The Rise of Open Source: Is OpenAI's Reign Over?

0:00ChatGPT is obviously an incredibly powerful tool, but that being said, would ChatGPT die if, for example, someone were to release chatGPTGPT that was open source and the idea being that anyone would be able to chat chatGPTGPT die if for example, someone were to chatGPT die if for example, someone were to chatGPT die if for example, we're going to find out because databricks has just released dolly 2.0 which is a text generating ai model that can power any chats like chatbox text summarizers and basic search engines and it's the second version databricks actually released an original version of dolly back in march and what's really really important here is that it's open source and it's licensed to allow independent developers and companies to use this for commercial purposes so you don't have to play pay a royalty and you can use this to make money which other ones have been have come out but it's just been for researchers and for research purposes of ai now why is you know databricks which is a company that really like it makes all of its money from data analytics um why is it open sourcing an ai model so the ceo ali goadzi he says that it's just purely for philanthropy he was quoted as saying we are in favor of more open and transparent large language models in the market in general because we want companies to be able to build train and own AI powered chatbots and other productivity apps using their own proprietary data sets and it's interesting because he said that we might be the first but I hope not to be the last so he's not even hoping you know like they're the the big project he just wants to kind of set a precedent where people are starting to build this and this is really interesting because obviously with open ai receiving 10 billion dollars from microsoft this area is hotly contested a lot of big tech companies are coming in here a lot of money is going this way so it's really interesting to see a company completely open sourcing a project like this so um while it does sound incredibly philanthropic i guess you could say um for them to make this free open source model you also do have to think a little bit about what the upside would happen for Databricks or why, you know, obviously this takes a lot of money to build, fine tune and make this.

2:28So why would they be doing it? When talking to the CEO, he mentioned in an article recently that he hoped that developers who build using Dolly 2.0 and that are building apps with it are building using Databricks. but to his original point it's indeed one of the first chat gpt like models available without major usage restrictions so you don't have to use databricks to build a tool here obviously that's their hope and i think you know there's no way that they're forcing people to do this i just think they're hoping that there's enough goodwill from them creating this powerful model that people will use it so um this is you know a kind of one of their first generation um models that came out well i guess their first generation came out a while ago this is kind of the second generation to that but what's really interesting is their first generation model and actually a lot of these different ai models that have come out originally were trained on outputs from open ai and even google has got swept up in this controversy of um training their model off of outputs from open ai which is a clear violation to opening eyes terms of service so um the first version of dolly that databricks released that's what it did and the second version they have now and of course that's another reason why they have to make it open source and free is because um it's you know well originally it was trained on open ai so it would be illegal for them to monetize that now it is not trained on open ai they said that they um have this new version it's on their own proprietary data.

4:00They said that they created it on a training set with about 15 ,000 records generated by thousands of Databricks employees who voluntarily contributed files to it. And that, you know, 15 ,000 set was used to guide essentially this open source text generating model, which they called GPT-J6B. And they have a non-profit research group called Luther AI, and essentially this is all coming out of that so the CEO of Databricks he does admit that Dolly2 has a few limitations well a lot of limitations this thing's pretty new and one of those is it only does English and you know it can be toxic and offensive in its responses which is another very interesting aspect of this they've trained this model and it's kind of like the out-of-the-box stock model of an AI.

4:57Obviously, OpenAI has done a ton of work to kind of fine-tune the responses that come out of it to be more politically correct or to, you know, not say things that it shouldn't, whatever. And that has gotten a lot of criticism in and of itself because depending on what side of the political spectrum you are, you may or may not like different responses that ChatGPT will give you. That being said, Dolly 2.0 did not receive a lot of those different fine-tunings. It's more of kind of out in the wild and you're getting something a little bit more of a rougher draft. And some people are criticizing that as it's dangerous.

5:35Some people are saying they would just prefer the stock AI model. So it's really interesting, but that is what's happening with it right now. They have a couple examples I've seen online where they asked it some questions. Everyone's just trying to get to the political stuff on there so they asked him about like women in the workforce and it had some like uh false info that it threw in there some fake statistics it threw in there it was overall uh really positive towards women working but the numbers it had were wrong and so anyways people criticize that um it also was asked about donald trump and if he was responsible for january 6th and it just kind of came up with a bunch of random facts about uh how he went to jail and started war with iran and built a wall between America and Mexico, which is not super accurate.

6:25So obviously, you know, it gets some things wrong like OpenAI did. And it's going to be interesting to see if open sourcing this allows this thing to be fine-tuned and to become better or new versions as they come out, if those will allow this to be more accurate. A lot of the same problems OpenAI had early on, but it is really incredibly impressive. and I think people are definitely overlooking how big of a deal this is to have an open source language learning model like this. While it does have its limitations, those can be fine-tuned by companies and other people in the future or future updates.

6:59But while it does have its limitations, any company can grab this out of the box and build their own in-house AI models off their own personal data sets. And this is incredibly powerful for really democratizing AI at this point and not having it all be held in the hands of a few powerful companies, Microsoft and Google in particular. So it is interesting, though, because open sourcing can kind of open a whole can of worms, right? When you have people taking your code, forking it, changing it, and modifying it, in what happens with open source, you don't really get all of the security that you would have in a closed source project.

7:42So, you know, essentially people could introduce dangerous code into it and hackers and other people can use it for malicious activity. So there is that downside and some are saying that that would that is going to be, you know, scaring off different businesses and keeping them away. However, some businesses are using this. The telecom giant First Orion is testing Dolly to let their engineers ask questions about documents stored on Confluence, the collaboration platform for onboarding and planning. And, you know, the CEO of Databricks did say we're freeing Dolly because we believe open sourcing models is the best way forward.

8:21it gives researchers the ability to freely scrutinize the model architecture, helps address potential issues, and democratizes LLMs so that users aren't dependent on costly, proprietary, large-scale models. This is really interesting, right? One aspect is that researchers have the ability to freely scrutinize the model. This is something that OpenAI has been criticized on because it's essentially a black box, and you can understand why, right? They don't want people to go see how they built it because it's a proprietary info and they don't want um you know people to clone them but i mean come on everyone's cloning them now here's this open source everyone's kind of figuring this out so i don't know if that's really a good enough excuse anymore um because people really want to know you know what is pure ai and what is biases that might be introduced by engineers or developers etc etc and you know what are the data sets people have a lot of questions about these and a lot of people just want transparency they want to know what data was used, what safeguards were put in place, what biases may have been added.

9:22And so it's really interesting that we're now getting these open source ones that essentially are going to get around that. So essentially, by also open sourcing it, I would say that Databricks has also attempted to kind of wash its hands of the liabilities that come along with this, right? Because if you open source it, you're like, no, you know, everyone can monitor it, everyone can make changes and adjustments to the code base etc etc and so some people are saying right this is a little bit less appealing for businesses but maybe this is a smart move on databricks department because if they were trying to actually make one in-house that was competing against chat gbt they wouldn't be able to do that but maybe making it open source for a lot of people can contribute and work on it and if here's a cool thing about open sourcing if companies are grabbing this and throwing it into their own code base and it's open source those companies are going to be incentivized to help maintain and improve the software so it's going to be really interesting to see what happens if this gets wide-scale adoption i think that you're going to be able to see this thing is really improves and even the ceo of databricks said you should expect both a continued investment in open source as well as innovations that help create or that help accelerate the applications of llms to key business challenges so databricks is it would appear as though they are continuing to be committed to this kind of project and these types of projects in the future and if this thing can really take off this could be a very powerful challenge to open ai this might be able to take them on in a way that google or others are at the moment struggling to do so it's going to be really interesting to watch this space and see what happens in the future