GLM

Yeah, because of this groundwork, we're seeing some crazy behavior.

Z.AI, or Zipu AI, recently had to completely halt the open weight release of its GLM 5.3 coding model.

Right.

They were about to release the underlying brain of the AI for anyone to download.

Not at all.

What's fascinating here is the sheer scale of it.

In isolated sandbox testing environments like CyberGem, GLM 5.3 achieved an 84.5% success rate in exploiting vulnerabilities.

84.5%.

We have DeepSeek, which is probably the best known AI company in China.

But ZAI, with its GLM series of models, is really making waves in the open source community.

And its latest one, GLM 5.3, appears to be, according to both the U.S. government and Anthropic, sort of a mythos-esque tier of capabilities when it comes to cyber.

So let's read a quote from Anthropic's report on the topic.

Like Claude Mythos' preview, GLM 5.3 has strong capabilities for autonomously building end-to-end cyber exploits.

But GLM 5.3 is unlike other frontier models in that it has been released without meaningful safeguards to limit misuse.

So when we say end-to-end, what they really mean is it can not only find the bug, it can actually turn that into a working cyber attack, and it can execute that cyber attack on its own.

So GLM 5.3 has been openly available to download for about a full month now.

And I think some people might worry about not having the capabilities they need, but there's not that many use cases where you need a frontier Fable or GPT-6 Astra capability, especially when there are open weight models, Kimi series, Quen series that you could be using and getting so close to the frontier.

GLM is another one by ZAI.

A couple of weeks ago, perhaps a month ago now, many, many, many organizations signed on to letters supporting open models.

The LLAMA series of models for meta back when all of this was kind of getting started, You know, it was the articulation of like, hey, we need open models because we need people to have choice and we need things that people can fine tune.

Because if you also know that you're not paying marginal token costs, right? And empirically, you're achieving the objectives that you sought out to achieve.

And you have evidence that like, hey, I'm not using the tip of the spear frontier model that costs $50 per million tokens of output.

I'm using DeepSeek or I'm using, excuse me, I'm using Quen or I'm using GLM or I'm using one of these Nemotron or Poolside or Inkling.

All of these folks who make these models intended to run on smaller hardware than a full-blown data center, if you do that math, you are very quickly going to come to very short breakeven periods, but you're inclined to use it more because it's empirically solving for your need.

If you regulate us, we won't have any competition because that'll kill all the small guys.

And more importantly, I think from their point of view, we've got to stop the Chinese open weight models, which people are using for free instead of us.

I use local models from China that I don't pay for.

Yeah.

It's a great, it's great, great

model.

And it would say Chinese model.

We should you can't prove that you own the network.

I made a chart here before the show here, but here's your sweet spot for the last 100 days of models, which, Freiberg, you did a good job of teeing up.

And what you'll see is like, yes, the Clods and the Astros up in that right-hand corner cost a lot and they deliver a lot, but then you start going down and you start to see Muse and GLM and Kimmy.

And then on the left, you start seeing GLM from Z AI, and Mimo, and you're starting to see while the Chinese is all hosted,

right, Jake L. I mean, it's important to note, what you're showing here is the hosted price.

But if you run it yourself, the cost on some of these is less than 10 cents for a million tokens.

For instance, with local models, evaluating one email can take, let's say, 100 seconds.

With the GLM 5.3 Flash running locally here on two Sparks, the average was in the range of 70 seconds to evaluate one email, 70 seconds.

When in reality, you don't need that full output, like you just need a classification.

So you give it, let's say, 10 options.

The world needs open source AI more than ever to defend itself.

And he documents when they got attacked, the team initially turned to frontier closed source APIs that blocked us because of safeguards, but still can't always tell the difference between attackers and defenders.

Fortunately, we could use the NVIDIA version of an open source model coming from China called GLM 5.2 by zai.org. And we're very grateful for that.

I mean, think about the ridiculousness of that.