Open AI Going OPEN SOURCE? Higgsfield AI Video, Agent Swarms & MORE! AI NEWS

Hello everybody. Welcome back to another AI news roundup. There is a ton to talk about, as always with AI news. But yeah, just the past three days alone have given me almost too much to talk about today. I don't want to waste any time. Let's jump into it. So on March 31, OpenAI closed a brand new capital raising round at $40 billion, the total valuation being 300 billion. And according to CNBC, this is the largest private tech funding round in history. So if you think that businesses aren't funding AI to max capacity and continuing to increase that funding, you'd be wrong. And OpenAI is at the very top of the list. OpenAI wants to release a powerful open weight large language model. Open weight meaning anyone can download it and use it for free. Open weight meaning actually open source, you know, like in the name of the company, Open AI. It has been a long time since OpenAI actually released an open source model. So this is great to hear. I mean, I'm really excited for this. We'll see what kind of license it really has. At the end of the day. There might be limits on commercial use, etc. But regardless, an open source model coming from OpenAI is a good thing. And a lot of people are suspecting this is a reaction to the deep seq R1 effect. The open source large language model that kind of embarrassed OpenAI not that long ago, he states in this post. Before release, we will evaluate this model according to our preparedness framework like we would for any other model. We will do extra work given that we know this model will be modified post release. They're going to be hosting developer events to gather feedback and allow devs to play with the early prototypes. As Plany replies to this post More we want more open source. Keep going OpenAI because we will never stop asking for open source models. Also in relation to this, another post kind of jabbing at Meta AI. They said they're not going to do anything silly like saying you can't use our open model. If your service has more than 700 million monthly active users, they want everybody to be able to use it. At the end of the day, it really comes down to this model actually being useful and actually being special compared to the other open source competitors that are out there. We'll see how this goes, but I am very hopeful for the future. Continuing along with other OpenAI news, ChatGPT plus is now free for college students in the United States and Canada through May. So if you are a student, ChatGPT plus is free for you definitely get this while you can. ChatGPT plus has access to deep research advanced voice mode. Other AI tools in general could be applicable in some way, shape or form to school, but even outside of school AI is a game changer. I couldn't imagine living my life without having access to some AI similar to chat GPT Plus. Moving on to some even more OpenAI related news, apparently image creation in ChatGPT got an upgrade. I mean that's at least what they say. ChatGPT now thinks a little longer to make more accurate and detailed images. I'm not exactly privy here to what they changed or how they upgraded the native image generation, but we do have this from Sir Mr. Meow Meow image Gen got an upgrade. It looks like there is a selection tool available for some, so looks like for image removal tasks or image changing tasks if you want to highlight a specific area. Take this with a grain of salt, though I have yet to see this on my own Chat GPT Even more Chat GPT updates There is a Think Slider now that is accessible at the very least in the app form of Chat GPT where you can slide for a quick response. A little bit of thought behind your response, or a lot of thought behind your response. To be honest, for me, I like to be able to just decide on the model that I'm using at the end of the day, whether it has thinking capabilities or not. At least for now. I think it's really intriguing because Sam Altman said a while ago that GPT5 was going to be more like a new type of interface to interact with AI, where it's going to decide the model it's going to use for whatever task you give it instead of selecting specific models that you want to use. So maybe this is kind of leaning into that. Maybe they're just kind of test driving. This could be a little bit of a sneak peek at some ideas they're having for future iterations of the app that include GPT5. And to cap off our OpenAI news, you can actually get unlimited advanced voice mode OpenAI's actual advanced voice Mode through Microsoft Copilot. So if you're a huge fan of Advanced Voice mode and don't pay for ChatGPT plus, this is a free way to get access to Advanced Voice Mode through Copilot instead. And if you take a listen here, it's definitely OpenAI's advanced voice mode. What do you think? Could we make a good brainstorming duo? Let's get chatting. I can't wait to learn more about you. We can collab on research, creative writing, or brainstorming. This brand new research paper evaluated large language models in a three party Turing test. Essentially how this worked was participants spoke to both a human and an AI simultaneously to decide which is which. And obviously the communication format here is going to be text, so it's not like they're actually speaking words to like an advanced voice mode type model. No, this is all just typing. Regardless though, GPT 4.5 when prompted to adopt a human like Persona, was judged by the participants to be human 73% of the time, which suggests that it does indeed pass their Turing test here at least tricking humans into believing it is also a human. Little bit creepy. You can see the other benchmarks here though that GPT4 omni with no Persona is quite low at under 25%, but we kind of expect that because it's not pretending to be human. GPT 4.5 also does really poorly without a Persona, but quite a bit better than GPT4O, suggesting that model size and how well the model performs in real world tasks does make a difference in how well it can convince people it's also human. I don't know, I just saw this research paper, I thought it was pretty interesting and I'd like to share it with you guys. I if you want to learn more, links to everything as always are in the description all right, let's talk about Google's Gemini. As Min Choi says in this Post, Google Gemini 2.5 Pro is the best AI coding model right now. The general vibes and consensus from the community is yeah, Gemini 2.5 pro is better than Claude 3.7. Anthropic typically wins in coding, but now Google's really taking the fight back. So I'm just going to run through a couple of examples that Minchoi has rounded up here. Someone created a whole office simulation game with Gemini 2.5 Pro, which is pretty crazy. You see all these autonomous little NPCs in the office are moving around and working throughout the office space, which is pretty cool to see. Here's a pretty insane 3D physics simulation that Gemini 2.5 Pro built inside of 3Js, so that is wild as well. Here's a procedural building generator also made in 3js by Gemini 2.5. I'm showing you guys all the cool visual stuff right now, but I don't want to spoil the whole thread, so check out Minchoi's roundup. And of course, if you want to see more really insane examples from 2.5 Pro. But I want to talk a little bit more about Google's 2.5 Pro and a brand new model. Even that seems to be making the rounds today. First of all, Epoch AI tested Gemini 2.5 Pro for themselves and found that it really does have a top charting GPQA diamond score of 84% which exactly matches the result that they reported themselves. Which leads us to believe that, yeah, the benchmarks for 2.5 Pro are legit and are extremely impressive. You can see O3 Mini High is lagging a little bit behind 01 is also lagging behind 3.7, Sonnet, Deepseek, R1 and of course older Gemini. As Chubby points out, Google really did cook with this newest model. But there's another model. This model appeared on the LM arena for benchmarking this morning and it's code named Nightwhisper. It is confirmed to be from Google in the metadata on the website and some say it's a state of the art coding model, possibly a Gemini 2.5 coder, which would mean that it's even better at coding than 2.5 pro general is, which is wild to think because not only did 2.5 Pro come out of nowhere and essentially become the community's go to coding model, but now it's already going to be replaced by what, another Google model? I hope this is the case because AI coding has so much potential and way, way farther in potentiality than it does right now. If we have AI models in a couple of years that are too three, four times better at coding, I mean you would be shocked to see what people would do with that kind of technology. I think either way, Leo here, who has been testing it out, has noted that it feels worse at general tasks than regular 2.5 pro, but more polished for code possibly. Let's continue to talk about large language models. This is a diffusion based large language model. You see typical large language models like any model from OpenAI, Google, etc. They're all like these auto regressive AI models. This is diffusion, like an image generator. Yeah, a lot of image generators, except for the brand new Native 1 from OpenAI are diffusion based, essentially meaning they start from like a random data seed and then work backwards from that to produce whatever AI based result. To put things simply at the top here you can see the text starts with write a story that ends with finally Joey and Rachel get married. You can see the LLM starts with that piece Finally Joey and Rachel got married. Married and then fills in all of the Gaps over time in between. These diffusion based LLMs are really intriguing to me. This one is called Dream 7B and again it's very small at only 7 billion parameters. But it's actually the most powerful open diffusion large language model we've seen to date. But it gets even more interesting because this dream 7B model outperforms existing diffusion large language models by a large margin, matches or exceeds top tier autoregressive large language models of similar size on general math and code coding abilities, and even demonstrates strong planning ability and inference flexibility. So you think by switching to diffusion for large language models we'd see some sort of dip in other places for the gains that we see like the native text planning or editing. But actually no, this thing benches pretty strongly for its weight class, which is really intriguing. Here's an output synthesized in an arbitrary order, for example, you can see it randomly sort of starts to fill in the bits and pieces to give you the full result instead of of again going line by line infilling with an exact sentence at the end. That's what I showed you guys in the beginning. Another really interesting part about this being a diffusion model, we can actually adjust the performance, either changing it for more speed or quality by simply adjusting the time steps. Instead of building a whole separate model, we can just adjust the setting on this one to accommodate our need in that moment. I really hope that the big companies out there are working on diffusion large language models, because this could open a whole new layer and capability set for large language models that are just based on diffusion rather than autoregressive techniques. Alright, stepping away or a little bit away from large language models, there is a brand new paper out demonstrating that a woman who is mute, cannot physically speak, can actually now communicate just through her brain alone. No typing delay or sounds made. It just takes the neural intent and starts streaming speech out. Really incredible. Basically, mind reading is what this is. Obviously this uses AI techniques to accomplish what it does, but essentially the person thinks about what they want to say. This device reads the data from the brain somehow. Of course links are going to be in the description if you want to check out the paper for yourself, but it uses multiple language models here to actually turn the patterns getting streamed from the brain into genuine text. Just absolutely wild. Obviously you know there's like a pretty large plugin setup for this at the moment, but 10 years from now this could very well be portable and commercialized. People with disabilities all around the world are going to no longer have disabilities due to AI technologies and incredible medical Papers like this. So this is really awesome and I really thought this was worth sharing with you guys. Alright guys, let's talk about AI video because there is so much to cover. First of all, just a quick recap. Gen 4 from RunwayML released and it's pretty darn good. I've been messing around with it. I did a whole video on it. It's awesome. I think it's pretty close to competing with Google's VO2. I think it's probably lags behind VO2 a little bit, but it's good enough. And man, the generations are fast, crisp, with good movement and decent controllability. You can see this is a pretty fun prompt from Crystal Ball here where the guy on the sign here kind of jumps off and comes to life. He jumps down onto the ground, he starts walking around and I love the lighting and the way that the light interacts with the ground as he's walking around. I mean it really looks professional and realistic. Crystal Ball also notes that this was one shot first try with this prompt. But Gen 4 is far from the only new model on the block. Higgs Field AI has made its debut and it is quite a bit different from traditional video generators we've seen. Here are some examples of different camera angles that Higgs Field has been able to replicate or recreate with their brand new video generation model. You can see this is a dolly zoom in. So the camera is actually coming closer to the subject physically while widening its own lens. But of course this is all generated with AI. It recreates the effect pretty astonishingly. Like really impressively accurate. Honestly. They have the crash zoom in so a really close and fast zoom in onto a character's expression trying to emphasize their emotions. They also have stuff like the traditional dolly in and you can see the quality of the generations are also quite high here. I mean the people look great. Here's a dolly zoom out, the reverse of the zoom in. Or how about a 360 degree orbit around your character? Also pretty insane. Or a camera roll, actually tilting the camera crane over the head. Drone shots, which are actually really easy to do in most video generators. Head tracking, I mean, you name it, they've got it. And yeah, good detail, good generation quality. Overall competitive and state of the art. So I think you guys are probably starting to notice a pattern here. With the Higgs Field AI. It's all about camera control and camera movement. You can see you can even take a gen 4 generation and actually put it right into Higgs Field and then get a bunch of crazy camera Movements based off of that original generation. So these tools are competitive with each other to some degree, but they also work together in very interesting ways. Here is their official post. You can see all kinds of pretty awesome and crazy camera movements going on with this, but another piece of this is gore and blood is also allowed. This thing seems to be pretty darn uncensored, which is enough to swarm their servers on launch, which definitely did occur. I'm looking into doing a larger, longer video on this one when things calm down on their server farm. But here's just a short list of motion controls they have right now. All different kinds of cinematic shots, or even for things like music videos, sporting events, tentacles. I don't even know what that is. Basically, the model that they built here not only generates state of the art, high quality AI video, but it also is able to capture the camera movement from another input video. At least that's what I'm inferring. Now, none of this is open source yet. It's all through their website and they have paid plans, so that's the direction they're headed in now. But I'm hoping that they open source some of this technology in the future so other models and open source frameworks can develop that allow for intense motion control like this. Also related to AI video, Meta announced Mocha. It's a new model that turns text or voice into realistic talking characters. So I think all of these examples you're about to watch were originally sound clips from movies, but. But all of the video is generated by this new Mocha model. It listens to the audio and then generates the person speaking. So what we're going to be looking for here is the lip sync quality overall and you know, the video quality. Does it look good? What are you going to do? I got to follow my own plan, stay here, get my own thing. I said it before and I'll say it again. Life moves pretty fast. You don't stop. You are so beautiful and so smart and it kills me that you don't. Don't forget to check out my stories for those extra bonus codes that you're not going to want to miss out on and all. It's kind of funny to see like these random clips just get generated, but these people, these characters look very real. I can't believe the smoke coming off of this cigar and the way this guy looks around while he talks. It seems very human. It seems very emotive. I think this is some nice research from Meta relating specifically to lip sync and AI video generation. But. But I Do have to say there are a lot of other tools out there not only for video generation, but also for lip syncing as well. So I think this is definitely state of the art. I think this is high quality, but they're not the only players in the game by a long shot. All right, now we're going to move into some more rapid fire type of stuff. Cocktail peanut announces ComfyUI one click launcher now finally supports RTX 50 series. So if you've got a 50905080 or 5070 and you want to use ComfyUI in all of its local AI glory, it is finally fully supported, which is great to see. I've actually been waiting for this personally so I can finish up talking about the RTX 50 series and how it affects AI workloads. Thank you. Cocktail Peanut, a master of the open source arts. Here's a brand new model called Unidisc. It's a unified multimodal discrete diffusion model. The researchers trained this sucker from scratch on 250 million image caption pairs using a discrete diffusion objective. So this model has the benefits of diffusion models, but now in multimodal space, meaning it can also give the text as well. The compute versus quality trade off is flexible. It does zero shot in painting and editing with better control. And the great part is they're open sourcing everything. Their code weights and training data set are all open. They've also got a demo on GitHub to test out the zero shot multimodal inpainting, which is pretty cool. No fancy tricks here, it can just do it. They also show off some native editing and we're even able to edit a couple of faces. So nice work. Love to see open source models pushing boundaries like this. Check this one out. Next up, we've got a new AI agent that is lightning fast. Let's just go ahead and show you the demo clip here. We ask it to apply for a marketing job on LinkedIn and this is all real time. The AI goes to LinkedIn real quick. It's able to navigate at unprecedented speed, click around, type new things, navigate the website and apply for a job for you all autonomously. I mean this thing seems pretty accurate, able to complete the task, but the main thing here is the speed. It does this in, I don't know, about 25 seconds or so. It's able to apply for the marketing job. That's faster than a human would be able to do. For these AI autonomous agents, speed has been a really large hang up and this one seems to have solved it. To some degree. So what exactly is this sucker? Well, it's actually ace. This is real time. It's booking an Airbnb in only 15 seconds, goes right to the website, clicks around, is able to book it for specific dates in a specific location, all in 15 seconds. That is crazy to me. Apparently it's not just running large language models on user interface trees, but it's also doing ultra fast vision recognition and action prediction. As DD points out here, if large language models are able to utilize this new ACE general agent, this could be a serious game changer for how we end up using our computers. Now, getting it into the hands of everybody and allowing everyone to be able to utilize this in an easy way is another story, but it's coming down the pipeline whether you like it or not. Fast, reliable, autonomous agents are pretty much here, guys. Next up we've got Lindy Swarms. Also AI agent related news. I covered Lindy a couple of months back and that video actually did pretty well. So for those of you not familiar, Lindy is essentially an AI agent platform. It allows you to build these AI agent workflows and now they have have hundreds of things able to be done at once with Agent Swarms. It's brand new. They have 5,000 different integrations through Lindy and 4,000 different types of web scrapers flow. The CEO of Lindy here says that with Agent Swarms, Lindy can divide and conquer, creating hundreds of copies of itself for specific individual tasks and then perform them reliably at the same exact time, accomplishing massive tasks seemingly effortlessly and very quickly. I'd love to know how this actually works in practice. Agent Swarms are definitely something that we have been thinking about in considering with AI for a long time now. How expensive is this under the hood and how well does it really work? Don't know. We might have to do a full video diving into that. All right, now for some image Gen news from Midjourney. They just ended their relaxathon. They are preparing their server clusters for a V7 midjourney model launch. It's been a long time since we got a brief brand new midjourney model and I don't know, I'm not super hopeful that they're going to be able to compete with the likes of the latest native image generation from OpenAI that has insane prompt adherence and still fantastic visual quality. Midjourney is known for its aesthetics and detail and artisticness, I suppose, but is that enough to carry it? People have enjoyed the aesthetics of Midjourney for a long time now. And a brand new model isn't going to give you magically better aesthetics, right? Right. Unless somehow they figured that out. Definitely excited to check it out when it launches. Who knows, they might blow our minds here. This is really interesting. Higs Field, which we just talked about earlier, says that AMD now beats Nvidia in inference time for image to video generation. They claim 20% faster and 35% cheaper when using their HiGS Field DOP model. Not to mention the AMD card here. The GPU running all of this has a very easy setup in comparison. This AMD card also has increased VRAM allowing them to generate longer high resolution video without memory optimization. And not only that, this new AMD chip is cheaper as well. So this is a really intriguing competition coming from AMD against Nvidia for AI video workloads. I don't know, we'll see if this can be applied to other models and other video generators, but for now Higgs Field is leveraging it. I'd love to see more competition from AMD in this regard. Nvidia really has dominance in the generative AI space as a whole, so love to see it amd. Keep up the good work. Devin 2.0 is also here. It's now a collaborative planning partner works with you to scope out a task before execution. Devin Search helps you with questions about your code base. It's also able to now create detailed architecture diagrams and link to specific sources. So yeah, a lot of AI agent launches this week. Finally guys, to end this off, Quen 3 is scheduled to be released in the second week of April. Another big announcement and release for open Source. Hoping that Quinn3 here is super competitive for being open source and competes well with models at the same weight class. Talk about AI news, right guys, all of that was in like three days. Everything that we just talked about happened in like the span of three days. So AI is just blisteringly fast. This train isn't going anywhere. More money, more development, more building. It does not stop. If you want to stay on the pulse of this stuff, definitely get subscribed. Check out my Discord server as well for the real time live updates. Or you can also check out my Twitter as well. I retweet pretty much all of these different pieces and oftentimes give my own take and opinion on them. Thanks so much for watching folks and I will see you in the next video. Goodbye.