
companies and organizations.
And I just came back fromChicago and I had the
opportunity to finally watch theThinking Game.
The Thinking Game is a movieabout Demis AEs.
And his quest for AI forHumanity.
If you want, those of you whodon't know Demi, he's the
founder and still the chief ofGoogle DeepMind.
Well, when he founded, it wasjust DeepMind before Google
(13:19):
bought them.
I highly, highly recommendwatching this movie.
It is really, truly inspiring.
But it connects well to what wejust talked about because it
shows how AI learning from gamescan deliver in the end, very
significant results.
So before I tell you what thatis, two words about the movie
and about Demis.
First of all, the movie made mefeel completely worthless.
I mean, I really liked Demi'sapproach before the movie.
(13:41):
It always felt a lot moregenuine, realistic, and really
grounded in.
True drive to make the world abetter place with AI compared
with some of the other leadersof the other labs.
And this movie made me feelsignificantly more like this
about Demi.
He's absolutely incredible.
He's really driven by making theworld a better place with the
usage of AI and watching hisjourney and thinking how he
(14:03):
feels and how he drives otherpeople around him.
Made me feel completelyworthless compared to what he is
doing.
It is really inspiring, but toconnect the dots to this segment
of this podcast.
As you probably know, Demis wonthe Nobel Prize for alpha fold.
Alpha fold is a model that canaccurately predict protein
folding.
Now why the hell does thatmatter?
Because proteins are, if youwant the machinery of life.
(14:24):
Like everything we know as aliving, anything is built on
proteins that are foldedtogether, but trying to
understand how they're folded,basically, how they create a 3D
structure that then becomes lifeis a crucial component of A
understanding life.
And B, being able to understanddisease and designing new drugs
to solve every illness on theplanet.
Well, previously, before alphafold, it was extremely difficult
(14:47):
and very inaccurate tounderstand the structure of a
protein.
It would take years in a lab tofigure out one protein.
Well, alpha fold is a predictionmodel can predict protein
structures in minutes.
And get it highly accurate.
And they've used it to predictstructures of over 200 million
proteins and then open sourcedit so anybody can use it for
(15:08):
research.
So how does that connect togaming?
Well, in the path to getting toalpha fold, the previous
iterations that gave DeepMindthe capability to even go down
that path was gaming.
They have developed, many of youprobably know of Alpha Go, which
has beat Lee Seel and afterwardsthe Chinese World Champions in
the Game of Go.
so these attempts to have amodel learn through a gaming
process with clear goals andlearning on its own how to play
(15:31):
a game, were the building blocksthat allow them to later on
develop alpha fold.
And most of the bigbreakthroughs that DeepMind have
created, and there's a long listof them, came from teaching AI
to play games.
So going from being able to playa single, really sophisticated
game like go to, being able toplay any computer game on its
own is a very big step on thepath to a GI.
(15:53):
And this is what DeepMindreleased right now.
So that ends this topic.
Now what you need to do is goand watch the movie and we'll
switch to the next topic, whichis the arrival of GPT five one
and the other models around it,and maybe the imminent release
of Gemini three.
So OpenAI just released GPT 5.1,which doesn't sound like a big
spread from five.
It just adds a 0.1 at the end,but it's actually a dramatically
(16:13):
different model that has a lotof big, significant
improvements.
So on the high level, theyreleased two different models,
GPT five, one instant, which isfaster and more conversational,
and it has a warmer tone foreveryday tasks and conversing
and GPT 5.1 thinking, which isslower, but has much better
complex reasoning, math, logic,multi-step planning, and all
these kind of things thatthinking models are good at now.
(16:34):
What are the biggestdifferences?
Well, first of all, it is.
it is a good combination ofbeing a smarter and B, more
natural and human-like.
So it is better at instructionfollowing an accuracy.
It is more conversational andless robotic.
Tom, which was one of the bigpushbacks against GT five, and
it has fewer hallucinations.
The GPT five, all of these areawesome.
It also has adaptive thinking,so they dramatically improved
the amount of time that it needsto think for specific tasks.
(16:56):
So for simple things, it willanswer faster than G PT five,
and for more complex things, itwill think longer than G PT
five.
In both cases, providing youbetter answers, aligned with the
relevant amount of time thatneeds to be invested in actually
getting there.
It also comes with severaldifferent personalities that you
can select from professional,candid, quirky, friendly, and
efficient.
And the basic model is finetuned for being warmer and more
conversational.
(17:17):
and it even proactively offersto adjust the tone during
conversations depending on whatis actually happening.
It is supposed to be better atcreative writing, better at
coding with more tools that itcan use, and it has enhanced
planning and multi-step task, exexecution.
So on paper, a significantlybetter model than GPT five.
Now I didn't get a chance toplay with it enough again.
(17:37):
I just came back from oneworkshop and I'm shortly leaving
to the airport for another, but.
I can say a few things.
On a high level, better, promptcoherence is always a good
thing, especially as we'removing from basic, simple things
to more advanced and complextasks that we want AI to
achieve.
And building agents andmulti-step agents and
multi-level agents.
So following instructionsaccurately is a very big deal,
but also being able to optimizethe amount of time that the
(18:00):
model thinks is a big deal.
I've been recently using Claudea lot more than I've been using
ChatGPT and a lot more than I'vebeen using Claude previously.
So right now Claude is my numberone go-to tool.
I feel that it's better thanChachi pity in basically.
Everything.
I have only two annoying thingswith Claude, and one of them is
that I feel that it thinks waytoo long on some things that are
sometimes really basic.
Like when I ask a really complexthings, I understand, think as
(18:20):
long as you want and get back tome.
It actually dings you when it'sdone, so it's actually really
cool.
You don't have to actually sitthere and wait.
So what I've been doing recentlyis a lot of AI multitasking.
I have multiple tabs open, somewith Cha g pt, some with Claude,
some with gr, some with vibecoding tools and they're all
running in parallel and I'mjumping between them, uh, and as
they're thinking and doing theirthing, I'm checking the status
of another task, giving it myinput and moving forward.
And I find this to be extremelyeffective in being productive
(18:43):
and generating a lot more in agiven amount of time.
And yet Claude, in some cases,things way too long on stuff
that is very basic.
And if GPT five one can fix thatand maybe write as good as
Claude, that will be fantasticand it will give more points to
chat.
GPT.
By the way, the best tool rightnow from my perspective as far
as balancing speed andefficiency is grok.
(19:03):
Grok will spit out answersimmediately like the old models
did when it has the answer andwill think longer for more
complex tasks.
And I, again, I think from abalance perspective, they got it
right.
So if GPT 5.1 moves in thatdirection, that would be great.
From a personality perspective,I actually mentioned something
that I think about it on ourcommunity Friday, AI Hangouts.
Our AI Friday Hangouts is acommunity of people that care
(19:25):
about ai, learning about ai, andwe meet every Friday at 1:00 PM
Eastern.
More on that in a couple ofminutes, but yesterday in this
meeting, I shared that I thinkthat AI personality should adapt
to the task versus to theperson.
Because for brainstorming, Ineed a specific kind of
personality.
When I'm assessing strategicdecision in my different
businesses, I want a differentpersonality.
And when I'm asking fortactical, mundane tasks, I don't
(19:46):
need any personality.
Just do the thing that I'masking you to do.
And so if the model will learnto adapt to what I am doing and
understand which personality I'mlooking for at that time, that
would be the most impactful forme.
So when I share that, one of thecommunity members actually
shared with me in real timeduring the meeting, a link to
Ethan Malik's post on LinkedIn,so those of you who don't know
Ethan, Ethan Molik, he is aprofessor and he's been sharing
(20:06):
amazing insights both onLinkedIn and on x and on his
blog about ai and its actualreal day-to-day impact and
different use cases, ordefinitely a person you want to
follow.
And he basically said similarthings.
He said that OpenAI has aninteresting task and that 5.1 is
really trying to balancebetween, and I'm quoting people
who want to chat with an AIseeking a quirky old buddy
(20:27):
against pros that are craving.
Every ounce of smarts around thestuff that they need to do for
their business.
And he shared something verysimilar to what I thought, that
he thinks that AI should takedifferent roles in specific
instances versus having apersonality that is fixed for
specific individuals.
He argued, and I'm quoting whowants to talk to a cynic all the
time, but then he said that youactually would want somebody
(20:48):
cynic when you're trying to getreal hard feedback on something
that you need the feedback for.
The bottom line, it adds only a0.1 to GPT five, and yet it
seems to be significantly betteracross multiple different things
and while maybe the technology'snot a big jump forward, the
access to the technology becomessignificantly more effective,
which as I mentioned multipletimes in the recent few months,
the way we interact with thesemodels, the tooling around these
(21:09):
models will play a much biggerrole than the technological
advancements in these modelsbecause it just makes them more
helpful, which is what youactually need.
So whether the underlying modelis the same, but you can get
more out of it right now, thatis a big step forward.
Now, the release of GPT 5.1without any lead time to it, and
not too long after releasing GPT five.
is hinting together with a veryserious rumor meal on X that
(21:30):
Gemini three is imminent andit's potentially might be
released in the next few days.
However, other stuff happenedthis week that puts a question
mark to that.
So several different, extremelypowerful Chinese models were
released in the last 10 days.
Moonshot just released Kmi K twothinking, which is not just
another model.
It is a model that plays a verysignificant role in the crazy AI
(21:53):
race between the US and China.
So Kmi K two scores 44.9 onhumanity's last exam, which we
talked about several times onthe show.
It's 2,500 questions of thehardest questions that the
people who put this benchmarktogether could harvest from
experts across multipledisciplinaries.
44.9 puts it at the number onespot in the world on this
(22:13):
benchmark outpacing G PT fiveand anthropic CLO sonnet 4.5 and
it's currently number one andit's outpacing GPT five and
anthropic clo sonnet 4.5, whichare the most advanced models.
In the Western Hemisphere now,this new reasoning focused
variant of Chemic K two alsoexcels in coding and agentic
tasks and logical problemsolving, and it is showing very,
(22:35):
very clearly that China is notbehind in the AI race.
Did das, who is their partner atMenlo Ventures said, and I'm
quoting, today, is a turningpoint in ai, a Chinese open
source model is number one.
So this is toward, so we arevery close to the end of the
year and this is another deepseek moment, which is how we
started the year.
So if you remember in thebeginning of 2025.
(22:56):
The world was shocked by deepsix's.
First release a Chinese model,open source model that was at
par or very close to that withthe leading models from the West
at a significantly lower cost ofdevelopment and significantly
lower cost to use the model.
Well, this release of Chemic Ktwo is that on steroids.
It's actually better than theWestern models on several
(23:17):
different benchmarks.
Now, to put things inperspective.
On the LM Arena, on the overallchart, the one that combines all
the different aspects,including, uh, hard prompts and
coding and math, and creativewriting and instructions
following and all of that.
It is currently sharing theeighth point where it's number
one in math.
Number two in creative writing,and number three in coding.
And again, number eight overall,but it is not the only model
(23:39):
that was just released fromChina.
Two different models from Baiduwere released this week.
One is a thinking version ofErnie 4.5.
Which is using only 3 billionactive parameters out of 28
billion total parameters throughmixture of experts architecture,
which many of the new models aredoing.
But the fact that it can runwith only 3 billion parameters
enables it to run on a single 80gigabytes GPU, which is
(24:00):
something you can have on thecomputer in your house.
One of the things that excels atis visual reasoning, and it can,
as they're saying, zoom in andout on specific aspects of the
image, finding specific details.
It is very good at trackingevents across videos.
And even solving STEMphoto-based problems outpacing
GT five High and Gemini 2.5 Proin document and chart
(24:21):
benchmarks, despite using afraction of the resources that
they need to use.
But they also released Ernie5.0, which is the next version
with some additional tools thatcomes with it, and it's acing.
Several different benchmark likethe OCR, bench doc VQA and the
chart QA benchmarks.
Again, all in visualunderstanding, recognition and
comprehension of visual cues anddata analysis from both visual
(24:44):
and numerical information, whichenables it to do stuff that is
really, really important for anybusiness process.
So the ability to analyze data,visual data, and numerical data
is the key of running successfulbusinesses, and it can do that
better than the top models ofthe west.
Robin Lee, the CEO of Baidusaid.
When you internalize ai, itbecomes a native capability and
transforms intelligence from acost into a source of
(25:06):
productivity.
And I agree with him a hundredpercent if you understand how to
use AI effectively.
And now that they're developingthese extremely efficient models
that are practically free.
You can completely transformbusinesses as they run today
with the ability to analyze hugeamounts of data and make better
decisions on how you're gonnarun your business, grow your
business, operate your business,and so on.
Now, when I say they'resignificantly cheaper, here is
(25:28):
the cost of this model comparedto the models from the West.
So if you don't know all thesemodels, when you run them
through the API are measured bythe cost per million tokens.
GPT 5.1 price is one point$25for every million tokens for for
every million input tokens.
Basically what you type into themodel, the prompts that you're
giving, and it's$10 per everymillion tokens on the output.
(25:50):
The answer you're getting fromChad GPT, so a dollar 25 and$10
Ernie five, cost 0.000.
Eight$5 per input token and0.0034 for an output token.
That is almost 1500 timescheaper for input tokens and
almost three times cheaper onthe.
Output tokens.
(26:11):
But they're not the last modelthat we're going to talk about
that came out of China.
That is extremely powerfulZippo, which is a new company
that just launched a model inSeptember called GLM 4.6, just
claiming really high scores onthe leaderboard as well.
So as an example on the webdevelopment leaderboard on the
LM arena, GLM 4.6 is now numberfive.
(26:33):
Ahead of it it's just ClaudeOpus 4.1, Claude Sonnet, 4.5 GPT
five Medium, and another versionof Claude Sonnet 4.5.
Behind it there are Coin three,minimax, which is a Chinese
model.
N Gemini 2.5 Pro and Grok Codeand other tools.
So on numbers five, six, andseven on the web development, we
now have AI tools that aresignificantly cheaper than the
(26:55):
equivalence from the west.
As you remember, I told you lastweek that Jensen Huang said that
China is winning the AI race.
Now while he has a clear agenda,when he's saying that it is
currently unclear who isactually winning the race and
definitely not who is going towin the race eventually, but
there's one thing that is asclear as daylight and that is
when it comes to building highlycapable and highly efficient
models.
China is currently kicking ass.
(27:17):
It is developing models that aresignificantly more efficient
than the Western models, both inthe cost of creating them and
definitely in the cost ofrunning them.
And why does that matter?
It matters because while peopleare right now willing to pay a
premium for the most advancedmodel in the world.
This will stop sometime very,very soon because when the
models become good enough to doa task, then you start looking
(27:37):
at cost, right?
So maybe the other model, maybeClaude 4.5 sonnet is the best
coding tool out there right nowthat is obvious across multiple
benchmarks and on the EllaMarina, but it is 15 x and in
some cases a hundred x moreexpensive.
Is it worth it when you're goingto generate huge amounts of code
with it?
The answer is, it depends on theuse case, but it's definitely
not the correct answer everysingle time, which means more
(27:59):
and more companies are going tomove to Chinese models and use
them as the backbone of whatthey are going to develop next,
just because of cost.
Which leads me to the bigcritical question that is the
outcome of that, which is Geminithree really far ahead with a
big enough gap from theseChinese models?
Will it be able to compete withthem on quality and on cost?
(28:21):
Because if not, Google may delaythe release of Gemini three.
So if you are open ai, goingfrom G PT five to GPT 5.1 makes
a lot of sense.
It's just a 0.1, it's not a bigdeal and yet we got a better
model out there.
But if you are releasing a majormodel that is supposed to be the
backbone of your AI deliverablesin the next six months.
It cannot be behind open source,nearly free models from China
(28:42):
across more or less anyimportant benchmark.
And so I'm certain that a lot ofpeople inside of Google are now
testing their models and tryingto optimize their model.
And if it's not dramaticallybetter or at least close in
price to the Chinese models, Idon't know if they're going to
release it.
They may release a Gemini 2.7 orsomething like that in between,
and I obviously don't knowwhat's the level of readiness of
(29:02):
Gemini three.
I obviously don't know how goodit is.
I only, all I know is the rumorsthat I'm seeing on X, which are
rumors, but it will be veryinteresting to see how this
evolves.
And I will obviously report assoon as I hear what's happening.
Now there are dozens of moreimportant pieces of news this
week, but sadly, I need to go tothe airport and you can still
learn about all these news justby signing up to our newsletter.
So there's a link to that in theshow notes.
(29:24):
You can click on the link andsign up and get access to all
the news every single week.
By the way, it's not just thisweek.
Every single week there's newsthat do not make it into the
recorded version, and they allexist in the newsletter.
But I promised you moreinformation about our AI Friday
Hangouts.
So there is an incredible groupof AI enthusiasts that meets
every single week.
Friday, 1:00 PM Eastern, forover a year now.
And we are talking about ai,we're talking about big picture,
(29:45):
where the world is going, howit's gonna impact our society,
but we also talk very tactical,reviewing tools and use cases.
That everybody from thecommunity is sharing and showing
other people new tools, thatthey found new use, cases that
they implemented, discussdifficulties and other people
are trying to help solve them.
It's an incredible community,and I really want you to learn
more about this because I wouldlove other people to join us as
well.
And so I decided to give you aglimpse into what's happening in
(30:07):
the AI Friday Hangouts, and thatglimpse is going to come this
Tuesday as a mix of severaldifferent discussions that we
had in the Friday Hangouts, andit is going to be magical, I
think for some of you, becauseyou will see a great mix of
different kinds of discussions.
Some of them are highlytactical, explaining very
specific things and how theywork, and even sharing specific
use cases with prompts andeverything else.
And some are very big pictureand.
(30:28):
In both cases today and onTuesday, there's a link in the
show notes for you to come andjoin those Friday Hangouts.
It doesn't cost anything it'snot mandatory.
You just join the community andyou can join us whenever you can
on Fridays.
That's it for this week.
I hope you all keep onexperimenting with ai, sharing
what you learned.
If you are enjoying thispodcast, Please subscribe on
your podcast and platform, andif you're on either Apple
(30:49):
Podcast or Spotify, please giveus a review that helps us get to
more people and share it with afew people that you know can
benefit from this.
Just click the share button onyour podcast platform and send
it to a few people that you knowcan benefit from it.
I'm sure you know some of thesepeople.
That's it for this week.
We'll be back on again, a uniqueepisode on Tuesday.
And until then, have an amazingrest of your weekend.
Popular Podcasts
Las Culturistas with Matt Rogers and Bowen Yang
Ding dong! Join your culture consultants, Matt Rogers and Bowen Yang, on an unforgettable journey into the beating heart of CULTURE. Alongside sizzling special guests, they GET INTO the hottest pop-culture moments of the day and the formative cultural experiences that turned them into Culturistas. Produced by the Big Money Players Network and iHeartRadio.
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by Audiochuck Media Company.
The Brothers Ortiz
The Brothers Ortiz is the story of two brothers–both successful, but in very different ways. Gabe Ortiz becomes a third-highest ranking officer in all of Texas while his younger brother Larry climbs the ranks in Puro Tango Blast, a notorious Texas Prison gang. Gabe doesn’t know all the details of his brother’s nefarious dealings, and he’s made a point not to ask, to protect their relationship. But when Larry is murdered during a home invasion in a rented beach house, Gabe has no choice but to look into what happened that night. To solve Larry’s murder, Gabe, and the whole Ortiz family, must ask each other tough questions.