AI News - New Models From Google & OpenAI , AI Drama & Humanoids In Factories

So, let's take a look at all the air news you missed this week. So, let's get into it. So, one of the starter pieces of news that you probably may have missed is the fact that Nano Banana 2 is finally here. And so, this is Google's second iteration of their image generation model that you probably did miss unless you were paying close attention. It is of course free in the Gemini Pro plan, but most people wouldn't have realized this because Nano Banana has been pretty good since it release. If you are wondering what those changes are, take a look here. It just essentially just adds that real level of detail. So, you can see it adds advanced world knowledge, precise text rendering and translation, 4K upscaling, which they actually had before. And you've got aspect ratio control, subject consistency, up to five characters, and 14 objects. Now, that wasn't the only thing from Google. that was released by Google. You can see they released something else, and this is really cool. I will have a video on this probably tomorrow, but this is the Notebook LM cinematic overview upgrade and it is very good at creating videos and information that you probably want to use if you are

trying to make videos on, you know, a variety of different things. Now, if you aren't familiar with what Notebook LM is, it's essentially an, you know, educational tool that people use to learn a lot of things and it's really good. And so in this update, you essentially get cinematic video overviews, which includes animations and motion graphics. And I really don't know what Google is using under the hood for this because I'm not aware of any video model that can create motion graphics on demand like true AI video motion graphics, not those, you know, other ones that you've seen. And it's really good. It does take a very long time to generate the videos. I waited hours for a few. But I think if you're using Notebookm and if you have the pro version of the model, then try and test this out. And by I mean pro, I mean if you have Google Ultra plan, that's the most expensive plan, $250 a month or $200 a month. Not sure exactly on the pricing here, but if you have that plan, then you're essentially going to be able to access this feature. So it's not, you know, out for anyone else just yet. And I think it's because it does take a large amount of tokens. So, it will be interesting to see what people do create with this. Now, of course, that wasn't just it from Google. Google continued to release Amazing AI after amazing AI. You can see here that they also released Gemini 3. 1 Pro. Gemini 3. 1 Pro is Google's flagship Gemini model. And

remember, this is a model that is designed as a natively multimodal high reasoning system for complex professional and developer tasks. Now, you got to think about this model, okay? This model was released very recently and it at the time was of course the best model. And if you're wondering this model released, that model's released, what is going on? Of course, remember that the Google models are essentially models that are basically the best multimodal models. Most people don't realize that these models can take in video, audio, images, pretty much any kind of thing you can throw at it, it can deal with. Most other AI models, they just don't have that capability. And this is basically now even stronger. So the one that you kind of want to focus on there is the MMU Pro and it actually got bumped up to around 76. 8. So it's a pretty good natively multimodal model and of course that's because Google they're trying to build world models. So it's got, you know, a lot of things going for it. It's got strong reasoning thinking modes, better reliability. It's got longer, more structured outputs. It's got a large context window around 1 million tokens. As you know, it's got long documents, large code bases, long audio, long video. I mean, it supports function calling, search grounding with Google search. I mean, it's pretty allround uh really good model, especially if you are already in the Google ecosystem. So, if you didn't know that this was there, for sure go ahead and take a look at this model. Now, of course, if we are talking

about new model releases, it makes sense to talk about what OpenAI did. So, OpenAI, they have released their own model and their model is GPT 5. 4 Pro. So, 5. 4 Pro is currently the smartest model on the planet. And before you think, oh my god, I'm going to have to just change from Gemini 3. Pro to GPT 5. 4 Pro, understand that like it really does depend at this level what your task is. Every model is going to have a set of things that it is just inherently better at. That's why when I previously discussed Gemini 3. 1 Pro, I said it is a natively multimodal model. So if you're someone who's constantly dealing with images, videos, and audios that you need AI to use, Gemini is going to be your best bet. But GBT 5. 4 Pro, this is a model that just excels on the very edge cases of reasoning. So for example, Frontier Math, it is just completely dominating. For computer use, it is currently dominating. And if you're wondering about those super hard scientific problems, that is also where GPT 5. 4 Pro dominates. So I would say that if you're someone that maybe you're a scientist, maybe you're studying, maybe you're doing really hard technical work, GPT 5. 4 Pro is going to be that high performance max capability model that is going to really help out those who are doing high stakes professional and enterprise work. So if you're in that category, this is going to be the model for you. Of course, you can use whatever model that you think is right for your specific workflow, but as far as I know and from the research and all that I've seen, this is where GBT 5. 4 Pro fits into the mix. Now, you should have this already. This is available in chat GBT4 Pro and enterprise users via the API or just the standard chat interface. And so, yeah, it is pretty good. And OpenAI are not playing games when it comes to their model anymore. And another thing that they did do as well is they did also manage to fix their reasoning mistakes. So prior models were pretty annoying if you did use them. I know that GPT 5. 2 was just absolutely terrible when it came to standard conversations. So I would say that GBT 5. 4 Pro, as far as I've seen, it manages to fix those annoying inclinations. However, you will have to test this on your specific use cases, but I just think that it is much better now in terms of standard responses. Now GPT 5. 4 4 Pro wasn't the

only model released that there was. Another company, Microsoft, decided to release C-pilot tasks. So, they launched this at the end of February, and these are described as a to-do list that does itself. So, you describe what you need in natural language. Then, C-pilot plans and executes it working in the background with its own computer and browser across various apps and services, then reports back when done. Now the big shift Microsoft frames it as the start of a new phase. Conversational chatbots were essentially the first chapter of AI and today is the beginning of a second. Drawing distinction between systems that reply and systems that carry out work. So what can this actually do? Well, the real world examples include surfacing urgent emails with draft replies every evening and auto unsubscribing from promo mail. you know, tracking apartment listings and booking showings, combining a Monday morning briefings on meetings and calendar, turning a syllabus into a full study plan with practice tests and blocked focus time. Now, of course, it does have some guardrails. It is designed to ask for consent before taking meaningful actions like spending money or sending a message on your behalf. And you can review, pause, or cancel a task at any time. Now, currently this is in a limited research preview with a public weight list expanding over the coming weeks before a broader launch. This is basically Microsoft's answer to perplexity computer and claude computer use. It's the same Agentic AI wave, but this one is aimed more at the mainstream and consumer users rather than developers. Now, before I talk about Plexity Computer, which is a pretty cool tool, we have to talk about Microsoft. Okay? And you know, we just talking about what Microsoft released, but Microsoft is

actually tired of this AI term. So Microsoft has this term that is currently plaguing their Discord, which is called Microsop. So I'm pretty sure everyone and their grandma has heard of AI slot by now. And so currently micro slop has been the derogatory name for Microsoft. And this is mainly used by people that are annoyed at Microsoft's aggressive AI push into Windows, Office, and other products, implying the company is serving up low-v value AI slop instead of useful software. Now, Microsoft, what that means is just, you know, a combination of those two words, and they're basically saying that Microsoft are putting no effort into their products. Now, the term has been around for quite some time, but it recently just exploded because many people are like, "AI slop, Microsoft, it's pretty easy to make that connection. " Now, it gained even more traction after Microsoft banned the word on the official Discord server.

So, of course, you know what internet trolls like. They're going to continue to use that word in pretty much every social platform that they can. So, yeah, pretty crazy. Now, of course, it's time for Perplexity Computer. This is a generalpurpose digital worker that operates in the same interfaces that you do. It reasons, delegates, searches, builds, remembers, codes, delivers. It's capable of running workflows for hours, for months. It's not just a chatbot. This is a genuinely impressive digital worker. It essentially uses 19 AI models and this is pretty crazy. They use Claude Opus 4. 6 as the core reasoning engine, and that orchestrates sub agents with the best models. Got Gemini for deep research, nano banana for images, VO3. 1 for video, Grock for lightweight tasks, GPT 5. 2 for long recall. It's essentially a model agnostic orchestration layer that picks the right AI for each part of the task automatically. And when it runs into a problem, it creates sub agents to solve it. And each task runs in an isolated compute environment with access to real file system, a real browser, real tool integrations. This thing is very impressive. It's kind of like the perplexity version of openclaw. It can research, design, code, deploy, manage your project end to end all from a single conversation. It has memory. It can remember your past work. It can connect to hundreds of services and it runs securely in the cloud. It can put out live websites, financial analysis, data, you know, visualizations, build full web apps. I mean, it's pretty crazy. It does some of my work. Not all of it, but it does enough. I mean it's being compared to like I said open claw but the key is that perplexity is like the multimodal orchestration and of course it doesn't have any safety issues. Now remember this tool is of course pretty pricey I guess you could say because it is $200 a month on the perplexity max tier. It runs entirely on the cloud and you get you know 10,000 credits per month plus a onetime bonus of 20,000 credits. So it's pretty effective if you're you know a power user. I would say try this out. If you're a power user that is pretty non-technical, like you like AI, but you don't want to deal with like the CLI coding and all of those agents and setting that stuff up, this is probably perfect for you. And yeah, it is uh it really is useful and I think more people will use this as time goes on. Now, if you're watching this channel, you're clearly someone who follows AI seriously. I do have a newsletter where I share what I think is worth paying attention to each week, not just what's going viral, and it is free. Link in the description for top tier AI news. Now

if we're going to get into more company drama, we need to look at what happened to Anthropic. So, Anthropic have been in a basically a tornado because they have just been undergoing just crazy stuff and there is literally no other way to put it. I mean, if we look at what Enthropic has been in the last week, it has been the craziest time and they've literally set history for one of the things that happened to their company. So it all started with this post and didn't really all start there but that was part of it like the major thing that was the turning point and the fact is that Anthropic basically said to the government look we aren't ready to let you guys use LM for two things one mass surveillance and two autonomous weapons and so when the Pentagon submitted a final offer Dario Amade basically publicly stated that they cannot in good conscience allow claw to be used for these things and then Trump posted this on Twitter. Now remember this is the United States president and he basically said they're never going to allow a radical woke company to dictate how their great military fights and win wars. That decision belongs to the commander-in-chief. So, I mean, you know, the thing here, I'm not going to read into all of this anthropic slander, but the thing that he said here that was, you know, truly incredible was that he said that every federal agency must immediately stop using anthropics technology and there must be a six-month phase out period. And he's threatening civil and criminal consequences. And if Anthropic doesn't cooperate during the transition, he's thinking the full power of the executive branch could be deployed against them. And I mean it it's pretty crazy that this has happened in AI. I didn't personally expect this for at least maybe one or two years, but I mean currently they're still in talks, but the relationship isn't great between these companies. And this doesn't bold well for the future. I mean, you have to have a good relationship between the top frontier AI labs and the US government if you're going to move forward as a country. When you designate a company like anthropic a supply chain risk, you are just completely souring that relationship. And I don't know exactly how this will move going forward because Anthropic is still in discussions. And I do know that their model is incredible in just terms of its general reasoning. So it will be interesting to see exactly what happens here because they've set some crazy precedent. But yeah, it is a wild time to be an AI right now. Now

that wasn't the only drama. There was also Quick GPT and an estimated 2. 5 million people have stopped using chat GPT as the Quick GPT movement has gained traction. Now, if you're wondering why this movement started, there were several factors. And if we're being honest, this movement could have started back in 2023 or even 2024 because of all the things that OpenAI have gone through. Now, the original big spark that kind of sparked this entire wave of backlash was Greg Brockman's $25 million donation. So, there were FEC filings that revealed that Opening Eye President Greg Brockman made a $25 million personal donation to Mara Inc., a proTrump supporter. And now, that triggered the initial backlash, especially among the left-leaning creative tech crowd who made up a big chunk of ChatBT's early user base. Now, there was the ICE angle on top of this where the Department of Homeland Security AI inventory published in January 2026 revealed that Chat GBT was being used in an ICE resume screening tool and that kind of made things a bit worse. People felt that their subscription money was directly funding immigration enforcement and then of course you had the Pentagon deal and this is what made it go truly mainstream. You know, remember how we literally just discussed that Anthropic drew a line when it came to allowing its AI to be used for mass surveillance and fully autonomous weapons? The Department of War wasn't going to agree to that, but then OpenAI basically stepped in and signed the deal instead. And that was the moment that things basically just exploded. OpenAI essentially took the contract that Enthropic refused on ethical grounds. And CHP mobile app installs jumped to 295% day overday. And Sam Alman later said that we shouldn't have rushed to get this out Friday. The issues are super complex and demand clear communication. And of course, there's also the fact that chat typ has gotten pretty bad with the 5. 2 model. And that's damaging their reputation because a lot of people did use it for their daily workflow. And there were many celebrities behind this as well. You literally had people like Katie Perry saying that you need to sign up for Claude. And this is I think one of the first times that Claude gained the mainstream conscious. I even had friends reaching out saying, "Hey, have you used Claude? " It was pretty crazy. Now, of course, we had OpenAI firing an employee for prediction market insider training. This was also pretty interesting. If you aren't familiar with what's going on there, prediction markets are essentially a way to bet on the future. And those markets tend to have things about AI. They'll be like, which AI is going to be the top model for the month of March? Which AI is going to be released on this benchmark? Will it have this value? Will it have that value? So, OpenAI managed to find someone and they managed to fire them. Now, if we're going to continue into the AI drama, we can look at the massive story on the Quen team disbanding. So, if we look at

the full breakdown here on March the 3rd, 2026, Ling Yang Yan, known as Justin, the technical lead of Quen's team, formerly submitted his resignation, and the news spread, prompting strong emotional reactions. On the same day, Yu Bowen, head of post training, also departed. And Hubin, head of Quen Code, had already quietly left to join Meta back in January 2026. And so, if you're wondering why Quen's team has disbanded, their lab has essentially planned to break up the Quen team, moving away from a vertically integrated structure towards a horizontally structured system with separate teams for pre-training and post-training, text, multimodal, and other functions. And this directly contradicted Lynn's own philosophy which held that these teams should be more tightly integrated not separated. Essentially what you had here was just corporate restructuring clashed with the technical vision of the person who built it. Now let's get into some robotics news. We had FSME and so this is a memory system from Stanford that lets robot AI learn physical principles in real time without retraining the model. It's essentially giving robots the ability to learn by doing exactly as humans do. The core

problem that it solves is that VLMs, which are the AI brains controlling the robots, understand physics abstractly. They know what friction is, but can't predict how a specific ball rolls on a specific surface without trying it. FSM fixes this gap between the book knowledge and the real world experience. So, if you want to know how this really cool thing works, it uses a three tier memory system. It's episodic, so it stores raw experiences, what happened. It then generates testable hypothesis why it happened and then it promotes verified principles of what to do next time. Key insight is the verification before application. It tests a hypothesis before committing to it as a rule, avoiding the trap of rigidity followed by following outdated experiences. Now, raw experience retrieval had a 23% success and FSM's principled abstraction had a 76% success. A real world improvement over a 30inut session which was consistent and measurable. Now, the big picture is that this is essentially a robot that's developing intuition through trial and error. Same way that a human learns to pour a drink or stack blocks. And this is a direct step towards robot that gets smarter the longer they operate, which is of course a fundamental requirement for real world deployment. The fact that learn principles are human readable and transferable makes it practically useful and not just academically interesting. This of course wasn't the only robotics, you know, update. There was also physical intelligence, arguably my favorite robot company. And this is the

most wellunded robotics AI startup right now. This is a company that's backed by Bezos, OpenAI, Seoia, Kosler, and others. And their entire mission is building foundational models for robots the same way that OpenAI built foundation models for text. And what they just released was MEM multiskll embodied memory. And essentially they are combining short-term visual tracking with a long-term narrative in natural language. Their latest models can now maintain focus for up to 15 minutes long enough to clean an entire kitchen or prepare a meal from scratch. And if you're wondering how this works, the architecture splits memory into two distinct modes. Short-term visual memory uses an efficient video encoder to capture dense imagebased memory of the last few seconds. For the big picture, the model summarizes semantic events in natural language. Instead of remembering every frame of a door opening, it simply stores a note like I open the fridge door. And then this textual memory is updated via the chain of thought process as the robot completes each subtask. Previously, if the model failed to grasp an object, it might try the exact same failed strategy repeatedly. But with the memory, the robot exhibits context adaptation. In one demonstration, the robot attempting to pick up chopstick from an unusually low table. it was like failing and then it remembered the failure in the short-term you know buffer. It adjusted the approach on the fly and succeeded on the second attempt. So once again we're starting to see robots get memory. Now something

interesting is here. In early February 2026 Faraday Future announced the formation of FFAI Robotics Inc. and they've officially launched Embodied AI robots at the NADA show in Las Vegas essentially copying Tesla Optimus' playbook. So they have three different products. They've got a full-size professional humanoid starting at $35,000 plus a $5,000 ecosystem skill package. Then they've got the FF master and athletic action master humanoid starting at $20,000 plus $3,000. And then of course they've got the Aegis, the quadriped security companion robot. Now it's pretty interesting, but there's kind of a problem because they actually have just rebranded Chinese robots. Now, some people have noticed that these models are pretty similar to the AGI Bot A2 and X2 and the FF futurist shares specific specs with the AI bots hardware. So, I mean, it's pretty interesting to see what this robots company is going to do. They are going to be able to actually have robots that work and evolve from the Chinese counterparts or are they just rebranding them? I don't know. Robotics is a very interesting place. We haven't seen too many humanoid updates just yet. I mean, if you actually want to look at real humanoid updates, we had BMW deploying their first humanoid in their European plant, it was very interesting to see

how they were able to, you know, have this robot here. Actually saw this tweet. I'll leave a link in the description, but the robot is actually doing work. I mean, they're trying to explore all the different, you know, AI explorations in terms of how you can actually use these robots within the factory to get more done. I do think of course it's not there yet, but it still is interesting to see what we could have in the future when it comes to these humanoid robots.