Explore the forefront of AI innovations with AIAW News' latest update, AI News v50 2023!We kick off by dissecting the phenomenal new open source model, Mistral 7b, and scrutinize its promising implications for future projects like GPT-4. We don't hold back as we voice our concerns over the questionable Gemini AI demo, questioning its authenticity and discussing the ripple effects on the AI community.
Shifting gears, we're setting the stage for a fascinating deep-dive into Tesla Optimus Gen2 release, the realm of humanoid robots and their crucial role in automation. We unravel how amping up chatbots with planning and reasoning could lead to AI's next big leap. We'll discuss the OpenAI's Q-Star project, underlining the vital role of control in molding autonomous systems. We'll also ponder the potential of humanoid robots, in automating tasks in manufacturing units and even nuclear power plants, without the need for any major changes.
Stay connected with the AIAW Podcast for these exciting AI advancements at www.aiawpodcast.com or follow us on X @aiawpodcast.
Follow us on youtube: https://www.youtube.com/@aiawpodcast
Explore the forefront of AI innovations with AIAW News' latest update, AI News v50 2023!We kick off by dissecting the phenomenal new open source model, Mistral 7b, and scrutinize its promising implications for future projects like GPT-4. We don't hold back as we voice our concerns over the questionable Gemini AI demo, questioning its authenticity and discussing the ripple effects on the AI community.
Shifting gears, we're setting the stage for a fascinating deep-dive into Tesla Optimus Gen2 release, the realm of humanoid robots and their crucial role in automation. We unravel how amping up chatbots with planning and reasoning could lead to AI's next big leap. We'll discuss the OpenAI's Q-Star project, underlining the vital role of control in molding autonomous systems. We'll also ponder the potential of humanoid robots, in automating tasks in manufacturing units and even nuclear power plants, without the need for any major changes.
Stay connected with the AIAW Podcast for these exciting AI advancements at www.aiawpodcast.com or follow us on X @aiawpodcast.
Follow us on youtube: https://www.youtube.com/@aiawpodcast
It's time for AI News brought to you by AI AW podcast. I'm gonna, I'm gonna, I'm gonna persevere to take topics that is not LLM or generative AI oriented. I'm gonna push this, but I don't want to start. I want to start with the. You know what? What is the other news? That?
Speaker 4:Mistral yeah, mixed, mixed, mixed Mixed.
Speaker 3:Yes, that's basically what I know, because they released it as a, as a, as a torrent and just put it up on on X, and I think it's about as well as like here here you go, guys, which which is quite interesting and also very different from Yemeni when it comes to kind of marketing. I I for one, I was very I was struck by lighting when I saw the video from from Google on Gemini, but then I understood that they all did a lot of cutting and editing in that one, so let's continue with the mixture.
Speaker 4:I think it's interesting from a technical point of view in many ways. So, mr Alino, they released another, you know, 7b model, 7 billion model.
Speaker 4:Before that was really amazing and could beat the performance of a lot of other, much bigger models as well, and it came from France. European model is awesome and very great to see. Now they released another mixture of expert model, the mixed troll, which is eight models at once, and this is, as, as at least speculated, what also GPT4 is. So GPT4, they say, is also a mixture of expert model that is potentially eight or sixteen different models that have different experts in them. And now finally and I've been waiting for this for some time when will the first mixture of expert open source model come out and mixed? Mr Alino was the first one to do it. So I think this is amazing, and I think this, of course, is the future, to have this kind of not as simplistic as just more transformer blocks, but actually having a set of experts, where each expert focus on different topics and in that way it's much faster to train and, especially, is much faster to do inference on and the energy consumption is, and the whole consumption of how many.
Speaker 1:how you use the whole neural network to have what you need to light up is a very different architecture. Could you elaborate a little bit? I was trying to read up to understand the differences, but it's quite cool.
Speaker 4:I mean basically each token as is go through the network. You know this kind of out regressive it takes one token at a time and when it has one token it takes the next token. But depending on the token it goes to different paths in this huge network of, in this case, not seven billion, but like 40 ish billion, 45 billion I think it was. But it doesn't go through every parameter.
Speaker 4:So this kind of traditional dense models have to take all, use all the parameters for every token. Now it can just take like a 1 eighth of it. It's not really true, because it's not every part of the transformer block that is duplicated is just the feed forward part, which is the one that has the most parameter in it. So that is the most important one, but still a part of each block. As it goes from layer to layer to layer, as all these kind of deep learning networks do, they are routed depending on the token and that is much more efficient to do, more efficient to train and, especially, much more efficient to influence on.
Speaker 1:So the interesting thing is there's not only oh, there's a new open source model from Europe. It's actually it's the first in building expert models like this, like the architecture.
Speaker 4:Yeah, but a mixture of experts has been around since 2014. Actually, I think it was Ilya Citroenkever that is still driving open AI. At first, they wrote a paper on this, I think in 2014 or something, but put it out in the wild, so to speak.
Speaker 4:Yeah, the big open source one. And of course, ilya is now working in open AI and probably GPT-4 is a mixture of experts, since he first published about it, to my knowledge. So it's super cool that we actually have an open source version that is in potentially the same type of architecture.
Speaker 1:So it's not the bigger what it builds 7B. It's 8 times 7B. It's a slightly different architecture. It's cool.
Speaker 4:It's bigger, but still faster, smarter. It's bigger and faster, more efficient.
Speaker 1:Okay, that's one news. Then we almost slipped into the Gemini news and the Gemini news. We talked about Gemini last week, but there is a twist now. It was a little bit faked. What did you hear, david?
Speaker 3:No, it just saw on the news and saw that they actually edited that quite extensively. They haven't shown, as far as I know, how much they edited, but it wasn't that real-time interaction.
Speaker 1:We were blown away by the video. We were blown away by the video it takes Patrick like what is this?
Speaker 3:Now we're on a GI, but it's a bit disappointed, of course, but as far as I understand it, anders, you're the real expert here, probably, but still a better model than Chetrpiti.
Speaker 4:In some sense, we can speak about the MMLU benchmark, which is to be heavily discussed and questioned, which is the one that they beat on potentially. But leave that aside. I feel cheated, I feel annoyed because we spoke last week about.
Speaker 3:Gemini.
Speaker 1:I spoke specifically about.
Speaker 4:They are just drawing an image of a guitar and they have an audio, and then they were writing or drawing other stuff and suddenly they have an electric guitar and suddenly they have drums, and it was faked. It looked like it was real-time. Now it did produce the things, but instead of what the video suggests, which is that it's just drawing and it's in real-time doing stuff, it was not at all in real-time and they produced like text prompts as well and not just drawings. And they had this kind of other demo which was this kind of duck which they drew Exactly dead ones and that looked like it in real-time, as you just added some blue paint on it. But that was also not in real-time and they had textual prompts.
Speaker 1:I'm like why would you do that at this point in time? And Google has done this before.
Speaker 4:If you, remember the old duplex thing, it's a standard.
Speaker 1:We're all going to know about it and it's used to destroy trust.
Speaker 4:I think this. I know one person. If you destroy trust, if you get caught, but they get caught, and they got caught before with the duplex thing.
Speaker 1:They get caught. Steve Jobs, he didn't get caught.
Speaker 4:Yeah, no one you know who likes Apple. I do. We have a personal friend that worked with Gem and I and he's a very good guy that I know and I'm sure he wouldn't say this, but I speak in his words when I say that I think he's very annoyed with the marketing department of Google. I think they would never, as a scientist, in deep mind, produce a demo like this, but then they have some kind of you know, need to be superior to OpenAI and then they have the marketing working with the scientists and developers in a way that I think is questionable, and I don't think internally in Google they are really happy with.
Speaker 1:The benchmarks are impressive and we should, but we let's you know it's very back to okay. Let's see when we get it and see what it's gonna look like in real life.
Speaker 4:It's still super impressive with Gemini. They didn't have to fake it. I think I'm a bit disappointed. I feel a bit treated, at least. What are you about, you?
Speaker 2:Yeah, I think so as well. I mean, they don't have to. No, they don't have to.
Speaker 1:You don't have to Do you have any news?
Speaker 2:Yeah, I just look into why today and as so one of the noticed on Neko, the full bodyscan which Spotify founder came up with as one of the probably maybe one of the new unicorns from Sweden and seems to be very, very promising and popular service where you can go into.
Speaker 2:AI for healthcare, yeah so it's more like you can go in, have a full bodyscan scan, scan with use, make use of AI to prevent being sick basically could be skin diseases or whatever and I think it's an interesting use case in general when it comes to prevention. And I see it in other parts of the world as well, for one year ago actually, when they use been using it for the same purpose, but more for diabetes, and with very successful results.
Speaker 1:And we talked about this before the pod. I mean like the whole angle of prevention. Yeah, and you, we work in the public sector. The compounding effects of value and saving money is probably very hard to even calculate. It's massive.
Speaker 2:Of course, the prevention is the way to go where we can, of course, I mean just when it comes to wrong payments within the government. It's tens of billions each year and hopefully we can reduce that by the new agency. When it comes to payments, it's about like a core, but also using new technologies like AI.
Speaker 1:Interesting where this is going with. Daniel Lake has been quite successful so far.
Speaker 4:He's behind it. I'm sure they have some good tech and motivation, and he certainly has a network to make success, so I think it will be, successful. I actually put myself on my waiting list and I got the time, yeah, but I didn't have the courage to go through with it.
Speaker 1:Why didn't you do that? Because you were unhealthy right. Again some weight, yes, All right, one more news from me, and I'm not going to talk to you in the AI. Instead, I'm going to talk about Tesla Optimus Gen 2. And could we have some clips so we can have a look at it? So let's start with the video.
Speaker 2:It's very smooth.
Speaker 1:So it's the generation Optimus, Gen 1.
Speaker 4:This is the first generation one, so this is not the new one.
Speaker 2:This is March, yeah, but we'll see the fingertips soon.
Speaker 4:So Tesla, you know, is not only building cars, they're building a humanoid. This is what the humanoid looked like. It's not the Terminator, but it is a humanoid robot.
Speaker 1:They have improved the neck movements. One of the key things improved the next generation of the hands as well. Two-dolph actuated neck and you know the first time they had it on stage she couldn't walk right. The you know the the.
Speaker 4:It's all shaky and stuff.
Speaker 1:Shaky, and then you know the. The really remarkable thing is the speed that Tesla has done this. If you take sort of Boston Dynamics, I mean like they use slightly different, I think, technology in terms of AI and Tesla.
Speaker 4:Boston Dynamics does not use machine learning.
Speaker 1:Oh this is what I mean. They use a different technology, but they've been at the game for 10, 15 plus years. And here comes someone who kind of knows data and tech and you know when did they start? 2021? Later than that. I mean like because, because what we're looking at here, picking up an egg, it is really quite cool, right, and they are. They are just getting started and look, wait wait, wait for us to wait for the last bit. Dancers, of course.
Speaker 3:How much editor this is.
Speaker 1:Oh no, no, this was the funny thing this was the funny thing In the press release explicit no editing, no editing. They may have explicit in the press release to make sure this is real, you know. Ok, he loves to bash other people. Yeah, he loves to. That was a direct bash for sure you know how much it costs to buy it. What?
Speaker 4:do you mean?
Speaker 1:So far, but you can buy it. You can buy it, I can. The end is going to come. Yeah, what is the first? What's the list price?
Speaker 4:It's like a car. It's like a car.
Speaker 1:I'm going to buy one too. Man, for what purpose? Yeah, no, I'm just going to have it like hanging on Another beer, ok. Yeah, why not?
Speaker 3:This is cool for conferences when you come.
Speaker 2:You pick up all the best from this guy Perfect.
Speaker 3:But I mean, what's the idea of this? What do you have as a purpose?
Speaker 4:I can speculate, but I think for one it's surprisingly often many. Ok, let me backtrack a bit AI today. For one, what Chatbot is doing is basically just perception. What they want to add is planning and reasoning. This is not really what ChatBT has today, but they want to do it. Potentially, the Q-Star thing that OpenAI is doing is that thing.
Speaker 4:But they need another third thing, which is control. So if you take a self-driving car, that is not general, it's narrow, but still it has perception through the cameras, it has planning, and now with version 12, that's done through deep learning as well. And they have control, and also now in version 12, it isn't released yet it will also be done in a machine learning way. So they have all the three components to do something that is truly autonomous as well and acting in the real world, and that is something that very few AI systems actually do. So what we hope and I think Ellen is thinking with this is that Optimus, when it has the humanoid form which the rest of society is all optimized for, can do so many more things in an automated way, and I think initially it will be just manufacturing, because so many things there is optimized for having a humanoid doing things, but that can then be extended to whatever kind of task you want, and the humanoid form is surprisingly general and useful.
Speaker 1:So the bottom line is to go in and do tasks where you can then install a robot and automate without changing the environment. So much so if you think about, amazon has done an amazing job to robotize and AI, put AI in their warehouses, but they need to build from scratch like a warehouse that is optimized for how to have robots in them. So how can you put humanoid now at places where sort of you don't need to change the environment? You can simply put them where the human was. That is the general idea. And then you come into manufacturing. You come into if I take my old employer Wattenfell, we have many tasks where I sort of cleaning inside the nuclear power plant. So you're doing high risks, high unsafe stuff. So the bottom line with the humanoid is to be able to put a robot without changing the environment.
Speaker 2:I think that's the bottom line use case.