I Asked ChatGPT and Its New Open Source Competitor StableLM if AI Should Be Open Source. Here's What They Said.

0:00The episode of the AI Breakdown that you're about to hear first premiered on YouTube on Friday, April 21st. In it, we look at Stability AI's just-announced Stable LM, an open-source competitor to ChatGPT, and we ask both ChatGPT and Stable LM, should AI be open-source? At least one of their answers might surprise you.

0:24Spoiler alert, while ChatGPT did the thing that's so impressive of aggregating a ton of information and spitting it back really fast, Stable LM actually seemed to kind of have an opinion and even referred to itself as a scientist at one point. All right, friends. Well, back to another AI breakdown. Excited to have you here. ChatGPT has changed the world. I think we're all aware of that. If you're watching this channel, chances are good that you agree with that. And just to put some numbers around it, ChatGPT, after it came out in November, became the fastest growing app of all time. 100 million users in two months, less than two months, really.

0:59Five weeks, I think it was. which is four or five times faster than it even took TikTok to get to 100 million users. And that's sort of what has been driving so much explosive energy into this space. Now, of course, there are other AI inventions and innovations that are capturing people's imagination as well. I think that it was actually the combination of chat GPT plus these incredible text to image tools that were accessible to so many people like MidJourney, like Stable Diffusion that really made this AI's moment. So ChatGPT is here. I think we can all agree on that. But as it has become and started to work its way so deeply into our lives, into our careers, many people have started to notice or at least think about the implications of what is ultimately a commercial product being so deeply intertwined in the foundation of our lives and our economies, And one of those folks or one of those companies that has thought that maybe that's not the best idea and that maybe this type of technology, this type of large language model needs to not just be the province of one commercial company, but open to a whole different set of use cases and people in a very different or with a very different economic model is Stability AI, the company behind Stable Diffusion.

2:21Earlier this month, they announced Stable XL, which was their, or SDXL, excuse me, which is their new version of Stable Diffusion, which has even more power behind it for their text-to-image protocol. But they also just announced this week something that has people really excited, and that's Stable LM. So earlier this week on April 19th, Stability AI writes, announcing stable LM. We're releasing our first large language models starting with 3 billion and 7 billion parameter models with 15 to 65 billion parameter models to follow. Our LLMs are released under Creative Commons BYSA license. We're also using RLHF tuned models for research use.

3:02For those just getting up on the lingo, RLHF means reinforcement learning from human feedback. Now, lots of people took notice of this news. Paul Baum writes, Stable LM is tiny in terms of parameters, 7 billion versus 175 billion in GPT-3, but very competitive in terms of data, 1.5 trillion tokens. More parameters help, but far less so than data. 1.5 trillion is half of all humanity's public knowledge already. We're running out of data. So I wanted to give a chance for people to understand these terms that keep coming up a little bit more. So I flipped over to ChatGBT and asked, what's a parameter in AI?

3:40Just to explain it better for you guys here. The way that it explains it is in AI, a parameter typically refers to a value or set of values that are used to control or adjust the behavior of a machine learning model. These parameters are often learned by the model during training, and they can be used to make predictions or classifications based on new input data. For example, in a neural network model, the parameters might include the weights and biases that are used to compute the output of each neuron. These values are adjusted during training in order to minimize the error between the predicted output of the model and the actual output.

4:10In general, the choice and adjustment of parameters is a critical part of building an effective AI model. Choosing the right parameters can significantly improve the accuracy and performance of the model, while poorly chosen parameters can lead to inaccurate or unreliable results. All right, so that's parameters. Now, what is a token in AI? ChatGPT writes, In AI, a token refers to a sequence of characters or symbols that is treated as a single unit of input for a machine learning model. Tokens are commonly used in natural language processing tasks such as text classification, sentiment analysis, and machine translation.

4:43In NLP, a token can be a word, a punctuation mark, or any other meaningful unit of text. Tokenization is the process of breaking up a text into individual tokens, which can then be analyzed or processed by an AI model. Now finally, I asked how many parameters have each of the GPT models had, just to see where stable LM fits relative to the GPTs. GPT-1 was 117 billion parameters. GPT-2 was 1.5 billion to 1.7 billion parameters. GPT-3 was 175 billion parameters. And it's not on here because it was after this data was trained, but GPT-4 was over a trillion parameters. So obviously, from the standpoint of the number of parameters, StableLM is much smaller than GPT-4, but that's not necessarily all there is to it.

5:29Imad, the CEO of Stability AI, was doing an AMA and someone asked, do you agree with Sam Altman's view that larger models are not the way forward? Are you exploring any alternative architectures now? Imad says, I have always said that. We fund a range of alternative architectures beyond transformers. Now, in terms of going through our beloved threaders and seeing who had the best write-up of this, I think my choice goes to Sully this time, who wrote, Stability just released their new LLM. It's open source, has 7 billion parameters, and is entirely free to use commercially. And it's a massive deal that has the potential to change up everything in AI.

6:04Here's why. Okay, so here's the problem. Right now, there is only one company in the world that has an LLM, which is easy to use, accurate, works extremely well, is easily available, and is cheap. And that's OpenAI. Basically, any company using AI generation is running off of OpenAI. This creates a multitude of problems. Namely, one, you have one provider only. If they cut you out, you're screwed. Two, no control of the data that goes to OpenAI. Three, no AI moat, meaning someone can use their API and achieve similar results to yours. One thought is, oh, with enough money and talent, you can make your own LLM.

6:39Wrong. There's only a handful of companies in the world that have the talent and GPUs needed to train a model like that. Example, Google, Facebook, OpenAI, Microsoft, Cohere, Stability, Anthropic, DeepMind, etc. So what do you do? Clearly training your own model is out of the question. Well, you can fine-tune an existing one. No need for massive capital, researchers, PhDs, GPUs, etc. ChatGPT was actually highly fine-tuned for chat, and it made the 175 billion parameter model affordable to use. So how easy is it? Well, hard, but not remotely as hard as training a full-blown LLM. And we actually saw this, when a few Stanford researchers turned Lama to create Alpaca.

7:16I know, lol. They managed to create ChatGPT for less than$600. So WTF is Lama, and why isn't everyone making Alpacas? Well, Lama is Meta's own 65 billion parameter LLM that's open source. Anyone can fine-tune it and use it for free. Folks have managed to get a tuned version to run on their phone. But there's a caveat. It's not available for commercial use. That means you can't use it for anything related to your business, or you'll have a horde of meta-lawyers at your door. And this is why Stability's LLM is such a massive deal. It's open source with full commercial rights. Now a small team of engineers can fine-tune the Stability LLM model based on their needs.

7:54Right now it's in alpha, with 3 billion and 7 billion parameters both trained on 800 billion data tokens. The larger 15 to 65 billion ones are coming in the future. This means no vendor lock, complete control of user data, no commercial usage issues, fine-tuning for specific use cases, opens up a lot of things for companies wanting to build AI products and businesses. No need to rely on OpenAI for their AI needs. Overall, the implications are massive, and it's actually a lot bigger of a deal than most people think. It should be interesting to see how this plays out. Now, I think this does an awesome job of both touching on the commercial and innovation questions of having something like OpenAI and ChatGPT be owned by a single company, but also the general larger questions of ethics and just what society we want to be and how controlled we want these sort of technologies to be in the hands of a few different companies.

8:48However, I thought it would be interesting, given that we have these tools that are meant to help us make decisions by processing a huge amount of information, to actually just ask ChatGPT what it thought about open source versus closed models in AI. And then, hey, as a way to test this new stable LM, which is available to test on Hugging Face right now, why don't we ask it the same questions and see both how it compares in terms of its performance, but also how it might quote unquote think differently. First up, ChatGPT. I started with a really simple question. Should AI models be open source?

9:25ChatGPT responds, as an AI language model, I can provide you with different perspectives on whether AI models should be open source or not. The question of whether AI models should be open source is a complex issue with both pros and cons. Ultimately, the decision depends on various factors, such as the specific use cases, ethical considerations, and potential risks. Now, a quick pause here. For anyone who spent any time with ChatGPT, you definitely get this type of sense over and over again, that it will be aggressive about not giving you an opinion or an answer, but instead just giving you a set of facts by which you can make your own decision.

10:01So moving on, it says pros of open source AI models, collaboration and innovation, access and education, transparency and trust, and customization. So open source can lead to faster innovations. It can lead to a wider range of people being able to learn and use AI. It can lead to users actually trusting the models more because they can see what went into them. And it allows for people to create more customized or niche uses. Now, what about the cons? Well, the cons include misuse in malicious applications, intellectual property and competitiveness, quality control and fragmentation, and security and privacy.

10:36So open source models might be misused by malicious actors for purposes like disinformation, deepfakes, and cyber attacks. It might discourage innovation by companies because it's harder to monetize. It could lead to a fragmentation in the industry where there's lots and lots of different models running around, making it hard to build interoperable applications. and having more people have access to these tools might mean that there are security or privacy concerns. ChatGPT says,

11:14Now, obviously, I wanted a little bit more juice here. I wanted some opinion. And I said, how would you, as an AI, weigh the pros versus the cons? ChatGPT says, As an AI, I don't have personal preferences or emotions, but I can provide an objective analysis of the pros and cons. Here are some factors that can help guide the decision. Purpose and impact. Consider the primary purpose of the AI model and its potential impact on society. If the model's benefits, such as educational opportunities, collaboration, and innovation outweigh the risks, it might be worth considering open sourcing. Then it goes through security and privacy, control and governance, intellectual property and competitiveness, regulatory and ethical considerations.

11:53It sums up, weighing the pros and cons of open sourcing AI models is a complex task that involves multiple considerations. It's essential to strike a balance between openness, innovation, and responsible AI development. Now, I tried to work around its default setting to not give an opinion by saying, Okay, now I'd like you to imagine that you're an entrepreneur who's building a company to train a new large language model designed for general tasks. Would you choose open source or closed development? It responds, as an entrepreneur building a company focused on training a new large model, I would carefully weigh the pros and cons of open source and closed development.

12:26My decision would be based on, and you see it starts to go through the same things that it had said before, just reframed in the context of it being the decider. So open source pros, collaboration and innovation, market positioning, accelerated development, enhanced trust. Now, I should say that it does customize those answers a little bit for the context of it being the entrepreneur who's building this and making that decision. But they're still pretty close to what we saw before. Cons, likewise, intellectual property and competitiveness, loss of control, and it also identifies revenue generation as a challenging issue.

12:59Now, when it comes to closed development, the pros include the protection of intellectual property, the ability to control distribution, and maybe an easier time generating revenue. But the cons are that it limits collaboration and innovation, and it also might reduce trust and transparency. Given these factors, it says, my decision would depend on the company's goals and the specific use cases for the large language model. Now, I tried to hone it in on the question of risk. I said, in this case, do you think being open source would make the model more or less risky? It once again said basically it depends and identified increased risks, misuse and malicious applications, intellectual property and competitiveness, loss of control and fragmentation.

13:35But then it reframed exactly the same things it had said before in terms of decreased risks, collaboration and innovation, transparency and trust, community engagement and support. It then gave another sort of packaged up answer, balancing these risks and opportunities is essential, blah, blah, blah, blah, blah. and tried one more time. If you had to make a decision, again, imagining yourself as someone creating a GPT-4 style LLM, would you choose open source or closed? At this point, it's just repeating itself. As someone creating a GPT-4 style large language model, the decision between open source and closed development would depend on several factors.

14:06And at this point I stopped it. I finally asked, I'm asking you personally though. It says as an AI language model, I don't have personal preferences, emotions, or subjective opinions. All right. Now this is all pretty standard, kind of what you would guess, this is what ChadCPT is designed to do. It's designed to help you frame conversations so that you can make decisions yourself. That's why I think it's hit so well, is that it is kind of a natural thing. It makes us feel empowered to make better decisions for ourselves, rather than being some weird magical mystery box that just produces decisions for us.

14:37Now, I wanted to share all of this because it stands in pretty stark contrast to the way that the conversation with StableLM went. And that's not just because there was clearly less information that StableLM was drawing upon, although that's part of it. You'll see when I read this that it really is an entirely different type of engagement. So I write, and again, this is just the demo on Hugging Face. Should AI models be open source? It responds, what does open source mean? I said, can you look up the definition of open source software? And it says, okay, the OSI model defines open source as colon, and then it never finished.

15:18So I said, as what? Stable LM responds, a software system that can freely and openly be modified by a user without restrictions or limitations of any kind. Software that can be freely distributed and modified. I said, correct. So open source AI are software systems that have transparency rather than opacity about things like training models and the underlying data. Stable LM responds, that's what I figured. And transparency is generally seen as good for society. There are many examples of AI models that were originally used for good, like GPT-3. But people can also use these tools to train themselves to take over our jobs or do harm, like creating fake videos of themselves saying they are an expert in a field, or getting celebrities to buy themselves out of political trouble.

15:57So in a way, these open source models are both a good thing and a bad thing. They're helpful for society, but they could also be used for harmful things. Already, we have such a different type of engagement. I said, I agree with your articulation of the particular risks. What do you think the benefits of having open source AI models are? Stable LM responds, that's a really good question. And it's something that a lot of people struggle with. In a nutshell, I think the benefits are threefold. And then again, it got hung up. But you'll notice when I used words like, what do you think? It didn't instinctively respond, or maybe better said, it wasn't trained to respond.

16:32I don't have thoughts, as chat GPT is. Anyways, it got hung up after it said, I think the benefits are threefold. So I prompted it again, threefold. Yes, says StapleLM. There are benefits to open source models in terms of transparency because they are more likely to be audited. Open source models are more likely to be subject to peer review, which is very difficult, because the results of the model are up for debate. Open source models are also less likely to be disrespected by people who don't understand the underlying technology, because people who don't understand the underlying technology might be working on it.

16:59This is kind of a fear of open source AI, but also of highly automated systems, and it's a hard one to fight. So as a scientist, I think the benefits are actually fairly strong. You can see in this response that it feels much more opinionated and like it's allowed to have more of an opinion. It even identifies itself as a scientist, which isn't something that I plugged into it. Responded, given that you've identified the benefits of open source AI and the risks of open AI, do you think the benefits outweigh the risks or the risks outweigh the benefits? Stable LM says, I guess it's a really hard question to answer.

17:34It's really dependent on what you mean by benefits. But I think that a lot of people who work on open source AI would say it's a good thing. I think the benefits of having open source models are pretty clear. One, transparency. Two, open source models can be audited by peers, which is very difficult for closed models. Three, open source models can be modified without restriction by anyone, which is just the sort of thing that scientists are used to studying. So there's a lot of scientific work behind this question. Four, it can also make it easier for people to contribute to an open source project because anyone can review and edit the code.

18:06But I think that the benefits are probably much greater than the risks, and that it's a good thing overall. So summing up just this little exchange where ChatGPT would not commit in any way to an actual answer or an opinion, StableLM does. It says here that the benefits are probably much greater than the risks, and that open source AI is a good thing overall. Now, how much is this just the opinions of the people who created this and trained this model, which is an open source model, reflected in it, versus some emergent property of how it's interacting with data? I'm not really sure, but it's fascinating just how different it is than ChatGPT.

18:41And I think the takeaway is outside this specific question of open source AI or not, it's a reminder of just how different each of these systems is going to be. and what a wide band of possibilities they're going to create and open up simply by virtue of how they're trained and how they evolve. I don't think you can really say better or worse between ChatGPT and StableLM because it's just so different. Clearly, ChatGPT is drawing upon a much wider repository of parameters, and it's doing the thing that it does really well, really well. It's giving you the college essay version of what might go into an answer about open source versus closed AI.

19:18Stable LM, on the other hand, seems to be here at least grappling with the actual question that I was trying to get at. And maybe it's just because it's willing to actually answer on the terms that I gave it, but it makes it feel kind of a lot more human. Anyways, guys, hope you've enjoyed this exploration. I certainly found it fascinating. I'm very excited to see that Stability AI is doing this with Stable LM. I think personally, as you can probably tell, that it's super important to have open source models on top of all of these closed models. So we will have to see how this industry continues to evolve.

19:50But for now, I appreciate you watching. And until next time, peace.