OpenAI is FINALLY OPEN! (gpt-oss)

You have to remember that is the world's information in your pocket. Should we jump into Open AI? Yeah. They came with two big news. First off the press is GPTOSs which is uh two state-of-the-art open weight models. And you got 120 billion uh parameter model and a 20 billion parameter model apparently deployable on consumer hardware. We'll get into a little theory that I have coming up in a bit. And these are these are fitted with the Apache 2 license, meaning that you can reuse them, post- train them, do what you like with them, put them anywhere, and you know, OpenAI can't take a scent from you, which is really, really cool. the performance is on par with 04 mini and uh this is the 120 billion parameter one and the 20 billion parameter is kind of like 03 mini. So, we're still not we're not we're not talking like amazing amazing models here, but there's something you can download and run on most um hardware, particularly the 20 billion parameter model. I don't know the exact sort of size of that, but apparently you can run it on 16 gig of RAM. So, that's pretty interesting. My theory is that it could be what Apple intelligence becomes is my theory. but good for tool use, train of thought, reasoning and and healthbench as well. Again, thinking about the fact that these can run locally and it performs really well on healthbench, meaning hospitals can um run this on on their, you know, I don't know how deeply ingrained AI is in hospitals, but you imagine a power outage and you know, you're you're very reliant on AI systems or having a you know, um another perspective with an AI on someone's condition or whatever. uh getting good results on the health bench is a really nice place for it to to to be good at, right? Got adjustable chain of thought which is really nice. So um you can actually say give us give me some quick results or you can make it massive if you want some some deep thinking and all the rest of it. I guess guess more for the health bench sort of stuff. mixture of experts which as we learned last week is basically a way for the you have designated experts that basically um encapsulate a certain portion of the data set. So 100 billion parameters activates 5.1 billion parameters per token. So there's an expert dedicated to 5.1 billion parameters and it can it's like a specialist in that subject basically. So imp improves performance uh because again you're only activating 5.1 billion parameters on that 120 billion parameter model whereas the 20 billion parameter only activates 3.6 billion parameters. So you're getting that um increased performance you're getting that the responsivity and things like that. It's got 128 uh,000 contexts which I was doing some research today and you know when when GBT first came out GBT2 I think it was I think it had only 6,000 let's be real. Um Google talk about a million parameters but like I think context size is way overrated. I I'm just I'm thankful for the 128 basically which is on a free open source model so or open weight model. And what's really really interesting is they've actually tuned it in such a way OpenAI have tried to fine-tune it to do malicious things and is actually really good at not doing those things because that's the thing with this all this open stuff is that once you release it out there it's done then you you've released it. It's it's free for the world to do whatever they want with it. Here we go. This is the fine tuning that I I told you about. So state is state-of-the-art approaches for safety training. During pre-training we filtered out certain harmful data related chemical, biological, radiological and nuclear um data and during post- training we use deliberate deliberative alignment and instruct hierarchy to teach the model to refuse unsafe prompts and defend against prompt injection. So um again just that uh safety thing. Just lastly, there is a $500,000 bounty to if Red Hat challenge basically. So, if you can crack this nut and reveal vulnerabilities, OpenAI are offering 500. So, red teaming challenge. So, there you go. Get on it, boys. So, you get half a million dollars if you can break into this and then Mark Zuckerberg will probably hire you. Yeah. Yeah. For like a couple million dollars at least. Obviously open source and and open weight isn't a brand new thing but let's be clear like um open AI are good you know and they have humanity's best interest at heart whereas you don't know these open source and and whatever models you you we often don't know what they're you don't you don't know whether they're aligned in the same sort of ways again so we're all taking it at face value but um I've not had a chance to use it I've seen mixed responses basically on how good and how fast this model is. I don't know if you've dug into it at all or honestly no I'm uh I want to mention a few other things. So one the Apache 2 license. So what we are saying what we are seeing is that open AI's new model is more open than what Mark Zuckerberg has been putting out there. I've heard that with Llama you can use the models up to a set limit of users I believe like after I might be wrong about this but after like a million users or something like this you have to pay them or show that it's their model or something like this Apache 2.0 Oh, that that sounds great. Just looking at the benchmarks. Benchmarks are only good until the next model comes around and naturally the next model is going to be better. But the I think the interesting thing about this model is that it doesn't aim to be the best. It's just interesting to see how comparable it is to some of the other, you know, state-of-the-art models. And you've got the the the two open weight models here competing quite well. I mean, this is um this is without tools. Uh sorry this is with tools without tools with tools without tools and it's compete it's still not you know it's competing very well with 03 and 04 mini here. So it's just it's just by comparison really and it you know it it stands it stands its ground. The other thing to end on really is like these open models these locally runnable models particularly the 20 billion model like I've now downloaded that onto my machine you know and I've got it ready to go whenever I need it. And that you have to remember that is the world's information in your pocket. Like I know we've got the internet, but to have it in a in a in a system in a way where you can ask a very specific question about health about, you know, we talk about health bench and stuff like that. To have that in your pocket is just next generation. And I'd like to know what your thoughts are on this potentially being the foundation for Apple intelligence. Yeah, it runs on 16 gig, which is exactly what Apple Intelligence supposed well is supposed to run on or what they're trying to get it to run on. Um, obviously they've had a massive focus in privacy specifically, you know, health as well. I mean, Apple are all about health right now with with watches and so forth, and they seem very well aligned with Apple's values and principles and it's free to use. You know, that sounds great. But when you when we when we heard about um Open AAI talking about running their model on on a on a hardware piece like on a smartphone or a laptop I was actually thinking about something else. I was thinking about not Apple but Johnny IV and their own product that we haven't seen and we haven't heard anything about it. 100%. They are like optimizing for that. Yeah, 100%. Yeah, these things need to be offline. They're optimizing for their tool. This is part of a larger conversation on my show, Command AI, which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter.