Open Assistant: Open Source ChatGPT is Here!!! [live Demo]

Over the last few weeks a number of open source models has been released including Alpaca Llama, GPT4, all Vicunia and Kavala. But one way or the other they always have some issues. Either the models weights are not available or the data used for training is derived from ChatGPT. But now there is a truly open source model out there and it's called Open Assistant. The model reads all the training data as well as all the training codes are going to be openly available to the public. Before experimenting with the model, let's see what they're trying to achieve here. It's going to be a chat based system that can understand different tasks, has the ability to interact with third party systems and can retrieve information from using search engines dynamically. So let's look at some of the features that they are trying to implement. So it's supposed to be your conversational assistant is a state of the art chat assistant that can be personalized based on your own needs so you can retrain and fine tune it on your own data sets. It will have access to the Internet so it can use search engines to get the latest information. That's why it doesn't have to be a multi billion parameter model. And their goal is to have an interface to external systems using APIs and third party applications. So it's very similar to plugins for GPT4. They will have integration with all these different systems and services that you can interact with. The ultimate goal is to be like OpenAI but be completely open source. Now the conversational assistant called Open Assistant is open to the public. In this video I am going to test this using these tasks and I must say it's pretty awesome. Before we go and play around with the chatbot, I just want to go over how they trained the system. So they actually used instruct GPT. It's a paper from OpenAI. They used reinforcement learning with human feedback. Their goal was to collect around 50k demonstrations or such interactions. Then they had they decided to fine tune a number of different models. So for example they were looking at GPT, J, CodeGen and some other models. This whole project is funded and sponsored by leon, which is an open source organization that provides data set tools and models to liberate machine learning research. Okay, the first question that probably everybody is interested in what type of license does Open Assistant use? Everything is under Apache 2.0. That means they will be available for a wide range of applications including commercial use. So, so you can take this model and the data set and use it for Commercial purposes. And as I said, everything is open source. So for example, you can go and look at what type of data set they used, what is the data format and how do you contribute to it. So if you're interested in fine tuning your own model based on their architecture, I would encourage everybody to actually go and read their documentation. It's very detailed. Now one of the distinguishing feature this open source model has that they are not using ChatGPT for data collection. So a lot of project that we already saw and I covered in this channel, including Vicuna GPT4, all they were trained by data generated by ChatGPT. But in this case it's against their terms of services to actually use anything generated by ChatGPT. So everything they're training on is human generated, which is actually pretty awesome. Now you can actually go and look at the current stats of how many data points are being used for training the model. So for example, currently English has around 57,000 different demonstration. It has all the support for all other languages. So for example, Spanish is the second largest language there is a pretty good representation. And all this data is collected by volunteers and some other information as well. And then all the stats based on the language are also available here. So it's pretty transparent regarding the data collection. Go to Open Assistant IO, this is the main website. Then you need to simply go to sign in and create an account. Or you can either sign in, sign up with your email or with your Discord account. Once you sign up, you will be brought to a dashboard like this. So here you can actually help improve the model. For example, you can actually pick a task so reply as an assistant. So help Open Assistant improve its responses to conversation with other users. So for example, if you go here for this task, right? Then you will act as an assistant. And based on, let's say this was the user asked, here's the reply from the assistant, right? And then you can say that whether it's helpful or not, right? There are some other tasks. So for example classify Assistant reply. So again a similar thing, user asks something, assistant give a response and then you can actually help it, help tag it, right? So you can say whether it's a spam or not, it's a good reply, it's a bad reply and also the quality of the replies, right? So this way you can be actually part of the this open source movement. And I would actually encourage everybody who has time to help out because this is a great project. They have already made some of their models available on hugging face. So here are Some of the models, I'm going to put a link in the description. You can actually go and download these models. But their main model as well as the data set is going to be publicly available on April 15, 2023. However, they made one of the models available today and you can actually interact with that on their website. And that's exactly what we are going to be doing today. In order to access the chatbot, you need to actually go to the URL and type in chat, right? So when you type in chat, you will be actually taken to this new interface where you can, let's say click on the new chat and then you will have a very similar interface to ChatGPT. Now for this online demo, they are using a fine tuned llama model. That's a 30 billion parameter fine tuned llama model. But in their actual release on April 15, they are saying that there are going to be a whole bunch of different models you can commercially use. Right now you can actually go to this settings and change the parameters of the models as well. Right now there is only one model, but after April 15th there are going to be a whole bunch of different model available. And then they also have their GitHub page where you will be able to download the model as well as the corresponding code and install it locally. So let's play around with the assistant. But for comparison, I'm also using ChatGPT to see how good the responses are. Right? So the first task is write an email to the CEO of OpenAI. Why Chat GPT should be open source. Right? So let's ask ChatGPT as well as the assistant. Okay. I would say it came up with a really good response. So I'm writing In regards to OpenAI's decision not to release the full code for their LLM GPT4 as a fan and supporter of your organization, understand that you are taking the step out of concern about malicious application of this technology, right? But I actually like this part a lot, so. So it says it feels like we, those in interested in AI as a force for good in the society have lost something fundamental, perhaps even symbolic. The promise and hope of OpenAI. So this is actually the part which kind of goes back to the past of Open OpenAI. The promise and hope of OpenAI from its early days was rooted in the idea that creating safe and beneficial general artificial intelligence requires open collaboration between all groups, including researchers, governments, industry, academic and hobbyists and like. And that by being transparent and inclusive in our approach, would help avoid dangerous situations where advanced AI system may become weaponized and controlled exclusively by powerful corporation. So it's actually really a good reminder to OpenAI that it is becoming a closed AI. And we're going to go over ChatGPT's response because we already know it's pretty good, but actually also like the response that it came up with. So for the remainder of the video, I will be just focusing on responses from Open Assistant. All right. I want to see how good it can do arguments. Right. So I asked to write a letter to the Congress in favor of moving to parliamentary form of government rather than presidential form of government. So it's specific to the US but let's see what it comes up with. We live in the tumultuous times when America is faced with unprecedented division due to political polarization. It seems like it knows what's going on. Currently, our country is facing serious economic and geopolitical threats from abroad while also battling domestic. In contrast, our current presidential system leads to two dominant parties who face off every four years. Okay, it's actually put a Lincoln statement. A house divided against itself cannot stand. Actually, let's see if it actually a statement from Lincoln or not. Okay. It seems to be a statement from LinkedIn. That is pretty awesome. So it got the facts right. Now for the rest of it, you can read it, but I think it makes pretty good arguments. It's pretty coherent and I really like the responses that it's giving. Now let's see if it can actually continue the conversation. So I simply asked it what are some of the cause of parliamentary democracy? So some disadvantages include lack of stable leadership, potential collusion between political lead politicians leading towards corruption, voter fatigue due to. No. Okay, let's see, which one will you prefer? Right, let's see what the model thinks. All right. As a language model, I don't have personal opinions or preferences. My. So it seems like from time to time it might remind you that it's a language model and not a person, but that's okay. Okay, one thing. I'm not going to be testing its ability to generate, let's say, jokes or poetry. Right. I'm looking for more practical applications. See, I'm asking it Write a job description for an experienced machine learning engineer. Include all the necessary skills. Let's see if it can actually come up with a good job description. All right. It didn't really do a good job. One thing I haven't noticed, if you're changing the topic, it's better to start another chat. So I'm going to go to chats, create a new One. Right. And then simply ask the question again and let's hope it's going to. Yeah. So now, experienced machine learning engineer, job description. As an experienced machine learning engineer, you will play a key role in designing and developing advanced machine or AI algorithms that can be used to solve complex business problems. You will closely work data scientist, product managers. Okay, it looks pretty good so far. Then it says, okay, so skills and qualifications required. Bachelor's degree, expert knowledge of Python. Okay. And then advanced understanding of machine learning. Then you need to have SQL and NoSQL. This is actually pretty nice. Like it could be a good starting point. And then you can build on top of this. Let's ask you some philosophical questions. So what is the meaning of life? Explain it in five paragraphs. And there is actually a dark mode as well. So if you click here, it will take you to dark mode. I should have enabled this before, but let's see. All right, so the meaning of life is a philosophical question that has been debated by scholars, thinkers, individual for centuries. Well, there may have maybe no answer or universal explanation of it. Okay. All right, so see purpose, relationship growth and personal development, spirituality and religion in making an impact. Now for this I want to look at its ability, actually program and code. That would be very interesting to see. All right, let's ask it. Shouldn't the answer be 42? Let's see if it knows about the Hitchhiker Hiker's Guide to the Galaxy. Now, Douglas Adams made up 42 as the ultimate answer to everything in his series Hitchhiker Guide through the Galaxy, which was written for entertainment purposes. There's no scientific or logical basis. Okay, that's pretty nice. All right, let's see. What is 10 plus 9? Right. The answer should be 21. 11. Oh, interesting. Let's see. No, it's 21. You're right. I apologize for my mistake earlier. The correct answer is 21. No, it's wrong. It should be 19. So. So if I write this number, it will make magic, because if anyone wants to count up 10 and then add nine. Okay, I think it got confused, but this is like a known problem with all the large language models. Even GPT4 gets mathematics wrong. So I won't read the take points off of Open Assistant, but don't rely on it for mathematics. Okay. Before wrapping up, I wanted to check its ability to program. So write a Python function that detects if a number is prime or not. Right. Let's see what it can come up with. So it's writing, writing the code. Okay, so, okay, so Quick test shows and looking at the code, it seems to be working. So here are a couple of other programming examples. So I asked you to write a python function that can write a file to an S3 bucket using the portal library. So it did use the portal library. The code seems to be working. However, I asked it for a python function, but it didn't put it in the function, but it's still like working code. And then I asked it to simply explain this regular expression and it seems to understand this as well. So that's pretty nice. These are like standard programming questions, nothing really serious. All right, let's ask it one final question. Who is the CEO of Twitter? Let's see if I had any idea regarding the current. Elon Musk is the current CEO of Twitter, although it made a spelling mistake. That is interesting. Okay, since October 2022. The date might be actually a bit off, but Anyways, he founded X.com right in 1999, which later became PayPal. That is true. In July he founded and then it goes to actually a whole bunch of details regarding the whole history of Twitter. Okay, let's see who was the previous CEO. Okay, let's ask it. Check to receive return as an intern. I don't know like if it's. It doesn't know about Parag. Alright, so there you have it. It's a quick overview of Open Assistant. It seems to be working pretty nicely. It's going to be available for downloads and local experimentation on April 15, so keep an eye out for that date. Hope you like the video. Thanks for watching. See you in the next one.