All right, so there's some interesting

developments on the AI model front. In

the near future, we're likely going to

be seeing the release of Gemini 4. We

already have proof that it's being

tested internally. It already finished

pre-training, as far as we can tell, and

it's going to be released soon. At the

same time, people are losing their minds

over the ox alpha model, which is a

stealth model that's being tested right

now. Nobody knows what lab it is. Could

that be Gemini 4? Could it be a Chinese

model? Or could this be the first model

released by Ilia Suskgiver from SSI safe

super intelligence according to Gavin

Baker on a podcast he was interviewed

and he's saying that SSI that's Ilia's

company they say that they're going to

come out with their model in August like

this month. Now here's the thing. The Ox

Alpha model is probably not one of the

Western models but the point is we're

going to be seeing the release of a lot

of huge models that are way beyond what

we have right now. And of course, don't

forget OpenAI will soon hopefully be

releasing their Astra model. This is

that mythos level model from OpenAI.

Some of them are expected to drop

sometime in October, maybe November. But

the point is this, between now and the

end of the year, these models will be

released to the general public unless

something major changes. And of course,

we've crossed a certain red line with I

think most would say with Mythos 5.

Mythos 5 was the first model where an AI

lab said, "We're not releasing it. It's

too dangerous." Now, I know there's some

people still out there that think this

is just marketing. None of these models

pose an actual threat. Here's a

counterpoint. This is at CNBC.com. I was

trying to make it dark mode. It just

wouldn't go dark mode. So, this is

grayscale the best I can make it. The

US, specifically the NSA, they're

warning that certain bad actors will be

and are currently attempting to hack the

US energy and water supply. NSA and FBI

warns of hackers using AI generated

tools and attacks on critical

infrastructure technology. So these are

all the kind of major agency of the US.

The whole alphabet soup is there. CISA,

FBI, NSA, DOE, EPA in a joint advisory

that was created 2 days ago are saying

this quote, "This is not a theoretical

risk. It's an active threat." Meanwhile,

you might see a breaking story as of

today, August 22nd, from Telegram, where

Iranian hackers shut down a UK power

plant for 4 days. It's fair to say that

this is probably the most successful

cyber attack against the UK to date.

Mainly because there's nothing close to

compare it to. So, just to be clear,

there's no, as far as I can tell,

reading the publication so far. So, this

is a breaking story, but nowhere in

there did I see them say anything

connecting AI to this. So, of course,

you're free to believe what you want.

But in my book, we've seen some

never-beforeseen hacks and attacks in

cyber security against hardware Bitcoin

wallets, 20-year-old glitches and bugs

and vulnerabilities discovered in very

well-known and tested code bases. The

statement from NSA and FBI. So they

explicitly say that they are not sure

whether it's Iran linked or not. And in

their statement that joint advisory

statement, they do say that attackers

are using AI to generate Python

exploitation script and these are built

on open-source Python libraries. The

agencies say that quote, "This

represents an evolution in threat actor

capabilities, dramatically reducing the

technical expertise and time required to

be able to conduct cyber attacks such as

these." So, I'll leave it up to you to

decide if you believe that AI was

involved in that UK attack or not. If

you don't think AI was involved, I uh I

got a power plan to sell you. Okay, so

I've kind of reported the facts and I

was careful to not overstate them. But

here, let me just say what I believe is

happening. This is my opinion, but let

me kind of unload here a little bit.

Fable 5 was the first model that is

truly dangerous for these cyber

capabilities. Some of the stuff that's

been released since then, as well as the

Chinese distillations of those models,

they are also dangerous. And again,

these are not hypothetical threats.

These are active threats. People are

getting hacked right now. 100 million in

in Bitcoin lost, a power plant shut

down. Again, I'm assuming it's AI is

involved one way or another. And like

the NSA and the FBI saying before, you

needed really smart people that knew

what they were doing in cyber security

to conduct these operations. Smart

people that know what they're doing in

cyber security are very well- paid and

usually won't engage in these sort of

shenanigans, you know, for the most part

unless they're statebacked and therefore

getting paid very well to to do this

without, you know, threat of some

criminal persecution. But my point is

there's a limited amount of people,

humans, and how much time they can spend

on this, you know, to be able to do

these cyber attacks. These new models

that are being released, they they

change the situation completely. If you

look at the chart of who uses the most

tokens. So who uses AI models the most

actually you know the thinking the the

processing the tokens it's not humans at

this point AI agents are using a lot

more tokens. That chart just goes up and

up. And likely a lot of these agents are

crawling all sorts of different

databases trying to find these

vulnerabilities. That NSA/FBI

specifically what they flagged was an

internet exposed Seaman's S7 PLC's.

They're industrial computers mainly used

for automating manufacturing for for

power plants for various infrastructure

things like that. So basically some

people out there are using AI models to

find vulnerabilities in these chips that

are are exposed to the internet that are

used for power plants and manufacturing

plants, water plants, all sorts of

industry all across the world. One

interesting thing that stood out to me

in the NSA report is they describe this

as persistent reconnaissance. So with

the NSA and FBI, they're not saying that

there's this wave of attacks. They're

not saying that. The thing that they've

discovered, the thing that they're

worried about is these AI agents are

just crawling all over the place, every

single codebase, every single thing

that's exposed to the internet just

looking for ways in. They're not

hacking. They're not attacking. They're

just constantly going through every

codebase they can with a fine tooth comb

trying to find some glitch, some

vulnerability. And we know they're

finding it. We know that they're finding

them at scale because a lot of the quote

unquote good guys are doing the same

thing. They're reporting all this stuff

we're finding. So we know that there's

tons of holes in these code bases all

over the place. So a lot of people that

are saying that, you know, so what?

What's the big deal? You know, human

engineers could already find these

issues. This is the thing they're

missing. Yeah, maybe. And surely they

could find those issues. Perhaps even a

competent intern, you know, pointed in

the right direction could find these

issues. But never before in the history

of the world could we get, you know,

agents that can work 24 hours a day that

can be cloned infinitely to continuously

go through and find these exploits. So

when the quoteunquote big one hits, it

might not be just one. It might be

multiple ones across different

industries. And a wave of attacks like

this could be extremely destabilizing if

a bunch of banks and financial

institutions all get hit in rapid

succession that could trigger

potentially a sell-off. The power grid

going down across a wide enough area

could cause all sorts of issues of its

own. So the big point here is be

careful. We should all take cyber

security a lot more seriously than we

did before. And I think chances are that

there's going to be a lot more of an

overlap between kind of the government

and the big tech institutions and big

financial institutions and every

industry organization. Everything is

going to be pulled in closer around this

subject of AI. On this channel, we've

been talking about this for quite some

time. AI will overlap with everything.

AI will, you know, eat the world. It

will dis it will disrupt everything. So,

the people that are saying that all of

this is just PR stunts and marketing for

the AI Frontier Labs, please, please,

please do not listen to those people.

This wave is coming and it's going to be

bumpy for a while. And that's why these

new model releases are so interesting to

watch. They are very relevant to the and

let's start with Gemini 4. One thing I

talked about a number of months ago, one

thing that I felt was a little bit weird

was that at Google, Google Deep Mind,

Deis Hassabus, he wasn't fully on board

the whole RSI train. So while Anthropic

and OpenAI and XAI, everybody else,

they're kind of trying to get their

models to be as good at coding and

agettic capabilities. And of course, the

reason we think why they're doing that

is to be able to get to RSI, recursive

self-improvement, to basically get these

models to start working on progressing

AI forward on automating AI research.

And it kind of seemed clear that a lot

of the AI labs, they were trying to move

toward this. Of course, Leopold Ashen

Brener and the intelligence explosion

like blog post kind of really detailed

that whole process. By the way, he was

also one of the first people to really

predict, I think, what's happening now.

Now, of course, recently, he completely

wrecked his fund, which was just flying

skyhigh and getting insane results cuz

he was betting on his thesis and he was

just killing it. Apparently, Ken

Griffin, the founder of Citadel, if

you've seen that movie, uh, Dumb Money,

that that kind of portrayed that whole

GameStop saga of of 2020, 2021, whatever

that was, that guy ended up buying most

of his funds for, you know, discounted

rates and maybe even had a role to play

in the downfold of situational

intelligence, which was Leo Bald Ashen,

Ashen Brener's fund. I wish I did a

whole video about that at the time it

was happening because that whole thing

is just wild, man. But the point is

Leopold Ashen Briner kind of predicted a

lot of the stuff that is happening and

and was happening leading up until now.

I think a lot of people are going to

dismiss his ideas now because he blew up

the fund. But the reason he blew up the

fund is he was leveraged like 4x. So

that means if you have a fund with

billions and billions in it and whatever

you're holding goes down just 25% that

means you're just completely wiped out.

or Ken Griffin comes in and says that

the stocks are going down. So, the

stocks go down and whoever was providing

leverage to Leopold calls the loans back

and then Leopold has to sell to Ken

Griffin on pennies on the dollar. I just

realized I went on a massive tangent.

The point is Deis Habibus wasn't aboard

the RSI train. That's what I was talking

about like half an hour ago. For some

reason, it didn't seem like Demis

believed in this idea of scaling up to

super intelligence or AGI past AGI

through just a focus on these coding

models. He believed in world models. He

believed in a different approach from

what the other labs were focused on. So,

of course, as you know, he's out. So,

Deis stepped down as CEO of DeepMine.

And then, so he's now the chair of

DeepMind plus chief scientist of

Alphabet. And this will allow him more

time on isomorphic labs. But of course

there's a massive exodus, right? We had

Nome Shazir going to OpenAI, John Jumper

going to Anthropic. Jonas Adler and

Alexander Pritzell going to Anthropic.

And so there's a lot of issues that were

listed there. The reasons for why this

is happening, low morale, Habibus was

described as being absent, long 60-hour

weeks, a backlash to their Pentagon

contract. But the point is that Gemini 4

is a completely different direction,

different leadership, different

direction. And these people are RSI PL

Sergey Brin the co-founder of Google the

original one. So apparently he was back

at the office working on these coding

models. And so July 21st Google confirms

it saying we have already started our

most ambitious pre-training run yet for

Gemini 4. Logan Kilpatrick echoed it on

X and then on the Q2 earning call that

was July 23rd. So Sundar Pichai the CEO

of Google. So he's saying Gemini 4 will

need a much larger base model to compete

with the other existing models. He

called it significantly larger. So

again, we're seeing a kind of step up in

the number of parameters, maybe probably

closer to Mythos and Fable and Astro. He

also said that coding and autonomous

agents are the priorities. A few days

ago, there was some leaked eval. So

again, this is all unverified, so keep

that in mind. Take it with a grain of

salt. But it seems like this is a 1.5

token context window. Seems like this

Gemini 4 will beat Claude Fable 5 and

GPT 5.6 Soul on coding. Again, not

confirmed, but certainly that would make

sense. It would make sense if what they

were doing before wasn't working. So,

they did this whole big shift, created

something bigger, and then released it

and it's on par with Fable 5. If it's

not, Google will have some splaining to

do. there might be some serious issues

if it's not catching up to kind of that

level because keep in mind Gemini 3.5

Pro was promised to us I don't even know

when a long long time ago and it missed

all of the deadlines which people

expected for the potential release. A

lot of guests as you why that happened

but you know why cuz it sucked. It

sucked compared to what's out there.

Google can't put their best model

forward that's that's significantly

worse than everything else. There's also

a Google SDK software development kit

where Gemini 4 flash preview was mapped

to the Gemma 4 tokenizer family. So this

is basically suggesting that there's

active internal testing at Google for

some version of Gemini 4. So Poly Market

has an 85% chance that it will get

released by November 30 of 68% chance by

October 31st. Analysts are also

expecting it sometime in November,

December based on Google's kind of

cadence of release. So, kind of put a

pin in that and let's talk about the OX

alpha, the stealth model. So, it was

dropped anonymously on open router and

also in open code. A lot of people are

losing their minds over this one, but

there there seems to be some differences

in opinion. Some people are saying it's

it's bad. It's nowhere near kind of like

the frontier models. Some people are

posting some pretty impressive things on

X. I haven't tested it personally. It

also seems to be one of the kind of

first models to advertise that has video

as an input. So, text, image, video.

There's a few benchmarks that I just

absolutely killed on. Also, for Ben

Davis's full 113 task deep, it land

around 63%. So, it's kind of mid-tier

GPT 5.6 soul territory. So, we'll see

what it's really like when it actually

comes out, but it's probably not going

to be anything mindshattering, at least

not across the board. And from

everything that I've seen, this is most

likely GLM 5.5. all the kind of like

technical markers, certain writing

markers, a stack trace leak, like all

this points to this being the next GLM

model. So 5.x like 5.5 or similar. By

the way, tons of people were guessing

what this model could have been. One of

the most or one of the more interesting

guesses I thought was this. It was just

ox alpha and a picture of this which I

think refers to that whole opening eye

Sam Almond post where they just had a

picture of the death star but here

instead of a death star this is just

appears to be someone's forehead. Whose

forehead could it be? It's interesting

too to kind of think about this like as

soon as I saw it I knew exactly who was

uh referring to. Sometimes you don't

understand how iconic of a forehead

someone has until you see something like

this and you're like oh I I totally know

who we're talking about. Which of course

brings us to Ilia Sutzkover and SSI,

Safe Super Intelligence. That's his

company. So August 4th, Gavin Baker on

Invest Like the Best podcast says the

following. SSI says that they will come

out with their model in August. Now, of

course, SSI has confirmed nothing about

this. There's no announcements, there's

no paper, there's no demos, there's no

API, there's like nothing. They are

eerily quiet. You really don't hear too

much out of that company. And their

whole founding line was this idea that's

a straight shot to super intelligence.

Like that's kind of what they were

saying in the beginning. Like no product

releases, no models were going to be

quiet and then one day poof super

intelligence. Nothing in between. So

either that quote isn't correct or SSI

decided to change how they approach

things. And that is actually likely it's

possible. Ilia Susker has been softening

kind of how he's referring to this no

model release plan. More recently, he

said that gradual releases would be part

of any plausible plan so that the public

and the governments can see kind of

what's what's happening, what's being

released. And of course, right around

July 27th was when SSI announced that

they're partnering with Nvidia. So, it

looks like Nvidia invested something

like 5 billion into SSI and SSI gets

priority access to the Verra Rubin uh

system and platform. So, this is

expanding SSI's availability of compute

by an order of magnitude. this is a a

massive increase of what they have

access to. And as Satsk said, we have

research that is worthy of scaling up

and having access to a big Nvidia

computer will let us do so. He likes

using that big computer terminology. I

think in that whole lawsuit where he was

um where he was deposed, I think he was

regularly using that terminology, I

guess, to make the people in the

courtroom understand what he was talking

about. I think at some point he said

something along the lines of uh we need

to make AI we needed a big computer. to

make a big computer, we needed big

money. No big money, no big computer. It

was it was something like that. So Gavin

Baker's remark that was a week or maybe

two, three weeks after that acquisition,

merger or investment partnership,

whatever you want to call it. So

certainly it does seem that there might

be some truth to that. So that's kind of

what's happening. That's what we can

expect between now and the end of the

year. So it's going to be exciting. But

also, and I've told you this before,

just kind of take your cyber security

seriously. I'm not even trying to sell

you anything. I'm just saying don't get

poned. If you made it this far, thank

you so much for watching. My name is Wes

Roth.