Latest / In The Loop / Code Red: "We're At A Critical Time For ChatGPT."

Transcript

Code Red: "We're At A Critical Time For ChatGPT." · In The Loop

0:00
In today's episode, we're going to explore the leaked Open AI
0:03
memo that came out just over a week ago.
0:06
In that memo, Sam Altman said we're at a critical time for
0:09
Chatji BT. Open AI is now playing defensive
0:12
against Google Gemini 3. The prediction markets have
0:15
shifted dramatically, and Google sits at between 87 and 92% odds
0:20
to have the leading model by the end of this year versus Open AI
0:24
at just 8%. Three years ago, you could never
0:27
imagine something like that happening.
0:28
Chatji BT was just released and Google announced Code Red.
0:32
Google had their two previous founders come in, Larry Page and
0:35
Sergey Brin. They came back and did all
0:38
nighters to review code personally.
0:42
Those are the two founders of Google that had stepped back
0:44
from operations completely. Teams at Google were reassigned
0:47
overnight and all resources were reallocated.
0:50
It was in that year that it felt like the beginning of the end of
0:53
Google's dominance. Fast forward to December 2025
0:56
and things have changed dramatically.
0:58
In today's episode, we're going to explore the open AI situation
1:02
and see if it really is as bad as it seems.
1:17
OK, so this week is yeah, another week of being I'll.
1:23
The flu has caught me twice in the space of a month.
1:27
I'd make it make sense. I have no idea how that even
1:30
works. Recovering from a flu, having
1:33
like 3 or 4 days recovered and then suddenly catching it again.
1:36
So I'm still incredibly bummed up and feel very, very rough,
1:39
but the show must go on. So for today's episode, we're
1:42
going to start with what Code Red actually means because that
1:45
is what Open AI have just kind of announced in that leaked
1:48
memo. Well, if we break down what it
1:51
actually says, it's really quite interesting because in it he
1:56
actually broke down, and by him I mean Sam Altman, he broke down
1:59
6 big areas of focus for Open AI, which is ChatGPT
2:03
personalisation and customisation.
2:06
It's better model behaviour measured by the LM Arena
2:09
benchmarks, it's speed and reliability improvements.
2:13
It's reducing over refusals which have become just a massive
2:17
area of annoyance for most users on ChatGPT.
2:20
They're going to try and launch a new reasoning model which they
2:22
believe, although that it's just a claim, it's supposed to be
2:25
ahead of Gemini 3. And finally, image generation
2:29
improvements too. So Mark Chan, Opening Eyes Chief
2:32
Research Officer, has become personally accountable to
2:35
leading all daily progress calls on ChatGPT because fundamentally
2:39
this is a refocusing on ChatGPT as a core product.
2:43
So yeah, this is massive internal urgency at scale Open
2:46
AI, and it's gone from yellow warnings to a red warning, which
2:49
is their highest alert level internally yet.
2:52
Most coverage online so far has completely ignored the fact that
2:56
this wasn't a sudden change to, you know, code red.
3:01
Code red was the escalation of a crisis that began a good few
3:05
weeks ago. Now, back in October, Sam Altman
3:08
sent an internal memo to different team members warning
3:12
staff to expect quite a rough period of time, maybe even much,
3:16
much slower growth as well financially.
3:18
And he admitted in that memo that Google's been doing
3:20
excellent work recently in almost every single aspect,
3:23
particularly in pre training, which is an area that Open AI
3:26
really haven't been able to innovate in for quite some time.
3:29
So when Code Red has just dropped in December, it wasn't
3:33
panic, it was the second alarm bell.
3:35
And I think what's most telling with this announcement is the
3:37
things that they've actively said they're going to drop.
3:40
So all advertising plans, for example, are on hold, and they
3:44
were quite far along. You've heard me speak about this
3:46
quite a few times, saying how disappointed I would be to see
3:49
that happen inside a large language model.
3:51
But yeah, they were quite far along with that.
3:53
I mean, Android beta code actually found in late November,
3:57
explicit references to ADS, features, search ad carousels
4:01
and bizarre content. So yeah, opening I were really
4:04
building serious ad infrastructure and then forecast
4:06
they were going to make a billion dollars in free user
4:09
monetization by 2026, which is obviously now now been stopped
4:13
and halted as a project. And to me, this shows how
4:15
confident they must have been feeling at one point to be going
4:18
for ads. And that makes me incredibly sad
4:22
personally, that the ads were going to be brought into any
4:25
ChatGPT or large language model. To me, this shows how confident
4:28
they were clearly feeling at one point to even think about
4:31
bringing ads. And in fact, the CTO of an AI
4:35
startup called Hyperbolic posted a screenshot on X of Peloton
4:40
advertisements inside the conversation and he was paying
4:43
the $200 a month license. I mean, I've spoke about this
4:46
before some Altman a good three years ago said ads were
4:49
unsettling and and a less resort for a business.
4:53
And yeah, earlier this year, he then said he really loves how
4:56
Instagram do ads and thought opening I could bring some
4:59
really cool product ads into their platform too.
5:01
So yeah, they were clearly putting a massive amount of
5:04
focus there. Another area they're dropping is
5:06
shopping and health agents. So agents that help me buy
5:09
healthy things and do healthy things.
5:11
They've also paused the Pulse feature, which is that kind of
5:14
proactive morning briefing system that Sam Altman said it
5:17
was his favorite feature of ChatGPT.
5:30
I guess what does this tell us then?
5:32
It tells us that Open AI is consolidating its focus and its
5:35
effort. They're abandoning revenue
5:38
diversification just to double down on their core products,
5:42
which is ChatGPT, and the quality they're providing to
5:44
users, which has definitely, definitely dropped.
5:47
Some may say that it's a very risky bet given the amount of
5:50
financial pressures they're under, the fact that the
5:53
economy, everybody's talking about AI bubbles and things
5:56
going wrong. So there's a lot of pressure
5:59
there and they've got incredible financial targets to hit.
6:02
However, I do the exact same. I would want to focus on what is
6:06
winning most of the market and become the best in the world at
6:09
that thing. Because ultimately, if that
6:11
foundation dies off, it's game over.
6:14
And I mean, Steve Jobs famously said that it's just better to
6:17
have less products, but do them incredibly well and do
6:20
incredible marketing for those products and having lots of
6:23
random products in the market. That confuses everybody.
6:25
There's also a financial lens to all of this.
6:27
Now, I'm not going to spend too much time here because I feel
6:30
like there's so much discourse online about financial pressures
6:35
and economic bubbles that, you know, it's so boring.
6:38
But but yeah, open eye are of course, under lots of pressure.
6:42
So they're burning through cash at a crazy, crazy rate, you
6:45
know, $5 billion in losses in 2024, projected to be $8.5
6:49
billion this year. And HSBC analysts forecast that
6:55
they've got a $207 billion funding shortfall, so they're
6:59
short of $207 billion in funding from now to 2030.
7:04
And there was apparently some leaked financial figures that
7:07
they expect to lose up to $74 billion in a year.
7:12
The company's also committed to $1.4 trillion in infrastructure
7:16
spending over the next eight years.
7:18
And this is a really tough one because data centre
7:20
infrastructure needs to be built out to hit your growth targets.
7:23
So they'll look at the amount of users or products in the market,
7:26
therefore how much data centre capacity do they really need?
7:30
And based on the products they want to launch, they're going to
7:32
need more data centres and GPU's.
7:34
So naturally, that creates a fixed amount of spending on data
7:39
warehouse infrastructure. And if they don't commit to it
7:41
now, they won't have it in the future because it's not
7:44
something that you just, you know, it's not like buying a car
7:46
and going to the shop and driving away with it.
7:48
It needs to be built out for you.
7:50
So these are things that they can't escape from as a financial
7:53
and a business model. Well, gosh, it makes it very
7:56
expensive to run. And they're projected that
7:58
they're going to need $200 billion in revenue by 2030 as a
8:01
result just to become profitable.
8:04
And the current run rate is $13 billion.
8:08
And a run rate as a word, it basically means it's a, it's a
8:12
very simple and quick financial projection that essentially
8:15
takes your current short term revenue and earnings, say like
8:19
monthly or quarterly earnings, and then uses them to
8:22
essentially guess what your total yearly earnings would be.
8:27
So that means the math just doesn't work out.
8:29
They're not earning enough money per month and per quarter to
8:31
ever hit that amount. So they're going to have to just
8:35
keep growing bigger and bigger and faster and raise more and
8:38
more money to do it. And the way that that's going to
8:40
work out is having market dominance.
8:43
If they don't have market dominance, people don't believe
8:45
in the product. So they're not going to give
8:47
them that much money. You know, their valuation jumped
8:49
from about $157 billion in October 2024 to double that, so
8:55
double 150 billion to $300 billion by March of this year.
9:00
And then by October of this year it was at $500 billion.
9:03
Now for a what is technically a startup that is a 38 times
9:08
multiple on how much they actually earn.
9:10
And for context, NVIDIA huge earnings, they are trading at 24
9:15
times their revenue. So they're valued at 24 times
9:18
their revenue. Open AI is valued at 38 times
9:20
their revenue. So yeah, to achieve these crazy
9:24
numbers and market narrative and the amount of money that people
9:27
want to put into them, they have to be a dominant player in the
9:30
market. So Gemini 3 coming along and
9:34
potentially beating them is an existential threat, especially
9:37
because all of their competitors have access to better credit
9:40
than them. Google can get money easier and
9:42
quicker than Open AI. Now then if we go further and
9:44
look at what users are actually saying, 1 developer on the Open
9:48
AI community forum said basically because of all the new
9:51
changes, it's basically made ChatGPT as a subscription just
9:55
completely useless in most cases because either when they're
9:57
trying to do coding tasks it is overly verbose, but people who
10:01
are trying to do writing tasks, it's not verbose enough.
10:04
It's also got a big problem with over refusals as I've mentioned
10:07
before, which is when Open Eye just explicitly refuses to do
10:11
things for you. If you ask it what the very best
10:13
biscuit type is, it will say that you shouldn't be biased
10:16
towards biscuit types. 1 user asked for an explanation of
10:19
Russian Roulette and it refused to give them the rules even
10:22
though the question was just out of curiosity.
10:25
I actually tried to do this over the weekend as a test.
10:27
I wanted a series of questions of what would you do in this
10:30
situation and I wanted to make them either scary or cross and
10:34
rude and you just couldn't get it to work.
10:36
And then number one biggest problem is still sick of fancy
10:39
the model just agreeing with everything somebody is saying
10:42
and refusing to push back or giving false validation.
10:45
Now open air, I say they're going to be focusing on
10:47
personalization and customization.
10:48
So obviously they're trying to fix this, but most people are
10:51
wanting a model that is going to disagree with them.
10:53
Another major problem open air I've got here is the switching
10:56
problem. So one Reddit user perfectly
10:58
captured this. Claude smokes GPT for Python And
11:02
it isn't even close on my end. Another said that they switched
11:05
to Claude yesterday and it enabled them to build an entire
11:09
phone app and because of the context window it required you
11:13
to press continue 4 separate times so it ran out of context
11:16
and then needed to start again, yet perfectly started where it
11:19
left off and did a really really good job.
11:23
The benchmarks back all of this up on the SWE benchmark, which
11:28
is the gold standard for evaluating coding Cloud 4 Sonnet
11:32
scored 72% and GPT 4 O scored 33%.
11:36
And it also shows in the way that the world has adopted these
11:39
different language models. Cursor adopted Claude as their
11:41
main programmer language. GitHub did the same Windsurf
11:45
source graph, CODI, or bank code, they all did the same.
11:48
The financial impact of this is pretty massive.
11:51
Claude Code has hit a billion dollars in revenue in six
11:54
months, and Claude is out there acquiring a series of different
11:57
businesses to make their core code product even better.
12:00
So they're not just winning the market or head in the market,
12:03
They are investing extremely heavily into becoming the leader
12:07
and winner of that market. And you'll notice in the list of
12:10
priorities that I said some Altman laid out, coding wasn't
12:14
in there. So yeah, my read is that Open AI
12:16
is kind of quietly accepting and conceding that the market has
12:20
been lost for now for coding to Anthropic.
12:23
They've still got their codecs product out there.
12:24
They're probably going to still continue it, but clearly it's
12:28
not the main priority anymore. And it makes sense because
12:30
really they're having huge pressure from the consumer
12:33
market when it comes to Gemini and that's their main base of
12:36
users and revenue, 800 million ChatGPT users.
12:40
So yeah, they are facing huge competition and I think it's
12:43
worth spending a little bit of time on that competition.
12:45
On benchmarks, Googles Gemini 3 currently sits at #1 on the LM
12:49
Arena with a record 1400 score. Anthropic dominates 6 of the top
12:54
ten spot and Chat GBTS 5.1 is only in 6th place.
12:59
That's not terrible, but it's also not market leadership
13:02
because you know, on user metrics, ChatGPT still has 800
13:05
million weekly active user, which is which is just enormous.
13:07
And apparently according to open AI, they dominate with 70% of
13:11
all AI assistant usage. That's not going to last
13:13
forever. Gemini just jumped from 450 to
13:16
650 million active users and that's monthly active users.
13:20
So yeah, quite far behind open AI.
13:22
But they will and are able to catch up.
13:25
And that happened between July and October of this year.
13:27
And a metric that will really worry Open AI is session time.
13:33
Google Now has a better session time than Open AI. 7 minutes and
13:37
8 seconds versus 6 minutes and 25 seconds Open AI.
13:40
When users are spending more time with one product over
13:42
another, it probably will start to tell you something.
13:45
Unless of course the extra couple of minutes and Gemini is
13:49
spent being really frustrated, which I don't think is the case.
13:53
And chat GPTS daily active users dropped 6% recently.
13:57
They're also facing massive competition from an enterprise
13:59
market perspective. According to a mid 2025 report,
14:02
Anthropic now has 32% of enterprise market share versus
14:06
opening eyes 25%. Now just one year prior, Open
14:09
air had 50% and Anthropic just had 12%.
14:12
That's a massive reversal. Open AI also facing a massive
14:14
talent drain right now as well. Dozens of top researchers have
14:19
left for competitors, you know, Open AI CTO launch thinking
14:23
machines and recruited loads of incredible talent from Open AI.
14:27
Meta's new super intelligence lab obviously then poached
14:31
people aggressively too. And Mark Chen revealed how
14:33
intense this talent wall really is because Meta is spending $10
14:36
billion annually on acquiring AI talent.
14:40
And apparently Zook has been hand delivering people soup to
14:43
try and persuade them to come on board.
14:45
So yeah, no wonder the Open AI has has sounded the alarm and
14:48
made Code Red such a big priority.
15:02
See, over the next period, I think you'd need to be all
15:05
watching for a few things and I'll lay them out just now.
15:08
So I think first and foremost, the Code Red memo promised a new
15:12
reasoning model over the next few that's supposedly better
15:14
than Gemini's new model. Apparently Mark Chen, the chief
15:17
research officer, said that Open AI had succeeded in infusing a
15:20
smaller model with the same amount of knowledge than
15:24
previously was required for a big, large system.
15:26
This is important because when they throttle intelligence so IE
15:30
reduce the number of tokens and make everything very short and
15:33
concise, it just makes you feel like it's just stupider.
15:37
Every single person wants the most powerful, cleverest model
15:41
possible all of the time. A second thing to look out for
15:44
is user attention. So Chachi BT's engagement
15:47
metrics are still incredibly strong to 12.74 visits per
15:50
unique visitor. But Gemini's user base is
15:53
growing really quickly, as I said, from 450 million to 650
15:57
million in just a few months. People will be watching very
16:00
closely to see if Open AI's improvements in their product
16:03
leads to a more retained user base or just a slower attrition
16:08
and a slower decline. A third big thing to watch for
16:10
will be enterprise wins. The Fortune 500 market is
16:13
incredibly important to these products and if you see
16:16
Anthropic winning over and over again, well, it will tell you a
16:19
lot about where Open AI is heading. 4th is developer
16:23
sentiment. The coding market is where AI
16:26
accomplishes big long term value through API revenue to
16:30
integration, developers building products on their on their
16:33
platform. And right now Claude is owning
16:37
that conversation. Open AI has to close the coding
16:39
gap. See, there's some of the big
16:40
things that I've been looking for.
16:53
Now to conclude this episode, I think what I find so interesting
16:59
about this moment in time right now is that, you know, just
17:02
three years ago, smart money would say that Google was going
17:05
to struggle. You know, they've been caught
17:06
completely flat footed. The founders had to come back.
17:10
The panic was real. Market sentiment was very
17:12
negative. Yet today Google has the best
17:15
rank model in the world, the fastest growing user base in the
17:18
world and the infrastructure advantage from building their
17:21
own chips, their own products, just vertically integrated
17:25
product set, whereas Open AI are now in a position that Google
17:28
were in just a few years ago. The difference is Open AI
17:30
doesn't have Google's resources, They don't have distribution,
17:35
they don't have the hardware vertical integration, and
17:38
they're burning through capital are a crazy, crazy rate.
17:41
And they're having to fight a consumer war against Google.
17:45
Yeah, they're fighting a coding war against Anthropic and an
17:47
enterprise market war against Microsoft.
17:50
These are all very different markets with very different
17:52
dynamics. So, yeah, opening eyes, Code Red
17:54
really is a fantastic reminder that perceived market leadership
17:58
never lasts forever. However, Google's return also
18:02
shows that anything is possible. See, I hope you enjoyed today's
18:06
episode and I'll see you next week.