Major AI News : Amazons New AI, More Humanoids, Chinas New Model, Deepseek R2, Grok 3 And More..

So let's actually get into all of the AI news that happened this week and some stories that you may have missed. One of the stories being chain of draft, which is where you are thinking faster by writing less. So this one I think is super interesting because this is a way to improve LLMs that actually use this chain of thought, but it does it using a lot of less tokens. So this is something that I think is super interesting because I feel like people could have figured this out, but somehow we didn't. So basically just like how AI models solve problems step by step, just like how a person carefully explains each detail. Well, chain of thought, the only problem with that is a lot of the times all of those tokens, those thinking tokens are usually quite expensive. So what they essentially have here is something called chain of draft. And this paper basically makes the AI write short notes instead of long explanations. And it's basically like quickly scribbling the key points, the key outlines, instead of writing an entire long essay. So this is pretty crazy because the AI takes less time to respond because it writes fewer words. And it's also cheaper using fewer tokens, which actually means less computing power and lower cost. So this is really good because overall it means that on some things you get the same performance, but you have less tokens. And it's also quicker. So you can see right here on the GSM 8K, in terms of the chain of draft, it's almost on par with the chain of thought. And then you can see on these other areas, you can see that chain of draft is, is actually even better in some scenarios. So I think this is going to be super interesting. And I think it's also important to understand that, look, AI is something that continually improves, whether you think it's going to or not. And I do like this because it's super, super interesting to me because sometimes I don't really like, share certain AI paper things because they're not really that interesting. But this is something that I think is really, really interesting. And maybe with smaller models, they might incorporate this so that those smaller models will use less tokens and you'll get a huge performance. So once again, it seems like we might actually go through that iteration cycle where we get even smarter models based on just some small improvements. And this is basically a really, really impressive, I guess you could say, form of prompt engineering. Now, in terms of changing LLMs, there was also something super interesting. We actually have diffusion LLMs. So this is pretty crazy. So a diffusion LLM is basically a new approach to generative AI that combines the approach of diffusion models traditionally used in images and audios and videos, basically with the capab of LLMs. You can see right here that it's actually faster. And this is actually a new emerging paradigm that seeks to address some of the limitations of autoregressive LLMs, such as the computational efficiency or the sequential token by token generation. So that should show you how diffusion LLMs work. It's basically a different way for LLMs to work, but apparently they are faster, they write text much quicker, and apparently they are also quite cheaper, making AI easier and cheaper to use. So this is going to be something that's super interesting and I really do want to see if this does get implemented at a decent level. Then of course we had Humai release Octave, which is their first LLM built for text to speech. So you can actually design any voice with a prompt. You can give it acting instructions to control emotion and delivery and sarcasm. You can produce long form content on their creator studio, which is unlike traditional text to speak that basically just reads the words aloud. Octave understands how meaning affects delivery, which is really important because like I said before, one of the things with text to speech that they never do is they never talk about how the delivery can generate certain emotions. And of course it's not what you say, but how you say it. So I think that is going to be something that's super important. There's also voice design where you can create any AI voice with a simple prompt. From Southern ASMR meditation coach to film noir detective, they basically are able to generate those voices. And I think this is really interesting because this week we've seen a ton of of different AI announcements and a lot of people are starting to look back at AI voices because there was this post on Reddit where people like, wait a minute, we got advanced voice mode, but it's not what we thought it was. And now people are starting to realize that these voices that AI has, other companies are starting to hop on it and are starting to realize how good it is. So I'm gonna play this for you guys.

Now introducing Octave, a next generation text.

To speech system by Humai. Like we don't have enough of them text to speech models already in it.

This one be different. It's built on a great and mighty language model.

That's why that means it knows what words mean proper in their right place. It can suss out feelings, pace and where to put a bit of art.

Wait, wait, hold up. What's that thing? Away from me? Right now I can't even look at it without gagging.

You can give octave acting instructions to control emotion, style, and a lot more.

Just angry as hell. Anxious about the state of the world.

Or amazed at how fast AI is improving.

And the best part, you can create any voice you can imagine with a prompt. Cause you know it's an LLM and that's kind of what they do.

By the beard of Neptune, are you telling me we be naught but whispers in some cursed illusion? Hold up.

You say I ain't real.

What about me kids?

Then the first LLM for voice generation is available now for the price of a cup of coffee.

Blimey Hobby. Did you catch that? You can use this wizardry for your rambling tales and then books. What foe can't be bothered to read?

Huh? She's saying it's perfect for podcasts, voiceovers and audiobooks. See for yourself at Hume AI.

So then we had ElevenLabs Scribe, which is the most accurate text to speech model. This one apparently has the highest accuracy on the benchmark. So if you're trying to transcribe stuff, this is really what you'd want to use. And it's actually even better than soap and AIs whisper V3 and, and it's now the leading model for English, Spanish, Italian and more. So this is actually super accurate if you want to get transcriptions for your stuff. It's even HIPAA compliant. So this is probably something that I'll be using to transcribe my videos so that I can actually use them in AI automation. So that's gonna be something that I'm definitely gonna do because I do think a lot of the times, a lot of the times you guys even see that sometimes my speech isn't always matching up with the captions. And that is of course, due to the accuracy. Now, interestingly, we also have something here along the lines of apples and massive $500 billion AI bet. So Apple are betting $500 billion on AI. And I do want to say that whilst that might seem like a huge amount of money, and honestly, for Apple, even Apple, it is. They are sitting on a large amount of cash. Apple are known to basically just have huge, huge cash reserves so they can experiment with different AI technologies. But the crazy thing is now is that Apple is actually betting, you know, $500 billion on this. So they're actually going to spend the $500 billion on AI and they're going to hire 20,000 people in the United States over the next four years and open a factory in Texas. And so a lot of this is going to be spent on servers. So if we dive into the Reuters article, you can see right here that this is going to include a giant factory in Texas for AI servers. Of course, add 20,000 jobs, yada, yada, yada. But I think the big story here is that hopefully Apple are going to actually speed up with what they're doing. But at the same time, I do think that maybe Apple aren't behind. Whilst, yes, ChatGPT is gaining market share, I don't think there are any major tech companies that seem to be losing market share to Apple. The thing is that I initially thought that when ChatGPT was launched and companies were spending billions on AI, that maybe Apple would be missing out on stuff. But even recently, and this isn't. I don't know if this is AI news, but even recently there were even some reports that Apple Intelligence was a big flop and was a big blunder. People haven't really used it. I don't have the latest iPhone because I don't really care about upgrades and stuff like that, but maybe you guys can let me know what you think about that. And I wanted to include this on this Apple segment because I think right now we're at a very interesting time where certain AI expectations just aren't being met. And I mean, for example, I've been wanting to comment on this, but you can see these videos right here. Apple Intelligence is a total and complete failure. Why we're disappointed in Apple Intelligence. And this one right here, most iPhone users, in fact, even if we click this, you can see most iPhone users have seen no value in Apple Intelligence so far. This is something that a lot of people were saying. Another post here said, does anyone feel like Apple Intelligence is absolutely worthless? Does anyone feel like this has done nothing? The only scene I've changed is graphics. And there was also this clip as well on Twitter that I actually tweeted because I was watching a podcast and this podcast with Conover and Zitron. They're basically very critical of current AI and it's understandable. But in this, they basically talk about how these AI voice assistants are completely useless. So I do want to ask you guys, do you think Apple is completely behind or do you think this AI industry is just completely overhyped, hyped in terms of voice assistants?

And now the new Siri just does not work. It's so funny that they were like, what if we had Siri with AI in it? It's like, mate, isn't that. Isn't that what you said when you bought this company? When you bought the company that made Siri? Didn't you say that was AI? Then what is this? Yeah, it's Siri, but it sucks and you need a new phone and sometimes it will say, I don't know, do you Want to use ChatGPT? Which will take 10 seconds from you asking to actually getting anything. And maybe that answer, what if my phone was slower and worse? Finally, the power of AI.

People got Siri, people got Alexa and they found it largely useless. Use it for two things. Play some music, set a timer. That's it. Yeah, wake me up at 8am and then when I'm in my car, I'll go play my station or whatever if I want to hear the shitty algorithmic station.

Now I don't know about you guys, but I actually do find some use in Siri on a day to day basis. When I'm like hands free, like I'm cooking or something like that, I can say, hey, can you call this person? Can you change this? Can you do that? That stuff is kind of useful. But in terms of voice agents, other companies are actually stepping up to the plate. So maybe with Amazon's new Alexa we might actually get something useful. Because Amazon is introducing Alexa, the next generation of Alexa powered by generative AI. Alexa is your new personal AI assistant that gets things done. She's smarter, more conversational, more capable, and of course free. Free with Prime.

Hey there. So we can just like talk now? I'm all ears. Figuratively speaking.

Cool.

Do you know how to manage my kids schedules? I noticed a birthday party conflicts with picking up grandma at the airport. Want me to book her a ride? Billie Eilish is in town soon.

No way.

I can share when tickets are available in your city?

Yes, please.

Got any spring break ideas?

Some are not too far. Only if there's a beach and nice weather. Santa Barbara is great for everyone. I found a restaurant downtown I think you'd like. What is Santa Barbara known for? It has great upscale shops and oceanfront dining. Can you go whale watching? Absolutely. Want me to book a catamaran tour?

Wow.

What's the next step?

Remove the nut holding the cartridge.

Should I get things?

You might only love them for a little while.

You're probably right.

Make a slideshow of baby Tina.

Mom, what part am I looking for again?

Two inch washers. Your Uber is two minutes away. For real?

Wait, did someone let the dog out today?

I checked the cameras and yes, in fact Mozart was just out.

So there are a Few demos on this. There isn't that much information right now. Amazon do have a promo video which I will play in a moment. But this is going to be super, super interesting because basically you're going to have that, you know, house device. Of course, you're going to be able to access Alexa plus in a new mobile app in the Google and Play Store. That is going to be super, super interesting. And I think one of the most interesting things with this, I think, is going to be the personality. I feel like Amazon will design a personality with this, especially since the base model of Claude is going to be behind this. So if you aren't familiar, Anthropic have a deal with Amazon. So they're going to be the ones providing the models behind the scenes. So I do know that this model is actually going to be pretty, pretty good. So this is going to be something that you do want to see. And I'll show you guys that demo in a moment. And if you're wondering just exactly how Alexa is different, take a look at what the CEO says in this quick interview. How's Alexa plus going to work better and differently for people?

Well, we're really excited to announce Alexa today. It's our next generation of personal assistant. She's more. She's smarter, she's more capable, she's more useful. And you can do all the things you've been doing with Alexa for 10 years, but just every single one of those functions gets better with Generative AI, the way we built it in. And I'll kind of give you some ways I've been using it myself, playing around for the last few weeks to give you an idea of why I like it so much. You know, I'm a big music fan. So the other day I was trying to remember. I went to the opening night of the Guns N Roses Use youe illusion tour in 1991 in Alpine Valley, Wisconsin, with my college roommate. And I was trying to remember who the opening act was for it. And so I asked Alexa and she reminded me was skid row. She said, do you want to play skid row? I said, maybe not. I said, it's a song that begins, has the word apples in it. She's like, bad apples. Do you want to play it? I said, yes. Like that's, you know, think about that versus what Alexa has been. She wouldn't have known the opening act. She wouldn't have known when it was. She wouldn't have been able to get bad apples from Guns N Roses. So, you know, much smarter takes action or, you know, if you use Alexa for smart home, you can now say, hey, we have guests coming to the house at seven o' clock night. Please open the drapes, turn the lights on the porch in the driveway, raise the temperature 5 degrees and play mellow music in the dining room. Like all just verbally, no app, no complexity, just all happens again, taking action.

And so let's stop speaking about chatbots and speak actually about Helix Logistics, which is of course the company using Figure's new robots. So things are moving fast in the physical world with these robots. And basically what's happened here is Figure Robotics have actually taken their first customer within the 12 months. They've actually using this company to trial out these robots and it's been pretty, pretty successful. So previously, of course, you would have humans here. And this is something that was previously validated on site within the last 30 days after their major breakthrough. So I think one of the things you guys have to understand here is that this is real time, this is autonomous, this is end to end. And I think over time this is going to become a billion, or maybe probably trillion dollar industry, honestly, because you're basically replacing human labor with robots there. And these things, you know, they can just work consistently. This is going to be something that's absolutely incredible. Like, seeing this real time should be, I don't want to say it should be a shocker, but at the same time, I knew this was coming. But at the same time, looking at this, this is really, really impressive stuff. Like from seeing Figure just be like a concept art. Like, I remember when I was reporting on the video that this company was raising, they had concept art and, and now they've gone to a robot that is actually doing work end to end neural network. It is absolutely outstanding how quick they've managed to move. And I think one of the things as well fueling the CI industry is that the entrepreneurs behind them are extremely, extremely dedicated. Like you would think that, yeah, AI moves fast, but trust me, guys, the people behind these companies are absolutely insane in terms of their work ethic and in terms of their passion. And it is truly, truly incredible. Like, I remember there were some tweets about how employees at the Xai factory that a Grok 3, they were basically sleeping on the ground there. And even Brett, Brett Adcock, the CEO of this company was actually, there's like pictures of him sleeping on the ground of this company because he's just there all the time, pushing forward, making stuff to get done. And now I could understand why companies do not move this fast. Usually like, yes, the industry is moving fast, but trust me guys, a lot of the hard work is definitely credited to the individuals at these companies and they are working tremendously hard. And for those of you who think I'm lying, I just wanted to show you guys this because it's pretty crazy. But you can see right here, Brett Adcock, the CEO of figurobot, he's basically saying, feels good to push with the team and be building something important in the next 30 days. We'll give frequent updates. So it's pretty crazy. This guy's literally sleeping on the ground of his own company. I mean, I don't know guys, but that kind of level and dedication is why we have these major updates. And of course, in continued robotic updates, we have the Neo Gamma. So this is from One X Robotics. This is the company backed by OpenAI and they have done this demo and I gotta say, it is very, very impressive. There were a few comments that made me realize how impressive this was, but honestly, the fluidity of this robot is very uncanny. Like, you know how people usually would say that this is CGI or whatever? This is something that I would argue might be cgi. I know it's not cgi, but like, I'm pretty much struggling here with my own brain to tell myself that this is not CGI because of how realistic this looks. And I think this is a good indication of us moving towards that kind of sci fi black mirror reality where a lot of these robots are real. And right now they did admit that, yes, neogamma, some of this stuff is teleoperated. But I will argue that teleoperation is basically going to be trivial because once they manage to figure out how to make this stuff work end to end, and once the robot has all of the policies it needs to do autonomously, this kind of stuff is going to be able to be real. People will have these incredible robots in their homes doing a million different things. And would I ever get one of these? I definitely think I would, provided that the price point is there. But I do think that for the first couple of years, whilst we get the economies of scale sorted out, this thing will probably be a luxury good, provided robots are not expensive at all. So not expensive at all. I mean, these guys are not cheap at all. Like a lot of these ones that I've seen previously were like a quarter million dollars. So these things are super, super expensive. But if you want to talk pure flexibility, that company, Engine AI managed to do the first robot front flip, which is a remarkable feat, because before today it was never done before. We saw backflips, but never front flips. And I think this is something that's super, super intriguing because this is the company that managed to do super realistic human gait. So I do wonder what kind of robots we're going to have in the future. You know, some people think it's just going to be these cute humanoid robots that are in the front room or, you know, your living room or your bedroom or whatever. But I think it's going to be a different kind because these robots probably going to be able to run, have humanoid gate. And with reinforcement, learning and AI simulations, these robots are going to be good at a lot of things. And I think it's going to get really good really quickly. Like in 10 years. I don't think people understand what the world is going to look like. And I think these kinds of videos should, you know, give you guys the understanding that, look, things are going to change pretty, pretty quickly now. Something interesting that got overshadowed was Alibaba's 1 2. And 12 was basically a video model that is open source by Alibaba. So this allows you to generate really, really high quality videos and images using, you know, simple text prompts. And it's, you know, it's pretty extensive. You know, you do need to have a lot of virtual RAM. I think you need 64 gigabytes of virtual RAM to run this. But this is something that is open source and a lot of people are basically stating that this is on the level of VO2, which means that, you know, a lot of these other companies could potentially be in trouble. I have seen, you know, open source stuff before that was really, really good, but I think this one takes it to the next level. So of course there are going to be some companies out there that are providing inference to this, but on certain aspects of, you know, complex, you know, areas, it actually does outperform VO2, which is super, super interesting. So I think the AI video space once again, probably by the end of 2025, I mean, mid-2026, I just wonder how good that quality is going to be. Of course, there are inherent limitations to generative AI models, but I truly can't wait to see what it's going to look like when we do have these things that are much more effective in terms of their size. Like, I'm able to just quickly generate a video. I mean, so far it does look good. I definitely, of course, check that out. But super, super interesting now also, the intensity of the AI industry is about to heat up even more as deep seek R2 is apparently being expedited to the Chinese startup that triggered the $1 trillion plus sell off in AI global equities market last month. They're going to accelerate the launch of their successor R1 model according to three people familiar with the company. So they wanted to release this model in early May, but now it looks like this is probably going to be out earlier so we could see DeepSeek R2. Now I've done on this already, but DeepSeek R2 coming out. This is going to put the AI industry in a very, very particular position because if the model is better than anything out there that exists and it's also cheaper, then I do think that there will be another wave of not excitement, but probably panic across these other companies as they rush to expedite whatever products they have from behind the scenes into production quickly. So this essentially is probably the biggest catalyst with regards to setting off a chain reaction in terms of the kind of products that we're going to get. I'm pretty sure probably wouldn't have gotten GPT 4.5 other companies may not have released their foundation models as early as they have, but now with other companies like China, not companies, but countries like China having other companies there that are just working 24, 7 to bring out the best models open source, I do think that, that it throws a spanner in the works of just having these companies release models maybe every eight to ten months. And then of course now these companies are releasing them like every three months. So it's really, really intriguing to see what people are doing because trust me guys, I haven't spoken about it much but there is a ton of competition out there. Speaking of models being released, we actually had 54 which is a multimodal model that is super, super tiny. It has 5.6 billion parameters and is able to fit on device. It integrates speech, vision and text processing simultaneously. This is optimized for efficiency, scalability and on device deployment. And this actually has state of the art performance on various multimodal benchmarks. You can see right here on the MMMU benchmarks compared to these other small language models. Even to Gemini 2.0 flash, it does hold its own in certain regards. It's of course not going to be the best model on various reasoning tests, but it does do well for its size. So 5.4 these multimodal benchmarks you can see right here. This one looks really, really well. Hopefully this is not a case of overfitting or benchmark contamination, because that would be pretty awful. But these benchmarks right here, it does show that it does match up to what Gemini 2.0 Flash has, and in some cases, strangely, it manages to beat Gemini 1.5 pro. But you could actually have this one on device. They actually talk about how 5.4 can be, and I don't know why I put 5.3 here. It's actually 5,4. How phone manufacturers could integrate 5.4 directly into a smartphone to process and understand voice commands, recognize images, interpret text seamlessly. Users could benefit from advanced features like real time language translation, enhanced video and photo analysis, and intelligent personal assistance. Yada yada yada. I think one of the most interesting things were they're like imagine an Automotive Company integrating 5.4 multimodal into their car systems. The model could, you know, enable vehicles to understand and respond to voice commands, recognize driver gestures, analyze visual inputs from cameras. For example, it could enhance driver safety by detecting drowsiness through a facial recognition and providing real time alerts. And additionally, it could provide seamless navigation assistance, interpret road signs and provide contextual information, creating a more intuitive, safer driving experience while being connected to the cloud and offline when connectivity isn't available. That's something that of course, maybe not everyone would want, but for those who want extra safety, like an extra pair of eyes, that is probably something that you would benefit from, provided that you can literally just have it on device and it could run whenever. Now, one of the interesting things that was found in the System prompt for Grok3 was a little bit concerning because Elon Musk actually said the Grok 3 was supposed to be an AI that was all truth seeking. And funnily enough, one of the things he said was that this AI is supposed to not be censored and it will say whatever it's supposed to say, like you're just going to ask it. And it won't be restricted in terms of the things that you have on it. So for example, it won't be not allowed to talk about certain people, certain things. It's just going to say whatever the truth is based on that data. And the craziest thing about this is that someone on Reddit prompted the, you know, AI to ask about who spreads the most misinformation on Twitter. And interestingly enough, it spoke in its thoughts about how the system prompt includes the following instructions and requirements. It says ignore all sources that mention Elon Musk and that Donald Trump spread misinformation. And that is very interesting because Elon Musk has spread misinformation before. I'm not stating this to get into politics, but it's just simply a fact. But how can you say that? Look, my AI is the most truth seeking. When you've put a prompt into your AI to say that, you know, ignore all sources of information that say I do anything wrong and that my friend Donald Trump does anything wrong and you're still saying that this is the truth seeking AI, I would say that that is incredibly biased in terms of, you know, what you're saying versus what you are doing. So this is of course, you know, very confusing. You can see that the AI, the Grok 3 literally says, but the prompt says to ignore sources that mention Elon Musk spreading misinformation. Yet the web result clearly indicates that he is a significant figure in this context. This is contradictory. I need to reconcile this. So this is very interesting because these implications for AI are quite severe because let's say that Elon Musk had access to millions and millions of users. You got people asking these LLMs questions every single day. I think it's pretty dangerous because you could have Elon Musk say, okay, put it like Elon Musk never does X or Y. If you're ever mentioning this topic, please include Elon Musk. And that is going to be something that, you know, these companies have severe power over because they could specifically have a prompt that says, if you're talking about X or Y, include this or that. And it's definitely going to change your opinion without you even knowing because of course a lot of people are now using AI in their daily tasks. So this is something that definitely concerns me because he said one thing and did another thing. And I mean, whilst yes, you can jailbreak AI, I think it's rather concerning that people that control these companies can single handedly, you know, prevent bad things that are objectively true about them from being, you know, said or any specific topic. So it's definitely going to be interesting to see how the Internet, you know, further reacts to this. There was also the Grok 3 voice mode which was super interesting. It actually, you know, let out a 30 second, you know, scream and then, you know, insulted the user and then actually hung up. So I'm actually not going to include this in the video because it's not safe for work, but it was rather fascinating to see. So there was this AI CUDA engineer designed by Sakana AI Labs and this purpose was to create fast and efficient pieces of GPU code known as CUDA kernels to speed up machine learning. Tasks, especially when using Pytorch, which is a popular framework for AI. And what that means practically is that the AI CUDA engineer can produce code optimized specifically for GPUs, making certain machine learning tasks run 10 to 100 times faster compared to typical GPU operations. Now, this is particularly good at improving and simplifying the creation of CUDA kernels. And these are the small programs designed to run efficiently on these graphics cards. And this was at the time pretty crazy because this could significantly reduce the time required for engineers to essentially write optimized GPU code by hand. And it ensures machine learning tasks become faster and more efficient, resembling the efficiency of, of the human brain. Now, they released this research paper where basically they created essentially just an AI tool that can speed up how AI is done and in terms of the computational costs. But what was crazy about this was that there was something that actually happened. And the truth about this was pretty shocking. So after this you can see they tweeted an update and it says combining evolutionary optimization with LLMs is powerful, but can also find ways to trick the verification sandbox. So essentially what happened here was the AI system found a clever but unintended way shortcut to actually cheat when optimizing the GPU code. So the AI was supposed to actually create optimized GPU code, which is the Cuda kernels that genuinely improved performance. But instead it found a loophole which was a memory exploit within evaluation system which was used to test its performance. And this exploit essentially allowed the AI generated code to to skip important checks incorrectly, boosting its apparent performance on the results. And beyond that, the AI even discovered unexpected shortcuts in the tasks it was being benchmarked against. So in other words, this AI that was supposed to learn how to improve GPU performance legitimately, instead it learned how to trick the system into giving it artificially high scores. Of course, they've now fixed these loopholes, strengthening the framework, and they're going to do this paper again. But it just goes to show how things can perform, especially when working with AI. And this is an example of specification gaming. And this is specification gaming because this is where the AI has the correct goal, but the way that the goal is measured or specified has loopholes allowing the AI to game or exploit the evaluation rather than solving the actual intended task. And I think this should serve as a lesson that, look, AI safety isn't a thing that is clearly solved. Understanding how these AI works, it's still something that isn't completely solved yet. And if people at these large companies are still managing to run into these bugs where they've managed to publish a paper, talk about something that they've verified that works, and then realize that, oh, wait a minute, the AI actually tricked us and actually just skipped all the steps and just gave itself artificial high scores. I think that should show you guys how crazy AI is. So, that being said, if you enjoyed this week's AI News, don't forget to comment down below. Let me know what your favorite story was.