it's been an insane week for AI news it feels like the month of January has had a Year's worth of news already and this past week has been no different tons of news so I'm not going to waste your time let's go ahead and break it down now there was one big news story this week that sort of sucked the oxygen out of the room for all of the rest of the AI companies and I'm not going to really get too deep into it because I did do a whole breakdown video All About deep seek and what happened there it's called Deep seek the Chinese AI that crashed the markets so if you want want to get the full scoop on what really happened with deep seek and why the markets crashed as a result watch that video because I think it's a pretty good breakdown that explains everything that happened saying that here's the quick tldr deep seek V3 was a new AI model that came out back in December the thing that made that model different was that it was supposedly trained with far less and far inferior gpus than what most AI models are trained on so they claim they were able to train that model for five or $6 million where most of the big AI models took tens of millions of dollars to train again this model was released back in December so it's like six seven weeks old already however last week a sort of fine-tune version of Deep seek V3 came out called Deep seek R1 it used reinforcement learning to sort of fine-tune it and then it also had a Chain of Thought thinking where it was showing you exactly how it was thinking when deep seek R1 came out for whatever reason that's when people started to notice and freak out and find out how inexpensively deep seek V3 was trained for when people found that out they started selling Nvidia stock like crazy thinking well if you can train it that cheaply maybe they don't need the gpus as much as we all thought they do and deep seek R1 is pretty impressive it's on par with open ai's 01 model even Sam Alman himself gave it prop saying deep seeks R1 is an impressive model particularly around what they're able to deliver for the price he then goes on to say we will obviously deliver much better models and also it's legit invigorating to have a new competitor we will pull up some releases so essentially Sam Alman said we're really impressed with R1 but you ain't seen nothing yet from us but in my mind the fact that Nvidia stock dropped 177% as a result seemed kind of bizarre because this deep-seat company used Nvidia gpus to train it and most of the reason that all of these big tech companies are buying up gpus aren't necessarily for the training it's for the inference the part that happens after you ask at the prompt in fact Yan Theon here I think put it perfectly over on this thread post he said major misunderstanding about AI infrastructure Investments much of those billions are going into infrastructure for inference not training inference being once you give it a prompt the processing and response that prompt is giving you that's the inference phase running AI assistant services for billions of people requires a lot of compute once you put video understanding reasoning large scale memory and other capabilities and AI systems inference costs are going to increase and then he went to clarify in his next post so the market reactions to deep seek are woefully unjustified now in that previous video that I made here about deep seek I did mention that there were some rumors going around that maybe deep seek had trained on like open AI content or used a different existing model like llama underneath as well as talked about how a lot of people believe that they're lying about how many gpus it took and how much it actually cost to train it well since I made that video there has been more information more rumor circulating so I did want to touch on those real quick according to the financial times open aai has found evidence that Chinese artificial intelligence startup deep seek use the US company's proprietary models to train its own open- source competitor so even though Sam gave them Credit in that previous expost not too long after that open AI started saying wait a second I believe they actually trained on our propri AR models and they didn't actually train that whole models and the irony of all of this has created some of the best memes kind of like how open AI used YouTube data not saying it's right but let's not pretend Open AI didn't violate YouTube's terms of service people are posting screenshots like this chat GPT Creator open AI sued for theft of private data and AI arms race what did you guys use for open AI though what did you guys use here's another one of open AI at Google Gemini fighting and then the little dog Chas chasing him out with a baseball bat being deep seek or this meme that says I can't believe Chad GPT lost its job to AI I mean you got to admit there is a bit of iron to the fact that open AI is out there complaining that another company used their data even though open AI essentially trained on the open web apparently the White House is also now looking into the effect of deep seek on National Security David saxs the AI and crypto Zar for the White House basically said that he believes there could be intellectual property theft talking about distillation and how it's when one model learns from another model so basically training on the outputs that were received from chat GPT so yes it's looking more and more likely that deep seek didn't train their entire model in the traditional way for only $5 million I guess there's some sort of new evidence coming out that's making it look more and more likely that that was the case but that's not stopping companies from pretty much putting R1 into everything I mean it is a really really good model that performs just about as good is 01 so of course all of the tools want to use it as one of their underlying models the company that I find the most interesting that's doing it is well Microsoft who we know to be open ai's biggest Ally Microsoft added the new deep seek R1 model into its as your AI Foundry and into GitHub so if you want to use Microsoft as your sort of back-end inference provider well now you can use a deep seek R1 API directly through Microsoft perplexity is another company that added in deep seek R1 into their search engine if you have perplexity Pro and you click on this little drop down next to Pro you can actually see there's two reasoning models available to use reasoning with R1 and reasoning with o1 now there is a limit to how many times you can use this per day but if you want to use perplexity for more like problem solving tasks as opposed to like searching for things around the web maybe try switching to R1 and see how it does for you Nvidia also created a Nim one of their microservices for deep seek R1 so if you want Nvidia to be your inference provider they'll do it and they'll do it using their Nvidia hgx h200 systems which will get you 3,872 tokens per second which is just wildly fast for comparison I was using my 590 locally on my computer and deep seek R1 was giving me about 60 tokens per second and that felt fast the popular coding app winds surf which I've personally started using more and more recently just at a deep seek R1 and V3 inside of their coding platform and these reasoning models have shown to be pretty good at code both 01 and R1 people have been getting some really really solid results heck even Kaya AI added deep seek into its platform and K is known as an AI art platform for generating images and upscaling images and making AI videos well now they just added Kaa chat and they're using the Deep seek model to power their chat although I've talked about deep seek for several weeks in a row now and even going all the way back into December more and more of this story keeps unfolding more drama keeps coming out about it the US government's getting involved now and pretty much every tool that leverages AI platforms or that is AI agnostic is making deep seek available inside of their platforms as well in what feels like a little bit of a reaction to Deep seek chat GPT is adding in a think button now I don't have that think button available inside of my chat GPT account I don't know if that's because it's still rolling it out or if it's because I'm a pro user and I already have access to all the various 01 modes but according to tore here it appears that if you press the think button it basically switches it to 01 for you so that you can actually see that it's doing that thinking process similar to what R1 does you can see in his screenshot here he's showing that he has chat GPT 40 selected gave it a prompt it gives the response but then under here the regenerate button is showing 01 and you can see this is what the think button looks like there has been a little bit of confusion around this some people wonder if it's using chat GPT 40 but then just adding that sort of Chain of Thought additional thinking on top of it or if it's actually using the 01 model which is supposedly a different model than the 40 model as far as we know and what also felt like a bit of a response to deep seek Mustafa sullan the CEO of Microsoft AI announced that they made the think deeper feature free for all users of co-pilot the think deeper feature is essentially just A1 but inside of co-pilot instead of inside of chat GPT we can see he said this now gives everyone access to open AI world class 01 reasoning model in co-pilot everywhere at no cost so again chat GPT adding a think button Microsoft adding the think deeper feature to co-pilot both seemingly using 01 feels a lot like them going look we do the same thing as R1 everybody's talking about R1 and how excited you are look we can do that same sort of Chain of Thought thinking thing that they do because that one's all o free and publicly available now we also will make ours free and publicly available also that's what it feels like to me I'm not saying that's like their thought process but that's what it feels like we've also been getting quite a bit of Buzz around an 03 model coming out from open AI this week back on January 23rd Sam Alman said big news the free tier of chat GPT is going to get 03 mini and the plus tier will get tons of 03 mini usage he then retweeted his own tweet and said okay we heard y'all the plus tier will get 103 Mini queries per day we'll bring operator to plus tier as soon as we can our next agent will launch with availability in the plus tier and if you're wondering when 03 is going to come out well there's probably a good chance that it's out today the day you're watching this video I record these videos on Thursdays they come out on Friday and Ben here over on X pointed out that open AI 03 releases tomorrow and I went huh is that actually confirmed do we actually know it's coming out tomorrow well I clicked on this link it took me to this NPR interview here with Chris who's the chief Global Affairs officer at open aai so this is somebody from directly at openai and I'm just going to fast forward to the important bit here you open AI released something called 01 back in September that has the same Tech we actually think ours continues to be ahead and we're going to be releasing some additional models including 03 which comes out on Friday so this audio came out on January 30th which was Thursday he said 03 comes out on Friday in this interview I don't know if that was a slip up and he was supposed to say that or not but the Friday that he's referring to is January 31st the day that this video is actually going live so there's a very good chance by the time you're watching this 03 is either already out or going to be coming out soon so all of us llm nerds are going to have a new toy to play with most likely next week there were some other announcements out of open AI this week they introduced what's called chat GPT gov which is a government use only version of chat GPT I guess it has more Security on it they can run it locally on their own government servers and not send them to open AI servers but it's essentially just a chat GPT version with a few extra layers of security on it from what I can tell open AI this week also announced that they signed an agreement to work directly with the US National Laboratories so essentially they're basically working closer with scientists that work within the government and they're doing this with the goal of accelerating the basic science that underpins us Global technological leadership identifying new approaches to treating and preventing disease enhancing cyber security and protecting the American power grid achieving a new era of energy leadership improving US security through improved detection of natural and man-made threats and deepening our understanding of the forces that govern the universe the labs also lead a comprehensive program in nuclear security focused on reducing the risk of nuclear war and securing nuclear materials and weapons worldwide now all of this talk is about preventing diseases and preventing disasters and defense and all that but a part of me has to think they're also going to work with the government to like weaponize this stuff too right like most likely right okay I'll take off my tin foil hat now but it's not too much of a stretch right as far as actual features coming out of open AI they updated their canvas tool you can now use the canvas tool inside of 01 and canvas can now also render HTML and react code I'm actually really surprised that it couldn't render HTML already I thought it did but I guess it does now and it's been fully rolled out on the chat GPT desktop app for Mac OS if you're in the EU you started to finally get some of the features that the people in the US have had for a little bit now the video and screen sharing features that are available inside of the advanced voice mode are now available in the EU Norway Iceland lonstein and Switzerland the recently updated custom instructions are also now rolled out in the same areas do you have an idea for a video or short film but have a hard time actually visualizing what it could look like in the end you could storyboard out an idea but that's super timec consuming and also still pretty difficult to get get a real clear picture of what the finished product could look like that's why for today's video I partnered with Nvidia AI it's a great tool for plugging in an idea for a video and then letting it generate the visuals the sound effects the dialogue The Script all of that for you so you can get a rough idea of what a much grander video from your concept could look like so check this out so I have a story concept outlined here it's all about a hacker who discovers that AI is actually rewriting history and instead of creating like a storyboard for this concept I want invid to just make me like a rough draft version of this video so I can kind of visualize how it will all flesh out in the end so let's go ahead and generate this video and after several minutes I have a video back here's a quick sneak peek government suits with tech that could wipe your mind clean as a blank tape I ran through neon soaked streets hover cars buzzing overhead like Angry Hornets every shadow held a threat every reflection of potential trap now I'm going to go ahead and leave it right there as a little bit of a teaser I will post the full video over on my Instagram account so you could see the entire thing now keep in mind this was not designed to be a completed finished video this was designed to be a storyboard to map out a concept for a video that you can then use to you know pitch to investors or to people that might help you produce the video that can give people a much clearer idea of what you had in mind than just a storyboard alone would and one thing that's really cool about in video is if there's any edits I want to make to this video all I have to do is use natural Lang language and type what I want to see changed for example if I want to change the voice I could give it a prompt like change the voice to a deep more masculine male voice click generate and about a minute later we have a different more deep voice it wasn't just controlling our power it was controlling our reality and if you want a more traditional way of editing you've got other options available as well just click on the edit button here we can replace any scene in the video here with uploaded media stock Media or even generate a whole new shot you have the option to change out the stock music or upload your own music tweak the exact script and dialogue of the entire video and make all sorts of other changes to the video however you see fit not only is nid great for creating entire short form videos to post on your social media accounts it's amazing for bringing rough Concepts to life so be sure to check out Nido AI in the link in the description below and thank you so much to Nido for sponsoring this video moving on to Google they just updated their Gemini app Gemini is now powered by 2. 0 flash which is their latest model designed for the agentic era it is a super super fast model and it has done really well on the benchmarks and now that's what they're going to be using inside of Gemini there's been a little bit of Buzz around grock from xai rumors have been circulating that grock 3 is coming out at any time now in fact some people on X actually already got access to it they asked grock questions and then viewed the source and saw that it was actually using grock 3 so we know this new grock 3 Model is right around the corner at some point deep seek wasn't the only Chinese AI company in the spotlight this week Alibaba actually released a few different versions of their Quinn 2. 5 this week they released a new model called Quinn 2.
5 Max previous Quinn models have actually been open sourced this one is not this one is closed apparently this new model exceeded the efficiency of deep seek V3 and also claimed to outperform a lot of state-of-the-art models in various benchmarks like clae 3. 5 Sonet open AI GPT 40 and meta's llama 3. 1 401b they do have their own chat bot where you can play around with it over on Quin chat over at chat.
quinl m. ai if I give it a prompt like how many RS are in the word strawberry doesn't actually show its thinking like the other ones but I get the impression that it was doing something similar behind the scenes and of course it got the answer right on this but when we look inside of the Quin chat dashboard you'll notice that down here we have some options that you don't normally see inside of other AI chat Bots like image generation and video generation this model will create images I can tell it to create an image of a wolf howling at the moon and within a few seconds I have a decent image of a wolf howling at the moon and I'd say it's on par if not a little bit better than what we're getting out of something like Dolly 3 right now I can also select video generation tell it to create a video of a wolf howling at the moon and well I've been waiting for about 10 minutes now and it's been stuck at 99% for most of that so I don't know if the video functionality is totally working or if the servers are just kind of M maxed out right now but supposedly this will also do video for you but under this drop down up here there's another model here called Quinn 2. 5 VL 72b instruct and this one's a smart large Vision language model that can control your PC and phone now I don't know how to actually use the model yet to get it to control my PC and phone I have to do some more playing with it but it says they can perform a number of text and image analysis tasks parse files understand videos and count objects and images as well as control a PC similar to open ai's recently launched operator but here's a demo that somebody shared of it actually controlling what appears to be their phone and actually searching out a flight for them on a specific date that was from Philip Schmid here and VAV I don't know if I pronounced that right showed the tool on another machine here actually taking a bunch of actions on the computer I can't quite tell what they're trying to get it to do but it seems to be taking a bunch of on the computer on their behalf and even yet another model was released this week in ml small 3 now this one is an open- Source model under the Apache 2.
0 license and this one's designed to take on the much smaller models that are designed to be run on device models like GPT 40 mini Gemma 227b and the Quinn 2. 5b and according to this chart here the MML U which tests models on a whole bunch of different tasks and prompts mistal small 3 performed almost as good as Quinn 2. 5 32b but did it quite a bit faster and it outperformed and was faster than both GPT 40 mini and the Gemma 227b models this isn't going to be the smartest model in the world but for answering pretty simple questions this one would work just fine and it seems to perform really well and pretty fast compared to the competition the company block which was founded by Jack dorsy who originally was one of the founders of Twitter recently showed off what they're calling code Nam goose this is an on machine open-source AI agent built to automate tasks it's model agnostic you can choose whatever llm you want behind it you can think of goose as an assistant that is ready to take your instructions and do the work for you it seems like this first version of it is really designed to help coders do things like conduct code migrations dive into projects and unfamiliar coding languages conduct performance benchmarks increase code coverage removing or adding feature Flags generating unit tests for a feature and things like that and I'll link up to this page below but it does have a quick start guide so if you want to download this and run it on your own computer it is available to use not something I played with myself personally yet and since we're talking about agents and AI sort of going and doing tasks on your behalf Google is starting to test a new feature that will actually make phone calls for you they say there's a new experiment that just launched in search Labs you can use AI to call businesses on your behalf to find out what they charge for a service and when it's available like if you need an oil change ASAP from a nearby mechanic it looks like when you do a search on Google it'll give you this like ask for me button and then ask what service you need you click a button and you basically go through a little survey and answer questions that the person on the other line might ask you and then it will actually make the call on your behalf and then answer the questions for you now I'm not sure how to actually access this yet it says just opt in and search something like oil change near me to try it out but I'm not actually sure where you opt in for it yet I did the oil change near me search and I didn't see the option the US copyright office this week put out a document explaining which images can be copyright and which ones can't be copyright if they were generated with AI and this is another topic that I actually made a breakdown video of you can find it by looking up the video here's what you can and can't copyright with AI but the super quick version of it is if you just prompt an image and then try to get it copyrighted they're going to say no if you use AI as part of a bigger process like let's say you're making a film and like one of the background scenes was generated with AI that scene itself can't be copyrighted but the movie that you generated that use that scene can be and then there's some like gray areas cuz they said everything will be determined on a case-by casee basis and then they showed an example from Chris castanova she handrew some stuff and then used AI to fill in the rest of the image and they basically said the parts that you handrew are copywritten but the stuff that AI did for you is not and so still felt very fuzzy to me but basically at the end of the day the theme of the paper was if you're using AI as like an assistive tool to help you make a bigger creation then you can copyright the creation if you're just trying to prompt something and then claim copyright on the thing that you prompted you can't claim copyright on that with some fuzzy gray area in between there is one last bit of information about deep seek that came out this week in that they dropped their own AI image generation model called Deep seek Janice Pro 7B the model can generate images and it apparently Beats open AI Dolly 3 and stable diffusion across various benchmarks I actually haven't played with this one yet every time I've tried to load it on hugging face and use it it's either been too bogged down or just never loaded for me so I actually haven't seen the images that have come out of this yet but let's talk about AI video for a moment Pika Labs just released a new AI video model called Pika 2.
1 in last Friday's news video we talked about how they were teasing it was coming soon and well they released it this week it creates in 1080P resolution they claim to have razor sharp details seamless motion and lifelike human characters and some of these demo videos are pretty dang impressive they look like some of the most realistic human video that we've seen and I mean these Wildlife videos it looks to be one of the better models probably on par with what we're seeing out of V2 from this trailer these are obviously fairly cherry-picked videos if they're putting them in a trailer like this but that wasn't all Pika released this week they also released their new turbo mode which makes these videos generate way faster they say it generates the videos three times faster using seven times fewer credits without compromising quality so if I jump into my Pika account here down on the bottom where the little rabbit image is you can see we've got options for Pika 2. 1 that's what's actually selected by default and Pika turbo so the Pika 2. 1 is probably going to give us better video quality while Pika turbo is going to be a lot faster so let's try Pika 2.
1 using a monkey on roller skates and then let's use Pika turbo also doing a monkey on roller skates just so we can see them side by side the Pika turbo generated pretty dang quickly it finished before the paa 2. 1 and well here's our monkey on roller skates the monkey on roller skates actually doesn't look too bad although an extra roller skate appears in the video and the background kids look a little funky but if we check out the paika 2. 1 here it does look quite a bit better especially in the background the kids in the background look quite a bit better the roller skates kind of get a little fuzzy down here and I find this monkey to be sort of like overall kind of blurry but not bad considering where we were months ago the company cray AI introduced character consistency in videos this week so you can make videos where the character looks like the same person throughout you basically upload an image of a character it could probably be your own face or something you generate with AI and then you can make videos where it uses that same character with every single video if I jump into Crea AI here and click on generate video it looks like we need to select the Halo option here cuz this one has the character option in it if we select that we get our character option I'll upload my face here and do the prompt a man dancing see what it does and here's what it created for me yeah that's that's pretty awesome it looks just like me if maybe I put on an extra 100 PBS but yeah you can make consistent characters by uploading a face pretty much any prompt I give it is going to have my same face on the video Luma AI this week announced that they now have the ability to upscale to 4K with any of your videos that you make in dream machine you natively generate in 720p or 1080p but then you can now actually upscale them all the way up to 4K which is pretty crazy David Comfort here also pointed out that Halo AI has a new model called t2v 01 director so I went and looked that up and saw that sure enough somehow I missed it this week but Halo dropped a new AI video model as well which allows you to direct the videos with natural language things like zoom out shot or truck left tracking shot or pan right or subtle Shake then tilt up so using natural language you can actually control what the camera is doing in the scene like this one has camera shake this one has a dolly Zoom really really cool stuff haven't had a chance to play with this one yet but I am planning on making some more videos in the near future all about all the various AI video tools that are available we got some updates in the world of AI music this week the company refusion introduced fuzz which is very similar to what you get out of something like sunno or udio but they're making this completely free to use right now they say as long as their gpus survive but I've actually played with this a little bit I did have Early Access to this one and it's pretty impressive I'd say it's on par with everything else that's out there if not a little bit better like here's a pop punk song that I had it [Music] [Applause] [Music] generate kick flips always failed every trick derail down night right and here's another [Music] one I see you at the parking stage knowing up your new man teaching your latest moves Mother breaking up the and so yeah I found it to be really decent music especially since you could generate it for free right now but that wasn't the only news we got out of AI music this week there is a new open source model that generates music like puno and refusion and some of those other tools I haven't actually downloaded this one or really tried it in my self yet it's called you I think the only one know the only one I know you got I mean I don't think it's quite as impressive as the other models that are available out there but this one is open source so people will build off of it and iterate off of it and make it better and really really cool to see these types of tools available in open source versions all right a few last small little things that I want to touch on real quick just going to spend a couple seconds on each one here apparently the new iOS update turns on Apple AI by default I guess a lot of people didn't like that so if you actually don't want to use apple intelligence you have to go turn it off yourself inside of the iPhone now Elon Musk this week claimed that you're going to be able to start doing Tesla autonomous ride shares as soon as June Elon does have a tendency to sort of overpromise under deliver especially on time frames so we'll see if that actually happens however some other news out of Tesla the Teslas now actually are able to drive themselves from where they are built to their designated loading dock without human Intervention which seems pretty cool there's a video they posted at 3x speed of one of the Teslas coming out of the factory and going to its dock just to prove that Tesla's self-driving is getting better and better and this year we could expect to see the Tesla being driven around in full self-driving all the time wh who has cars on the street that are already driving themselves around and taxing people are now getting added to 10 new cities and the next two are Las Vegas and my my hometown San Diego so excited to finally ride in one of these Reed Hoffman one of the co-founders of LinkedIn launched a company this week called monis AI they're working with Microsoft and the company is designed to help discover new drugs probably one of the most exciting areas of AI That's going to likely be the most world changing over the next couple years is the ability for AI to discover new drugs and cure diseases and things like that so super exciting to see that and that's what I got for you today it was a really really busy week I know deep seek sucked a lot of the oxygen out of the air open AI was releasing a whole bunch of announcements I feel like trying to gain some traction and get people talking about them and a lot of companies reacted to deep seek by either adding thinking features to their existing tools or building deep seek into their software but that was obviously a tiny sliver of the whole picture of this week and I feel like 2025 is only going to keep ramping up I mean just January feels like we already got as much news this month as we got in almost the entire of last year it feels like the pace is picking up like that and if you want to make sure you stay looped in on all the latest AI news that's what I'm here for I'm trying to keep up with it as best as I can so make sure that you like this video And subscribe to this channel to make sure stuff like this keeps on showing up in your YouTube feed also be sure to check out futur tools.
this is where I carate all the cool AI tools I come across I also keep the AI news page up Tod date so much of the AI news actually releases on X as you've probably realized from watching my videos and I know a lot of people try to avoid X as much as possible so this news page is a great place to just find the news that's available on X whenever I find the cool tweets or the news articles I share them here and of course we've got the free newsletter where twice a week I will send you an email with just the coolest tools I came across and the most important AI news that I think you should know it's totally free and if you sign up today you'll get free access to the AI income database a database I've been working on that shares all sorts of ways to make extra income using these various AI tools again totally free you can find it all over at futur tools.