Well hello there so recently there's been a mystery online

there's a new ai model that's been kind of making the waves on

threads and x basically the entire ai community and this

model is called ox alpha so in this video i'm going to tell you

what ox alpha is and how you can use it and also kind of the

theories on where this ai model is from so let's get right into it

now if you're new to this channel my name is john i'm a senior

staff software engineer at meta and on this channel i go over

tech news ai content and a bunch of tutorials on doing agentic

coding so if you're interested in that please subscribe a lot of

you guys are not subscribed i think like 90 80 percent of you

guys are not subscribed but just watch all my stuff do me a

favor it was my birthday recently subscribe all right with that

intro out let's get right into it so wtf ox alpha a quick rundown of

ox alpha is that it's actually a stealth model it's basically just

got released anonymously a few days ago, I think like August

20th or so. Here are the high level specs. It has roughly

a million, you know, context token. It has like an output

maximum of like 131 tokens, which is pretty large. It supports

multimodal like input and videos, but it only generates text,

which is like interesting. It doesn't do image model output.

And right now it's totally free. And that's kind of the very

interesting

like point. I think the most interesting take on this entire model

is the marketing, like the stealth model launch. I think the entire

industry and the whole AI community, they really like this

interesting like approach. There's some story behind it,

in my opinion. But yeah, so as I was saying, it's free,

it's mysterious, and it has a lot of capacity right now,

20 trillion tokens. So one of the interesting takes here is that

because there's so much free tokens available during this like

launch window, is that it has to be one of the major labs.

That's kind of like the only places

where the capacity can come from. So people are thinking

maybe it's Gemini, you know, from Google or Grok.

Elon's been hinting at the new like Grok model, like I think 4.7

or some Chinese lab that is getting a lot of support from the

government or something, right? It got really good usage

right away. I think it had 260

something K unique users. And I think OpenCode like reported

that it was like number two on the high, like traffic wise,

like 5.4% of OpenCode traffic, which is really large.

And as you can see right here, it like spiked up and then

it dropped. And kind of like interest. I think a lot of people are

like testing it in and out here and then people kind of

lost interest.

And here is OpenCode's posts on like kind of the,

like capacities. This really went viral, which was

really interesting. But the real story, the thing that made this

like really insane beside it being like free to use, I think was

some of the initial benchmarks. And I think a lot of people ran

some of the DeepSuite benchmarks and it got like 80%,

which is like incredible. And people are kind of losing their

minds on this. But it turns out,

turns out that it was actually just a very small data set.

Only about eight out of tens has scored that much. And when

they did like a larger data set, it actually scored on the 58.4%.

So if you look at this chart, if it's 56%, it would say right around

here in between Opus 4.8 and QEM3 Max. It's definitely

nowhere near the frontier, right? But if it was 80%, it would be

better than Opus 5, which is like pretty good, right? Sold in

Opus 5. One important thing about the whole initial numbers

was that it wasn't really an official benchmark that was run.

It was just like community run. There was some people just

running it on their own. And then that kind of came out with a

lot of like weird results. But it really helped all of this mystery

behind it. And like kind of the initial small sample size made it

like really pop and people were like kind of freaking out about it.

This is a guy named Ben. He actually is one of the like early

posters about it. He was like claiming that it was like 80%,

but after like a lot of testing, it was a much smaller.

And, you know, he kind of like went through it a bunch, giving a

lot of takes that it has better voice than Cloud or HGPT, has like

decent design handling. It runs code for a long time, but it's like

very slow and it has some weirdness. But yeah, so a lot of

people were seeing these like initial takes and it just turned out

it was like completely wrong. And that's kind of where I think a

lot of why the initial usage was like really high,

but then it kind of like tapered down now.

And eventually people like tested it on their own using open

code or like open code to like open router or something

like that. And they, everyone kind of came to the similar

conclusions as I did. It's not really frontier level at all. So there

was a lot of takes like this where it basically said it scores along

Kimmy 2.6 and it's like about as good as Luna, maybe worse.

And here's like another take they're saying it's right around

like little above luna maybe and like around terra nowhere near

the frontier at all that's kind of high level where it is a lot of

people are kind of like guessing where things are because i

think the interesting thing is that gemini still hasn't released

their like latest models there's been delays

and you know grok

4.7 or so the elon was hinting at that so there's like these like

kind of moments that people are so waiting on and they

thought it was this and because of the initial kind of like wrong

benchmark scores from the community people are like jumping

their gun and you know claiming that oh the only people that

can do this kind of stuff is people with a lot of compute and you

know google and only like the major labs have

compute worth to spare one of the like other really interesting

things that i

found while doing my research on this topic is that there's like

this whole side

of people using a lot of like interesting engineering techniques

to try to like reverse engineer get the models to like reveal

like what model it is but yeah so there's all these like interesting

things like something called tokenizer

and also like java class path to try to like get the information to

be released i found these two articles that was really

interesting i really recommend you guys to take a look there's

this article by asim sharin and essentially it used an

uncensored like ai model to like hunt down what ox alpha was

and basically it's saying that it found it to be glm5 there's a lot

of interesting strategies here but essentially at a high level this

is kind of a rough shot of what that i think he used quen model

but basically he just had to go

and like hit these models multiple times

and try to like compare with different models and

basically, here's like, kind of like a rough outline. And there's

like things like tokenizer differentials. And then there's like,

you know, kind of like looking at the emojis and things like that,

and trying to like match it with different how the model outputs

on this normal and trying to like reverse it to some other model

that would like kind of match its behavior. And this is kind of

how a lot of the watermarking stuff is going to work in

the future, where it's mainly like a stats based where you're

going to be able to try to know based on how the

model answers, what that model is like, what the original

source model is like, if you distill the model, it will have a similar

characteristics and behaviors of

like the model that it is still from. So that's kind of the strategies

is doing. And basically, at the end of all of this, what he came

down to the conclusion of this author is that it's the JLM five

from z.ai. And there were some other people that making the

same claim using like different techniques, like this is so

AI generated, but like, the model will never say that it's like a

JLM model, but it's servers apparently did. And based on

this data, the server is just servers where it's coming from the

origin server is from JLM, there's nothing conclusive here.

But like from what I researched, and like kind of the leaks that

I've seen, and people doing these kind of different research,

it really does feel like it's a Chinese model, and most likely JLM

because I think like Quinn has already released a

model recently, the sick as well, and also Kimmy as well.

So like, all the major ones that I know of have already

released something. So I think JLM is the only one. I believe

we're waiting for. But yeah, it's still unknown. We don't know

anything else besides,

like what I've been saying. So yeah, so the highest theories is

that is JLM is probably the strongest fit, maybe is Gemini,

I've seen some random posts, there's like engineers who

work at. Google are hinting at something.

But I think honestly, for that, I think they're hinting at the models

that's been delayed, I'm sure like they're saying it's gonna be

the best and same with Elon, you know, it's like fits the timeline

of what he was saying for the next rock model. So that's the

only reason but I think from all the other data, it definitely feels

like JLM.

Okay, editor John here, minor update as news just got released

that the OX alpha model actually is. JLM 5.3 flash.

So yeah, I guess I was right. But yeah, it sucks because I was

like working on this.

Anyways i think there's some good takes

on this video that's not just about the model itself and the

mystery but yeah i hope you guys enjoy the rest of the video so

that's basically it i think the glm model is the ox alpha most

likely we'll see if i'm right if we'll see i'm wrong maybe it's a

completely new

like model that distilled from glm i don't know it could be

anything but

besides the model itself i think one of the interesting lessons

and i've been kind of feeling this lately that the ai war this whole

like industry it's being driven so much by social media i think

there's so much attention that is happening with these and

people essentially move from model to model like very quickly

and then they switch sides very easily like i think even a few

months ago

i think around when fable came out originally like clarko was at

the peak in my opinion of

very favorable public opinion, but right now because of the

cost of Fable being so expensive and I don't know, I just feel

like some of the sentiment has been

shifting to maybe use, you know, open AI models.

Sol is my favorite model right now. Also, like if you remember a

few months ago, like Google was

like top dog again, but right now they slipped a little bit,

you know, they haven't launched their latest model.

There's been a huge shift, a bunch of people leaving Google

right now, you know, Jeff Dean left for Discovery Loop, took a

lot of the great AI researchers and distributed

systems engineers. Also like Demis, the head of like Gemini,

now he's like Alphavis chief scientist instead. So like, we don't

know what happened there, but there's like a move and. I think

Sergey's coming back to lead Gemini. This industry is insane.

It's like, it just moves so quickly. It's changing so fast and

people have different opinions all the time. Like everyone sees

me as like the cloud code guy. I made a really popular,

cloud code video, but I'm a codex guy right now.

I use codex app the most. I still use cloud code. I still love

all of these tools, but you know, my daily driver right now

is codex. And then this whole stealth model job, I think is really

interesting because it's like a very fun

marketing strategy. Basically

this model released would have been perfect if it actually

delivered on the benchmark. I think from a

marketing perspective, they did really

good, but it didn't meet the benchmarks. So it kind of fell flat on

its face. The AI war is going to be won based on social media,

as funny as that is. Also how good the model actually is.

The media pushes people to use things, but they stay for the

model quality. So you kind of need both. But yeah, I hope you

guys enjoyed this video. I'm going to cover more news, AI tech

news in this channel. So don't forget to subscribe. Recently,

I started posting on threads and the X a lot more, and I did

an AMA

and I answered like 20 something questions or so that people

asked me. I don't know why, but people seem very interested in

like takes from some random like questions that only like L7

senior staff engineers could answer, I guess, I don't know.

I'm just like a normal dude, but I guess people were very

interested because that post really blew up. Check that video

out and let me know in the comments below what kind of other

videos that you would love to see on this channel. But until I

see you guys on the next one, bye.