Meta STRIKES AGAIN! New AI DEVICE, Microsofts NEW Model PHI-3, Adobe Firefly 3 STUNS! And More

So, with another crazy day in artificial intelligence, there were some key stories that many people did miss that you're going to want to see. So without further ado, let's take a look at some of the most pressing stories. So Meta's smart glasses, the Ray Bans actually now have AI in them. If you aren't familiar with what these glasses are, they're basically like what the Google lens was supposed to be. And if you don't know what that is, it's basically glasses that have a camera embedded in them that can take high quality pictures and you can do a whole loads of of cool stuff with it. Now, I think this was a very obvious step for Meta, considering the fact that they just released their new AI tool. And I'm truly excited for this update because I think this actually goes to show what is going to be coming in the future with AI device. Now, I know a lot of people have been skeptical to try out these glasses and I have actually tried them out myself. And I have to say they don't disappoint. I mean, when we look at generative AI and just a personal assistant, we, we actually wonder what the form factor is going to be. And one of the key things I've noticed with the Metaglasses is that they aren't actually weird, they aren't actually awkward, they fit pretty well. And as someone who's seen them on other people as well, they don't look like something that's out of place. And the best thing about them is that you don't actually have to get used to wearing a new and different product. This is something that people already use. Of course, some people do use sunglasses more in hot countries, but you can actually get prescription lenses fitted. So these can be your everyday glasses. Now, they didn't have the AI update before today, so if you do have a pair, you can actually access this update. But it is important to note that currently this is only in early preview. So if you're like me and you do have the glasses and you're in a country like the uk, they haven't actually rolled it out worldwide yet. But I'm guessing that when it is rolled out completely, this is going to truly change the game because like I said before, the glasses were already pretty, pretty decent. But this is arguably going to be one of the easiest devices for the average person to use in terms of actually getting to grips with an AI system that's non invasive. And all that means is that this is just naturally going to be there. So you'll Basically just have glasses that potentially in the future will always have an AI in them. And maybe a camera is just going to be that something that people get so used to that it's something that's now expected. Now I think this will really, really take off in the next two to three years when we do get AI systems that have much lower latency and the quality completely increases. For example, when maybe some influencers start to take video calls or they start to record certain pieces of content, I think that is when the spiral up is going to truly be there. Because as you know, social media, virality trends, all it takes is one big trend and then we have thousands and thousands of people and even millions and millions of people getting in on this new AI trend. And I think the only thing that's currently stopping widespread adoption of this is of course the latency between talking to an AI system and actually getting that response. And I do think, like I said before, within the next three years that's not going to be a problem at all. And once that isn't a problem, I'm pretty sure that lot of people are going to be utilizing this. And it does show an interesting trend because maybe this could be the future of AI form factor and other companies like OpenAI and Humane might actually switch their form factor to compete with what's Meta doing. And I definitely do have to say that putting glasses with a camera and all sorts of technology that they do have inside is definitely a pretty hard thing to do. Now, when actually talking about AI devices, something that you should all know is, is that Rabbit actually recently had their pickup party in New York and this is the first live unboxing of the Rabbit R1 device. Now, why am I even talking about this? Well, number one, it's quite similar to the Meta AI glasses in the sense that it's an agentic AI platform. And the fact is that right here we have a monumental moment because it's a landmark event where we actually have an AI device that many people are actually excited to use. Now I, I think this is going to be truly interesting over the next couple of days because that's when this thing is going to start getting into the hands of different tech reviewers. And I think that's when we can start to see the wider reviews of this, especially compared to Humane's recent AI pin, which was honestly just a blunder in the tech industry because I don't think they listened to initial reviews of what people were stating about the device. And then of course, it led up to other larger Creators stating that they just think that this device has a lot of potential, but at the current price point, it just simply isn't worth it. Now, the reason I'm actually talking about this device again is because the live demo actually showcases what this can really do. And because it's live, you can't really fake it, and you can see all the flaws of this software. Now, I'm going to play a couple of minutes of this because honestly, like, genuinely, I was actually surprised at what this device could do, because we all know that when there's a product demo from even an AI system to a simple AI pin or device or whatever it may be, we're always usually skeptical because we think that they just want to sell us their product, and then by the time we order it, it's not going to live up to the hype. But honestly, a live demo where he's showcasing all of the abilities of this system do honestly surprise me in the fact that it does live up to the hype in terms of what it's acclaimed to be.

Check this out. Check this out. So now the vision camera, or the camera works with analog scroll wheels too. So I can just literally do that. It flip back and forth. Try one more time. It flips. Okay, so let's. Let's just do that. What's in front of you? One. I'm counting. That's sub two seconds after. Let me see. Okay, that's pretty fast. Or should I say this is the fastest? All there, guys. All right, so I post all my demos on Discord all the time. You guys know that, right? And thank you. It's been like. It's been like, I think over a hundred demos I posted during the last 104 days, but I saved one for you guys for today. Okay, Just for today. This is a simple spreadsheet that my colleagues hand drawn.

And.

And. Come on, I can do better than that. You guys know that, right? Again, I want to show you something much more interesting. Okay, so let's just point right into this. So let's do something like, for this spreadsheet, transcribe it and swap the color and number column.

Taking a look now.

You don't need to point it, by the way.

Okay, Just check your email for the attachment.

Wait, what? There's an email?

Come on, get on.

Guys, check your. Check your watch.

Check your watch.

Okay, I want to make sure this is not a spam, because it's called Rabbit. We're gonna spam a lot. Okay, I'm gonna put this down for a bit it looks like it says here's the revised spreadsheet you asked for and it came with a little nice attachment in a CSV file. Let me just down. Guys, R1 is real. Yeah.

So I mean this isn't a paid promotion or anything. I think that this actually goes ahead and shows us where we're headed in terms of agentic AI. And this is just an on device AI that is simply just taking the inputs of what you're looking at and then it's doing something really quickly. And honestly, doing that email like within, I think it was around six to five seconds. That is pretty, pretty quickly. Like I actually can see multiple use cases for that. So I got to be honest, they've done something pretty cool here. Now what this also gets me excited for is what probably OpenAI is building, because as you all know, OpenAI is literally like a year, two years ahead of where any other company is. So it seems that maybe the industry in some areas are quite further along with some companies than others. So I'm pretty excited for this and I'm honestly wondering what the reviews are going to be like in the coming days since to see if they're as positive as this live one is. Next in AI news, we actually had OpenAI finally release some research surrounding the security and safety of AI models and this one isn't as exciting, but it still is pretty interesting because we do have a scenario where large language models and these AI systems are consistently being jailbroken through prompt injections and other attacks that allow people to simply bypass the restrictions. And essentially the paper is called instruction hierarchy training LLMs to prioritize privileged Instructions and essentially explores the issue where as long as you have an input to the AI system, it's eventually going to read that and use that. Now of course the problem is, is that the LLM systems actually treat all types of input. For example, the system messages from developers, the user messages and third party content with equal priority, which make them susceptible to malicious prompts. And to counter this, the authors actually propose an instruction hierarchy for LLMs, which are essentially a framework where system messages have the highest priority, followed by user messages and then the third party content. So this hierarchy is designed to guide LLMs when conflicting instructions are present, prioritizing higher level directives or disregarding or refusing to comply with lower priority potentially harmful instructions. And this paper introduces a method for automated data Generation to train LLMs on this hierarchical instruction following behavior by simulating different types of attacks and training the models to respond appropriately, ignoring lower priority malicious instructions the approach aims to significantly enhance the robustness of LLMs without actually sacrificing their general capabilities. And evaluation results actually suggest that the models trained with this method are more robust against various types of unseen attacks, indicating improved safety and reliability in real world applications. But I'm going to say what everyone's thinking. OpenAI, can you please finally release your updated model because I'm sure the entire industry is waiting for whatever it is that you have up your sleeves. Now Adobe actually finally released their Firefly 3 model and if you don't know, Adobe have a model that is quite similar to midjourney and other image generation capabilities. But the problem was with that model is that it wasn't that good, so many people didn't use it. And considering the fact that it's baked into the Creative suite of tools, this was something that definitely needed an improvement. Now finally they've updated this model to version three, which actually looks very, very effective and you can create higher quality images which photorealistic details, improved mood, improved lighting. You can also expand images and people have been doing some very, very interesting demos with this stuff. There's a Twitter thread here that explains exactly what's going on. You can see a wildlife photography shot which looks stunningly realistic. I would argue that I still don't believe that this is AI generated even though it completely is. And then of course compared to Mid journey, these still all look real to me. I know I'm living in a delusional land where I think those are still taken by a photographer, but I guess my brain can't handle the fact that this is AI generated. Now of course you can see they also released this video where you can see all of the small updates that they've made to the platform. This update I think is rather important because I use this from time to time to explore different creative ideas for many different purposes. And I think having a model that can actually generate stuff that does really work is rather effective because I think like I've said before, whilst yes, generative AI is good when you can build something that's rather effective, but one thing that many people are missing, and you're probably missing yourself, is that when they build these AI tools, it's only going to be used if it's actually available to the public in a format that they are used to. So for example, with AI generated images, now that it's in Adobe Firefly, we're definitely going to see rapid adoption of this because it's in a form factor that people and creatives are very Very used to and it's easier to use. And like I said before, even when they release text tools that beat things on the benchmarks, a plus 2 or 3 in the MMLU isn't that crazy if people can't access your model in a nice clean user interface that is very easy to use? And this is why I'm saying that we're about to start to see over the next two years. I think a more widespread adoption of these technologies is as they get baked into more familiar tools that many people use on a day to day basis. And if you wanted the comparisons between Firefly V3 and Midjourney V6, you can see the comparisons on these different models. And I think Firefly V3 in these prompts right here, you can definitely see the improvements. Although yes, Midjourney does take the cake in terms of photorealism, I think that Firefly 3 is definitely a big improvement on what we've seen before now. Now, I'm not including this to be an AI Dumo or someone like that, but I actually do believe that contrasting opinions are always important because they help the conversation progress. Because if everyone thinks that AI is this magical tool that has no downsides, then I don't think anyone's thinking at all. And I think whilst yes, the majority of the points in this video I disagree with, I think it's still important to have that as part of the conversation. And essentially, if you don't know who Gary Mark is, he's a notorious, I would say, AI critic that tries to point out all the flaws with LLMs and in some cases he does keep people grounded. But a lot of the time I would say that it is a bit nitpicky to just look at only the flaws of these LLMs and completely disregard the breathtaking capabilities that GPT4 and other state of the art models have given us. So just take a look at this video and then I'm just gonna talk even further about this.

You have some people that I think are more gullible that want to believe for whatever reason that this AI is great and they notice the successes and they don't notice the failures. Well, a scientist says you need a numerator and a denominator here. You need to know what percentage it's correct. You also need to know when it's wrong, how bad are those errors going to be. And so the reality is like everybody got super excited last year, but we are running out of improvements, at least for a little while. Things are slowing down and if you look at Most businesses, they all tried it out last year. Everybody subscribed to the hype. But most people are like, well, we're doing a pilot experiment right now. We're seeing if it works. Maybe we're running three pilot experiments, but you see very few people saying, yes, we use this every day and it's absolutely essential for what we do. The only people like that really are programmers. And there's some evidence that it's actually making the quality of their code worse so you can write faster, but sometimes there are errors and they're hard to debug and so forth. So we're still, I think, a ways from having the practical AI that we're now dreaming of. Like, last year was amazing. It brought AI to the public. People realized, hey, this might really help me. But at the same time I think it was premature and that right now the tools that we have are limited in what they can actually do. And we're kind of like fantasizing about some future world, but the tools we don't have, the tools we have right now are limited.

And I think part of this is correct in the sense that, like, yes, the tools we have right now are limited, but I think that's all going to change in the next 12 months. So I would say just put a pin in this idea for now. For those of you who are, you know, wanting to contend with this opinion, because as we've seen before, AI development is one of the fastest moving fields. And if it's anything like we've seen before, and as we've taken a look at many different statements, you know, regarding GPT4, GPT5 Frontier models and what people are working on, I think it's probably going to be an even more shocking 12 months in terms of the AI acceleration, because we now know that there are three frontier labs just all going at it in terms of trying to maintain the front spot.

This is what a lot of the research that's focused on synthetic data is focused on, right? So if you, if you, if you don't do this, well, you don't get much more than you started with, but it actually is possible by injecting very small amounts of new information to get more than you started with. If we go back to systems of eight years ago, so if you remember AlphaGo, which was a system that was used to play Go, note that the model there just trains against itself with nothing other than the rules of go to adjudicate and those little rules of go, that little additional piece of information is enough to take the Model from no ability at all to smarter than the best human at go. And so if you do it right with just a little bit of additional information, I think it may be possible to get an infinite data generation engine. This is what a lot of the.

Research and I think that that concept is particularly fascinating because being able to get an infinite data generation engine could definitely lead to some kind of, you know, self improvement feedback loop where we get an AI that generates synthetic data, we feed that data back in as long as it's verified and checked. And then of course, you know, there are training runs and stuff like that that do take a lot of time, but you know, that would just be, I guess you could say exacerbated by the training run times. But of course over time that time would continue to collapse as the GPU efficiency increases and as we get bigger and bigger data centers that could handle more compute. So I think this interview is actually insightful because it does cover quite a lot of stuff. He talks about the next generation of AI models and where we're really going to go. And from actually paying attention to Dario Amodei in his interviews, I think it's clear that there's still a decent amount more that we can squeeze out of these models in terms of capabilities. I mean, we still haven't cracked personalization, we still haven't cracked long term memory, we still haven't cracked AI agents in the sense that they're able to go out and perform a task very, very effectively. I know we recently just saw Rabbit do that, but I, I still think that there's a whole new area where once it's truly explored, I think it's going to be the next level of AI that people are truly surprised by.

Models is completely crazy. My good friend Yann Lecant thinks it's the right thing to do. He thinks we're all going to be fine, we'll keep in control of these things. They won't have any goals of their own, they won't have any desires of their own. I think it's very, very dangerous to open source them. I think it's like open source and nuclear weapons happens. So I think I would love governments to forbid companies to open source big models.

So the clip that you just watched there was of Jeffrey Hinton talking about Yann Lecun and open source AI. And I gotta say, this clip and open source AI has received so much backlash. Now I agree with this clip for several reasons, but I'm gonna explain what most people just, I think are missing. The point here, and the point here is that he's not stating that an open source ecosystem is bad. That's not what he's stating at all. And I've seen a lot of the replies to this tweet stating that, you know, he's a D cell, he's, you know, awful, you know, he should do this or whatever. And it's just like all he's saying is that when AI systems get to the point where they could literally destroy the entire planet and companies have that, okay, they've built a model that could, you know, do a catastrophic effect. It could teach an average person to how to, you know, from scratch, you know, go to Walmart, go to wherever, build a biological agent or do something that could cause the most maximum amount of damage possible. Okay? He's just saying that when a system gets to that capable level, or when it's autonomous or where it's able to, you know, acquire resources in the physical realm, he's saying when it gets to that level, you just probably shouldn't open source the AI. Now I think that makes sense because if that is open source, guys, I have no idea what kind of world we're going to be living in because trust me, whilst, yes, society seems relatively safe now and it's been getting more safer as time goes on, and I know that doesn't seem like it because the news puts on everything awful going on in the world, that's how they get clicks. But the point here is that we're living in the safest time in history and I think people forget that. Like if we have a completely open source model that is capable of doing absolutely anything, it only takes a few hundred people to completely upend society. And I know you might be thinking, okay, we'll have AIs defending that. I think at the level where AIs could completely cause a catastrophe, I just think that that's completely irresponsible. Now if they're going to open source something that is on the level of GPT4 or something that's on a standard level that helps developers and helps people build whatever applications they want to build, I think that's completely fine. But at that level where it's at level four risk, these companies are simply never going to open source them yet even release them. OpenAI have previously stated on their blogs that they're literally never going to release a system that could help people develop novel biological agents. They literally stated that in their paper and it's something that I've read. So it's him stating this is just him stating the obvious. He's just stating it out in the open. So I'm not sure why there's such a backlash. Now there was this tweet that actually does, you know, dive into it and it completely dives into some of the reasons. It says model developers and providers are able to fine tune their models and implement security guardrails on the API to avoid producing dangerous outputs. But with open source models, these guardrails are simply not present or can be trivially removed. And there's no known method to ensure that these models, once released, are, are not able to be further fine tuned to produce bad outputs. In other words, once the model's weights are released, there's no way to ensure that it cannot be used for dangerous purposes. And remember guys, whilst yes, you might live in a society that is completely safe and you're not going to do anything bad with it, maybe you just want to use it for yourself. There are tons and tons and tons of people out there that will use much more advanced, sophisticated systems than the one we have today to do unspeakable things. I mean, do you guys really want another Covid situation? And I'm sure that many people were affected by that situation so negatively and it was something that was so hard to contain. So we do have to think about the actual real world scenarios because it only takes one bad scenario for everything to be completely awful. And whilst yes, people are like oh, we should open source it, we should open source it, I think you don't want that because like I said before, if they have like it, like let's say, for example GPT8 and let's say they open source GPT8 which is never going to happen. Okay, but let's say they did okay and something awful happened. If that actually happened, that would be completely worse because once something awful happens then that sets the precedent for hey, hey, hey, we need to be be much more stringent on what kinds of AI systems can be built, how many chips are going to be there, and I've got a video coming on on how certain bills that are being proposed could potentially change the landscape. So I don't think what he's suggesting is completely awful. I don't want random people to be able to develop novel biological agents. Yes, I'm trying to support a healthy developer ecosystem, but I think with systems like GPT7 and GPT8, I don't think it makes sense for those companies to open source that technology. But, but I do think it makes sense for them to open Source, the research where they're talking about preventing those systems from doing any harm. And I think those are two different completely. Distinctions like this don't actually help the discussion because he's stating that regulators to ban open source AI models is just going to get him backlash. But that's not what he's stating. He's just stating that advanced AI models that can do anything shouldn't be able to do that. So I think it's important to always make that distinction because if there's one thing you have to remember about being on platforms like Twitter and on the Internet, and it's that bad news always spreads faster than good news. Now something that was pretty crazy was Microsoft released Phi 3. Now if you don't know about the Phi series, it's essentially really, really tiny models that Microsoft have been working on. And essentially these models are very, very effective at what they do. And they've also been focusing on really, really high quality synthetic data. And so far we've seen that these models are surpassing things in every benchmark at similar sizes. And the reason that this is actually so surprising is because it was only a couple of days ago that we got llama 3 at 8 billion parameters. And we can see that Phi 3 mini at 3.8 billion parameters is actually doing better in the MMLU and in the HeLa swag. And that's a model that is 3.8 billion parameters, which is pretty much half the size. And you can see at the GSM8K it's doing better. And this is pretty crazy. Okay, like this right here is, I guess you could argue a GPT 3.5 level model, but it's 3.8 billion parameters. And the 7 billion parameter model is pretty much on par with GPT 3.5, which means that we could potentially in a couple of months, you know, have GPT 3.5 or even close to GPT 4 level performance on our phones. And that's going to be something that as we continue to see with the FI series, be continually increased. And honestly guys, this is something that surprised even me because I knew Phi 3 was going to be even better. But these benchmarks are really, really incredible, especially when it was only a couple of days ago that Llama stole the show in terms of what was possible. Ethan Mollick did actually do some demos of fi. You can see that it's able to do some really interesting things. He says, give me a Porter's five forces analysis of a company that teleports cheese into your mouth and then you can see it says the threat of new entrants. So here's Porter's five forces. It's essentially something that you use to add analyze a business's viability when entering a certain market. You can see it says the threat of new entrants into the cheese teleportation industry is relatively low due to high barriers to entry. The technology required for teleporting cheese is complex and expensive, requiring significant investment into research and development as well as patents to protect intellectual property. Additionally, the regulatory environment, yada yada yada. The point is, is that surprisingly, having a model that's so tiny be able to give you really concise and insightful answers is, is going to present a really, really interesting future for where we can have almost instant responses and insightful responses streamed instantly to us for a vast amount of different situations. Now, whilst I do actually talk about the post labor economy quite a lot on the second channel, which is essentially where I talk about how the future of AI is going to upend the economy in terms of taking a large majority of the jobs, I think situations like this, it does make sense, number one, because I think the people that do this kind of hate this. Although yes, you're interacting with customers. I've seen a million different videos on TikTok where people are just simply getting videos taken out of them. Like people are trying to do pranks for TikTok and stuff like that. And I think it's pretty disheartening to see that because as someone just trying to do their job, it must be awful to have that happen to you where you're just trying to do a job and then someone's literally just trying to make a joke out of you at work. And the reason I think this is also really cool as well is, is because certain people don't speak all languages. So this definitely is an access thing whereby different people from different countries are able to easily access this technology. I mean, imagine you're on holiday and you are talking to an AI system that instantly translates your language and can instantly understand your order, which is truly going to break down the barriers in terms of being able to connect with other cultures and people. So just take a look at this.

Welcome to Wendy's. What would you like?

Can I have a chocolate Frosty?

Which size for the chocolate Frosty Medium? Can I get you anything else today? No, thank you.

Great.

Please pull up to the next window.