Stargate AI Plan Premature and Exclusionary
Key Points
- The proposed “Stargate” AI infrastructure plan prematurely declares OpenAI (backed by SoftBank and Oracle) the winner, ignoring the continued competition from Anthropic, Meta, Google, and emerging model makers.
- Critics argue that crowning a single winner undermines the dynamic AI landscape, where numerous companies are rapidly advancing with new models, synthetic‑data generation, and innovative compute strategies.
- The plan is based on a 2023 architecture that assumes ever‑larger GPU clusters and massive data sets, a paradigm that recent research shows yields diminishing returns and overlooks newer efficiency‑focused approaches.
- It neglects the shift toward inference‑time compute and continual‑learning models that enable smarter AI with far less hardware, rendering the proposed massive data‑center strategy increasingly outdated.
Full Transcript
# Stargate AI Plan Premature and Exclusionary **Source:** [https://www.youtube.com/watch?v=L4TT6OAtuS0](https://www.youtube.com/watch?v=L4TT6OAtuS0) **Duration:** 00:06:19 ## Summary - The proposed “Stargate” AI infrastructure plan prematurely declares OpenAI (backed by SoftBank and Oracle) the winner, ignoring the continued competition from Anthropic, Meta, Google, and emerging model makers. - Critics argue that crowning a single winner undermines the dynamic AI landscape, where numerous companies are rapidly advancing with new models, synthetic‑data generation, and innovative compute strategies. - The plan is based on a 2023 architecture that assumes ever‑larger GPU clusters and massive data sets, a paradigm that recent research shows yields diminishing returns and overlooks newer efficiency‑focused approaches. - It neglects the shift toward inference‑time compute and continual‑learning models that enable smarter AI with far less hardware, rendering the proposed massive data‑center strategy increasingly outdated. ## Sections - [00:00:00](https://www.youtube.com/watch?v=L4TT6OAtuS0&t=0s) **Critique of AI "Stargate" Initiative** - The speaker contends that the newly announced trillion‑dollar "Stargate" AI infrastructure program prematurely crowns OpenAI as the inevitable winner, ignoring the rapidly evolving competition from Anthropic, Meta, Google, DeepSeek, xAI, and other model makers. ## Full Transcript
Stargate is out it is both a pretty
terrible TV show for the 1990s and also
a half a trillion dollar infrastructure
program that was just announced all
about
AI I have some real
questions the issue with Stargate as far
as I can tell is that it crowns a winner
before the race is
over it says open AI is going to win the
game they're going to win it funded by
Soft bank and Oracle is going to build a
data centers obviously for those three
players that's
great even Microsoft gets in on the game
they're happy to they're partner of open
AI Nvidia of course is supplying the
chips the problem is that there are a
lot of other players in the game and it
is not clear how this reshapes the race
for them they're not giving up anthropic
is not giving up I know I just did a
video on them but they're not giving up
and meta's not giving up Google's not
giving up the Stakes are too
high and yet here we are crowning a
winner and I'm not even getting to the
shifts that we've seen with model makers
like deep seek entering the scene or how
x. a is coming on quickly with
Incredible gains uh huge compute
clusters and so when I look at the
problem space and I say to myself this
is a hugely Dynamic situation there's
lots of model makers they're all
competing how does it make sense to have
only one model maker get in on this
project I don't think it
does and I think it exists that way
because this was Sam Altman shopping a
deal for this kind of a data center back
it feels like almost a year ago like it
was like 10 11 months ago and then it
died down and now it's back so that
brings me to my second issue with
this this is a
2023
architecture that they are describing
not a 2025 architecture and I don't know
why that they are like what that doesn't
make sense we've learned so much this is
such a dynamic space it changes so fast
so I'll explain what I mean
2023 we thought that we had to have ever
bigger clusters of gpus to train on ever
bigger data sets in order to make these
mod
smarter we thought that in early 2024
too that's why this thing is talked
about as having 10 million
gpus
well the thing we discovered is that at
the end of the
day you can have all of that compute but
there may be diminishing marginal
returns just for
pre-training you have issues finding the
data unless you're generating it
synthetically which we've made some
progress on but that's a big scale up in
synthetic data production if that's what
we
do as as Ilia famously said in November
of last year long after Stargate was
first kind of kicked around we have one
internet right like we have one internet
siiz data pool we've used it
so I think the reason why that feels
weird in that context is that this is a
architecture that fundamentally
assumes this sort of older Paradigm for
how trained AI models and the new
paradigm the one that's unlocking
continual progress that everyone's
excited about it's not mentioned
inference time compute is a very
different Paradigm it allows you to run
multiple threads simultaneously is what
happens when the model thinks frankly
Gemini dropped a version of that
yesterday with their new update to flash
2.0 thinking it apparently I haven't had
a chance to even try it yet apparently
it's on par with o
Pro
so the model makers are continuing to
compete they're competing on different
architectural
standards and Stargate is sitting here
like with this 2023 structure and
everyone's saying it's going to be a you
know vaccine for cancer this and that
well maybe
but it's a weird way to go about it now
and it makes me wonder if we've seen
this much drift in the way we do AI in a
year because we're learning so much and
this thing takes four years is this just
going to feel outdated by the time we're
done with it it
might it
might and that kind of comes back to the
goaling like in other major
infrastructure projects that America has
undertaken we've had very clear goaling
you go to the Moon you bring back the
astronauts it's super clear by the end
of the decade they even had like a
classic timeline on it
fine this is not very clear it's like
yeah we'll do some answer
stuff okay what what does done look like
what does good look like does this mean
that like also the defense department
will be using it maybe it's not really
clear who gets to decide how all of that
compute resource is allocated that's
also not clear does soft Bank decide
that I doubt it
so I have a lot of
questions as you can probably tell I
think it absolutely reshapes the race
it's worth talking about I put more
thoughts on my substack but at the end
of the day to me this is a project
that makes me tilt my head and think and
raises more questions that it answers
and everyone's sort of talking about it
as if it's a done deal it's obvious like
this is it I I don't know like I don't
think building the future on two years
ago architecture four years from
now is
automatically the win maybe like maybe
they just repurpose the compute I don't
know but it feels a little odd what do
you think