Altman Targets Model Overload with Orion
Key Points
- Sam Altman’s recent blog post outlines Open AI’s roadmap, mentioning the upcoming GPT‑5 and a previously leaked internal project called “Orion,” now slated for release as GPT 4.5.
- Altman criticizes the current ChatGPT UI for offering an overwhelming and confusing list of model options, arguing that intelligent systems should not require users to navigate a complex dropdown menu.
- He categorizes language models into three tiers: fast “reasoning” models that respond instantly, medium‑speed “thinking” models that operate under a minute, and deep‑inference models that take longer than a minute to yield more nuanced, higher‑quality results (e.g., 01 Pro).
- The emergence of “deep research” models—capable of extensive inference and producing detailed reports—has begun to surface online, demonstrating the tangible benefits of the third‑class, deep‑inference approach.
Full Transcript
# Altman Targets Model Overload with Orion **Source:** [https://www.youtube.com/watch?v=BA-TyC-ZTUs](https://www.youtube.com/watch?v=BA-TyC-ZTUs) **Duration:** 00:07:03 ## Summary - Sam Altman’s recent blog post outlines Open AI’s roadmap, mentioning the upcoming GPT‑5 and a previously leaked internal project called “Orion,” now slated for release as GPT 4.5. - Altman criticizes the current ChatGPT UI for offering an overwhelming and confusing list of model options, arguing that intelligent systems should not require users to navigate a complex dropdown menu. - He categorizes language models into three tiers: fast “reasoning” models that respond instantly, medium‑speed “thinking” models that operate under a minute, and deep‑inference models that take longer than a minute to yield more nuanced, higher‑quality results (e.g., 01 Pro). - The emergence of “deep research” models—capable of extensive inference and producing detailed reports—has begun to surface online, demonstrating the tangible benefits of the third‑class, deep‑inference approach. ## Sections - [00:00:00](https://www.youtube.com/watch?v=BA-TyC-ZTUs&t=0s) **Altman's Roadmap & Model Overload** - The speaker summarizes Sam Altman's new blog post detailing a bewildering array of GPT models in the app’s dropdown, his criticism of that complexity, and the upcoming launch of a streamlined Orion/GPT‑4.5 version that won’t be a reasoning‑focused model. ## Full Transcript
I'm here to tell you all about Chad gp5
and know that is not just clickbait we
actually have real news on it so Sam
mman published a blog post uh a couple
of hours ago that lays out his personal
Vision or roadmap for where they're
going next and yes it includes talk of
GPT 5 it includes what uh they call
internally Orion which got leaked all
over the internet in the fall and then
never materialized let me tell you about
it net net what Sam wanted to emphasize
in his blog post is that he has become
very concerned about the number of
models in the drop down and how complex
they are to use just to give you a
couple examples if I hit my drop-down
now in the chat GPT app I can see one
two three four five six models that are
visible and more models hiding in a
carrot plus a temporary chat icon that's
too many like if you can see 4 40 with
scheduled task 01 03 mini 03 mini High
01 PR mode and by the way none of those
is operator none of those is deep
research I have to go to chat gp.com to
use deep research I can't even use the
app right now even though the results
appear in the app it's a very very
confused
experience and so Sam is basically
saying this is not what we want the
magic of intelligence to be it should
not be this ridiculously complex picker
where you don't even know what the what
the heck you're
picking and so what he said was they are
going to release Orion or GP 2 4.5 it is
supposed to be very good but it is not a
reasoning model which means it doesn't
use inference time compute and I want to
be really precise about this because
people are kind of equating the cheaper
more widely available reasoning models
with deep inference reasoning models and
I think that we need to get better at
naming you have models like GPT 4 o
right now or Claude that just come back
with something very quick
you also have models that claim to be
thinking models but that do not think
for minutes at a time they think for
under 60 seconds generally I would argue
that uh the flash thinking model from
Google is a good example I think deep
seek is a good example uh there's a
number of other examples as well and
then you have models in a third class
that think for more than 60 seconds at a
time and do deep inference 01 Pro is a
great example there there aren't a lot
of other examples and it's unfortunate
that 01 Pro is not more widely available
because I think if people used it they
would understand there is that third
class and the results are definitely
different and meaningfully better if you
give the model that much time I think
the closest that people are starting to
see is when they're starting to see deep
research and starting to see the results
of deep research published on the
internet deep research is another
thinking model it takes a very long time
to infer and come up with something but
it's very compelling report and that's
leaking all over the internet now people
are seeing the results and are rightly
amazed so we have at least three classes
of model people are generally conflating
them into two I think that's not
helpful and what Sam is saying is 4.5 or
oion is really in that first class it's
a fast response model it's just very
very
good now that being said he has made it
clear that that is the last Model that
they are going to release as a separate
model
it will be named 4.5 probably and it
will appear in the UI and that will be
that since he gave a hint previously it
would be during the winter I think that
gives him about one more month to
release it and then he says he's moving
on what is he moving to chat
gp5 so he's keeping the GPT branding but
every other capability they have built
is getting wrapped under that name label
and brand which is a good branding
decision Chad gp5 will have deep
research Chad gp5 will have operator
mode Chad gp5 can think for minutes at a
time or think for a second it will
decide and it will decide based on the
task that you give it as someone uh on
Twitter said very snarkily chat GPT is
the friends we made along the way
basically it's all these models along
the way and we bring them together and
we wrapped them into one clean consumer
bow so the consumer can use them now
that all depends on a lot of really
smart interpretation of user intent so
we will have to see how well they do
with that history suggests they're going
to do pretty
well and in terms of timing he didn't
give a time but I'm strongly expecting
GPT five to come out uh probably before
the fall if I had to guess given their
Pace they've been doing major releases
every two or three months so I would
expect sometime in the summer is my
current guess and that is a guess I do
not have private unknowable information
it's just a
guess so there you go that's what we
know about the road map I am glad
they're making the effort but I do have
a fairly big question around 03 because
we were promised
03 I we were promised 03 Pro I don't
know where that is now I don't know
where it is so we will just have to see
um it may get wrapped in as GPT 5 and we
just have to tolerate that I thought we
were on a slightly faster time frame
with it but maybe what Sam felt is that
at the end of the day
it's too complicated for people to add
yet another number in model and so he
decided to change it one more tidbit for
you this is not about Sam mman it's not
about open AI uh I just learned this the
owners of ai.com which currently
redirects to deep seek it's not deep
seek I thought it was deep seek it's in
pardonable that I think it's deep seek
because it goes to deep seek so I
thought deep seek bought it that's
normally how domains work but no it's
apparently a for profit separate entity
run by run out of
Malaysia um and someone bought that
domain back in 1993 I wish i' bought
that domain back in 1993 and they now
make a mint renting the domain to
whoever pays them and it appears that
deep seek is currently using some their
five and a half million or whatever they
use to train the model to uh buy access
or redirects from ai.com to deep seek
so it's just a tidbit I have nothing
more to follow up on other than the
internet is weird and you can make money
in lots of ways and someone has figured
out out of Malaysia how to rent ai.com
for what is presumably a pretty penny I
am sure it is not cheap to do that so
there you go that's your tidbit lots
more on GPT 5 hope you enjoyed it