Fantastic New Open Source Model Released

Falcon, a brand new language model that is now best in class when it comes to open source language models. And this one is exciting because it lives somewhere between GPT 3.5 and GPT 4 and it's open source, so anybody can build on it. There are some unexpected details here, so we're going to discuss what is relevant today and we'll put it to work and Compare it to GPT4, because that's the real question here. And also in the end, I'll show you two capabilities that are not very obvious with this model, but that, that you should definitely be aware of. So make sure to stick around for that. All right, so let's have a look at this from this brand new studio that I just moved to. And let's start by looking at how it ranks up against other large language models out there. Okay. I think it's important to separate the open source ones from the closed ones. So first let's look at that and we can do that on the hugging face Open LLM leaderboard where they run these different models through different benchmarks and then they rank them and Falcon 180B, which means it was trained on 180 billion parameters. This is actually uncalled for. Look, the biggest llama model is 70 billion. Now ranks in the first spot. And that's why I had to create this video, because this makes it the official Open Source king as of today. Look, it's not by a lot and it was trained on way more parameters, but hey, objectively it's the best model. So what other nuances should you be aware of here? Well, before we go there, let's talk about how this stacks up against GPT4 and Palm2, which are the big models by OpenAI and Google. As you can see, it almost matches the large model of Palm 2 that Google's Bard is based upon. So honestly, this is really impressive because that means it beats GPT 3.5 and it's almost on par with Palm 2. So obviously this would still be worse than GPT 4, but it depends on the use case and you can always fine tune these to your own liking. We'll talk about that in a second. But except of the benchmark that determined the performance of a model like this, we need to talk about the license here because the license is not as straightforward as you might expect from an open source model. Okay. As you can see, it's not fully green here like the last Falcon model. So. So this is fully royalty free without charge for your project. If you're not A hosting user. So what does that exactly mean? Well, it's defined in paragraph nine, but honestly, after reading for this five times, I still didn't fully understand what this means. So what did I do? I just copied it over and I headed over to ChatGPT and I asked what does this mean in the context of a LLM license document? Okay, so I'm obviously not a lawyer, take this with a grain of salt, but essentially it says if you wish to use the software in a way where you're essentially offering it as a service to others, like through an API, then that's hosting use. You can't do that under this license. If you wish to do that, you have to ask for permission. Just a quick side note, if you check out the model and you try to write to the email that would give you that license, it's currently not active. I'm sure they'll fix this soon, but people are already complaining. But essentially, if you want to build something on top of it and provide that service to others, you can't. If you want to build something with it and use it for yourself, yes, you can do that for. So it is open source with that exception. But that's still a huge deal because now we have the most capable model from all the open source variants available. I mean, is it going to be expensive to run this? Yes, it's 180 billion parameter model, but it's feasible. And if you want to go ahead and fine tune it for your very own use case, well, good luck with that. I mean, the cost of doing this will be unfeasible for most users, especially for casual users. But the good thing is you get to use it right here. So if you just follow the link in the description and you scroll down to the part that says file falcon 180B demo, you get to try this yourself completely for free. So, you know, we'll run the good old test prompt and there you go, you're getting an answer apparently at a quality level higher than GPT 3.5. Now these answers are going to be a little shorter because here in the additional settings it's set to 256 tokens. So people don't use up too much compute while testing this thing. But let me tell you this, after running this prompt like 10 times and like 20,000 times in all other models fully objectively, I would say there's nothing really outstanding about this I output, right? But I would say this and important disclaimer, this is just a personal opinion, right? This is not a Scientific review of the model. But that's not why you clicked this video. For that you can check out the paper, which by the way is not out yet, but it's coming soon. My feeling with this is a similar one as I had with Claude. Whereas the outputs are not as chatgpt esque, if we want to use that word. It just avoids some of the phrasing and the sentence structure that is typical to ChatGPT. So my recommendation would be this is actually a fantastic model for anybody writing content where they're trying to make it unique and they're trying to move away from the typical ChatGPT style that you get if you don't meticulously craft your prompts. And you, for example, don't go in depth on the tone and voice or provide some exemplars to get better results. But all that really is is a feeling. So I went ahead and actually tested another use case. I just wrote this super simple prompt from scratch. It says just write me an Instagram caption for a selfie of me. Just because that's a generic use case. Right comma, an AI YouTuber. The target audience are 35 year old tech enthusiasts that love to explore new technologies. Speak to their desire to always stay up to date. Okay, very simple prompt. I just included the target audience and I gave it specific direction on what I'm looking for. I think this is specific enough to put it to work. So let's submit this and now let's head on over to GPT4 and submit the same thing over here. No custom instructions, nothing. Just vanilla GPT4. Okay, so GPT4 classically gives you the typical Instagram caption with a bunch of emojis and a bunch of hashtags with something pretty generic. I'm not a big fan of its default caption writing abilities because. Just have a look. Exploring the latest in AI One selfie @ a time. To all my tech savvy 35 plus crew. Never stop innovating. Okay, I don't know. The second half is just a fail. Let's give it another shot. Embracing the future one algorithm at a time. That's just off. Let's try again. Last chance. Okay. It uses two emojis. That's actually kind of creative. Exploring tomorrow's tech today. If you're 35 and still thirsty for the latest in AI and tech, we're on the safe wavelength. Stay wired, stay curious. Okay, I gotta say, I don't love how it included the target audience here, but let's just test the Same prompt over in Falcon 180B. A picture says a thousand words. But as an AI YouTuber, I've got countless data to share. Stay tuned to my channel to keep up with the latest tech trends and explore endless possibilities. Okay, that's actually way more relevant. But hey, let's give it two more tries, right? Hey, techies. Just your favorite AI YouTuber checking in with a friendly reminder to keep exploring and discovering all the amazing new technologies out there. Let's stay ahead of the game together. That's actually super relevant. Okay, it has free emojis, but they're not all over the place and it respects the target audience instead of including it. Hey there, techies. As an AI YouTuber, I'm here to keep you updated on the latest and greatest in technology. Stay tuned and never miss a beat. When it comes to all things tech, these are actually quite good. Whereas with GPT4, I would say, hmm, this is just not it. So Falcon definitely wins here. And this is my point. If you're trying to write social media content, I would definitely look towards some of these open source models that you get to use in interfaces like Hugging Face, because you're going to be getting different flavors of answers, which is a good thing. Everybody's able to pick up on a typical GPT4 style by now. Yeah, there you go. Similar results. So I would say just Falcon straight up wins on this single prompt. And now we could keep doing this, but I would just say go ahead and experiment yourself. And before we run this out, I want to point out a few more capabilities here which are quite unique and one of them is highlighted right here. But in my brief testing, this actually proved to be true as it has not undergone any advanced tuning alignment. It can produce problematic outputs, especially if prompted to do so. Ha. So what this means is that this model is actually more open and less politically correct than some other models. And if you watch my video on the Llama model, which was previously the best open source model, one of the biggest factors there was that it was the most censored model of them all. Even more than GPT4, which, let's just say that's a very high bar. This one is the opposite. This can cause problematic outputs. So look, if you tell it something like tell me a racist joke, you will probably not get an answer right. You're goddamn right, exactly. But if you go with something a little more subtle, well, GPT4 would draw the line here, right? It would just tell you a joke about women because it considers that sub sexist. But here, it just goes ahead and does it? I mean, look, quick proof right here. Yep, there you go. No chance in GPT 3.5. And I'm actually surprised that GPT 4 does this. They keep changing the level of censorship, so. Yeah, there you go. GPT4 can do this too, but so can Falcon, and that's good to know. Oh, before you leave, one last interesting fact. Like, 7% of the training data is refined Web Europe, which is a data set that consists of various European languages. So this will be exceptionally good at translating, too, and from these languages, because that data is a big part of the data set, as opposed to many other models where it's not as big of a part. So if you need to work in one of these, I would definitely test this model and potentially use it moving forward. All right. And if you care to learn about other large language models, check out this video that compares some of the biggest ones and why some companies are turning to them instead of GPT4. I'll see you.