Sam Altman FINALLY Reveals GPT-5 (GPT-5 Explained)
Title: Sam Altman FINALLY Reveals GPT-5 (GPT-5 Explained)
Author: TheAIGRID
Transcript:
so we actually finally got the road map
for GPT 5 and the reason I'm actually
really excited to be recording this
video is because for once we're actually
not getting a bunch of random
speculative information we're actually
getting the entire GPT 5 road map laid
out for us to completely understand so
let's not waste any time and dive into
exactly how GPT 5 will be deployed
because it does offer us some rather
interesting information that goes to
show how these future a systems will be
and I actually do agree with this and I
think this is how most AI systems will
be when it comes to Major chatbot
providers so you can see a few hours ago
Sam Alman tweeted that this is going to
be the official road map update for GPT
4.5 and GPT 5 and he says that he wants
to do a better job of sharing the
intended road map and much better job at
simplifying the product offerings he
said that you know he wants AI to just
work for you and it realized how complic
ated the model and product offerings
have gotten so basically what they're
talking about is the fact that if you've
ever used chat gbt before and you're
someone that is using it you know and
you've got a lot of money invested and
by that I mean let's say for example
you're paying for the highest tier chat
GPT currently is a complete mess now I'm
going to show you guys exactly what I
mean by that so when you come into chat
GPT it does seem pretty simple but when
you do this drop- down menu right here
you can see that unfortunately as
someone who has a different layout to
you probably do I have one two three
four five six seven gp4 mini and eight
gb4 and nine if you include the
temporary chat basically the problem
with this is that this causes analysis
paralysis and what this does involve is
that this involves users decisions so
for example when someone comes to use
the AI chatbot they genuinely don't know
which model is the best because You' got
GPT for for questions gbt 40 scheduled
01 03 mini High 03 mini 01 prom mode
you've got basically too many choices
and one thing that they constantly say
in product choices is that if you have
too many choices your product is going
to fail and in this instance for chat
GPT what they want to do is they
actually want to simplify this entire
thing here because this is far too
complex for the average person to use as
someone that knows exactly what I'm
going to be doing for every task for me
this is just fine but for the average
person this is actually a detriment and
isn't actually a good thing so they're
actually going to change this in a
really interesting way which I will talk
about later and this is where they
actually talk about how they hate the
model picker as much as we do and want
to return to Magic unified intelligence
so I do think that one of the approaches
that they're going to take to sort of
return to this unified machine
intelligence is through the router llm
so basically this was I wouldn't say a
framework but this is basically how you
set up llms in order to basically make
things a lot easier so so you would have
your single system prompt your user
prompt actually not the system prompt
your user prompt so for example you
might put what's the weather today and
then there's basically a router that
analyzes every single prompt that is put
in so this router will look at your
question that you put in it will analyze
the cost the latency and basically it
will look at your question and think
okay does this question need to be
routed to a really difficult model or
does this question you know need to be
routed to a super easy model then I can
answer within 2 seconds if it is a hard
question it's going to be put towards
the most intensive llm the one that's
going to think for very long periods of
time and if it's something very simple
like you know which came first the
chicken or the egg it's going to be
routed to a very simple model and then
of course finally you get a system
response now this is something that they
have tested before maybe not opening ey
but I do know other companies have
tested something like rout llm before
and it has gone very well considering
the fact that I think it dropped cost by
around 85 % and you can read more about
this at this blog post route llm an open
source framework for costeffective llm
routing and basically just talks about
how in their routing setup they focus on
the case where there are two models a
stronger more expensive model and a
weaker but cheaper model and their you
know aim was to minimize costs while
achieving high quality between routing
through two models and you can see right
here that they said a top goal for us is
to unify the O Series models and GPT
series by creating systems that can use
all our tools know when to think for a
long time or not and generally be useful
for a wide range of tasks so like I was
talking about the previous framework
route llm it's quite likely that we will
have a situation on our hands where we
do get a system that is basically all in
one and it's quite likely that we're
going to get future models as an
all-in-one system so whilst yes there
might be you know parts of the system
that are O3 mini and 03 you know high or
whatever it seems like we're basically
moving towards one giant model or you
know orchestrated set of llms that
basically provide you with this unified
level of intelligence and it's quite
likely that you know depending on your
tier you will have access to you know
those levels of intelligence so maybe if
you're paying $200 a month you get
access to the model's entire brain all
the you know Advanced reasoning
capabilities but if you're just a free
user you know you can maybe query it
about one or two things a day maybe just
10 to 20 queries a day um and that is
going to be how the model is probably
deployed because it's you know a lot
easier to do that rather than having six
different models and then when you know
different companies come out with
different models they're like okay this
model is clearly better but this is
probably going to be a lot better for
opening ey in terms of their branding
and they also state that you know in the
chat GPT and you know in their API
they're actually going to release GPT 5
as a system that integrates a lot of
their technology including 03 okay and
they said that the no longer going to
ship 03 as a standalone model so it's
clear that GPC 5 is probably going to
absorb 03 as just part of its you know
architecture and it's going to be I
guess you could say part of that model
for quite you know some time so it's
like gbt 5 is probably going to be a
system that compromises of maybe 03 01
you know different parts of different
questions which is going to be super
interesting to see how that you know
entire system works and something that
they also said here that Sam Alman said
that I think Super interesting is that
he said that GPT 5 as a system okay he
says it's going to integrate a lot of
our technology and I don't know if this
includes everything that we currently
have because there was one time that
samman said that you know this year they
are going to be releasing something that
nobody has even thought of yet and I
think that is going to be super
interesting so I'm guessing that GPT 5
is probably going to be also you know a
computer agent I'm guessing that they're
probably going to wrap the name
underneath that because that is what
makes sense so that does intrigue me a
lot because I'm thinking what on Earth
you know uh you know is like the final
version of GPT 5 going to be cuz I know
that the model is really smart but they
also did say that you know they're going
to ship the next you know ship GPT 5
five they also said that you know we
will next ship GPT 4.5 the model we
called Orion internally as our last
non-chain of thought model so what they
mean by this is that the last Model that
they basically did the old method of
training where you know you basically
collect a bunch of data put all that
data in the model train it you know then
have the post training the fine-tuning
and all of those things that you have
for a standard model output like a GPT 4
a GPT 40 all of those things basically
are going to go in the last Model so
it's clear like I said before we are on
a new paradigm of training these models
and it's clear that the GPT 4.5 is
probably going to be in my this is just
my best guess it's probably going to be
on the level of claw 3.5 Sonic but
probably a lot better than that probably
not a lot better probably marginally
better because CLA 3.5 son it is just so
good but I'm guessing it's going to be
qualitatively better because you know
whilst you know other models do surpass
Claude 3.5 on benchmarks internally when
you use a model like clae it just seems
so much smarter and it just always seems
like it actually knows what's going on
so I I'm thinking the opening eyes aim
is to probably just Edge out claw 3.6
Sonic and I think they will be able to
do that with GPT 4.5 which is the model
they called Orion internally which they
knew was smart but I'm guessing the
reason why they didn't deploy this
before which most people were wondering
um and they actually spoke about this in
a article before was basically the fact
that the model it just cost too much in
terms of inference so they couldn't
serve the model yet but we do know that
as time passes inference consistently
drops if you remember yesterday's video
we spoke about how Sam Alman he speaks
about how like literally every 12 months
the price of llms drops by around 10x so
it's quite likely that in the past few
months okay since they've managed to
train this model and do all the safety
testing the price of inference has
dropped to the point where it's going to
be reasonable so that that they can
release this model now they if we do
talk about you know the estimated time
of arrival when is this model going to
come we have weeks and months this is
what chubby said any ETA for gbt 4.5 or
gbt 5 so for gbt 4.5 it will be likely
coming in the next coming weeks so I'm
suspecting by the end of March that we
would have GPZ 4.5 and for gbt 5 I'm
guessing maybe by the end of Q2 or maybe
by the end of Q3 but that is just my
very best guess so for that in terms of
time frames if you want to think okay
one of these next models coming out that
is going to be that aspect of that area
so for me that's what I do think in
terms of where that is going to be now
of course for free users they actually
talk about you know the free tier of
chat GPT is going to get unlimited chat
access to GPT 5 at this standard
intelligence setting note standard
intelligence and this is of course not
the high intelligence so there isn't
going to be you know like 03 mini high
so it's not going to be accessing the
long chains of thought that are at the
threshold of you know the most
intelligence that we could currently
access but it does seem like this is a
move because with deep seek they
basically just gave away you know crazy
level intelligence at a fraction of the
cost and it basically made it more cost
effective for people to use deep seek
than anything else which is really
really intriguing so you know we can
also see that they also talk about you
know the fact that plus subscribers are
going to be able to run GPT 5 at a
higher level of intelligence I'm
guessing having the model to think more
and of course tap into those long chains
of thought that provide you with deeper
insights to your queries and it says Pro
subscribers will be able to run gbt 5 at
an even higher level of intelligence and
these models will incorporate voice
canvas search deep research and more so
I think what openi are doing are
basically saying look if you want access
to our free models you are going to get
that for free but if you want access to
you know voice canvas search and deep
research you're going to have to
probably play the highest highest level
which is probably around $200 a month
and I know that does seem quite
expensive but I think what openi are
banking on is that every single time
they release something they're basically
just going to add it to GPT 5 as a
system so overall as the system gets
smarter and smarter and smarter what
you're paying for essentially Goes Down
and Down down if that does make sense so
yeah it's going to be really interesting
to see how the pricing Works opening I
have discussed paper usage pricing
because of course $200 a month is
absolutely insane I don't think anyone I
know personally could get that much
value out of the model but it definitely
is something that people will be doing
so I also do think that something that
was super interesting was will you be
smarter than gbt 5 and this is where
samman actually talks about the fact
that you know he doesn't even think he
will be smarter than GPT V how many
people feel smarter than GPT
4 okay how many of you think you're
still going to be smarter than gbt
5 expecting more hands here
um I don't think I'm going to be smarter
than GPT V and I don't feel sad about it
because I think it just means that we'll
be able to use it to do incredible
things and you know like we want more
science to get done uh we want more we
want to enable researchers to do things
they couldn't do before this is the of
this is like the long history of
humanity um it does feel a little
different this time because of what this
can enable but if scientists can do
things because they have like a crazy
high IQ tool and they can focus more on
figuring out the right questions to ask
address things quicker do their search
space faster uh that's just a win for
all of us so we're thrilled to get to so
with with that being said what do you
guys think about gp5 are you excited I
think I'm excited for the unified
intelligence I think this is something
that actually makes sense unifying these
models into one entire thing just makes
so much sense in terms of the work
you're trying to do and the Simplicity
of the model there was a famous example
in marketing that you guys might not
actually know about but just take this
as a business lesson where there was a
jam store and basically with this Jam
store they had two stores they had one
that had 24 choices of jam another one
that had six choices of jam and then in
the 24 choices you think that because
there were more choices more people
would be buying because they're offering
more services but only 3% of people
actually bought but when they had six
choices a jam 30% of people bought and
basically this talks about when you have
too many decisions you don't actually
pick anything and it's the same thing
with AI when users have chat GPT and you
have you know GPT 401 03 mini 03 mini
High the problem is is that you have too
many choices when we go over to Claude
we have maybe three models when we go
over to Gemini we have maybe one or two
models that we're going to be using and
in fact Gemini has probably twoo too
many models now but the point is is that
like clae everyone knows it's clae 3.6
on it that's the one model you use and
it's like with deep seek you know you
have the thinking model that's the basic
model that everyone uses so with this
you've basically got too many choices so
it's quite likely in the future these
are going to be unified into one big
model SL architecture so with that being
said hopefully you guys enjoyed the
video and I will see you guys in the
next one