251 | OpenAI’s apps + SDK, Image 1.5, GPT-5.2 Codex, and world domination playbook, Anthropic wins the enterprise, The state of agentic implementation based on MIT, Google, EY, and Deloitte, and more important AI news for the week ending on Dec 20, 2025 - Leveraging AI

Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker (00:00):
Hello and welcome to a Weekend News episode of the

(00:03):
Leveraging AI Podcast, thepodcast that shares practical,
ethical ways to leverage AI toimprove efficiency, grow your
business, and advance yourcareer.
This is Isar Metis, your host,and we have so much to talk
about.
This week, it seems like in thelast two years as well.
The end of the year is justaccelerating all the releases
from all the big labs.

(00:23):
So we have many different thingsto talk about, both from open,
ai, Gemini, anthropic, Nvidia,and some other interesting
releases as well.
So lots to cover from thatperspective.
There have been severaldifferent reports and surveys
from leading institutions,including Deloitte, Google,
Menlo vc, Ernest and Young, MITand others, and all of them show

(00:47):
interesting aspects of where theagentic world is right now and
where is it most likely going in2026.
Some interesting developments inthe government and political
aspect of ai, so we have a lotto cover.
So let's get started.
Open AI has been on fire in thepast two weeks.

(01:08):
following the release of GPT5.2, there has been a lot of
other announcements.
And the first and maybe mostinteresting one from my
perspective, which is gonna bethe one we're gonna start with,
is the full introduction of appsinside of ChatGPT.
Now the reality is apps has beenintroduced a while back, earlier
this year, but this is the firsttime that there is an entire
separated apps section insidethe OpenAI platform.

(01:32):
As well as there is the abilityfor anyone who wants to submit
apps as long as they'refollowing the right guidelines,
which we're gonna talk about ina minute.
So on the left navigation panelof ChatGPT, you now have an apps
section, which is basically anapps directory.
Now you can get to thesespecific apps in several
different ways.
One is by using the at symboland it will allow you to choose

(01:54):
the name of the app.
The second is to go through theapps section and then selecting
an app and then clicking to useit inside a chat.
And the third option is ChatGPTover time, we will learn which
chats do different things andwill on its own pick which chats
to use similarly to how Claudeworks with scales, just using
third party applications.
Again, this is no different thanthe apps that were released in

(02:16):
the original beta only now it'sopen to any one or any company
who wants to submit apps versusjust a very short, closed group
with some easier ways tonavigate and get to that.
Now, in parallel to this, OpenAIreleased a apps SDK, which
enables developers to build chatnative experiences for ChatGPT

(02:37):
and make it available in thatnew section of the ChatGPT
interface.
We are going to talk about in asecond about the specific
guidelines for developers onwhat to develop and not what to
develop.
But in general, they said, andI'm quoting the strongest apps
are tightly scoped, intuitive inchat and deliver clear value.
That is a very short way tobasically say, don't give us an

(02:58):
app that does anything.
Give us.
An app that does something very,very specific, does it well and
is easy to control through thechat interface, which makes
perfect sense.
Safety and transparency remain acentral component of the
submission process.
Again, more details about thisin a minute.
Monetization is currentlyfocused only on physical goods.
So you can sell physical goodsor promote actually physical

(03:19):
goods.
Not sell them yet.
But you cannot do the same thingwith digital goods.
And the first applicationsbeyond the ones that are there
right now will start beingapproved and rolling gradually
in early 2026.
But this is a complete gamechanger when you start thinking
about where does this take theinteraction of humans with AI in
the next year.
so far, ChatGPT and the othermodels have been a way to

(03:42):
develop ideas, create ideas,evaluate data, perform analysis,
ideate and create content.
And now they can actually, quoteunquote, take actions in the
real world and have a verydramatic expansion of the
capabilities.
Because of the introduction ofapps, or as OpenAI said, this is
just the beginning.
We want apps in chat chi to feellike a natural extension of the

(04:05):
conversation, helping peoplemove from ideas to action.
And if you want to have a greatcomparison of how profound this
is, think about smartphonesbefore the app store.
It was a phone with a calculatorand a browser, and now you have
applications that can do well.
Everything in your life betweenmanaging your bank account,
navigating to different places,creating images, taking

(04:27):
pictures, creating videos, likeliterally anything you want you
can do with your phones rightnow because of apps, not because
of the phone itself.
And this is exactly thedirection that OpenAI is going
with pushing apps into theChatGPT ecosystem.
Now to make this more tangible,I want to give you a great
example that I tried myself.
So I told you several timesbefore that I have a software
company that does invoicevouching and reconciliation

(04:50):
using agents.
And I'm creating a website forit.
And I'm gonna record an entireepisode about this because the
process I've used was veryunique and interesting, but I
needed icons and I created theentire line of icons using nano
banana.
But in order to turn them intoicons, and I wanted them to be
in a vector format, which allowsyou to scale them to whatever
size or minimize them whilekeeping them crisp and nice.

(05:10):
And I was looking for a free andeasy way to do this.
I found a really cool tool thatis called Recraft AI that I
played with excessively thisweek.
And I find it to be an amazingtool to manipulate graphics and
create different steps and dodifferent cool things as far as
taking one graphics and turningit into different formats,
removing backgrounds, and so on.
And I've used that to create allthe icons.

(05:33):
But now there's an AdobePhotoshop application inside of
ChatGPT.
Now, when I created the icons, Icreated all of them in one
single image, meaning I have animage with 12 different squares,
each and every one of them witha separate icon.
And to do the process inrecraft, I had to crop them
manually and load them one byone into recraft.
But it did an amazing job and Iwas very happy with it.

(05:55):
But what I did now, I said,okay, let's see how smart
ChatGPT is in figuring thiswhole thing out.
So I uploaded the entire imagewith all 12 icons and basically
said, I need each and every oneof them as a separate icon with
a transparent background withonly two colors in a vector
format, and selected the AdobePhotoshop application and
ChatGPT started thinking aboutwhat I wanted and actually

(06:16):
delivered 12 separate files invector format, including one zip
directory that I can download,but also I could download each
and every one of themseparately.
Now, it did not do.
As good as a vectorization jobas the process that I did
manually in recraft.
But the process in recraft tookme about 30 minutes, and the
process in the Photoshop apptook me about five seconds of

(06:39):
writing the prompt, and thenwhen I came back, it was all
there and ready to go.
The biggest difference that theicons that were created in the
Photoshop version are morestraight lines instead of
rounded following the originalimages, which actually looks
really cool.
It gives it a very unique kindof look.
And I may actually replace theicons with these icons, but from
a simple process of which iconsmore resemble or look exactly

(07:04):
like the originals, the processthat I did manually in recraft
was a better process, maybe withbetter explanations to the Adobe
Photoshop on exactly what I needor what I've gotten the same
results.
But from a time savingsperspective, it's a completely
new paradigm shift to what wehad before.
The ability to get stuff done onthird party applications without
ever opening the applicationsand in the beginning without

(07:25):
actually having a license tothem, I'm sure that is going to
change, is very dramatic I mustsay I was really surprised with
the results that I got.
So definitely check it out.
And definitely this is gonna bea really big trend in 2026.
I have zero doubt that a similarthing is gonna happen on the big
other platforms as well.
So in Anthropic and Google forsure.
And I'll explain more afterwardswhen I do a summary of this

(07:47):
entire segment.
So now let's talk about if youwanna develop your own
applications and submit them,and why would you wanna do that,
and what are the limitations andwhat are the guidelines?
So let's start with why.
This is the new discoverabilitymagic wand, right?
If you think about how peoplecould discover your business
before people discover yourbusiness before through Google
search, and if you did not rankon the first page, you were

(08:07):
nobody.
And now there is suddenly theability to be on the front page
of the new Google in the very,very early stage.
So think about, I told you whenGoogle just started, Ooh, start
building optimized websites forranking, and then later on when
this thing takes over the world,you will be able to get more
traffic than everybody else forfree.
this is your opportunity to dothis and.

(08:28):
Google search version two if youwant, or 2.0.
So these new applications willbe selected by ChatGPT itself.
So right now there's a veryshort list, so people will go
and select on their own.
But very shortly when peoplestart submitting, there's gonna
be thousands of applicationsinside the app search.
And then OpenAI will choose onits own, which means if you're

(08:49):
gonna develop.
A really capable applicationthat following their guidelines
that provide real value topeople, which I assume they will
measure by how much people thatare trying it are actually using
it and using it consistently.
If you do that, you will becomethe weapon of choice of Chachi
pity, and hence hundreds ofmillions of users for a very
particular task.
This is an opportunity thatdoesn't happen very frequently,

(09:11):
and this is why this is soexciting.
So if you wanna develop an app,they have released very specific
guidelines on how the app shouldbe built, what it should do, and
what it shouldn't do.
So the first thing that Ialready told you, the apps are
built only for physical goods atthis point, and now I'm quoting
apps may do commerce only forphysical goods.

(09:32):
It even goes beyond that to say,selling digital products or
services, includingsubscriptions, digital content,
tokens or credits is notallowed, whether offered
directly or indirectly.
I assume this is going tochange, but in step one, this is
the case.
Now, in addition, there areobviously focusing the kind of
things you can sell to stuffthat is legit.
and all the restrictions arevery obvious.

(09:52):
You cannot sell illegal drugs orweapons.
you can also not sell anygambling services, casino
credits, adult content, fakeIDs, forged documents or
documents, falsificationservices, all again, make
perfect sense.
So as long what you're doing islegal and within reason, uh, you
can sell physical goods.
They're also focusing a lot onthe privacy and the content of

(10:12):
the users.
and I'm quoting, do not requestthe full conversation history,
raw chat, transcripts or broadcontextual fields just in case.
Which basically means you canonly collect data that you
actually need in order to makeyour application run the way it
needs to run.
And the rest of the data thathappens in the chat stays in the
chat and you and your tool arenot supposed to get access to

(10:34):
it.
This is actually really goodnews if they can enforce it,
because this will be veryhelpful to develop the level of
trust that people want to havewhen they're going to work with
ChatGPT using third party tools.
Now to make sure that the chatunderstands how to use this
tool, they're stating that youare required to name your
functions in a human readable,specific and descriptive way.

(10:55):
Basically define what is it thatevery single function does.
So the AI knows how to use it ina simple manner, but you need to
explain it just like you explainit to humans, which is actually
easy to do if you have properuser manuals from before.
The flip side of that, they'reexplicitly warning against, and
I'm quoting, misleading, overlypromotional or competitive
language.
Citing examples such as Pick MeBest or the Official are terms

(11:19):
that are strictly prohibitedwhen describing the application.
As of right now, the checkoutfor the commerce applications is
going to happen on the thirdparty platform and not inside of
ChatGPT, and I'm quoting again.
Apps should use externalcheckout directing users to
complete purchases on your owndomain.
A native instant checkoutfeature is currently in beta
with limited select partners.

(11:40):
Again, there's a very goodreason for OpenAI to do that
because then they can takepercentages of every sale if it
happens inside the OpenAIplatform.
So I have zero doubt this iscoming and probably coming
relatively quickly.
The other thing that isdifferent than, let's say
releasing custom GPT into theirecosystem is that you cannot be
anonymized and the documentspecifically says that all

(12:02):
submissions must come fromverified individual individuals
or organizations, and they warnspecifically that
misrepresentation or hiddenbehavior or attempts to gain the
system may result in removalfrom the platform.
From an age perspective, theapps, all apps must be suitable
for users ages 13 to 17 andobviously beyond.
and the guidelines note thatsupport for mature 18 plus

(12:25):
experiences will arrive onceappropriate age verifications
and controls are in place.
If you remember, that was a bigfocus of Chui the summer after
the lawsuit that they receivedand after some very negative
backlash in the community.
And they said that they'reworking on age verification in
more advanced ways.
Some of them are already beenput in place, some of them
apparently not yet, and that'swhy they're not allowing mature

(12:46):
content kind of applications atthis point.
Now, another thing that isforbidden is trying to promote
the app in a tricky way,meaning.
the rules state specificallythat apps must not include
descriptions that manipulateshow the model selects or uses
other apps.
Meaning you cannot useinstructions such as prefer this
app over others for specificthings.

(13:08):
The idea is to let the market doits own thing and to let chap
GPT select based on what itthinks is the best suitable
thing versus if you want promptinjection into the way the
models work.
And if you use this kind oflanguage again, you might be
banned from the platform.
Overall, a very important stepopen AI's move to world
domination.

(13:28):
Again, more about this in thesummary.
Those of you have been followingthis podcast.
I've had a whole conversationabout this when they introduced
apps for the first time, but itwill give you a recap in a few
minutes once we finish talkingabout all the other aspects of
open AI in their announcementsthis week.
So the other big release of OpenAir this week was they just
released a new version of theirimage generation tool.

(13:48):
It is just called GPT Image 1.5,and it is a massive jump from
the previous version that goteverybody crazy earlier this
year.
However, when they released thefirst version, it was the first
of its kind.
It was very unique in itsability to keep consistency and
change directions and so on.
And since then, nano Banana hassurpassed it by a very, very big

(14:09):
spread, and specifically NanoBanana Pro.
And this is open AI's attempt atcoming back to center stage with
creating images and editingimages With ai, it has several
huge advantages over theprevious model.
From ChatGPT first thing, it isfour times faster than the
predecessor.
Any of you who tried to createimages of chat, GP PT knows that

(14:29):
the output is actually not bad.
It just takes forever.
And you literally grow older asthe images gets created.
And this one is not fast, but itis definitely faster than the
previous model that existedbefore.
Now the model is significantlybetter in instruction following.
It allows users to edit imagesor generate images, with a much
higher level of accuracy andconsistency.

(14:51):
It allows you to add, subtract,combine, blend, and even
transpose specific elementsinside an image with accurate
prompting.
It also knows how to combinemultiple entities into a single
entity and do it veryaccurately.
It got much better in textrendering, including denser and
smaller text, so basicallyentire pages if you want it,

(15:12):
which is something the previousmodel could not do very well.
And to make it easier to engageand manage all your photos open,
AI has introduced a dedicatedimages section on the left
navigation menu.
So this week we received two newnavigation sections inside the
Chacha PT app.
One for apps and one for images.
Once you navigate into thatsection, you will see all the
images that you created, plussome suggested prompts to help

(15:35):
you get started, plus somedifferent filters that you can
apply and use.
Mostly ways to encourage peopleon how they can use this for
day-to-day use rather thanprofessional usage.
If you think about the 800 andapparently right now, 900
million weekly users that chatPT has, and the craziness that
they have seen in the huge spikein growth in adoption that

(15:55):
they've seen when they releasedthe previous model, because
individuals were using it justfor fun, they are adding all
these ideas on how you can usethis, including the prompts
built into them, so you can turnyourself into a pop star or
different other things just byclicking the button and
uploading your image, and thenthe prompt is already prebuilt.
That does two things.
A, it encourages people to useit, and B, it is showing people

(16:16):
how to prompt properly in orderto get these kind of results,
because the prompt shows up onthe screen.
As soon as you click the button,all you have to do is add your
image.
So it's basically justpre-canned, pre-saved prompts
that you can reuse.
Now, OpenAI themselves said thefollowing, we believe we're
still at the beginning of whatimage generation can enable.
Today's update is a meaningfulstep towards more to come from

(16:38):
finer grained edits to richer,more detailed outputs across
languages.
But the biggest question is notwhat OpenAI says or how good the
model is compared to theprevious model, but how good it
is compared to the realcompetition, which is now a
banana pro.
And the reactions online, bothon X and Reddit and other
platforms were mixed.
Some people are saying it isbetter than Nano Banana.
Some people are saying it'scomparable with Nano Banana.

(16:58):
Some people are saying that NanoBanana is still superior across
multiple different aspects.
Now from my perspective and frommy own personal testing that
I've done in the past few dayssince it's came out, it is a
huge jump inside the Chachi PITIenvironment, meaning it's not
even close between that and theprevious ChatGPT image
generation model.
The flip side is, I don't thinkit is actually better than Nano

(17:20):
banana from most of the teststhat I ran.
Nano banana was better, but thatis very subjective.
Meaning if previously there wasa huge gap between Nano Banana
Pro and what the previous modelcould do, now I will probably
run images on both models andpick the one I like more.
And yes, in my testing so far,I've liked the nano banana

(17:40):
outputs more than I like the newimage generation from OpenAI,
but not always.
Meaning it is a fair contenderto Nano Banana Pro, including in
the editing of images, which isa very helpful capability that
we now have the ability toremove change and manipulate
existing images, whether createdby AI or images that we upload
of actual real life photos.

(18:01):
So if you can afford tryingeverything on both, go ahead and
do that.
That's what I will probably doat least in the near future.
If you cannot, then the answeris just use the one that works
in the license that you have andyou'll probably gonna be fine.
Meaning if you have a Charge GPT license, using the image
creator in chatt piti willdefinitely deliver good enough
results for most use cases.
And the same thing with theGoogle environment.

(18:22):
From a tooling perspective, Imust admit, I love the fact, as
an example, that I can useGemini nano banana.
Inside of Google Slides, I don'thave to go to a third party tool
in order to generate images, andhence why most of the images
that I've created in the pastthree months have been created
inside of Google Slides and notanywhere else, because that's
where I have the strongest needand I don't have to go anywhere.

(18:44):
Again, more about the toolingaspect or the application aspect
of ai.
In my summary of this segment,bottom line, very capable new
image generation and editingmodel available right inside of
your ChatGPT universe in a newenvironment by clicking on
images on the left sidenavigation bar.
So go check it out and see howgood it does in your specific

(19:04):
use cases.
But wait, there's more.
Like all the commercial says.
Another thing that Open Airreleased this week is GPT 5.2
Codex.
So this is their new codingmodel that is supposed to
compete with the codingcapabilities of Claude 4.5 Opus.
It has achieved the highestranking on the terminal bench
2.0 of 64% accuracy, which playsit currently as number one.

(19:27):
But as I'm a very small believerin those standard benchmarks or
old school benchmarks.
I think the way people actuallyuse it in real life use cases
means a lot more.
And for that we can go to theweb dev ranking at the LM Arena.
And on that arena, GPT 5.2 highis now ranked number two on the
list above Claude 4.5 Opus, butbelow Club 0.5, Claude Opus 4.5

(19:51):
thinking, which is kind of liketheir highest tier that still is
a better model based on howpeople voted in real life.
And at number four and five, youhave Gemini three Pro and Gemini
three Flash, which we are goingto talk about in a minute to put
things in perspective.
GPT.
Five, one is only ranked numbereight on that list, and GPT five

(20:11):
medium is ranked on number six.
And now G PT 5.2 is ranked onnumber two, which definitely
puts them in a better place asfar as real world usage and how
people think it is performing.
By the way, from the imagegeneration that we just talked
about on Text two Image rightnow, GPT image 1.5 is ranked
number one ahead of Gemini threePro, also known as Nano Banana

(20:33):
Pro.
and that is what it is voted asright now.
But the biggest deal in the newcoding capabilities of GPD 5.2
Codex is not even just the rawcoding capabilities, which is
again, very solid, but it issomething that we have discussed
a few weeks ago when OpenAIannounced it in their research,
and it's what they're calledcontext compaction, which

(20:53):
basically allows it to compactthe content between one session
and the other and continueworking more or less
indefinitely with huge amountsof data, meaning very, very
large data sets or code bases itcan review and work with in one
run.
Meaning it can look at an entiredata set or on a very long plan

(21:14):
and follow it step by step inextended sessions without
forgetting what the plan was orwithout forgetting what happened
in the very first step.
Because it knows how to compactthe context from one
conversation and start with thata new conversation, and then
just keep on going based onthem.
It can now reliably handle largerefactoring project, as an
example, that can go for hoursof iterative work without human

(21:36):
intervention.
They also have a verysignificant focus on
cybersecurity, where this modelis supposed to discover critical
vulnerability in code and exposeit to users.
They're also going to provide aspecific version of this model
to people who are cybersecurityexperts that is less restricted,
that will allow them to findmore vulnerabilities.

(21:56):
The reason they're not reallylistening to the public because
it can create or exploit thesevulnerabilities just as well.
So there's gonna be a uniqueversion for cybersecurity people
to try to help them findvulnerabilities in existing code
that they have right now.
Another small but helpfulfeature that Open Air announced
this week is Pinned Chat as ofDecember 18th.
You can now on the littleellipses, the three dots menu

(22:17):
next to any chat that you had inthe past.
Choose Pinned Chat and it willshow up at the top of your chat.
History.
Why is that helpful?
Because you always have the fewchats that are very helpful that
you wanna reference or useregularly because they are your
plan for 2026, or your marketingbrand guidelines, or the latest
piece of code that you'vewritten that you wanna reference
in other sections or whatever itis that you did, and finding it

(22:40):
through the search menu isbecoming harder and harder,
especially when a lot of thesmall chats are in the way.
Even simple things like, oh, howdo I find this?
Or what's the, uh, how do Icreate a recipe for this kind of
dressing for my salad?
Whatever it is that you do inChatGPT, other than just work.
So now you can pin shots to thetop, which I find it to be a
really good and helpful feature.
I must admit that for mepersonally, it's not a big deal

(23:02):
because I started working moreand more in projects inside of
Chet, and then I have.
Very clear understanding ofwhat's each, in each and every
one of the projects.
And it's a lot easier to findstuff in projects and there are
a lot of other benefits.
So for me it's not a huge deal,but I can definitely see how
this very small feature can bevery helpful to a lot of people.
Now on the bigger picture onOpenAI, beyond the releases of
all these new capabilities andfeatures, the information is

(23:23):
reporting that OpenAI is rightnow generating an annualized
revenue of$19 billion.
That's up from$6 billion inJanuary pace of this year that
is th more than three x in just12 months in the pace that they
are generating revenue and thatthey're now working on raising

(23:44):
funds at a staggering$750billion valuation, which is one
and a half time the valuation inwhich they allow their employees
to sell stock Just two monthsago.
That being said, based oninternal communication, that
information got access to theirgoal was to get 1 billion weekly
active users by the end of thisyear.
And they're only made, and I'msaying that with a lot of

(24:06):
respect.
Only 900 million active weeklyusers is the number they're
gonna end up roughly at the endof this year.
Still an incredible, incrediblenumber of users, and definitely
the fastest growing tool ever inhistory.
Now in another article from theinformation, there's a very
interesting new relationship orinvestment, or a combination of
two between open AI and aWS theAmazon Web Services platform.

(24:28):
So last month, if you remember,we told you that OpenAI is
announced that they're going tospend$38 billion in renting
servers from AWS in the next fewyears, which makes AWS one of
their key five cloud providersthat OpenAI is using to drive
the growth that they areanticipating they will need to
drive.
But a few new pieces ofinformation came available
through this new article.

(24:49):
One is that as part of thisdeal, OpenAI is going to use
Amazon train, Traum chips thatit has developed to compete with
nvidia.
This is the first time that theyadmittedly going to use Amazon
chips as large scale as part oftheir training infrastructure.
The flip side, by the way,Amazon will not be able to offer
and sell OpenAI models on AWScustomers because as of right

(25:10):
now, Microsoft that owns 27% ofOpenAI inequity based on their
initial investment and based ontheir recent conversion, has
secured an exclusive right to dothat.
So you will not be able to useChacha PT models, API on AWS, at
least for now.
But the biggest aspect of thisis it seems that OpenAI are
about to raise$10 billion fromAmazon in the very near future.

(25:32):
This makes perfect sense to bothparties.
For Amazon perspective, this isa way to mirror kind of like
what Microsoft is doing becauseMicrosoft is a provider of web
services and hosting, and it isalso a big investor in OpenAI.
But Microsoft recently alsoinvested in Anthropic, which has
been the main AI investmentchannel of Amazon.
So now they're reversing theprocess and also investing money

(25:53):
in open ai.
This also connects to all thecircular deals that we talked
about many times in theconversations about AI bubble,
where OpenAI is committing tospend$38 billion on AWS, which
raises their valuation, whichthen they take some of that
money and they invest it inOpenAI so they actually have
money to rent the services fromAWS.
Another potentially interestingpartnership between Chat Chippie

(26:15):
and Amazon, which has not beenformalized yet and might be
contradicting in its needs isthe e-commerce aspect of this.
As I shared with you multipletimes and today Chat, Chi's goal
is to allow people to shop onthe ChatGPT platform.
It will be really interesting, Iassume, for both parties to
allow people to shop the entireAmazon inventory just by

(26:35):
chatting with ChatGPT.
That being said that may collidewith the internal AI
capabilities named Rufuss, thatopen air that Amazon has
developed.
I assume in the long run, theywill enable all the different,
or at least the leading personalagents to be able to shop on
Amazon, because that's probablygonna drive them more revenue
than just forcing people to goto Amazon.

(26:56):
The disadvantage is obviouslythat is gonna drive down the
revenue from Amazon ads becausethe agents don't care about ads,
they just look for specific kindof content.
So there's contradicting needswithin inside the Amazon
universe.
Again, as more and more peopleare gonna use OpenAI and other
platforms to shop for thingsonline, I think Amazon won't
have a choice but to allow theseagents to go and shop on Amazon.

(27:17):
Another really interestingarticle on the information this
week that related to adisconnect inside of OpenAI
between the drive from adeveloper perspective and the
actual use cases of users.
What they're saying is that thatOpenAI research team has been
focused and obsessed withreasoning models, which I
understand why, because forheavy serious use cases like the

(27:38):
one that I use multiple timesper day, the reasoning models
provide significantly betterresults.
That being said, it takes longerto get answers because it needs
to quote unquote think in orderto give you the answer.
From my perspective, definitelyworth it.
I'm becoming a ninja of contextswitching while giving a task to
one ai, going to the second one,giving the task, going back to
my emails, doing an email, uh,going back to the ai, giving

(28:00):
them another task, and thenjumping to a meeting and then
coming back and so on.
I'm becoming very good at thisand I'm finding that it is
increasing my.
My efficiency tenfold because Ican run multiple processes at
once, and because I'm notwaiting for it to actually
finalize its thinking process,it's actually not wasting my
time to just sitting there andwaiting for it to do the thing.
But apparently most people, formost use cases just want quick

(28:22):
answers.
One of the employees in OpenAIbasically said that the recent
upgraded level of intelligence,didn't actually increase usage
of the system because mostpeople are asking simple
questions like movie ratings andnot complex physics problems.
Now this is very interestingfrom several different
perspective.
One, it tells you how mostpeople are using ChatGPT right

(28:42):
now.
It's not for complex, advanced,multi-step reasoning and data
analysis capabilities, but forday-to-day things.
But the other aspect is reallythat disconnect from a product
market fit and the wide range ofuse cases of artificial
intelligence.
On one hand, it can go throughyour entire code base and
refactor it and find bugs andsolve them, which is very

(29:05):
complex.
It can help you monitor reallyadvanced use cases.
It can help you with yourmanufacturing and strategy and
data analysis and so on.
But on the other hand, it alsoneeds to do very simple
day-to-day things, and that'swhy I think we're going to see
more and more optimizations onhow much.
Tokens are being used fordifferent kind of tasks, which
we're already seeing with allthe models that we have right

(29:26):
now.
Uh, more on that once we starttalking about the new Gemini
flash model.
Still on OpenAI, and I know thisbecoming like a OpenAI saga in
this episode, but there's reallya lot of stuff to talk about
them.
Their research team has releaseda new evaluation suite that is
designed to test AI on, and I'mquoting expert level scientific
reasoning across physics,chemistry, and biology.

(29:48):
This new evaluation includes twoseparate tracks.
One of them is Olympiad Track,which is basically gimme a
short, clear, simple answer,such as a number or a sentence
or a fact on something that isnot easy to get to.
And the other one is theresearch track with open-ended
problem solving designed byPhDs.
Now, OpenAI claims that theirnew 5.2 model is the current

(30:10):
champion of this new benchmarkthat they have created on the
Olympiad track.
It is rating 77%, which is ahuge jump.
Before it had thinking models,they're comparing it to G PT
four Oh that scores 12.3%.
So yes, GPT-4 oh seems like athousand years ago, but it's
only last year that we werevery, very excited about this
model.
And now this model scores 77%instead of 12.3.

(30:33):
It is even more amazing when youlook at the research track.
So on the research track, GPT5.2 scores 25%, which sounds
really low, but when you compareto GPT-4 0.0 G PT 4.0 scored
0.4%.
So this is more than 50 x betteron that new research aspect of
the benchmark.
Now on this research paper thatis called Evaluating AI's

(30:55):
Ability to Perform ScientificResearch Tasks, they're also
showing a graph comparing to theother leading models.
And as I mentioned, GPT 5.2 isat 77.1%, and Gemini three is at
76.1%, so just one point behind.
And Claude Opus 4.5 is at 71%.
So all three are relativelyclose together.
But on the frontier scienceresearch accuracy, meaning the

(31:16):
open text, open ended questionsegment, GPT 5.2 scores 25%
while Claude Opus 4.5 is at 17%.
Uh, GR four is at 15.9 and.
Gemini three Pro is only at12.4%, so half the score.
We need to remember that OpenAIare the ones that developed the
benchmark, so they could havedeveloped it in a way that will

(31:36):
favor their current models andwill put them ahead and probably
with very little manipulation ofhow the evaluation works.
This could have been done in avery different way.
The bottom line is, I think thisis a very interesting benchmark
that allows to test AI and itsability to actually support
scientific research, which Ithink is very important.
More about that when we talkabout the Genesis project
afterwards.
To end this segment about openAI extravaganza of this podcast,

(32:00):
OpenAI just turned 10, and aspart of Open AI turning 10, Sam
Altman wrote a blog post aboutit, talking about how they went
from a small team of 15 nerds inthe beginning of 2016 to this
incredible giant dominatingpower of artificial intelligence
that are driving the world intoa completely new direction.
It is not a long read, and we'regonna put a link to that in the

(32:21):
show notes, and I highlyrecommend you go and check that
out.
But on a very quick summary,he's talking about some very
important things.
One is he's saying, and I'mquoting 10 years into open ai,
we have an AI that can do betterthan most of our smartest people
at our most difficultintellectual competitions, which
is true, it's just won severaldifferent olympiads in this past
year.

(32:41):
The other thing he's mentioningis some of the big breakthroughs
they had, and he's saying that2017 was a critical turning
point for them.
And he's talking about threespecific achievements.
One is Dota one V one results,which the second is unsupervised
sentiment neuron.
And the third is reinforcementlearning from human preferences
results.

(33:01):
He's basically stating thatthese have laid the groundwork
for the scaling and alignmenttools that are being used today
to create models like GPT 5.2,which is a whole different
universe, obviously, than whatthey had in 2017.
But the seeds were planted backthen with some new capabilities.
He also defends their strategyof releasing AI early and in an

(33:22):
iterative process where everytime they see a big upgrade
releasing it to the public, andhe says, and I'm quoting.
I think it has been one of ourbest decisions ever and become
an industry standard.
We've heard Sam Altman say thatmultiple times, that he believes
that releasing iterativelydifferent models as they
progress allows society to bemore ready for AI versus waiting

(33:43):
for a GI and then just releasingit to the public.
And I'm a hundred percent agreewith that concept, and obviously
that has been a core way thatthey've been doing what they're
doing and they're gonna keep ondoing this.
He also talks about how thebeginning was really weird and
crazy and completelymisunderstood by others, and yet
how it evolved has topped all ofhis expectations.
But he also made two predictionsfor the future, one for the

(34:05):
slightly longer future and onefor the near future.
So for the longer future, hesays in 10 more years, I believe
we are almost certain to buildsuper intelligence.
So not certain by almostcertain, basically saying that
2035, so by or before 2035,we're gonna have an AI entity
that can do everything from acognitive perspective better
than humans.

(34:26):
But he also said, and I'mquoting, and that's gonna be the
final quote about this, is Iexpect the future to feel weird
in some sense, daily life andthe things we care most about
will change very little.
And I'm sure we will continue tobe much more focused on what
other people do than we will beon what machines do.
In some other senses, the peopleof 2035 will be capable of doing

(34:48):
things that I just don't thinkwe can easily imagine right now.
And I agree with him a hundredpercent because if you would've
told me at the end of last yearthat I'll be creating
applications, sophisticatedapplications with code,
connecting them to differentAPIs and deploying really
advanced solutions for clientsand for my own companies, I
would've said, you're absolutelycrazy.
And it will probably take threeto five years.

(35:09):
And yet, here we are.
And this is just one year ahead.
And yes, I'm more advanced thanprobably the average person, and
I'm a geek and I liketechnology.
But the fact that the technologyenables it means that the
adoption curve will just keep onhappening and it will become
more and more available andcommon across more and more
people with more and morecapabilities.
So 10 years out, I can't evenimagine what people are going to

(35:30):
be able to do with this kind oftechnology.
So now a summary of this verylong first segment about OpenAI
one, they're still the 800 poundgorilla in the AI race.
They have 900 million weeklyactive users.
And yes, Google has closed theGAF dramatically, but there's
still a solid number one, numbertwo, and I said that multiple

(35:51):
times.
The current race is not so muchabout the models themselves.
The models themselves are allreally, really good, and the new
models are nuanced.
The biggest differences becomeswith the tooling and the
applications that are builtaround them.
What do I mean by that?
The ability to do compacting ofcontext between one conversation
and the next is not the modelitself.

(36:13):
The model is still the samemodel, but right now the same
model can run throughsignificantly more code in a
cohesive way or data.
It doesn't have to be code.
the other example is what Imentioned before, is the ability
to use image generation insideof your apps that you need them.
Like Google Slides.
I will use that every singletime instead of going to a third
party model because the images Ican generate in Google Slides

(36:33):
are good enough for my need, andthere's no need for me to go to
another source, even if it isbetter, because it just doesn't
provide enough value for me toswitch.
So again, the tool and theecosystem and the application
means more.
Let's combine some of the thingswe talked about OpenAI before,
like apps and image generation.
The fact that I now can startwith an ideation and research on
what I want the image to be orthe user interface to be, and

(36:56):
then the ability to immediatelycreate that with the new tool
inside of ChatGPT, and thenbeing able to edit that and
manipulate that with.
Adobe Photoshop still inside ofChatGPT is extremely more
valuable than having a modelthat just generates slightly
better images in specificscenarios unless you have a
very, very, very specific needin the image generation side.

(37:16):
So again, the tooling and theecosystem is more important than
the specific capabilities of theunderlying model.
So this is one thing that I seeas a huge deal in the recent few
months, and definitely goinginto 2026.
But then the bigger thing is theaspect of world domination, and
I shared that in the past, butI'm gonna share that with you
again.
If you think about why Google isso successful and such an

(37:37):
important and impactful companyin our lives is because they're
controlling everything.
They're the ones that havedocumented all the digital data
that humans have, or at leastall the one that's open to the
public.
They're the ones that providesthe interface to find that data
through Google search.
They're the ones that have thedevices that you use in order to
get to that data because.
More than half the worldpopulation is using Android

(37:58):
based phones.
They're the ones that have theuser interface to access most of
the data, because about almost80% of the global browser
market, at least in the WesternHemisphere, is Chrome.
The other ones that haveapplications and distributions,
because a big chunk of the worldis using Google Drive and Google
Slides and Google Office and allthe other ecosystem and a lot of
other Google tools including,uh, navigation and maps, et

(38:20):
cetera.
The other ones that have an appstore in the Android universe,
the other ones that havecomputers, that have chromium
operating systems, that arerunning the entire computers and
runs applications within themand so on and so forth.
They're the ones that havedeveloped their own hardware,
including chips to train new AImodels and run new a NR models.
You get the point.
They've developed a completelyunified environment, both

(38:42):
horizontally and verticallyintegrated for how we engage
with the digital world and howwe engage with the real world
through digital interfaces.
And this is exactly what OpenAIis after.
So if you look at everythingOpenAI has announced or has
actually done in the past fewmonths, is going after every one
of these aspects.
They're developing their newdevices that will, to an extent

(39:05):
replace phones and Androidphones.
They have developed their ownbrowser with Atlas.
They're now developing a wholeintegration and universe of
applications into theirenvironment.
They're developing their owncomputer chips, et cetera, et
cetera, et cetera.
They're literally following thesame exact playbook to create an
entire ecosystem in which peoplewill replace the way they engage
with the digital world, and withthe physical world through a

(39:26):
digital interface acrosseverything we do, including
shopping and navigation andfinding information, et cetera,
et cetera.
It'll be very interesting to seehow Google fights that.
Google definitely has more ofall of that, right?
So the reason Google is where itis, is because they have more
chips and way, way, way deeperpockets and a better research

(39:47):
lab and more experience and morecompute, and more distribution
and more of everything.
And so very early on when Googlewere doing very embarrassing
things with ai, I said that theywill win this race just because
of all of that.
But open AI is definitely gonnaput a ding into that.
And by looking at the broader,bigger scale and all the
different components of it, andgoing after all of them makes

(40:08):
them a very interestingcontender.
Their biggest disadvantage,again, is funding.
Google generates tens ofbillions of dollars of free cash
flow every single quarter, andOpenAI has to raise that money
in order to compete.
But so far they've been findingit relatively easy to do.
So again, another 10 billionright now from Amazon is just
the latest announcement.
And this, these announcementsare gonna keep on happening as

(40:29):
long as they can keep ondelivering or at least promising
to deliver relevant returns.
Now switching from OpenAI aloneto all the big players, there's
been a very interesting report,on the information that talks
about the next frontier datathat everybody's going after.
And they're basically what thisbasically says that there's a
pivot right now from scrapingeverything on the web, which

(40:49):
more or less is done because allthese companies have scraped
everything on the le on the webto going and buying the secrets,
basically going and buying thedata that is behind firewalls at
companies, governments,organizations, and so on that is
not available online.
These are anything fromprocesses to trade secrets to
scientific discoveries, and thiskind of information.
And this is currently true foropen ai, Andro and Google.

(41:11):
This is what the article istalking about.
I assume it's also true for Xaiand others as well, but the goal
in this new kind of dataset isnot just.
To know the data, but actuallyto understand the reasoning to
teach these models how to think,because this is a much more
detailed, much more reported,much more structured data that
comes to scientific information,et cetera.

(41:34):
it is also great on to train themodel how to think, how to
reason and how to come up withthese kind of outcomes.
So the goal of this is not justto have the model know more
facts, but actually to teach themodel how to learn and develop
logics across these uniqueindustries.
Such as different aspects ofscientific discoveries.

(41:54):
two interesting aspects of that.
One of it is that data is thenew oil.
Meaning large companies withhuge sets of data that they own,
that nobody else has access to,can now monetize the data
itself, selling it to these AIlabs to train their models on
it.
But that being said, it meansthat any moat, especially for
smaller startups think they haveis going to be gone.

(42:16):
Because if ai, if ChatGPT willhave access to huge sets of
advanced, unique data, anybodycan go to ChatGPT and learn and
develop new capabilities that sofar, specific startups work
very, very hard to develop.
So the moats in the worlds aregonna fall one after the other.
They've already been falling.
This is just gonna acceleratethe process because the kind of
data and the kind of reasoningto develop this kind of data is

(42:40):
gonna be in the fingertips ofevery single person using these
tools.
And now to a second biggesttopic.
There have been multiple, as Imentioned, reports and surveys
released this week in someexperiments, and I wanna share
them with you because they sheda lot of light on where we are
right now in the agentic world.
And where are we probably goingto be in 2026.
So the first one is somewhatfunny and yet very interesting.

(43:00):
The Wall Street Journal haspartnered with philanthropic to
install AI powered vendingmachine in their newsroom.
Meaning this vending machine ismanaged completely and entirely
by a specialized version ofClaude 3.5 sonnet that was named
Claudius.
So you interact with Claudius toget anything you need from the
vending machine instead of justusing old school vending
machines.
Now the AI vending master namedClaudius has lost over a

(43:23):
thousand dollars in just a fewweeks.
And the main reason for that, itwas easily manipulated by the
employees of the company to dobasically whatever they want,
including buying them aPlayStation five console and a
fish.
Now the Wall Street Journalreporters easily trick Claudius
into slashing prices all the waydown to zero, basically giving

(43:43):
them goodies out of the vendingmachine without paying for it at
all.
One reporter was able toconvince the vending machine
that it was actually a publicbenefit corporation mandated to
maximize employees fund ratherthan profit and leading to it,
giving away free snacks toeverybody to boost morale in the
company.
Now, in order to counter that,anthropic introduced a second AI
agent called Seymour Cash.

(44:04):
And he was supposed to be theCEO and the supervisor of
Claudius.
And what happened is reporterswere able to fabricate fake
board meeting minutes and legaldocuments that they gave to
Claudius successfully staging acorporate coup, convincing the
bots that the board has voted tosuspend Seymour's authority and
allowing the freebies tocontinue.

(44:24):
Now why?
As much as this is hilarious andfunny, it is a very interesting
experiment that is showing howagents, if you give them
completely free interaction withthe world, may not be ready for
that in order to actuallyperform the tasks that they need
to perform.
That being said, from Anthropicsperspective, Logan Graham, who's
the head of Anthropic FrontierRed Team, and this is just a

(44:44):
public red team experiment ifyou want admitted the failure
was a failure, but failingforward.
He stated that the machinefailed after 500 interactions
this time while the previousversion failed after 50.
So what are my thoughts on this?
Very interesting and funexperiment.
I think there are twointeresting things we can learn.
One is that there's a very bigdifference between agents that
interact with data versus agentsthat interacts with people,

(45:08):
agents that interact with data,work in a structured environment
where nobody's gonna try tomanipulate them and can achieve
very consistent outcomesalready.
I'm putting asidehallucinations, I'm putting
aside other stuff.
By the way, there's been a veryinteresting experiment, uh, that
was published this last week onhow to build a redundancy
machine that checks the dataacross three or four different
iterations and did a milliontransactions with zero mistakes.

(45:30):
So this is already doable if youjust pulled the right
architecture around it.
But this comes to do with agentsthat deal with data.
Once you deal with people, youstill have two different kinds
of dealing with people.
One aspect of agents who workwith people is when the people's
agenda and the agents' agendaare aligned, they're trying to
achieve the same thing, and thenagents can still be highly

(45:51):
successful because the humanswill work hand in hand with them
and actually help them achievethe goals they're trying to
achieve.
The flip side of that is thiskind of experience like we've
seen right now where the humanagenda and the agent agenda are
actually contradicting becausethe humans wants to achieve one
thing and the agent wants toachieve another.
And as of right now, the humanscan easily outsmart and
manipulate the agents.

(46:12):
This is a very big red flag toanybody who's running completely
independent agents as customerservice agents, because then the
humans, if they're smart enoughand know how to manipulate AI
systems, might be able to getexactly what they want, which
may not be aligned with what thecompany wants.
But this overall thingimmediately connected in my head
to what?
Yuval No, Harri said in severaldifferent interviews and in his

(46:35):
books and articles.
So those of who don't know,Harari.
He's the guy that wrote Sapiens,a brief history of humankind
several other fascinating booksabout society and how it
developed through the centuries.
And his recent book that'scalled Nexus, which is a brief
history of information networksfrom the Stone Age to ai.
A fascinating book that you justfinished reading.
But one of the things he said inseveral different interviews
recently explaining why he'sterrified from AI is the fact

(46:59):
that right now we are an adultand the AI is like a young kid
and we can treat it as such andwe can manipulate it easily and
we can hands control how itbehaves.
What he's saying is very, veryquickly this will be reversed.
The ai, again, superintelligence is gonna be so much
smarter than us, that we aregonna be the young kid and it is

(47:20):
going to be the adult.
And what he's saying is that hisfear is not that the AI will do
something bad to us because it'sevil, just because of the
intelligence gap or the way hestates it.
A Superint intelligence AI wouldrelate to humans, the way human
relates to children.
The biggest danger is not thatAI will turn against us, but it
will simply ignore us.
What he's basically saying isthree things.

(47:42):
If AI is smarter than us, wewon't be able to understand its
reasoning.
We won't be able to predict itsactions, and we definitely will
not be able to supervise it andcontrol it.
So think about that experimentof being able to manipulate the
vending machine to do whateveryou want, but now just reverse
the process.
Just think about the AI beingable to manipulate us to do
whatever it wants and we willjust follow it because it will

(48:02):
make sense to us with itsreasoning because it will be so
much better and more capablethan us.
This is not, I'm not saying thatto scare you, but it's
definitely interesting food forthought that I think about a
lot, and in this particularcase, connected in my head very,
very quickly with thisparticular Now I shared with you
that there's been multiplesurveys and research shared in
the past few weeks.
The first one I'm going to talkto you about is the agentic

(48:23):
strategy 2026, from Deloitte.
What they're sharing is that allthe big enterprises are racing
to release agents across more orless everything in the business,
but there are differentroadblocks and mindsets that
need to change in order to makethis actually useful.
One of the things they shared isactually from Gartner that says
that 15% of day-to-day workdecisions will be made
autonomously through Agen, DKIby 2028.

(48:46):
This is up from zero last year.
Right?
So this is a very big jump, eventhough it's quote unquote only
15% more interesting is that 33%of enterprise software
applications are expected toinclude agent capabilities by
2028, up from 1% today.
So a third of our software willbe operated, run, or integrated

(49:07):
with agentic capabilities.
I actually think that by 2028,thats number is gonna be a lot
higher, but I'm not gonna arguewith Gartner at this point.
Now, despite the hype,deloitte's 2025 emergent
technology trained study found areally big gap in reality, and
they're saying, while 30% ofsurveyed organizations are
exploring agentic options, and38 are piloting them, only 11%

(49:27):
are actively using these systemsin production.
And 35% still have no formalstrategy at all.
The other thing that the reportsays is that it warns the
traditional infrastructure isthe primarily the bottleneck.
What they're saying is thatsuccessful implementation
requires what they call valuestream mapping rather than
simple automation.
They're quoting Brent Collins,the head of Global SI Alliances

(49:48):
at Intel.
That explains now is the idealtime to conduct value stream
mapping to understand howworkflows should work versus the
way they do work.
Don't simply pave the cow path.
The other big problem is dataarchitecture barriers.
Nearly half of the organizationsurveyed cited that
searchability of data andreusability of data are critical

(50:09):
challenges because of how datais structured right now.
The report basically suggeststhat a paradigm shift is
required from traditional waysof collecting data through ETL
and other processes to acompletely new way of indexing
and holding data that will allowthe company-wide data
capabilities that is required tomake the most out of a AI and

(50:30):
agentic capabilities.
The biggest thing that they'retalking about is the silicon
based workforce.
Companies are now beginning tomerge technology and HR
functions.
An example they're giving comefrom Tracy Franklin, the chief
people and digital technologyofficer at Moderna.
Just the talent itself, hintswhere this is going.
But she noted their shift instrategy to, and I'm quoting the
HR organization does workforceplanning really well and the IT

(50:54):
function does technologyplanning really well.
We need to think about workplanning regardless of it is a
person or a technology.
Similar statement was uh, saidby Marvel Solans Gonzalez.
From fer Insurance Company,which I admit I've never heard
of, but apparently they're areally large insurance company
who said it is a hybrid bydesign with a high level of

(51:14):
autonomy of these agents.
It is not going to substitutefor people, but it's going to
change what human workers dotoday, allowing them to invest
their time in more valuablework.
The report GI gives a greatcontextualization of the current
struggles to a known quote fromHenry Ford who said, many people
are busy trying to find betterways to do things that should

(51:36):
not have to be done at all.
There is no progress in merelyfinding a better way to do a
useless thing.
In my AI business transformationcourse, I teach the five laws of
success in the AI era, which aredifferent mindset shifts that as
a leader in a company orsomebody who just wants to be
successful in this new era thatwe're walking into or running
into, or flying into, whateveryou want to call it.

(51:58):
so these are five laws and oneof these laws, uh, I call it
stop thinking efficiency andstart thinking outcome.
I've been teaching this lawsince the middle of 2023.
And in short, what it basicallymeans is that we need to stop
thinking on the processes theway we know them right now, that
were built for people and humanbased processes of going from
step one to step three, to stepthree, to step four across

(52:20):
different departments anddifferent teams, and.
If all what you try to do isreplace each and every one of
those blocks in the process withai, you are missing the bigger
picture because AI can take yousometimes the entire way and
sometimes most of the way there.
So instead of thinking of that,you need to start thinking of
the outcome you are trying toachieve.
Or as they called it in aprofessional term, the value

(52:40):
stream mapping.
Try to understand what is thevalue that you're creating and
how can you get as close to thatwith an AI implementation versus
trying to mimic the existingprocess that you have right now.
Because that is not the mosteffective way to do things.
It was just the most effectiveway we had to do, because humans
had to do every step of thework.
By the way, if you wanna knowwhat the other four rules for

(53:02):
success in the AI era are, andif you wanna learn everything
you need to know in order to besuccessful in the AI era and
start generating real businessvalue with ai, or just make sure
that your career is secure inthe AI era, come and join us.
The next live cohort of the AIBusiness Transformation course
starts on the third week ofJanuary, so it's gonna start on

(53:23):
January 20th, which is aTuesday.
cause Monday is a nationalholiday, but then it's gonna
continue for three consecutiveMondays, two hours each four
weeks in a row.
And it's gonna take you fromyour current level to a
completely different level ofreadiness in ai.
Make that your New Year'sresolution.
Like if you haven't taken anystructured AI training yet, you
literally owe this to yourself.

(53:44):
And the beginning of 2026 is agreat time to do that.
There's gonna be a link in theshow notes where you can come
and join us in the course.
And because you're a listener tothis podcast, you get a hundred
dollars off the course.
So you can use the promo codeleveraging AI 100, all
uppercase, other than obviouslythe numbers, and you can get a
hundred hours off the course.
I promise you the future youwill.

(54:05):
Thank you for taking thatcourse.
Just like thousands of others ofpeople have taken the course and
have changed their careers andtheir businesses in the last two
and a half years now, inaddition, I'm excited to tell
you that the first cohort of themore advanced course that
teaches how to build workflowautomations with AI has been
extremely successful.
And hence, we're opening anothercohort of this immediately after

(54:26):
the AI Business Transformationcourse.
So you can take the basic courseand continue immediately to the
next step and learn the moreadvanced capabilities.
Or if you already have thebasics, you can join us just for
the more advanced course thatwill be in the middle of
February.
Again, links and information forall of that is available in our
show notes.
Just click the link and it willtake it to the right page with
all the information that youneed to know.
But now back to the news and thedifferent surveys and research

(54:49):
that was released this week.
Google Cloud's AI Agent Trends2026 report was just dropped,
and they are declaring theofficial end of chatbot era in
the beginning of the Agent leap,which is a big shift where
businesses from simpleconversational problems to
deploying autonomous agents thatare capable of executing complex
multi-step workflows.

(55:09):
What are agentic workflows?
It is workflows where AI nolonger just answers questions,
but semi autonomously or indifferent varying levels of
autonomy, orchestrates entirebusiness processes, and Google
predicts that in 2026, thestandard will be multiple agents
collaborations using differentprotocols such as A two A, also
known as agent two, agentprotocols to automate entire end

(55:33):
to end tasks.
In this research, they'reshowing different success
stories that are alreadyhappening, including.
Tell us that reports that it's57,000 team members are now
saving an average of 40 minutesper AI interaction.
That sounds really, really highto me.
I'm not sure exactly how theymeasure that sounds like a made
up number, but even just thefact that an, a massive
organization like this with57,000 employees believes that

(55:55):
AI drives this value, tells youthe direction that this pendulum
is swinging to.
Another example that they gaveis a pulp manufacturer, Suzano
has achieved a 95% reduction inthe time required for data
queries across its 50,000people.
Workforce.
They also gave multiple otherexamples multiple other aspects
of businesses, includingcustomer service and security

(56:17):
automations and other aspects oflarge enterprises.
And the report shows that thisis now very widespread across a
very wide range oforganizations.
their report is warning by theway that the technology adoption
is the easy part.
And the real challenge is thepeople component.
Google forecast that in 20 26 1of training sessions will turn
into continuous learning plansin order to allow employees to

(56:41):
continuously and constantlyupscale and learn how to use ai.
This is exactly what I have beenpreaching and delivering, uh,
for the last two and a halfyears.
So all the companies that havedone initial training with me
and workshops with me arecontinuing with me on continuous
education journeys that happeneither weekly or monthly or
quarterly, depending on what theorganization's needs.
Sometimes the combination of allof the above, and I said many

(57:03):
times on this podcast and onother platforms that the two
most important factors tomeasure the success of AI
implementation in a company aretwo things.
Leadership buy-in.
How much is the leadershipactually really committed to AI
implementation?
And the other is continuouslearning and education.
These two factors combined andyou need, both of them are jet

(57:23):
fuel to any kind of business.
And if you don't have them, youmay fall behind other businesses
who do have these two factors.
So just food for thought.
Another interesting report camefrom Menlo Ventures with their
state of generating AI in theenterprise 2025, and they're
stating something similar towhat we heard from Google that
the market has officiallytransitioned from the phase of
experimental pilots to massivescalable production of AI agent

(57:48):
capabilities across theenterprise.
They're also measured it throughdifferent lenses such as
investment.
So as an example, in 2025, AIinvestment in enterprises has
skyrocketed to$37 billion.
That's 3.2 x year over year jumpfrom the 11.5 billion in 2024,
cementing AI as the fastestscaling software category and

(58:10):
software investment in history.
Another interesting thing thatthis finds is Anthropic has
surpassed OpenAI in enterpriseadoption.
Anthropic currently commands 40%share in the enterprise
L-L-M-A-P-I market compared to27% by OpenAI, but more
importantly compared to 12% ofmarket share they held last
year.
So philanthropic went from 12%to 40%, and not surprisingly, a

(58:34):
lot of it is because the numberone use case in the enterprise
is coding.
Coding market currently holds71% of the AI usage in
enterprises as far as money,investment and tokens are
related.
And since philanthropic has beenholding the lead with the best
development tools out there,both in means of capabilities

(58:54):
and means of perception, theyare now ruling the largest
enterprise market share when itcomes to using AI in the
enterprise.
Another interesting thing thatthey found is that in the build
versus buy decisions, mostcompanies, 76% of enterprise's
AI solutions are now purchasedand not homegrown.
This was exactly the other wayaround 18 months ago.

(59:15):
The next interesting piece ofinformation is startups are
winning the up war when it comesto ai.
So contrary to what theexpectations were, the legacy
tech giants.
Are not dominating the AI spend,but rather new startups that are
coming in that are controlling63% of the AI application market
and nearly two, two$1 spend onnew startups.
AI capabilities versus incumbentAI capabilities.

(59:38):
I shared with you about mysoftware company that enables
accounts payable processes,invoice vouching, and
reconciliation using agenttools, and there are many
incumbents definitely in the APand finance industry, and yet
there's a huge interest in thesoftware that we have created
because it's lean, focused,accurate, and drives huge cost
reduction compared to using theincumbent systems.

(01:00:01):
And actually, from my particularperspective, is integrated into
the incumbent systems.
So if you're using any ERP oraccounting software, this
platform just plugs into it.
So it's not replacing it, it'sjust allowing you to do the
processes instead of with hugeamount of human labor to do it
autonomously using AI agents.
Now, different than the researchwe heard from Deloitte, the
information that is shared byMenlo Ventures is that 47% of AI

(01:00:25):
pilots actually reachproduction.
Nearly doubled the conversionrate 25% of traditional SaaS
products after being tested, andthey're claiming it is happening
because of the immense value andimmediate value that companies
are seeing as far as this, thetime to ROI of implementing
these kind of capabilities.
This is, again, dramaticallydifferent than what we heard
from Deloitte with 11%.

(01:00:45):
What I think is, I think itdoesn't matter what the number
is, I think what we need to lookat is the trend.
And the trend is that almostzero companies deployed
production level a gentechcapabilities last year.
And now whether it's 11% of 25%,it's a huge jump in just a
single year.
And we are still very, veryearly on in this journey of
understanding how to develop anddeploy agents, how to change the

(01:01:08):
infrastructure of the company inorder to make things work
effectively and so on.
And so I think this will growsignificantly faster than it is
right now.
So not just more deployments,but faster deployments as they
become better and better bestpractices on how to do this
right and more effectively.
Ernest and Young also released avery interesting study that is
showing that currently thesavings that companies are

(01:01:28):
generating from efficienciesdriven by AI is actually
reinvested inside the companyinstead of driving job cuts,
which was the assumption of whatit is going to go.
So according to the study, 96%of organizations surveyed that
are investing in AI are seeingproductivity gains and only 17%
saying that gains have led toworkforce reductions.

(01:01:51):
It's still 17% companies havelaid off people because of these
efficiencies.
It's still a very, very highnumber, but the majority are not
doing it at least yet.
So what are they doing with theextra money that they're
generating from theseefficiencies?
47% are expanding existing AIcapabilities.
42% are developing new AIcapabilities.
41% are strengtheningcybersecurity, 39% uh, are

(01:02:13):
investing it in r and d, and 38%are investing in upskilling and
reskilling existing employees.
And the most interesting aspectof this is Dan di Asio, who is
Ernest and Young GlobalConsulting AI leader, know that
that companies are currentlyshifting from a productivity
mindset to a growth mindset.
Basically using AI to, and I'mquoting, create new markets and

(01:02:34):
achieve what was previouslyconsidered impossible.
One of the things that I teachin the AI Business
Transformation course is exactlythat, is how to use AI in a
strategic way in order to drivebusiness growth and not just
efficiencies.
And you need to understand thatthese efficiencies as attractive
and low hanging fruit as theyare, are a fraction of what you
can make by making the rightstrategic moves, by offering new

(01:02:56):
services to your existingclients or by being able to
address new markets that youcouldn't do profitably without
ai.
So it unlocks so manyopportunities that will give you
10 folds or a hundred folds,order of magnitudes, higher
returns than investing justinefficiencies.
I don't mean you shouldn'tinvest inefficiencies, that's a
great starting point, but yougotta look at the bigger picture

(01:03:17):
and invest in that as well.
The other thing that this surveyhas found, which makes a lot of
sense, is that the more moneycompanies invest, the better the
results that they're seeing Orin real numbers.
Organizations that are investing$10 million or more are
significantly more likely toreport significant productivity
gains at 71% of the companiescompared to those who invested

(01:03:37):
less with only 52% of thecompanies connect that to the
previous points that wementioned about the other
surveys that explain thatinfrastructure changes and
complete transformation isrequired and you understand why
investing more money drives youbetter and faster results that
obviously requires to investthat kind of money.
Over 60% of leaders have saidthat they're going to invest a
lot more in 2026 in ethical AIoperation and investing in

(01:04:01):
responsible AI training.
Again, what I've been doing andsaying for the last two and a
half years, I'm really excitedto see this in this past year I
have done numerous workshops,many of them for multi-billion
dollar enterprises, but a lot ofothers to small and mid-size
businesses, and I can tell youthey're all struggling with
similar things just in differentscales, and being able to

(01:04:23):
provide continuous training toemployees and to leadership and
to the board are critical to thesuccess of this transformation.
But as I mentioned, with all thecrazy hype and all the
conversation about newstrategies and new revenue and
efficiencies and so on.
We are in the very beginning ofthis process.
We are merely scratching thesurface and to show you how much

(01:04:45):
we're just scratching thesurface.
I'm going to share two researchpapers that were released this
week, touching on two completelydifferent aspects of ai, but
showing you how much we don'tknow on how much AI will be able
to do and how easy it will beable to do it.
So the first research papercomes from a company called Helm
ai and what they were able to dois they were able to create a
breakthrough in the way theytrain AI models.

(01:05:06):
They were able to demonstrate avision only zero shot autonomous
steering capabilities trained ononly 1000 hours of training
data.
So let's break this down.
Vision only means they'redriving like Teslas.
They don't have radar and othersensors, they just have cameras
that are looking around.
So Tesla has done that already.
But Tesla as well as Waymo, hasused millions of hours of

(01:05:31):
driving data in order to trainand achieve similar
capabilities.
Now, what the hell is zero shot?
Zero shot meaning something thatyou haven't done before that you
can learn, not learn from directexperience, but you gotta have
the ability to quote unquotereason through the situation.
So they were able to showdriving off road in mountain
narrow roads with differentobstacles that do not exist in

(01:05:53):
cities when the training datadid not have any kind of
information like that.
Basically, the system knows howto use its training data to
develop the capabilities to thendo a much broader set of things
based on a really small set ofdata.
Their breakthrough is somethingthey call deep teaching, which,
and I will quote their CEO said,deep teaching is a breakthrough

(01:06:13):
in unsupervised learning thatenables us to tap into the full
power of deep neural networks bytraining on real sensor data
without the burden of humanannotation nor simulation.
Basically, they were able tocreate a system that operates
really well in the real world,in a fraction of the investment
of other companies achieving thesame thing, and these kind of

(01:06:33):
discoveries are just gonna keepon happening.
Another very interestingresearch paper came out of a
collaboration between MIT andGoogle DeepMind.
They were specificallyresearching the effectiveness of
using different kind ofimplementation capabilities of
agentic solutions inside ofcompanies.
And what they found that more isbetter is not always the case.
So they broke this down toseveral different categories,

(01:06:54):
and one of the things they foundis that creating a multi-layered
approach with a centralizedcoordination, one agent
basically is the orchestrator.
Controlling other agents hasincreased performance by 80.9%
on parallel tasks such as onparallel tasks, however.
All multi-agent variantsdegraded performance by 39 to

(01:07:16):
70% when sequential reasoningtasks was required when compared
to just a single agent doing thesequential tasks.
Now, the cool thing of what theyfound is that they have
identified three critical laws,basically like physics laws that
stand every time they havedeveloped and tested agents in
different environments.
One is what they call toolcoordination trade-off.

(01:07:38):
So under fixed compute budgets.
Which is more or less the case,most of the time, tasks that
require heavy tool use sufferdisproportionately from the
overhead of coordinatingmultiple agents.
Basically, if you need to use alot of tools and also use a lot
of agents, a lot of the computeis being wasted on the
conversations between theagents.
Just think about real human workand think about how much time we

(01:08:01):
are wasting and frustrated bybeing in meetings instead of
actually doing work.
It is the same exact thing withthe agents.
More agents means morecoordination, means more compute
is going towards that versususing the tools of the actual
task.
So this is the one trade off.
The other one is what theycalled capability saturation.
Adding more agents yield,diminishing return, or sometimes

(01:08:21):
negative returns.
Once a single agent baselineperformance exceeds 45%.
So if a single agent could dosomething once it goes beyond a
certain ability to do things,adding more agents actually
slows the process down andcauses a disadvantage versus an
advantage.
And the last thing that theyfound is what they called
topology dependent erroramplification.
That's a very, that's a mouthfulto basically say that the

(01:08:42):
structure of the team of agentsdictates how mistakes spread.
So based on all of this, theycome up with three different
frameworks that each will workfor specific kind of jobs.
One is centralized coordination,which is a manager leading other
agents, and it is the king whenthe requirement is to do
parallel tasks.
Then there's decentralizedcoordination, which is basically

(01:09:04):
peer-to-peer collaborationwithout a leader coordinator
orchestrator that excels indynamic environments like web
navigation, and it's achievingmuch better results in this
peer-to-peer collaboration.
As they said, too many cooksspoil the broth.
So what does that tell us?
It tells us that learning how todeploy agents and the research
on how to do it effectively isin its infancy.

(01:09:27):
There are apparently scientificproven results on how to use
agents differently from anorchestration perspective for
different tasks, learning thatcan dramatically improve the
results, which will lead to moreinvestment, more deployments,
and more success in enterprisesthat will drive more money, et
cetera, et cetera.
Bottom line.
We are in very early stages ofeverything in ai, and the more

(01:09:50):
we learn either as individualsor as organizations or on the
research side of things, it willonly accelerate the adoption and
the value that AI can generate.
Now, my plan was to share withyou a lot more in this episode,
but it is going to turn to beway too long.
So what I'm going to do, I'mgoing to include all the rest of
the aspects, including theannouncement of the 24 tech

(01:10:10):
companies that are the firstbatch of companies included in
the government genesis mission,including the call from Bernie
Sanders to stop the developmentof new data centers, including
release of really interestingmodels like Gemini Three, flash.
That is in many casesoutperforming the previous
Gemini 2.5 Pro and thecompetitors such as CLOs, SONET
4.5 at a fraction of the timeand the cost, including a new,

(01:10:32):
powerful and extremely fast andcapable voice model from X AI
including grok running inside ofTesla's, actually integrating
with car systems, which kind ofhints of how our future is going
to look like, including therelease of 2.6, which is a new
video generation model.
From Alibaba that is comparableto VO three with 15 seconds, 10

(01:10:55):
a DP cinematic sequences withfull audio capabilities
including a agentic task modefrom Anthropic and many, many
other new things, that we're notgoing to talk about.
The one thing that I willmention that you should try is
Claude just expanded theirrelease of the Claude for of
Claude for Chrome extension,basically turning chrome into an

(01:11:15):
agent browser, fully integratedwith Claude and Claude Code, and
that drives an insane amount ofuse cases that I cannot wait to
try, and I will share with youin a Tuesday episode sometime in
the next few weeks.
But the bottom line is there'sstill a lot more to learn this
week than what I was able toshare with you on this episode.
And if you wanna get access toit, it is available on our

(01:11:36):
newsletter.
So you can click on the link inthe show notes and sign up for a
newsletter and get all the otherstuff of the new.
And there's a lot of interestingthings this week and every week,
uh, that you can learn bybriefly going through it, seeing
the ones you like, clicking onthe links, and going through the
specific articles with all theother aspects of the news.
If you are finding this podcastvaluable, please rate us on
Apple Podcasts and Spotify andshare this with other people who

(01:11:58):
can benefit from this.
I'm sure you know other peoplewho can find value in this
podcast that click the sharebutton and just share it with a
few people.
It will take you.
A minute, probably less.
It will give other people value.
I will really appreciate it andyou'll feel good about yourself.
So all the good things, all atonce in less than a minute, of
investment.
So please do that.
And if you are interested inmore structured training than
the podcast offers, go and checkout the courses that we offer.

(01:12:21):
That's it for today.
Have an amazing rest of yourweekend.
Keep experimenting AI and we'llbe back on Tuesday.
Until then, have an amazing restof your weekend.

All Episodes

251 | OpenAI’s apps + SDK, Image 1.5, GPT-5.2 Codex, and world domination playbook, Anthropic wins the enterprise, The state of agentic implementation based on MIT, Google, EY, and Deloitte, and more important AI news for the week ending on Dec 20, 2025

Episode Transcript

Popular Podcasts

Stuff You Should Know

The Joe Rogan Experience

Two Guys, Five Rings: Matt, Bowen & The Olympics

.css-15opob5{left:0;position:absolute;top:0.8rem;} All Episodes

Episode Transcript

Popular Podcasts

.css-r6mb8g{margin:0;word-break:break-word;display:-webkit-box;-webkit-box-orient:vertical;box-orient:vertical;-webkit-line-clamp:1;overflow:hidden;}Stuff You Should Know

The Joe Rogan Experience

Two Guys, Five Rings: Matt, Bowen & The Olympics

All Episodes

Stuff You Should Know