Reflection 70b Controversy is PROOF our Perspective on LLMs is wrong.

This whole situation has me incredibly frustrated. By and large, here's what's happened recently. Matt Schumer, who is well known in the AI space for his work on small open source large language model projects, is announcing a fine tune of Llama 70B called Reflection 70B and is claiming that it is the world's top open source model. And he claims that it's trained using Reflection tuning, which is a technique developed to enable LLMs to fix their own mistakes. T during inference this thing was open sourced, the weights were available, and I mean right off the bat that's a huge claim. The best open source model, it's also 70B which is significantly smaller of a sized model in comparison to leading cutting edge models like GPT4O or Claude 3.5 Sonnet. So this appears to be a really big deal to the community. Then people download the model and it is severely underperforming. It is not performing as expected whatsoever. Then Matt Schumer announces that there was supposedly sort of an issue with the weights and now new weights have been uploaded so now it should be better. And then people test it and it is a little bit better. And somehow in the midst of all of this, Matt Schumer released an API where you could use the correct version of the model, but then users on Reddit etc figured out that it was apparently a wrapper for Claude 3.5 sonnet with prompting techniques applied. And obviously that's like really concerning because that would insinuate that Matt Schumer lied. And by the way, we still have no answer or conclusion on that whole API situation. But the model is here so we can test and use the model ourselves. And it does appear to be a real 70B sized model that is fine tuned to do this reflection inference. The community still can't figure out if Reflection 70B is actually a really good model or if it's just because you're prompting it correctly or better. Now first of all, context is incredibly important. So let's learn a little bit about the creator of Reflection 70B. Matt Schumer is the CEO of this company, Hyper, Right AI that has produced products that I have covered on this channel in the past. In fact, I kind of know Matt Schumer. I'm well acquainted with him. I've DMed him and I've spoken to him before. But Hyperrite is considered to be a pretty decent AI product at the end of the day, not everyone agrees with some of the decisions that they've made in the past on how they've built up their platform and done pricing etc etc, but overall decent AI platform, pretty well respected. I think from the market's perspective it's successful. In fact they actually released self operating computer which is an open source framework to enable multimodal models to operate a computer based off of vision, which is pretty incredible. And again this was a big deal when it released as open source. I thought this was fantastic for the community and in fact it was fantastic and still is fantastic for the community. And I mean not only this, if you take a Look at his GitHub, he has quite a big decent amount of popular repositories that use large language models. Prompt, engineer, author, researcher, journalist. And I mean this all fits in, right? Because hyperrite AI is all about maximizing large language models to write better. I mean that's kind of the core of the product. So it's not surprising. But he is releasing this as open source repo so other people can build off of him. He has been considered to be a builder in the open source community, a contributor, and maybe he's not a large language model researcher like the people that work for OpenAI or Anthropic, et cetera. So he doesn't work on that level, but he was still a pretty prominent community figure. So when he's announcing that he's releasing this model that's supposed to be benchmark topping, everyone is taking it seriously, including me. I think because he is like a prominent community member and people trust him. This is not adding up the fact that Reflection 70B's official API that was presented is Sonnet 3.5 under the hood. Yet obviously the Reflection model that is now released and we can test and access today is, is a completely separate reflection 70B model. So because all of this was going on and I wanted answers, I made a tweet to try to pry some information hopefully out of the community and see if anyone has anything to say that can bring light to this situation. For this whole thing to be a sham, as Brian here points out, he'd have to be in collusion with the Glaive founder. And if he's a grifter, what the grift exactly? How is he making money off of what he's been doing, as Marco points out from AI and design. And I don't want you guys to see this as me making excuses for him if it turns out that in fact it is all sort of a sham or a scam. I'm merely pointing out that the facts are not adding up here at all. There's more of a story to be told in terms of the whole API situation and we just don't have the information to it. So I can't draw any reasonable conclusions. And I'm not ready to write off Matt Schumer entirely, especially given his past. And like I said, we do have access to real weights, we do have access to a real model that we can actually now use, and I've been using it. So before we dive into what I have learned so far, I have a quick message from Today's sponsor. I know lots of sponsors in videos lately, but you have to understand that this is literally the reason I can do this job full time is sponsors. And today's sponsor is a pretty darn great one. I want to share something that could be a game changer for you if you want to dig a little bit deeper into the weeds of how AI actually works. Enter Brilliant. This is the ultimate learning platform for not only tech enthusiasts, but AI aficionados as well. And you might be thinking, well, I can learn about AI anywhere. But what sets Brilliant apart is its unique approach to learning and mastering topics. And by the way, their list of topics includes machine learning, neural networks and data science. All very crucial pieces of modern AI technology. Imagine learning about AI not just from reading or watching videos, but by actually building and manipulating AI models right inside your web browser. Their courses are designed by AI experts from top tech companies, prestigious universities, and this ensures that the learning you're getting, the knowledge is directly from a good source. One pretty mind blowing course is their neural network series. You start by creating a simple perceptron and gradually build up to complex multi layer networks. By the end you're training your own AI to recognize handwritten digits. But of course, Brilliant isn't just about AI. They offer a wide range of STEM courses, probability statistics, computer science. I mean, if you really want to get deep into the technical side of AI, Brilliant's kind of a no brainer. Brilliant's also great at fitting into your schedule. Their courses are filled with bite sized daily problems to keep you active even when you don't have the time to delve fully into the courses. So if you're ready to dive deep into the technical aspects of AI, visit Brilliant Dot or click the link in the description to get a 30 day free trial. Plus, if you decide you like Brilliant, they have a 20% offer for the annual subscription. Trust me, your future AI expert self will thank you. Huge thanks to Brilliant for sponsoring today's video. Now back to your regularly scheduled content. Well, thank you everyone. Welcome back so today I'm going to be testing the model llama 3.170 B reflection on hyperbolic Labs. They are serving the new weights that he released on hugging face at FP16. So this is not the API that Schumer distributed himself. That was apparently according to Reddit, actually Sonnet 3.5. So this is not Sonnet 3.5. This is the real reflection 70B. And let's take some notes and impressions as we use this. Okay, first of all, we have some basic settings here. We've got max tokens, temperature, which is maybe a little low, but not too low, top P, etc. These are all pretty normal settings. I don't see anything weird there. We've also got a system prompt that was distributed by Schumer himself. And to achieve the correct benchmark results, you need to be using this system prompt according to Schumer. And you're going to notice something as soon as you interact with the model. Literally. Let's just give it something super basic like, hey, what is your favorite color? It starts off with this thinking. It notes that it's a simple question that doesn't require complex reasoning. However, it'll still use straightforward approach to provide a suitable response. Understand the question favorite color. Consider the nature of AI As a machine, I don't have personal feelings. Formulate a response, explain that you don't have a favorite color. And as a machine, yeah, I don't have a favorite color. Seems legit. Seems like it's going through a little bit too much work because, hey, it's a simple question, but it's giving us an acceptable response that makes sense. With a basic question like this, if we go over to Chat GPT and send it the same thing, it's going to give you a very similar answer. But the model's differences really start to come out when you break things down into more complex questions. How many Ls are in this sentence? This is something easy for us to figure out, but for a large language model that uses tokens, these are usually incredibly difficult. Thus is the famous Strawberry R's prompt that I think you guys have seen before, where ChatGPT consistently can't guess how many Rs are in strawberries. I'm gonna make my own sentence though, because, you know, that's been floated around a lot lately. People are worried that that Strawberry question is in the training data and it's gonna it how many L's are in this sentence? Lemons make me really happy and I love them. Let's go Buy some lavish lemons since we are goated humans, folks. 1L, 2, 3L's, 4L's, 5L's 6. 7L's total. Send it off to the large language model. So to solve this problem we need to count the number of Ls. It recognizes that it has to do so in that way. And hey, in this case it actually got it right. Output there are seven Ls, but weirdly, if we go back and look at the point where it actually counts them, it does break the whole sentence down properly as you might do this. But for some reason it thinks there's two Ls in lemons somehow. Really? Which actually does have two Ls has none. Love is correct here. Let's is correct another. Two Ls for lemons at the end there and one correct L for lavish. So somehow it got the correct answer by doing the wrong math. Now let's give this problem to ChatGPT. ChatGPT gives us the wrong answer and does the right math, which is even more mind boggling there. It's like they're reversed from each other. So technically Speaking here, yes, Reflection 7B is better in giving us the right answer, but the math is completely wrong. I mean, what am I supposed to make of this really? Well, if we go back and look at the original post, we can see that Matt Schumer is apparently working on a 405 billion parameter version of this model. So that might maybe get the math correct every time and also give us the correct answer. The thing is, the reflection tuning is clearly doing something. We just can't tell if it's actually super valuable or not or is even more valuable in comparison to just giving a model a good instruction prompt on how to actually just think out while it's inferencing. Reflection 70B is a model that is essentially fine tuned with a system prompt. It is fine tuned to reflect every single time. So it's going to do it no matter what chat. GPT might be able to reflect if we give it a good system prompt. So here's what happens when you actually try that. Here we are with a System prompt in GPT4O on the API. It is designed to essentially mimic exactly what goes on with Reflection 70B. Except it's not fine tuned into the model. It's simply a system prompt. Chain of thought reasoning. We have a thinking section, reflection section. We'll give it that same exact lemon problem. And keep in mind, it is inferred. It is assumed that GPT4 Omni is a larger model than reflection 70b comes up with a plan, breaks down everything, and is able to count the L's correctly, actually doing the right math to get to the equation too, and it gets the correct answer. 7 we reflection tune A70B model, which is smaller assumedly than GPT4O to reflection tune. It gives us, at least in this example that we saw today, the correct answer, but using the wrong math, essentially counting the L's wrong still. And maybe it just got lucky, but it did do the whole thinking and reflection portion. Chat GPT with no system prompt. GPT4O with no system prompt gives us an entirely wrong answer, yet is still able to count up in its counting procedure correctly. And then when we just take a reflection prompt, a system prompt, and apply it 2gpt4o, it's able to do everything correctly. So we get correct checking of all of the letters and the correct answer. And this leads us to the question of what actually is the difference here? I think the community is being a little bit disassociated right now because they don't understand the difference between fine tuning and doing a system prompt. As these models scale up and as they get larger, the better and better they will become at actually adhering and following system prompts because they're literally being scaled up more parameters. AI has been scaling. I mean, that's why GPT4 was better than GPT3. It had way more parameters. Sure, there's some other little nuances in there, but in general it's scaled up. So the only reasonable conclusion that we can draw Here is that reflection 70B may be better or it may not be better. Reflection tuning might be better than just giving it a good system prompt. Maybe the 405 billion parameter model will be better because it has reflection fine tuned into itself, than just giving the larger GPT4O a system prompt that does reflection. What I really think is hilarious about this whole thing is that it's actually a big reveal. And the big reveal here is that large language models actually are incredibly powerful if you can prompt them correctly. People are learning that, you know, maybe, maybe the large language models aren't as bad sometimes as we make them out to be when we're judging them and when we're benchmarking. And maybe our benchmarks are all off anyways, because as it turns out, oh wow, you can make the large language model a million times better for whatever scenario you want it to be in just by, by prompting it correctly, just, just based off of a prompt. Of course, we can do fine tuning as well, like with reflection, but you can also just prompt it. Prompting is where all the power has been all along. And it seems really obvious in hindsight because the large language model doesn't do anything unless you prompt it. And it can't do any. Any of the work that you've been doing with a large language model comes from your prompts. Whether it's able to infer things or not based on your prompt is actually kind of in your court to a degree. I mean, the smarter and larger the model is, the better it's able to just kind of pick things out of your sentence and infer what it needs to do for you. And that's nice because then we can just lazy prompt it. But at the end of the day, prompting is, is everything. And Matt Schumer actually does have something that he said in regards to this. He says that everyone has been sleeping on applying prompting techniques to models natively. Reflection was just a first attempt to show this. And he could be right. He could absolutely be right. Applying it to the model more natively by fine tuning it in there from the beginning instead of doing it after the fact via a system prompt could yield better results. And he also stated that prompting models in the system prompt to behave in the same way as fine tuning them looks like it works, but in reality the model will often make mistakes it wouldn't have otherwise made just to fulfill instructions properly. He also says they had to develop a good training strategy to enable the functionality. And again, this, this could be true. We don't know if this is true or not. I'm not willing to take Matt Schumer's word for it. I am waiting for that 405B model to appear. The reflection tune, 405B. I would think that as the models scale, they get better at understanding system prompts and actually adhering to them. And for example, it becomes more like a fine tune the larger the model is. But we really don't know because we're waiting for this 405B reflection model. Without having two models trained in the same way of different sizes, we can't really compare how the size affects the fine tune. I mean, he's been developing it, so we assume that he would know more. But now apparently he's lied about this whole API situation with Claude. So there's this huge distrust in the community. Like I said, this is aggravating. It's super aggravating to me. But I think it's kind of a wake up call for the AI community. Surrounding LLMs. Now that they're a lot more mainstream, we can start sussing out what actually makes the value inside of the large language model. And that's a good thing because we're going to learn a lot about large language models. I think we're going to learn that fine tunes are good for certain circumstances, but I think we're also going to learn that if you need quick and dirty solution to do a problem that just appeared in your life, you might want a model that is more broad and can then be system prompted to essentially complete your task. Specifically, it might depend on how specific the task is. It's a learning game. It's weird because large language models are like this thing that the human race has built that we have to explore. It's very, very strange. I mean, that's how AI is. That's, that's one of the main characteristics about AI. It's like this invention that we have to, to learn about. Normally when you invent something, it's like, oh, you know, this controller, for example, we know how it's built, we built it to function in a certain way, so it's fine tuned by us to do a certain task. And large language models are the same way. But also it can do a bunch of stuff that we don't know about it because it trained itself. We didn't actually manually train it, it trained itself because it's AI. It's super weird. I actually just saw this reply to my tweet today. I really do agree with this. The whole idea that it was a grift or a scam from the beginning is pretty silly and it obviously doesn't really make a lot of sense. The two options here are he started off thinking he had knocked it out of the park. We know that's absolutely true because he said it's the world's top open source model. Something went wrong and he tried to cover it up. Or two, he started off thinking he had knocked it out of the park. Something went wrong and he hasn't tried to cover it up. And obviously this is just in regards to the whole API situation. So now finally at the end of the video, I'm going to tell you what I think has happened. This is just based on all of the information that I've gathered and my belief on how things went down could be wrong and I hope we find out. First things first, it is obvious Schumer clearly trained a reflection tuned model. This was all done internally. I don't think that he got a whole lot of feedback from outside sources who don't do a lot of prompt engineering or aren't deeply involved with AI. And I think he truly believed and still believes that this is the absolute best way to train a model to solve a lot of current benchmarks. And he could be right. He then released the model with incorrect weights on hugging face. Somehow something got screwed up and when he made his initial announcement, he really hyped it up. So he's freaking out at this point. I think that he applied a prompting technique to Claude 3.5 via an API and then just said, here, use this model. This is reflection 70B. So in that case, yes, he did lie. I think that he probably did it in a panic and he'll probably have to apologize if that is the case. I also think that there's a possibility that someone from his team set that up. Maybe there was some miscommunication down the line and somehow that was set up instead of the correct reflection 70b API. I really don't know. But that's giving him sort of the benefit of the doubt. If it turns out that he was fully aware and he did lie just to kind of like cover it up while he fixes the actual weights. I don't. I think that's not really a smart choice and we'll deal with that as the information comes out. I do think that he really did reflection tune this model and I do think that it's, it's has its value. I think that this model is still valuable. I don't think it's better than current state of the art ones because it's 70B, but I do think it's going to be better at certain tasks. I just don't know how much better it's going to be in comparison to just doing a system prompt. I think that we really need to look at this as a wake up call to just investigate system prompting a lot deeper and start to categorize it and actually run benchmarks using different system prompts to see what they do. And I think we need to develop new benchmarks as well. We're thinking about LLMs all wrong. And that's, that's the moral of the story. We are making definitions of they are and what they can do and they actually are quite outside of those definitions. And AI is like Magic man. It's very powerful stuff. I'm not surprised that we don't really understand it. Thanks so much for watching. I hope that this had some clarity for you maybe on this situation. I hope this kind of helps you understand where things are going. The reason I want to talk about this is not because I like to talk about drama. I really wanted to talk about this because it reveals something about large language models to us, and it's going to be sort of an alignment tool. We can look at the situation as an alignment tool for us as humans on how we should perceive AI and how we can move forward using it. But yeah, huge thanks to all of you guys. If you have any information about this, if you want to contribute in any way, feel free in the comments. We'll see how the rest of it plays out. And I am very excited to try that 405 billion parameter model that is reflection tuned and compare it to the heavyweights of the AI world. See you in the next one.