In the Loop AI

OpenAI released GPT-6.2 Astra right before Anthropic's IPO, and it backfired. The reviews were brutal, and the company that came out of nowhere to take the number two spot in the AI race isn't Anthropic. It's Google. Ejaaz gives his honest take on Astra, what the testers found (Every, Theo, Artificial Analysis), and what OpenAI might really be going after: math, science, cybersecurity and government contracts.

Then the comeback: four Gemini 4 models in five weeks (Argon, Carbon, Hydrogen and Neon), why Neon is beating Opus 5.5 at a fraction of the cost, and why Google's rebuild means the race isn't closed yet, though the window is closing. Plus Nvidia's first quarter without Jensen Huang: a record $115.8 billion, the neocloud financing question, why the stock still fell, and what the chip crackdown on China means for GPU prices.

==
FOLLOW EJAAZ ON TWITTER:
https://x.com/cryptopunk7213

==
Timestamps
00:00 OpenAI Fumbles, Google Rises
02:10 Honest Takes on GPT-6.2 Astra
05:23 What OpenAI Is Really Going After
08:34 Google Takes the No. 2 Spot
09:22 Four Models in Five Weeks
13:04 Anyone Can Still Catch Up
14:32 The Window Is Closing
16:58 Nvidia's First Quarter Without Jensen
18:31 The Neocloud Question
20:32 Why the Stock Still Fell
21:46 China, Chips and Cheap GPUs
23:31 Final Thoughts

#ai #google #gemini #openai #nvidia

==
AlphaSpace by Yahoo Finance: A Professional-Grade Investment Platform Built for Everyday Investors.
👉 https://finance.yahoo.com/about/promos/gold/alphaspace/?ncid=100003571

Creators and Guests

Host
Ejaaz Ahamadeen

What is In the Loop AI?

Exploring the frontiers of Technology and AI

In an unexpected turn of events, OpenAI has completely fumbled the lead in the

AI race after releasing their brand new model, GPT 6.2 Astra,

which was meant to be a kill shot attempt at Anthropic right before their IPO,

but it ended up doing the complete opposite.

The reviews for GPT 6.2 has been none other than absolutely appalling and terrible,

with people calling it an absolute flop.

But what's perhaps more surprising is who came in out of the shadows to replace

them as the number one or even number two AI leader.

It isn't Anthropic, it's Google.

Yes, the same company that has been behind the entire AI race for the last 12

months came out with not one, not two, but three brand new models in the last

three weeks, the most recent of which

beats GPT 6.2 Astra and even competes and beats Opus 5.5 and Fable 5.2 from Anthropic itself.

But it's also a lot cheaper. So there are so many things that I want to dig

into here. Number one, what was OpenAI thinking? What was Sam Altman thinking?

Releasing a model that was this bad ahead of a very prominent IPO from one of

their biggest competitors.

Now it just seems that Anthropic is running away with it. But number two, what did Google do?

What did they change to somehow leapfrog the entire competition to almost take pole position?

And that's very much the story right now. I truly believe this.

It is between Google and Anthropic right in this very moment,

and both are going for the golden cup of recursive self-improvement and solving

cancer and disease, both of which Anthropic and Google have invested billions

and billions of dollars in.

So I'm going to get into all of this, as well as the downstream effects of all

the NVIDIA drama. It's resulting in crazy chip exports against China,

which is putting around 20 to 30 percent of NVIDIA's revenue in danger.

As well, of course, the biggest IPO that we've seen, Anthropic IPO'd yesterday,

and what's happened since then. Now, quick reminder, if you aren't subscribed

wherever you're listening, please do that and turn on notifications.

It helps us out massively. Right, let's get into the episode.

So let's start with some honest takes on GPT 6.2 Astra, OpenAI's brand new model.

Now, a few things before I get into it.

I've been playing around with it for a few days, so this is by all means not

an extensive review, but it's good enough to start speaking about what my initial reactions are.

And number two, I feel like OpenAI rushed this model out.

The reason why I think that is it's just a model that feels unfinished.

It rambles for way too long, and then when it does come out with something useful,

it is incredibly complex to understand.

And then when you ask it to explain the complex thing, it just gets into a spiral

loop of kind of talking to itself and not actually doing the work that you gave

it to. Now, compare it to its predecessor, GBT 6.1.

GBT 6.1 was performant. It was a good model.

Now, compare the two side by side.

And it's very obvious that one, the newest one, is so much worse than the other.

And it's not just me that's feeling this. A lot of people that had early access

have given probably their most negative review of GPT ever.

Theo just had a public rage quit, basically saying that the taste of the model,

the way that it kind of understands you, is just not there. When it comes to

coding, it is certainly in no way better than its predecessor model.

It's nowhere near as good as Fable 5.2.

And in often cases, it gets lost and doesn't complete the task,

long horizon task specifically. So typically, if you've been playing around

with this thing for coding, Yeah.

Any long tasks that you set it for, usually these newer models can work independently.

Now it doesn't seem to be the case. And objectively, when you look at the benchmarks

of this model, you typically expect to see it exceed in many cases.

In this case, it didn't really advance that much. Now, what are the good things?

Well, it's good at visual generation. So if you pair it with a tool that has

access to creating visual graphics, it is phenomenal at doing that.

Blender is a very popular example if you want to try it out.

Number two, when prompted very carefully and specifically, it is actually quite good at reasoning.

But I still personally find that the answers take way too long to answer.

And when I do, I'm given like pages and pages of an essay, which I just don't want to read.

In fact, that's been my frustration with GPT for a while now.

And maybe if you're using it for software engineering, which is not really something

I spend a lot of my tokens on, you're seeing something else.

But for me, for the majority of people that are using this for writing,

creative pursuits, or whatever that might be.

Isn't the best model. It just feels rushed, like OpenAI fine-tuned this model

to kind of get really impressive stats that would be on paper,

but actually the model doesn't translate very well.

If that sounds familiar, that's exactly what Google had been doing up until,

I guess, three weeks ago.

But the idea here is I think Sam rushed this to try and get it ahead of the

Anthropic IPO, to which my question then is,

why would you rush and risk your own reputation as a Frontier Lab right before such a

crucial, but not that crucial event where it could affect your valuation of

the company very significantly, especially if it's not a killing blow against your competition.

People reacted to this model and obviously didn't like it and gave you that

feedback. So surely you would hesitate before releasing something like this. But anyway, I digress.

The strategy behind OpenAI's model releases seems to be in danger.

And let me zoom about and tell you what I think is actually happening here.

I think over the last couple of months, OpenAI has tried to make a very important

decision around where they allocate compute to train their models.

Now, obviously, they're kind of in a head-on-head race with Anthropik,

their main competitor. And so they want to make sure that they keep up with

them in coding, but they also want to forge into new pastures,

greener territories that might give them their.

Give them billions and billions of dollars of revenue so what might some of

those sectors be well the obvious one is maths where open ai's internal models

has been spending a lot of time solving very esoteric

mathematics problems and is impressing a lot of researchers and scientists out

there the idea here would be,

if maths can translate into some kind of novel mechanical engineering problem

whatever that might be chemistry whatever that might be then open OpenAI will

be able to capture that new market in a way that Anthropic was able to do with

coding with Claude Code.

Now, the second sector that they're obviously going after is cybersecurity.

And actually, the most impressive cybersecurity stat from OpenAI is not from

their most recent model.

It's from an internal model that they haven't been able to release. It's called,

actually, I don't even know what the codename is, but they did reveal a very

important stat recently, which showed that it was two to three X better at cybersecurity

exploits than any other model out there, including Anthropic's most powerful model.

The idea here would be, huh, if they can get contracted by governments or enterprises

all over the world that want to protect their systems against a growing frequency

of cyber attacks, then they'll be able to earn tens of billions of dollars from that.

So that's the bullish case. And I'm giving them a lot of credence here.

Maybe they don't care about the retail market anymore, but I think that's a

mistake. And with GPT 6.2 Astra.

It just seems like it's a heavily throttled and restricted model.

For you and I, for the 99% of the rest of the world that is still trying to

figure out what the benefits of super intelligent AI is for each of us,

respectively, in our personal life with personal AI agents,

you'd think that that's a very lucrative market to tap into.

The consumer market is absolutely huge.

And what's ironic here is we're seeing a kind of changing of the guard.

Anthropic used to be very enterprise focused and then they kind of broke into

consumer. So they started with enterprise and went to consumer.

And they're doing a good job of balancing both now. Right now,

they're the best company that balances both.

OpenAI went from consumer and thought, ah, we need to double down on enterprise

and is going really hard on cybersecurity.

But unfortunately, that's at the trade-off of releasing a really poor model.

So I don't think it's over for OpenAI. I think they'll be able to come back.

I think this is a case of retraining and building a different architecture.

And I think they'll be able to do this over time. And honestly,

I think the model right now from when you're listening to this episode will

have improved massively over the next, let's call it three to four weeks.

But again, bear in mind, that sounds like a short period of time.

But in that time, you'll have a new model that's released from the likes of

Anthropic, or in this case, Google, which brings me to my second hot take.

You can credibly argue that Google has taken the number two spot right behind

Anthropic in the AI model race.

Now, it's crazy for me to say this because it was only a few weeks ago,

maybe a month ago, where I was saying Google is too far behind to catch up.

You have Anthropic and OpenAI, which have both shown internal research of achieving

recursive self-improvement, which is AI models that can basically automate the

AI research. So the feasible output of that is.

An AI company that can just kind of keep self-improving and run away from the

competition, and that Google hadn't caught up. In fact, the model that they

released at the time was some useless flash model, and prior to that,

they hadn't really had a major model release.

That all changed over the last couple of weeks, where they released not one,

not two, but four frontier models.

I think we had Argon, we then had Carbon, we then had Hydrogen,

and just today or yesterday, the time of recording, we have Neon.

Now, all of these models are frontier models. You can kind of think of Google

Gemini 4 Carbon. Let's call it just Carbon for simplicity's sake.

That's the smartest model. That's their Fable 5.2 competitor.

And in many cases, it competes really well. And what I want to point out specifically

is coding. Google has typically.

Been terrible at getting their AI models to be good at coding.

They've done the creative writing, they've done the graphic stuff,

but coding sucked and it required a very specific type of training architecture.

Google has been able to solve that. Carbon is a fantastic upgrade.

Now, going down below this, Neon, which is their most recent model release, is the mid-tier.

It competes with the likes of Claude Opus 5.5 and GPT Terra, the mid-tier models.

Now, what's fascinating about this is Opus 5.5 has been Anthropik's breakthrough

model after they had kind of like a slump.

And everyone that's used Opus 5.5 has reported nothing but good news.

Gemini's Neon review so far has been that it is absolutely fantastic,

on par, in some cases, better than Opus 5.5, but very importantly, a fraction of the cost.

Now, this comes with a disclaimer, which is Google has an introductory pricing

right now. It's super cheap.

And you'd expect that after this introductory pricing period,

it's going to get a little more expensive. So I'm keen who chooses which model

after that. I'm going to play around with Neon and give you my full thoughts

at a later episode, but it's just something to point out.

And then the third tier of Google model is something called hydrogen. Now, actually.

This is my favorite model. And it's not because I use it like a regular chat GPT or Claude model.

I use it in Google search. You see, when you use Google search,

you have this novel feature right now, which is AI mode. So when you type something

in Google search, you don't just get a list of hyperlinks anymore.

You get a brief AI summary, and then you can choose to kind of engage with that AI if you want.

And I found myself really enjoying the Google summaries, the AI summaries recently.

In fact, if you look at Google's recent quarterly report, you'll see that their

search volume, their search revenue has started to decrease slightly.

It's still increasing, but it's at a slower rate.

And that's because more people are using AI mode. They're enjoying this AI version of Google search.

Now, up until the release of Google Hydrogen, it was using a crappy old model.

Now it's using Hydrogen, and it is just a fantastic model.

It speaks like an expert, but not too intimidatingly. You can understand it,

and you can have a very concise, verbose conversation with it that leaves you

feeling like you've just grown more intelligent, if that makes sense.

So it's a fantastic model. And of course, it's their cheapest model as well.

So the way I feel about this is if you have like a bunch of stuff that you want

to get done that isn't, you know, doesn't require the highest intelligence or

isn't a complex problem, Google

Gemini, the series of models are a really good alternative right now.

And in many cases, it might hit around 90 to 95% of Fable 5.2's capabilities,

but for a fraction of the price.

The only other instance that we were able to really see this was in a lot of

the open models that China was putting out. It's this new world where Gemini

took up the rear end of the AI race and somehow they were able to leapfrog to second place.

OpenAI, meanwhile, has made some, in my opinion, poor strategic choices,

especially with the release of their latest model.

And they need to kind of get back to base and kind of re-communicate what their

actual vision is for putting AI models out in the world today.

Now, one last positive spin on this entire thing is if Google can be so far

behind and somehow pull it out the bag. And listen, I understand that Google is a big company.

They have a lot of money. I know that they have all the data and all the props

and resources. But the fact of the matter is since Gemini 3 or 3.1 Flash,

they haven't come out with a really good model for over a year.

And that's very embarrassing for a company that big. And so I thought,

you know, if Demis has left, which he did, he retired as CEO of DeepMind.

And if Jeff Dean, who is one of the key technical people on the AI team,

left as well, who is left at Google to kind of like allocate resources?

There's obviously a cultural problem.

Well, the good news is they've apparently turned that over, which means that

any company that is in the background right now, I'm talking about SpaceX AI,

which I still love. I love Grok, but the fact of the matter is that model hasn't caught up yet.

I'm talking about Mark Zuckerberg and Meta, who has recently brought up the

rear and kind of like really pushed what the cheapest, most competent model

could look like with their, I think their smartest model is called Wonderful.

Still a very good model, not Frontier just yet, and Google's kind of beat them now on both fronts.

But if you're one of these companies that has a lot of money,

that is willing to hire the research talent and rebuild, not too scared,

this is very important, not scared to rebuild your entire business.

Team architecture design of the AI model from scratch, you can still potentially

pull off a miracle and come to the front. But I will say that window is closing.

And I fully believe this. I made an episode, I think last week, last Friday, titled.

Well, it was talking about AI 2027, which is a very famous AI predictions report,

which ended up being the most accurate predictions report, and they talk about

the stage of AI where you have AI self-improving.

And by that point, all you need to do is just spend as much money as you can,

throw as much compute as the thing, and let the AI improve and build itself.

When we get to that stage, which is estimated to be around sometime in Q1 of,

2027, so I'm talking about four to six months from now.

When you reach that stage, that AI can effectively just run away with it,

the number one position, because then it just becomes a matter of how much capital you have.

There are very few companies that will be able to compete in that world.

It's Google. It's OpenAI that's raising all this money. It's Anthropic that

just raised over a hundred and, I think it's $107 billion in their IPO,

$87 billion from the public.

And then we had an extra $26 billion coming from a private placement.

Then you have Google who has, you know, one of the largest cashflow generating

businesses. You have Meta.

Maybe you have SpaceX, but there's very few players that can enter after this.

Yeah, you have Microsoft, but Microsoft isn't really putting out any competitive frontier models.

So it really is a select few.

And the alternative to this, if you want a bit of hopium, is I guess the open source ecosystem.

But getting the funding, getting the talent, paying that right talent,

designing the right tools has obviously been a struggle.

Putting the NVIDIA scandal aside, if you look at what's happening with Chinese

labs prior to the entire scandal, they weren't able to catch up over the last

four months to any of the frontier models. Why?

Because Anthropic shut down distillation and OpenAir is doing the same thing.

So you could argue that a lot of the momentum right now is with the West and with Western AI labs.

It's not necessarily what I want in this world. I believe that there should

be an abundance of AI models that are freely accessible, that have differing

opinions, tastes, ideas,

can train on your private data, all so that there's optionality and we aren't

just reliant on a very select few.

But right now, that latter vision is what seems to be playing out.

Now, speaking of NVIDIA, here is the latest. They had their quarterly earnings

report, which was a milestone event post-Jensen removal.

Colette Grass, who is the interim CEO and also CFO of the company,

gave her take of the entire landscape and how NVIDIA is doing.

For all intents and purposes, the earnings was a record blowout.

I think they made $151 billion over three months, which is the most money that

any company has actually made in three months.

Very, very impressive. What I will say is the bulk of that was still under the

leadership of Jensen Huang. Worth noting, you know, Colette Grass,

I think, had only been in the position for two weeks, so there wasn't really

much that she could do to influence it.

But what's more important is what the future looks like.

And she made a few comments. Number one, she said, or she implied rather,

that momentum has never been better with NVIDIA.

Customers are still buying GPUs. And most importantly, we're still on track

to deliver all those GPUs.

And it's not a surprise to me because NVIDIA works that way.

They work in long-term contract deals, which means that all of the revenue that's

going to be made over the next three months, arguably, for the next quarterly

earnings, had already been locked in.

It's just cruise control for NVIDIA, honestly, for the next six to 12 months.

You're only going to really see

the test case of whoever the new CEO is of NVIDIA and how good they are.

After 12 months. And that's just the fact of the case and how the market is reacting to it.

So Colette kind of gave confidence there, but then people had a question mark

around the whole NeoCloud financing.

Because of course, NVIDIA has committed up to $500 billion, that's half a trillion

dollars of finance backing for a lot of these NeoCloud providers.

These are providers that buy NVIDIA GPUs, stand them up, and then sell them

to hyperscalers or labs like Anthropik and OpenAI.

So the question mark is, is NVIDIA still going to be risk-on like Jensen Huang was as a leader?

Colette gave a mixed response. She said, basically, listen, we are committed

to helping this space grow.

For the big neoclouds, CoreWeave, Nebius, Iron, we're good.

We will honor that as long as you guys are able to, one, find the money to buy

the GPUs that we're selling to you.

Number two, build the data centers. and number three, bring those GPUs online.

Then we've got you on the financing backing. If there ends up being lower demand

for this thing, we'll cover up to 25%. That's where the $500 billion came from.

But it's important to point out that if they don't satisfy any of those three

requirements, these NeoClouds, or rather NVIDIA, isn't obliged to pay up the money.

So really, the amount of money that they have to commit, this $500 billion commitment

is kind of like a nothing bug. It's not really that high, in my opinion.

I mean, you listen to this and tell me if you have different information,

but that's my interpretation of it.

Now, where the danger lies is for the smaller neoclouds. What she basically

implied in her speech was that the smaller neoclouds are going to struggle to

get the backing from NVIDIA, at least in new deals going forwards.

Jensen was very gung-ho. He was very risk-on. He was like, we need to build

and grow this entire space. We need to push out open source.

I don't know if Colette is going to have the guts to be able to do the same

thing. I think she's going to play it more conservative. And indeed,

maybe if you're an NVIDIA investor, you want a little more risk off at this

point after a very scandalous couple of weeks.

But what I will say is this will impact the market going forwards.

Now the reaction, NVIDIA stock has not been performing well.

You would expect record blowout earnings to kind of pump the stock.

Kind of go up pre-market, but overall it kind of declined and wiped out all

the gains, especially after Colette's comment.

So I think people feel a loss without Jensen as the leader. And they believe

that, you know, scandal aside, he is the only competent person to lead NVIDIA

right now. That could change if a new CEO is brought on and they're very competent.

But the truth of the matter is you're going to have to wait six.

Certainly six, max 12 months before you actually get to see how NVIDIA is truly

performing under new leadership.

It might be a slow death if it's a bad CEO and it's run by committee,

but it could also be a change of guard. And NVIDIA is a fantastic company.

The talent that they have is remarkable. Alone, it wasn't just Jensen Huang

that was required, but he was a meaningful part of it.

He had a lot of supply chain deals. He had a lot of connections with TSMC,

who is now saying, listen, hey, NVIDIA, all our other customers,

you've got to pay more upfront because we're not dealing with Jensen anymore,

that could end up being problematic.

And when you look at where a bulk of NVIDIA's revenue actually comes from,

it's from Asian data centers, the ones that are under scrutiny right now for

potentially selling chips or access to those chips to China.

Now, there's two sides of this. I'm not going to dig into this anymore.

I've done enough episodes on this. But the takeaways here is that was China

accessing these chips illegally.

One version of this is it wasn't illegal. It was technically lawful,

but kind of looked down upon.

You see, Trump and the admin in general has approved a bunch of GPU types in

older NVIDIA GPU generations that they can get access to. And so, you know.

They've got access to this. But there are some chips that live in the gray area,

in between the old chips and some of the breaking edge new chips that they potentially

may have had access to. And so the question mark is, was this legal?

Should this be allowed? Or should it just be a free market accessible to all?

We know which one or which side of the camp Jensen was laying on.

But then the other part of it is, have they got access to frontier chips that

they shouldn't technically get access to?

And how many data centers in Southeast Asia were giving them this access.

Remember, the number has been around $10 to $15 billion worth of chips to China.

That's a meaningful portion of revenue quarterly that NVIDIA can earn.

And roughly the mark is between 10 to 20% of NVIDIA revenue comes from the Southeast

Asian data cloud. So what I'm starting to see now is a lot of secondhand compute

come onto the market, flooding the supply.

And so GPU costs or prices to be rented are going down. Now,

I think optimistically, this is going to be temporary because I think it'll

just get bought up by Western labs and hyperscalers who are still dying for more compute.

But it is a bit of a treacherous situation and we have to see how it ends up playing out.

So yeah, that's the quick and dirty on the most recent headline kind of themes

and trends that I'm seeing in AI for this week.

I meant to spend a bit of time talking about the Anthropoc IPO,

but I'm sure you've got so many other outlets for that. So So,

you know, if you guys want to hear another episode on that, let me know.

But, you know, I'm sure you've got all your information on that already.

But yeah, just a crazy week. I am done with all the drama.

It has been a crazy, crazy two weeks,

with all the Jensen stuff, with NVIDIA shedding around $2 trillion worth of

market cap, coupling that with like the world's largest and biggest IPO,

where Anthropic raised the most money that any company has ever raised on the public markets.

And they're going for one of the biggest visions ever, which is RSI,

and we might actually be on the verge of it within four to six months.

It's hard to wrap my head around. A lot is going on, but...

What you can guarantee is that I'll hopefully still be here unless,

you know, Fable 6 replaces me anytime soon.

But I'll be here talking about it and giving you my honest thoughts about it,

my honest opinions. And listen, you guys might disagree with it.

And I actually like that you guys disagree with it. It helps kind of train my

thought process and kind of fill or plug gaps that I miss. But I like getting

these episodes in the hands of people that, you know, truly enjoy the content.

It's not about the views of the numbers.

So if you're here and you're listening to it right till the end.

I appreciate you. Thank you so much for listening and I'll see you on the next one.