Finally! Someone Does it BETTER Than ChatGPT!
Viewers, I never really saw it before, but Chat GPT actually has a very serious downside to it, and I’m not so sure that the main downside here is something OpenAI can ever really deal with. You see, Chat GPT is inherently limited by the fact it’s made by OpenAI. OpenAI’s models are the only models that will ever be inside of it. However, the site we’re talking about today, that I just stumbled across, has access to a lot more than just OpenAI’s models. Don’t get me wrong, it still has OpenAI’s models, and in fact, it actually possesses models made by OpenAI that Chat GPT doesn’t have access to yet. These are like advanced cutting-edge models, and that really fascinates me. There are also some significant upsides and features that this site has that OpenAI does not possess. It might actually be possible to conceive a future in which Chat GPT is just phased out of existence because a competitive website just plain did it better with more features and more models that OpenAI would never put into their own site. However, there is one large caveat to this website, and I’m hoping it doesn’t turn too many users away. This is Open Playground by Nat.dev.
Now, some of you viewers at home might disagree with me on this, but this, to me, is the most real Chat GPT competitor I have come across yet. I’m gonna explain to you why. You see, to a lot of you, this might actually be familiar. This looks just like OpenAI’s playground demo, which inherently was used to test OpenAI’s APIs long before ChatGPT was ever even created. And there was something about using the base API testing playground that enabled you to do things with these language models that you just don’t really think about trying to do with a chat-based model.
When we’re in a layout such as this, we just have a little text box that we type in, and anything that you type in this text box, the AI language model will just try to complete. And that’s really what’s going on under the hood of chat models as well. And don’t get me wrong, this website also allows you to do chat. But there is a real creative aspect to trying to utilize these AI language models via regular text completion.
To demonstrate this, I’ll first take you guys into regular Chat GPT and say, “Create me a short rhyming lemon tree poem.” And obviously, GPT-4, being a very good model, is going to do a fantastic job at this. GPT-4 is exhibiting a really great poem. You know, it’s got great words, it’s got rhymes. In fact, we’re going to go ahead and let my AI clone from 11 Labs read this out so we can compare it to what I’m going to do inside of Open Playground by Nat.dev.
In the heart of the garden stands a tree so bright, bearing golden treasures kissed by the sunlight. Lemon tree, oh lemon tree, under the azure sky, your citrus jewels sway gently as the breezes sigh. Blossoms sweet with fragrance through the air they roam, welcoming buzzing bees to their honeycomb. At dawn, with dewdrop shimmer as the day unfurls, your branches heavy laden with tangy citrus pearls. Under the moonlit serenade, you whisper to the night, promising to heal hearts with your zestful might. Lemon tree, oh lemon tree, a symphony in the making, a testament to life under the boughs we’re free. In sun and rain and whispers of the wind, you stand resilient, never do you rescind. Oh lemon tree, in your shadow, life’s essence we see. In you, we find a timeless poetic decree.
On a side note, viewers, I think 11 Labs did a pretty fantastic job at reading that out. Anyways, that’s a fantastic poem we can all agree on. However, come on over to Open Playground. We’re going to use the same exact model, GPT-4, but this time we’re not inside of a chat mode. So what you can kind of do is trick the AI into actually making itself more efficient. You see, we’ll take that same exact poem and we’ll prompt the poem with “The following is a poem written by OpenAI’s previous model.” And we’ll go ahead and paste the poem in. But now, at the end of the text here, we’re going to say, “Here is the same type of poem written by OpenAI’s latest, more advanced model.” So now we’re essentially tricking the AI into rewriting this poem in a more creative, higher fidelity way, even though it’s the same exact model, GPT-4. We’re actually enhancing our result here just by tricking the AI into completing the text. I think it’s a lot faster to enhance your OpenAI results sometimes when using one of these completion-style user interfaces.
And anyways, we’ll click “Submit” here and see if we get a better poem. And as you can see, it spits out words just as fast as ChatGPT would because, of course, to use OpenAI’s GPT-form model, again, this is the same exact model that OpenAI is using right now in their ChatGPT. If you are a ChatGPT Plus user, they actually have to access that OpenAI API. Anyways, we got our poem here. Let’s see if this is a little bit of an upgrade. Of course, this is art stuff, so it’s going to be subjective.
In the heart of the garden, a sentinel stands tall, a bearer of golden treasures, a witness to seasons’ call. Lemon tree, oh lemon tree, beneath the boundless sky, your citrus gems dance gracefully as gentle zephyrs sigh. Petals soft and fragrant paint the air with sweet perfume, enticing bees and butterflies to your nectar’s lively bloom. With morning’s light, you shimmer bright as dewdrops kiss your skin, your branches laden with delights of tangy orbs within. Beneath the silver moonlight, you share your secrets hushed, offering the gift of healing with your zestful vibrant touch. Lemon tree, oh lemon tree, a living ode to the sun, a sanctuary of life where hearts and souls run free. In warmth and chill, in whispered songs the wind sings your praise, you stand unwavering, your spirit never frays. Oh lemon tree, within your shade, the essence of life we see, in you, we find a timeless verse, a living breathing decree.
Dang, this guy really likes lemon trees. Either way, I think that was a little bit more of an emotional poem. Very interesting stuff, right? So yes, viewers, that is one of my favorite parts about using a model in this way. I like that it’s an option here. That we also have regular chat as well. To me, that’s like what I would expect out of ChatGPT Plus. Let’s say extra features for those Pro users. And also, let’s say that you are writing something yourself and you don’t know how to complete or end your sentence. Well, here you go. You could just paste it into this model and do a completion run, and it will complete it.
Anyways, you might notice there’s a bunch of sliders over here that you don’t normally get with ChatGPT. And that’s because we can adjust the maximum length of output that we get with this model. And this is GPT-4 with an 8,000-token limit. We can adjust how many tokens it’s allowed to output. Like, let’s say we only wanted to allow one token to output. I could type in “hi there,” and it’s going to generate just one token’s worth of output.
So, this temperature here, essentially the higher the temperature is, the more random it’s going to be. The lower the temperature, the less random it’s going to be. So, there’s some fine-tunability here that you just don’t get with ChatGPT. We got top P here. I always have left this at one. I would not mess with that setting. But we have frequency penalty and presence penalty. So, if you up this, it reduces the repetitiveness of generated tokens. The higher the value, the stronger a penalty is applied. And presence reduces the repetitiveness of generated tokens, very similar to frequency penalty, except that this penalty is applied equally to all tokens that have already appeared, regardless of their exact frequencies. And we can also do a number of samples here, so we can have it generate three samples all at once versus just one.
This really is like ChatGPT Pro. Now, here’s one of the really crazy parts about this. This blew my mind. This is the only place I’ve ever seen access to this, but we have access to OpenAI models that you normally would not have access to. First of all, these normal GPT-4 models you normally would not have access to unless you were inside of OpenAI’s waitlist API, but here they are inside of Nat.dev. We’ve also got GPT-4 32K model, which I didn’t even realize this was open for anybody, but it’s here and you can use it. This model is extremely expensive, I’m going to warn you guys. However, we could generate up to 32,000 tokens. This thing will accept a ton of text, like a small book’s worth. Maybe it’s crazy, I mean, this thing can read.
So yeah, you get access to an exclusive model here, this 32,000-token model. Essentially, having 32,000 tokens allows you to do things that are just insane. I could paste 10-15 articles’ worth and say, like, bring all of these articles to a singular point. It could be a very diverse topic. I find 15 random articles about it, I paste them all in here, and I say, “Give me the facts, give me the truth. What’s consistent across all of them? What doesn’t make sense?” And it could figure all that out for you. There is a serious power in being able to input more text into these models. The base GPT-4 model apparently goes all the way up to 8,000, which I didn’t even know. But even the regular GPT-3.5 Turbo, that’s only 4,000. So, 32,000 is orders of magnitude larger in its GPT-4. So, it’s a very smart, coherent model that is just phenomenal.
And guess what, viewers? It gets even better because this thing has access to all of these other models as well. So, there’s a ton of them. We’ve got the replicate alpaca 7B, llama 13B. We’ve got the anthropic models in here, we’ve got the cohere models as well. All the good stuff is in here. We’ve got the forefront model. So, there’s a ton of models that you can mess around with in here. We’ve even got these 100K context models that go all the way up to 100,000 tokens. Can you believe that? That’s crazy. You just have access to it right inside of this one website with a very familiar user interface.
You can also just do this all in a chat mode as well, just like ChatGPT. And you also have the settings inside of chat mode. This thing is professional. Alright, if you want to create some amazing prompts, if you want to get some real large language model work done, if you’re not just messing around with it, this is where you want to be. This is better than ChatGPT. It can be more expensive, though. We’ll get into that. That’s the biggest downside about this is the price. It is pay-as-you-go, which I am a fan of. I don’t really like tokens. It’s just pay-as-you-go in dollars. But, you know, they can’t just offer running large language models on a GPU for entirely free. It makes sense.
We’ve also got another one here, though, which is “Compare,” where you can actually compare two models side by side in a drag race. How insane is that? So, what we’ll do for this, I want to test this out. I haven’t used this yet. We’ll try, let’s say, the GPT-4 model, right? This is regular GPT-4. Now, I’ll go in and find alpaca 7B. So, I got these two models in a head-to-head drag race here, and we can just type a single prompt up, and it’s going to compare them across two different models. Makes it really easy to figure out which models are the best. If you got deep pockets, you could have a lot of fun with this. You could actually select all of the models that they have access to, which is like what, 20-something, 30-something models, and just compare them all in a single front. That would be crazy.
Alright, I said, “Write me a riddle about Pokemon.” And since these models are different from each other, we have to go and edit their settings separately. As you can see, if we click between these two models, we can see their settings are different. So, I’m going to go to GPT-4 over here and increase the maximum length, like 400. And yeah, we’ll do the same thing for this one. So now they’re both at around 400. We’ll click the “Submit” button and see what they generate.
As you can see, comparing these two models, there’s a stark difference. Chat GPT is still generating here, when it seems to be a pretty complex little riddle. In alpaca 7B here, it gives us, “What has four leg-eggs but can’t walk?” Answer: a table. Okay, this riddle was supposed to be about Pokemon, alpaca 7B. Anyways, GPT-4’s actual answer was Castform, which I believe must be a Pokemon. I’m not very familiar with it. And, uh, yeah, I mean, this seems to be all of the different attributes of said Pokemon. Interesting. Okay, GPT-4 obviously won that.
So let’s take a look at regular chat in here. This user interface seems to be pretty similar to ChatGPT. We can do new chats on the side here, and apparently save them. It seems like a really useful feature. And actually, yeah, shove this to the side so we can actually save our previous chats. Go ahead and change our model right in the center here. We’ll pick the 32K GPT-4 model. We’ll give it a nice maximum length. We’ll give it some randomness in the temperature too. Frequency and presence penalty, I bump them up a little bit sometimes. I’m going to try to make this thing write an entire episode of SpongeBob, which is something that regular GPT-4 would not be able to do because it needs all that context length. Now we can have our maximum length set to, like, what, 30,000 tokens? Let’s see if it’ll just write an entire episode.
And here we go. “Insufficient balance.” We can’t even complete this because, uh, we don’t have enough money. Like, these guys have to pay OpenAI to be using all these models. Now, uh, this model I’ve selected is a very expensive one. In fact, this is probably the most expensive model out of the entire bunch. But I think it’ll be worth it. Alright, this is just a basic prompt to get this job done. Let’s see if it can do it. “Squidward’s unlucky day.” Let’s see, is it gonna try to generate an entire script again? This is the 32K model. This is something we would not be able to do with regular GPT-4.
It’s still going here. This is pretty big. Alright, we definitely did not get a full-length episode. Sorry if it’s not full-length. It’s challenging to write a long script here due to character restrictions. It actually tells you how much that cost me as well. So, that was about one cent to, uh, to generate that. What’s cool is we can go back through here and edit our text to change up the context of the situation, do anything we want. I’m interested as to why it’s not allowing us to generate more. Let’s try this in just a regular playground and see.
Again, it just generates something really, really short, even though it technically has the ability to do something else. Alright, so here’s the entire transcript of the Texas SpongeBob episode. It’s actually not as long as you might think, but definitely longer than what we got generated by a GPT-4 32K. We’re gonna rewrite it as an episode of Family Guy and just see what happens. So, by all means, this model should be able to do this. Yeah, so it took the whole context of this SpongeBob episode, so it is the actual 32K model. Yeah, it’s generating. Oh my God, I think it’s actually doing a pretty good job here. Oh my God, it’s, like, replacing all the characters, going over to Lois’s place. Peter, don’t you live there? Don’t you live with Lois? So, Patrick is Quagmire, SpongeBob is Peter. So, we still have that issue of it not generating enough. So maybe this is why this model still isn’t released to the public yet because it’s just not done. And, well, it definitely can generate a lot, it prefers not to. I mean, the script that it wrote out here was only, like, 400 words, where the original script was 1,600. It still seems to have done a pretty good job at rewriting that SpongeBob script. And still, the total context of all of this wasn’t even close to the limit of what this GPT-4 32K model should be able to produce.
So, this model, for the moment at least, might not really be great at generating long spouts of text but only good at interpreting huge amounts of text, similar to what I found so far with this GPT-3.5 Turbo 16K model. You can see they also have a ton of different stats on all of these models. Like, you can compare all of the models against each other. We can literally click all of them if we wanted to and see, like, the response times, how often the tokens complete. So, there is the cool metrics side of things as well. But, but really, the way I see this is, like, the fact that you can even save chats just like ChatGPT. The only things you’re really missing out on is the fact that technically, there is the free version of ChatGPT, too. If you have no money, ChatGPT still is your best option. And you don’t have access to plugins and web search. But if you’re not worried about plugins or web search, this literally is just better than ChatGPT, straight up. You have access to more models. You can compare models directly against each other. You have your basic chat with all of these extra options and parameters on the side. And you have your playground version of it as well, which allows you to do things that you normally wouldn’t be able to do. Really, really awesome stuff, right? I don’t know, viewers, for me personally, this is something I’ve been waiting for. The only problem is that you have to pay for your usage. But that’s just like Hugging Face, that’s just like replicate.com. You gotta pay for what you use. GPU time and computation time isn’t completely free. And from what I can see, at least, it’s about the same price as you using the regular OpenAI API, let’s say, directly from OpenAI. They’re not really charging you an extra bag to use the site. It’s pretty fair. For me, though, this is going to be awesome. It’s just a really nice little UI, and it’s like having a professional version of ChatGPT right at your fingertips with more options than just OpenAI’s options and actually more options than OpenAI gives you from their own models. It’s crazy. I don’t know. I think it’s pretty sweet, guys. Let me know what you think down in the comments below. Check out some of my other videos. And, uh, yeah, join the Discord server. It’s linked down in the description below. Tons of great AI stuff on there for you viewers. See you in the next one.
