HUGE Gemini 4 Carbon LEAKS, Claude Fable 6, GPT-7 Bel, Google CRACKED RSI?! & More! AI NEWS

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: HUGE Gemini 4 Carbon LEAKS, Claude Fable 6, GPT-7 Bel, Google CRACKED RSI?! & More! AI NEWS · Channel: WorldofAI · 31:03 · Watch on YouTube

Topics: AI agents, Software engineering with AI, AI models, Local and open-source models

Summary

The video rounds up a wide range of AI news, beginning with unverified claims that Google is testing a Gemini 4 Carbon checkpoint and may be approaching recursive self-improvement, alongside rumors of future models from Anthropic and OpenAI. It covers new products and features from Google, Microsoft, OpenAI, Qwen, and Anthropic, including image generation, decision-making, coding agents, and multi-agent workflows. Other stories include Chinese companies developing AI models, reported misuse of Anthropic’s startup program, and GPU access for Google AI subscribers. The video closes with experimental projects using AI-assisted coding to bring Nvidia GPU support to macOS and explore PS5 software on PC, while emphasizing that both remain works in progress.

Key points

Unverified reports say Google is internally testing Gemini 4 Carbon, which one employee reportedly compared to Claude Opus 5.5 for coding; the comparison has not been independently confirmed.
The video discusses speculation that Google, Anthropic, and OpenAI may be making progress toward recursive self-improvement, while noting that the claims remain unconfirmed.
A rumored Anthropic pre-training run may be for a future model beyond Fable 5.5, while OpenAI’s rumored model codenamed Bell is linked to recent mathematical research results.
Qwen Image 2.1 Turbo generates and edits images using eight denoising steps, and its model weights are openly available.
Microsoft introduced Decision 1, a model designed for fast, structured decision-making tasks.
OpenAI announced Codex composer predictions and a Windows sandbox mode, and made it possible to create Dots from the ChatGPT mobile app.
Anthropic’s managed agents dynamic workflows coordinate multiple agents; the company reported that its multi-agent workflow found more planted bugs than a single agent in its tests.
The video reports AI model projects from Chinese companies including miHoYo, Xiaomi, and Meituan, and alleged misuse of Anthropic’s startup benefits program.
Google AI Pro and Ultra subscribers reportedly gained access to Nvidia GPU resources through Colab, subject to availability and compute limits.
Experimental projects use AI-assisted development to pursue Nvidia GPU support on Intel-based macOS systems and PS5 software compatibility on PC; neither achievement means broad compatibility or full PS5 emulation is solved.

Chapters

Start	Chapter	Summary
00:00	Gemini 4 Carbon and RSI rumors	The video opens with unverified claims about Google’s Gemini 4 Carbon and speculation that Google, Anthropic, and OpenAI may be approaching recursive self-improvement.
04:06	Google’s Gemini checkpoints and self-improvement	The host expands on reports about Carbon, its rumored coding performance, and claims that AI may be helping Google accelerate model development.
08:15	Anthropic and OpenAI’s rumored next models	The video discusses an alleged new Anthropic pre-training run and rumors connecting OpenAI’s Bell model to mathematical research and AI-assisted research.
10:51	New models and coding tools	Stories include a Gemini AI Studio ultra mode, Qwen Image 2.1 Turbo, Microsoft’s Decision 1, and new Codex features.
14:58	OpenAI updates and a free model	The host covers Codex composer predictions and Windows sandboxing, creating Dots on mobile, Devon access through ChatGPT plans, and Solar Mini 4.
19:40	Anthropic agents and China’s AI competition	The video reports on Anthropic’s multi-agent workflows and Claude Code Projects, then discusses Chinese companies entering AI development.
23:49	Startup program and Google GPU access	The host covers reported abuse of Anthropic’s startup benefits program, GPU access for Google AI subscribers, and an Opus 5.5 game demonstration.
27:27	AI-assisted macOS and PS5 experiments	The final news segment looks at experimental Nvidia GPU support for Intel-based Macs and a project exploring PS5 software on PC, both developed with AI coding assistance.

Quotes

feels like Opus 5.5 for coding — 00:30

a cryptic expose and faster development cycle doesn’t necessarily confirm that Google has achieved the true RSI claim — 07:44

the model weights are openly available on hugging face — 12:23

the multi- aent workflow consistently found 66 out of 70 bucks in all three runs. — 20:43

This doesn’t mean full PS5 emulation is solved — 30:01

Concepts and entities

Gemini 4 Carbon, Gemini 4 Argon, Claude Opus 5.5, GPT-7 Bell, recursive self-improvement (RSI), Qwen Image 2.1 Turbo, Microsoft Decision-1, Codex, Claude managed agents dynamic workflows, Claude Code Projects, Google, Anthropic, OpenAI

Full transcript (original language)

00:00 What an unexpected turn of events cuz just yesterday we were talking about Gemini 4 Argon and some of the latest updates surrounding the model. But now it looks like Google is already working on something potentially even bigger. According to the new insider claims, Google’s latest internal dog food checkpoint reportedly code name Gemini 4 Carbon is already compared to Claude Opus 5.5 when it comes to coding, which is an outstanding claim. One Google

00:30 employee also allegedly tests the model set carbon [music] feels like Opus 5.5 for coding. Although they admitted they need to test it more. Think about that. Gemini 4 Argon hasn’t even launched and Google is reportedly already testing out another model internally. Now knowing Google and how often they struggle to deliver on their expected timelines, I’m taking this with a grain of salt until we actually get our hands on these models. It’s all speculation. But that’s

01:00 not even the craziest part. As this news was circulating on X, another claim emerged suggesting that Google may have already achieved recursive self-improvement. According to the unverified insider reports, Google has either cracked RSI or is getting extremely close. And some are pointing to Google’s rapid progress with the Gemini Flash series and claims that its models have gone from roughly GPT5 level capabilities to something that is comparable to Opus 5.5 in just months.

01:30 Now, rapid improvements doesn’t necessarily prove that RSI has been achieved, but if these claims turn out to be true, the implications would be enormous. And something to even add on to that, one of Google’s employees did even validate that claim on X, which is kind of funny. But moving over to Enthropic, things are getting really interesting cuz according to another unconfirmed report, Enthropic reportedly started a new pre-training run about a week ago. And this supposedly isn’t Fable 5.5, but rather a future model

02:02 that could become Fable 6 or potentially something even bigger as Enthropic looks to compete with OpenAI’s rumored next generation model, which is the GPT Bell. And speaking of OpenAI, insiders are also claiming that Bell, which might have another internal code name, is behind the impressive mathematical research results that OpenAI has recently published. There are even suggestions that AI assisted research or RSI may have contributed to those breakthroughs. Again, none of this has

02:33 been independently confirmed. But here’s the bigger picture. We’re hearing claims that OpenAI, Enthropic, and Google could all be approaching major breakthroughs in RSI. If even part of that turns out to be true, the next few months could be absolutely insane. And that’s just the beginning of today’s AI news. We’ve also gotten the new Quen image 2.1 Turbo model, which can reportly run locally. Google is also introducing the new Gemini reasoning modes, including ultra mode, which is potentially preparing for Argon. And Microsoft’s new decision one

03:05 model is out for fast decisionmaking, which is somewhat of a duplicate of Jev. On top of that, we’ve got major codeex updates, cloud code improvements, a new free AI model, and two absolutely wild developments involving AI assisted software engineering. First off, we have developers who have demonstrated experimental NVIDIA GPU support for Mac OS, which is potentially allowing compatibility with Intelbased Hackintosh systems to run modern RTX graphics

03:36 cards, which is nuts. Another one is any PS5. This is an experimental project trying to get PS5 games running directly on PC. But the craziest thing is is that this was all fully done with AI assisted breakthroughs in software development which got to this level of developing these two systems and flows. And there’s so much more to cover. But before we even get to that guys, it would really mean a lot to me if you guys can go ahead and like this video and please subscribe cuz majority of you guys aren’t actually subscribed to the

04:06 channel. So it’ really mean a lot to me if you guys can do that. But with that thought, let’s just simply dive into today’s upload. Let’s start things off with this new Gemini 4 Carbon model because it looks like Google is already testing something potentially even more powerful than Gemini 4 Argon, which hasn’t even officially released yet. Essentially, Business Insider today published a new blog post that is suggesting that Google is internally testing a new Gemini 4 checkpoint that is cenamed Carbon. Now, this is potentially rivaling up against the

04:37 clawed opus 5.5 when it comes to coding. And according to this report, one Google employee who had tested Carbon had said that it feels like Opus 5.5 for coding tasks, although they admitted that they needed to test it out a bit further, make a definitive comparison. Now, what’s particularly interesting is that Google’s work on Gemini 4 Argon is apparently still ongoing, but employees are already testing new checkpoints. We’re hearing about several internal code names including Argon, Berium, and

05:08 Carbon, with the newest version reportedly being prepared for public release, which is Argon, actually based on the earlier checkpoint known as Barryium-B. Meanwhile, Carbon is supposedly a newer checkpoint that has already landed inside Google’s internal coding tools. So essentially, Google could be preparing to release Gemini 4 Argon while already testing a potentially stronger version behind closed doors. And look, I have to say this is definitely exciting, but a little frustrating to me because knowing Google, if they’re going to take forever

05:40 to actually get these models into the public’s hands, and you combine that with what feels like they’re almost 100% pre-launch nerf rate, Archon might end up looking pretty underwhelming by the time we actually see it. I mean, what’s the point of having these incredible powerful models internally if everyone else is already catching up by the time they launch it? Of course, there are still internal reports and we haven’t seen independent benchmarks confirming Carbon’s performance against Opus 5.5, but if these claims hold up, Google

06:11 could be making a comeback. If you want the best AI tools, workflows, and drops before everyone else, join my free newsletter with the link in the description below, which is completely free. And to top off this whole story, things got even crazier when we start talking about recursive self-improvement or RSI. Essentially, Business Insider report referenced discussions suggesting that multiple Google employees have been talking about RSI potentially playing a major role in the company’s recent AI progress. And what makes this more

06:41 interesting is that Google’s researcher had posted on X quoting this blog article simply asking, “Have you heard of recursive self-improvement?” Now, that is a pretty interesting thing to say, especially when all these reports are coming out about Google’s latest Gemini 4 checkpoint. We’re also seeing claims that some Google researchers believe recursive self-improvement is also well underway, potentially helping accelerate the development of new models. And honestly, looking at what’s happening with Gemini 4, as well as

07:13 whatever happened with the Flash series, I can see why people are connecting the dots. Think about it. Gemini 4 Argon hasn’t even rolled out publicly yet and Google is reportedly already testing new checkpoints like Carbon, which employees are informally comparing to Claude Opus 5.5 in coding. If Google is increasingly using its own AI models to help research, train, and improve the next generation of models, that could potentially explain why they’re iterating on all these models so quickly. Now, to be clear, a cryptic

07:44 expose and faster development cycle doesn’t necessarily confirm that Google has achieved the true RSI claim, but I genuinely think that Google could be making significant progress in this area. And if these reports are accurate, they might be much further along than what we realize. But that’s not even the craziest part, guys, cuz we are hearing some pretty wild developments coming out of Enthropic. According to a couple of insider claims from a trusted source, Enthropic is reportedly starting an

08:15 entirely new pre-training run about a week ago. And what’s interesting is that this isn’t supposedly even for Claude Fable 5.5. We have been seeing a lot of leaks around Cloud Fable 5.5, which is going to be releasing soon, but this is not even for that model. This is something that is entirely different. Instead, it’s rumored to be something much bigger than Fable 5.5, potentially being COD Opus 6 or Fable 6, or an entirely new generation of models that are designed to compete with OpenAI’s

08:48 rumored next generation model, which is Bell. Now, for those who are unfamiliar, pre-training is essentially the foundational stage where an AI learns from massive amounts of data. So, we’re potentially talking about a much more substantial upgrade rather than another incremental improvement. And considering how powerful Enthropic’s recent models are, the idea that they’re potentially pre-training something even more capable is pretty exciting. But that’s not even all because we’re also hearing some pretty insane rumors surrounding OpenAI’s next generation model, Code

09:19 Name Bell, which some are speculating could eventually become GPT 6.5 or GBT7. And according to these insider claims, Bell, which may actually have a different internal code name, is supposedly this model that is behind the impressive mathematical research results that OpenAI had recently published. And here’s where things get really interesting, cuz the same rumors suggest that RSI may have played a role in achieving those results, potentially allowing OpenAI to use its own AI system

09:50 to accelerate research and develop increasingly capable models. And when you connect this with everything we’re hearing about Google’s Gemini 4 checkpoints and Enthropic’s rumored next generation training run, you start to see a much bigger picture. We’re now hearing speculation that OpenAI, Enthropic, and Google are all getting extremely close to achieving RSI or may have already made breakthroughs in that direction. And guys, just think about the implications of that. If AI systems can meaningfully accelerate the

10:21 development of their own successor, we could potentially see the pace of AI progress increase dramatically. And to be honest, this could be why all of these major labs have been saying pace the AI race. Maybe they do know something behind the scenes that we don’t know, which is why they keep on saying and reiterating that we should pace the future. But it also doesn’t make sense why they keep on releasing these models that are more capable every single month. Now, moving back over to

10:51 Google, we’ve got another interesting update that could potentially be connected to the upcoming launch of Gemini for Argon cuz essentially a new ultra mode has been spotted in development inside Google AI Studio Build with the description build with advanced skills and tools. This is a new mode that reportedly appears alongside plan and build as well as a previously discovered security review mode which is suggesting that Google is expanding the AI studio with more advanced development capabilities. And with Google’s Gemini

11:22 for Argon reportedly getting closer to release, there is speculation that these modes are being prepared to take advantage of its improved coding and reasoning abilities. Now, moving over to some open- source AI news. Quen has officially released Quen image 2.1 Turbo. Now, they had already released this image model, but this is a new accelerated image generation model that is built off of the previous version, the image 2.1. And this is something that can create and edit images in just in just eight denoising steps. Now,

11:53 what’s impressive here is the original Quinn image 2.1, which typically uses around 40 steps, meaning Turbo reduces the number of sampling steps by 80% while reportedly maintaining comparable image quality. And we’re not talking about low resolution images either. This is built on a 7 billion parameter architecture that is capable of generating detailed 2K images from text prompts while also supporting natural language image editing. So you can generate an image, add accessories,

12:23 change the background, or completely modify a scene just by describing what you want. And the best part is the model weights are openly available on hugging face, meaning developers can download and run it locally on compatible hardware. It also supports hugging face diffusers, which is recommended with eightstep sampling configurations, making it relatively straightforward to integrate into existing workloads. And say if you don’t want to run it locally, Quinn has launched a hosted API for both the models, the Turbo as well as the

12:54 original Pro model. And honestly, getting potentially comparable image quality with just eight steps instead of 40 is pretty significant in terms of efficiency as well as overall quality. And of course, fewer steps doesn’t automatically guarantee five times faster generation, but it’s still impressive. So it is something that is quite great to see especially it being open source. Next is where Microsoft is creating their own Jev alternative. This is where they have introduced a pretty interesting new AI model that is called

13:26 Microsoft decision one. Specifically designed for fast and efficient decision-m. It’s exactly like Jev where unlike traditional large language models that focus on generating text, this is a model that is optimized for structured decisionmaking tasks where the goal is to quickly evaluate information and determine the best possible action. Now, according to Microsoft, Decision 1 delivers impressive performance, reportedly outperforming both traditional large language models and other specialized decision models in

13:56 terms of response latency as well as decision quality on its evaluated tasks. And what’s particularly interesting is that Microsoft is already testing this model internally across several real world applications, including incident response, quality control, and even scientific discovery. So imagine an AI system that can rapidly analyze a production issue, determine the most appropriate response, or help researchers make complex decisions without relying on massive generally purpose language models for every single

14:28 step. If you like this video and would love to support the channel, you can consider donating to my channel through the super thanks option below. Or you can consider joining our private Discord where you can access multiple subscriptions to different AI tools for free on a monthly basis plus daily AI news and exclusive content plus a lot more. Now moving over to OpenAI as a part of their 28-day pledge to ship something new every single day. We’re now on day five and they have introduced a pretty interesting new feature for

14:58 Codeex. Essentially OpenAI has introduced composer predictions. This is something that is available in beta for Codex Pro users on the desktop app. But what this does is something that makes Codex predict your next prompt based on the conversation history project context as well as even how you typically communicate with the AI. And this is based off the memory that has been built over time. So instead of manually typing out every single instruction, Codex can proactively suggest what you might want

15:29 to do next, making the coding workflow much more smoother. And according to OpenAI, this was one of their most loved features during internal testing with predictions apparently being so accurate that they sometimes make you do a double take. And the best part is is that it’s included with pro plans without consuming your usage limits. It’s relatively a small update, but I can definitely see this becoming incredibly useful for all of us. Now, this update is more for codecs on Windows, but essentially OpenAI has introduced a new

16:01 sandbox mode that is built using Microsoft’s execution containers or MXC that is designed to make running AI agents on Windows faster and more secure. And this is essentially a sandbox that offers a faster setup, stronger network restrictions, and more granular file access controls. Meaning developers can give codecs access to specific files and resources without unnecessarily exposing the rest of their system. And you probably know this by now, giving an AI coding agent the

16:31 ability to execute commands and modify files autonomously is quite dangerous. So having stronger isolation helps reduce the risk associated with giving AI access to your computer. The only catch is that you’ll need a compatible Windows 11 device to take advantage of this new mode. We also got some fresh new updates for OpenAI dots. This is their always on AI agent that is powered by GPT6 Astra. And this first major update is that you can now create your own dot directly from Chat GBT mobile

17:03 app both on iOS and Android. Previously setting up these autonomous agents were quite limited, but now you can actually create them straight from your phone. So it’s going to make it easier to get started wherever you are. And considering that DOTs are designed to handle ongoing tasks and work more independently, bringing that creation to mobile is going to be pretty useful and obviously great for anyone who even uses dots. Remember Devon? Well, you can actually get access for it completely for free if you have a Chad GBT subscription that is going to let you

17:34 access the autonomous AI coding agent. Essentially, Devon, for those who do not know, is the first software engineer that they had marketed about 2 years ago now. And I was actually one of the first ones to cover it. And that was actually one of my first videos that blew up. But this is a pretty incredible cloud-based software engineer that is AI focused. And now Devon is going to let you connect your existing chat GBT Go Plus or Pro subscription. And it allows you to use GBT models inside Devon without needing a separate OpenAI API billing

18:06 setup. Instead, your GBT model usage will draw directly from your existing Chad GBT plans quota. So if you’re already paying for Chad GBT, you can potentially take advantage of that subscription while working with Devon and making it much more convenient for developers who use multiple AI coding tools. Now obviously it is something that is not getting you free access to Devon. Kind of is but not cuz you’re paying for your own quota at the end of the day. But still it is a little

18:37 misleading as to this announcement that they had announced but it’s a nice integration at the end of the day. A free model that I was talking about is solar mini4 which is completely free right now within client. It is a model from a Korean lab which is called upstage AI. I actually made a video with them a while back. Essentially, this is a 35 billion parameter mixture of experts model. But what’s interesting is that it only activates around three billion parameters at the time, making it significantly more computationally efficient. And despite it being

19:08 relatively small in terms of its active parameter count, it is something that supports a 524k context window and reportedly delivers up to speeds of 208 tokens per second, which is pretty fast. This model on artificial analysis intelligence index scores a 24 making it the highest scoring model in its three billion active parameter category. This is a model that is just one point behind Neotron 3 Ultra from Nvidia which uses around 55 billion active parameters. So

19:40 getting this model completely for free right now is definitely a steal. And it’s not meant for coding but it is meant for document parsing as well as processing. So it is something that can help you in that domain and you can try it out completely for free with incline right now. I’m not sponsored by them. I’m just trying to get you some insights on some free stuff. Moving over to Enthropic, they have launched claude managed agents dynamic workflows which is in public beta at the moment and it is introducing a pretty powerful new approach to multi- aent orchestration.

20:11 Essentially, instead of relying on a single AI agent, Cloud can now create a plan, distribute tasks across multiple agents, execute them in phases, and combine the results. And the performance improvements are pretty insane. In Enthropic’s testing, they planted 70 bugs inside a codebase containing over 116K lines of code. And across three runs, a single agent discovered 14, 15, and 27 bugs. while the multi- aent workflow consistently found 66 out of 70

20:43 bucks in all three runs. That’s a massive improvement. And what’s even crazier is that these workflows can reportedly orchestrate up to a,000 agents in a single run. Developers can try by configuring their agent with the multi- aent flag and asking Claude to execute a workflow. This could be huge for complex coding tasks, especially when multiple AI agents can work together, but not too sure on how it works with the rate limits. Another interesting update from Enthropic is Claude Code Projects, which is now

21:14 opened up access to everyone who has previously been on the pro and max weight list. Essentially projects for those who do not know lets you manage complex coding tasks through one ongoing conversation where Claude coordinates the work and runs individual tasks in separate threads simultaneously. So instead of manually juggling multiple coding sessions, it’s going to make it easier for you to help organize and execute different parts of your project in parallel. projects is still in beta, but if you haven’t signed up yet for the weight list, you can actually join the

21:46 wait list, and I’ll leave a link to that in the description below because Enthropic says that they’re gradually letting everyone in slowly as capacity allows. Moving over to China, the AI race is even getting more competitive as some pretty unexpected companies are now entering the large language model space. According to a new report from the information gaming giant Miho Moi Hoyo, sorry, I don’t know why I said Mihoyo, but the company that is behind Genchin

22:16 Impact, the smartphone manufacturer Xiaomi, and the food delivery giant Muton, they are all investing in developing their own AI model. What’s particularly interesting is that Mi Hoyo is reportedly wanting to become a top tier AI research lab within the next few years. While Xiaomi is obviously doing their own thing with their uh Mimo Pro models, but Muton is already releasing their own capable model in the next couple of months. Think about how wild

22:48 that is. It’s like McDonald’s suddenly deciding to compete with OpenAI and Enthropic by developing their own Frontier AI model while focusing on their fast food franchise. And honestly, Mi Hoyo is the one that gets me the most excited for cuz just imagine a massive Genchiniz open world where NPCs are powered by AI actually remembering your choices, adapting to your play style, and evolving alongside you instead of following predetermined dialogues. characters could create unique

23:19 interactions and storylines that developers never explicitly scripted. But overall, it just shows how far China’s AI ecosystem is expanding beyond just traditional AI companies. And I genuinely think AI powered NPCs could be one of the biggest innovations in something like the game scene. Now, moving back to Enthropic, remember how we recently covered the new announcement of their expansion of the Claw Startups program that is offering founders some pretty generous benefits? Well, essentially what happened was eligible

23:49 startups could receive a free year of Claude team up to $1,000 in API credits, exclusive partner offers, and access to office hours with Enthropic applied AI team. Well, guess what? Apparently, some of you DJs decided to abuse the entire thing. And according to reports circulating online, people were allegedly registering cheap $10 domains to pass as startups and claim the benefits without actually running any legitimate business. And apparently things got so out of hand that

24:21 applications reportedly skyrocketed by 21 times in just 48 hours. Now, what’s even crazier is that some users are claiming that their access is being revoked even after their applications were approved, potentially affecting legitimate founders as well. And of course, the exact scale of abuse and these rejections haven’t been independently confirmed, but seriously, Enthropic probably should have verified these applications before approving them rather than just potentially having to revoke access afterwards. I mean

24:52 offering thousands of dollars in benefits and insufficient verification that doesn’t seem right. What did he actually expect was going to happen? Moving back to Google, they have introduced some pretty exciting benefits for Google AI Pro and Ultra subscribers. Essentially, for anyone interested in running or fine-tuning AI models. Essentially, Google AI Pro now includes premium collab benefits with access to Nvidia’s 800 GPUs featuring 80 GB of VRAM, while Ultra subscribers can potentially access even more powerful

25:23 800 GPUs, which is kind of insane. And with that much GPU memory, you can reportedly run Gemma 4 31B and BF16 while it is fully fine-tuned with the Gemma 4 E4B performing Laura finetuning on Gemma 4 26B A4B and even Qura fine-tuned with 31B models all directly from your browser. And of course, GPU availability and compute limits still apply. So, it’s not unlimited access, but considering how expensive those GPUs

25:54 are, getting access through your existing Google AI subscription, it is definitely incredible. Small update that you might have missed today is that Opus 5.5 BAS mode has rolled out and you can actually access it now, which is awesome. Obviously, this is going to use up your tokens, but Enthropic Opus 5.5 gets around 285 TPS on fast mode and only uses two times your usage, which is 10 times better than Open AI’s recent speed upgrade, which is kind of insane,

26:25 but I believe it doesn’t even use your limits at all. You have to actually pay for it. So, my bad if I got that wrong. And guys, speaking of Opus 5.5, I wanted to quickly showcase this pretty cool game that I saw from a developer on X, which was entirely done by the Opus 5.5 on Medium Reasoning. This is just insane how far AI generated games have come, cuz this is something that is able to fully build out visuals, textures, gameplay mechanics, and it is able to

26:56 even polish those outputs even further. We’re getting to a point where AI coding models can potentially build entire playable experiences from a single prompt. Something that would have taken someone an incredible amount of years to actually generate full-on games. But the fact that I also wanted to emphasize is that you don’t always need to be using the highest reasoning efforts to get the best outputs. This is something that was done with medium reasoning. So take that into consideration. And lastly guys, I wanted to showcase two absolutely insane

27:27 developments where people are using AI assisted coding to tackle some incredible complex reverse engineering and software development challenges. Starting off with what I genuinely never expected to see, modern Nvidia GPUs running on Mac OS. Essentially, a developer going by Null Moth had released an experimental open- source Nvidia driver for Intelbased Macs as well as Hackintosh systems, reportedly allowing modern RTX graphic cards to work with Mac OS Seoia. But guys, do you

27:58 realize how crazy this is? We’re talking about running an RTX 5060 on Mac OS 15 with hardware acceleration, including support for desktop, Metal 3 applications, and potentially games. Now, the developer reportedly accomplished this by porting Nvidia’s open- source GPU kernel modules to Mac OS, building a Metal compatibility layer using Messa’s NVK Vulcan driver, and translating Apple’s Metal Shaders into instructions for Nvidia GPUs so that it can actually understand it. And what’s

28:29 particularly interesting is that AI assisted development, including cloud code, reportedly coded out all of these elements, played a huge role in debugging and developing parts of the driver. And now, before everyone gets too excited, this doesn’t mean that you can suddenly plug in an RTX 5070 into your M series MacBook and start running CUDA. This is currently an experimental project that is targeting Intelbased systems and CUDA compatibility which hasn’t been established by the demonstration 100% just yet. But the

29:00 fact that we’re seeing in modern Nvidia hardware acceleration on newer Mac OS versions is pretty remarkable. Now the second development might be even crazier cuz we’re talking about any PS5. This is an experimental project attempting to bring PlayStation 5 software and games to PC. Now, what’s fascinating here is that developers are working on translating PS5 system libraries and graphic instructions into something compatible with PC hardware, taking advantage of the fact that PlayStation 5

29:30 already uses an x86 64 processor architecture. And one of the demonstrations circulating online reportedly show PS5 system software is running on a PC through emulation rather than simply showcasing a PS5 game. Think about that. We’re potentially seeing parts of an entire console environment being recreated on completely different hardares and AI assisted coding tools were used in this specific scenario as well with cloud code reportedly contributing to the project’s development heavily. Of course, this

30:01 doesn’t mean full PS5 emulation is solved or that modern PS5 games are suddenly playable on any PC. There’s still a massive amount of work involved that would need to actually get that done. But what I find genuinely fascinating is that both of these projects is using AI coding tools which are helping developers tackle engineering problems that would have traditionally required enormous amounts of specialized expertise and development time from experimental GPU drivers to recreating complex console software environments. We’re seeing some pretty

30:33 wild possibilities emerging. And honestly, if this is what developers are already experimenting with today, imagine what they’re going to be building in the next couple of months. Now, that was truly a long video. This has to be my longest video ever. So, I really hope you guys enjoyed it and got some sort of value. Make sure you guys like this video and subscribe as it truly helps me a lot and helps me make these videos for you guys. It is where I spend a lot of time researching and obviously scripting, making the content, making the visuals. So, it would really

31:03 mean a lot if you guys can even just like this video because it really truly helps me out make more of these videos for you guys. But with that thought, guys, I hope you enjoyed today’s video and got some sort of value. Make sure you go ahead and take a look at the world of AI school community. Take a look at the world of AI benchmark as well as our second channel. Join our newsletter, join our discord, follow me on Twitter, and lastly, make sure you guys subscribe, turn on notification bell, like this video, and please take a look at our previous videos so that you can stay up to date with whatever is happening in the world of AI. But with that thought, guys, have an amazing day. Spread positivity, and I’ll see you guys fairly shortly. He suffers.