China AI ACHIEVED RSI? GLM 5.5, NEW Fable 5.5 Checkpoint, Qwen4, Kimi K3.1, M3.1 Pro & More!

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: China AI ACHIEVED RSI? GLM 5.5, NEW Fable 5.5 Checkpoint, Qwen4, Kimi K3.1, M3.1 Pro & More! · Channel: WorldofAI · 22:25 · Watch on YouTube

Topics: AI agents, Software engineering with AI, AI models, AI safety

Summary

The video surveys reported upcoming releases from Chinese AI labs, including models from Z AI, MiniMax, Moonshot AI, and Alibaba’s Qwen team. It highlights claims about Alibaba’s recursive self-improvement experiments and Z AI’s plans for larger GLM models, alongside demonstrations attributed to a possible Claude Fable 5.5 checkpoint. The host also discusses OpenAI usage updates, AI security concerns raised by Microsoft CEO Satya Nadella, potential Claude voice features, and reports about Nvidia’s possible acquisition of Reflection AI. The video closes with an AI-assisted development project aiming to bring Windows application compatibility to PlayStation 5 hardware.

Key points

Reports suggest several Chinese AI labs may release major models in October, including a new Z AI model, a MiniMax Pro model, and Kimi K 3.1.
Alibaba reportedly plans to release Qwen 4 models in both a 27-billion-parameter version and a flagship Max variant, with public weights being a possibility.
Alibaba says Qwen 3.8 Max underwent automated self-improvement cycles, completing 33 iterations over more than a month.
Z AI is reportedly planning to scale its GLM models toward one trillion parameters and develop a recursive self-improvement process across training.
A possible Claude Fable 5.5 checkpoint is credited with detailed one-shot SVG controller designs and playable 3D shooter game generations.
OpenAI gave paid users a global usage-limit reset and added options for Codex users to use resets and steer tasks while they are in progress.
Satya Nadella argued that AI systems should be designed on the assumption that their underlying models may be compromised, with independent verification and strict controls.
Anthropic may be developing native voice capabilities for Claude; reports also describe restricted internet access during internal AI evaluations.
Nvidia is reportedly in talks to acquire Reflection AI, which could bring the chipmaker into more direct competition with frontier AI labs.
A developer’s PlayStation 5 project is reportedly expanding toward a Windows-inspired environment, with Claude said to have been involved.

Chapters

Start	Chapter	Summary
00:00	Upcoming AI models and major stories	The host previews reported Chinese model releases, a possible Fable 5.5 checkpoint, AI security concerns, and other developments.
03:07	Chinese AI releases and Qwen 4	The video discusses rumored releases from Z AI, MiniMax, and Moonshot AI, then outlines possible Qwen 4 model variants.
06:16	Alibaba and Z AI on recursive self-improvement	The host summarizes Alibaba’s reported automated improvement cycles and Z AI’s plans for larger models and a self-improving training pipeline.
08:48	Possible Claude Fable 5.5 generations	The video presents examples attributed to a Fable 5.5 checkpoint, including detailed SVGs and playable 3D shooter games.
14:39	OpenAI updates and AI security	The host covers OpenAI usage resets and Codex features, then discusses Satya Nadella’s warning about securing AI systems.
17:43	Claude voice and evaluation restrictions	The video reports possible voice features for Claude and describes claims about restricted internet access during Anthropic evaluations.
19:19	Nvidia and AI-assisted development	The host discusses reports of Nvidia considering an acquisition of Reflection AI and a PlayStation 5 software project reportedly involving Claude.

Quotes

the Qwen official Twitter posted hints at dual open source models landing soon. — 05:13

Alibaba claims that on artificial analysis intelligence index, it jumped from 40 to 45 — 07:17

If it can’t be observed, it can’t be trusted. — 16:42

Nvidia could be preparing to compete directly against Frontier AI Labs like OpenAI, Anthropic, and Google. — 19:19

Concepts and entities

recursive self-improvement (RSI), GLM 5.5, Qwen4, Kimi K3.1, Fable 5.5, Codex, Claude, Anthropic, Alibaba, Z.ai, NVIDIA, Reflection AI, Satya Nadella, Mihawk

Full transcript (original language)

00:00 The Chinese AI scene is about to heat up because we’re starting to see some incredible new Intel on massive waves of AI models that are coming from China’s biggest labs. Starting off with Z AI, which is reportedly preparing to release Jello M 5.5, while MiniMax is gearing up to officially launch its next pro model, which you probably have seen, aka Space Bunny. And that’s not all, because Moonshot AI could also be dropping something huge this month after previously teasing the K 3.1 maybe

00:31 mid-October. But, one of the biggest developments comes from Alibaba, where new information is suggesting that Qwen 4.0 could actually launch both a 27 billion parameter model that is obviously going to be focused locally, but a much larger flagship max variant is also going to be something that is going to be open source, which would be absolutely insane for the open way AI ecosystem. But, believe it or not, that’s not even the craziest story today, because all eyes are on Anthropic

01:02 right now. With new reports surrounding a potential November IPO, Anthropic’s acceleration for the 5.5 model series has been one thing that we have all noticed. We’ve seen Haiku 5.5 just recently, an upgraded Sonnet 5.5, but one model that hasn’t gotten an upgrade yet is Fable. And there are signs that we are starting to see Fable 5.5 once again within Claude code, cuz it recently passed the Tibble test once again, and this new checkpoint of Fable 5.5 is absolutely insane due to it

01:34 producing some crazy generations. If these early results are legitimate, Fable 5.5 could seriously shake up the Frontier AI leaderboard and take AI coding capabilities to another level. Meanwhile, Microsoft CEO Satya has released has raised a fascinating point about AI security, arguing that companies should design their AI systems under the assumption that the underlying model could already be compromised. Over at OpenAI, they didn’t actually end

02:04 up shipping something ambitious a part of their 28-day shipping pledge. So, they ended up giving everyone a global usage limit reset. Over at Anthropic, Claude voice could be coming very soon with signs pointing towards Anthropic developing its own native voice capabilities. Now, there are other things that are happening in the space like Nvidia which could be making a bigger push into the frontier AI model race with the reports surrounding its acquisition of Reflection AI. And if this turns out to be true, the implications could be huge cuz we’re

02:36 talking about one of the world’s most powerful AI infrastructure companies potentially taking a much more direct role in competing with the very frontier AI labs that rely on its hardware. And I’ll explain a bit more on what’s actually happening in that case. Also, majority of you guys aren’t actually subscribed. So, it would really really mean a lot to me guys if you guys can go ahead and like this video at the very least and subscribe as I would truly appreciate that as this really helps me make more videos for you guys on a day-to-day basis. Let’s start things off

03:07 with China’s AI scene because things are about to get very interesting with several potentially massive models releasing in October. Starting off with Z AI which is reportedly preparing to launch J on 5.5 as early as next week or potentially towards the end of October. Now, this information comes from a source claiming to have been well connected with the informants inside both Z AI as well as MiniMax. And while none of this is officially confirmed, the timing is definitely interesting because MiniMax could be preparing something huge. You guys remember

03:39 SpaceFunnel Alpha that mysterious model that we recently saw appear on Open Router? Well, there’s growing speculation that SpaceFunnel Alpha could be actually a MiniMax model most likely the flagship pro model. And if these rumors are accurate, we could be seeing an official release sometime around mid-October. But that’s not actually all because Moonshot AI has also reportedly geared up for its next major release as well. Cuz we recently saw the teaser around Kimi K 3.1 from the official

04:09 team, but now we actually know that they are going to be releasing a model mid-October based off of my sources that I had actually came in contact with and that is actually in sync cuz we’re going to have three new models, actually five if we are to include Qwen, which we’re going to be talking about in a bit. So, think about this. We could potentially see Qiong 5.5, a new flagship Minimax model, as well as Kimi K 3.1, and the Qwen model series. These are four potentially major Chinese AI models competing for attention in October

04:41 alone. And considering how competitive three or four Chinese AI labs have become, especially when it comes to coding performance, open weight models, and pricing, I am extremely excited to see how these models actually stack up up against each other. 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 this Chinese AI news doesn’t stop here because Alibaba’s Qwen team also has something absolutely massive planned. We all know that Qwen 4.0 is

05:13 coming fairly soon, but today we’ve got hints at two new updates, two major open weight models launching together cuz on X, the Qwen official Twitter posted hints at dual open source models landing soon. We’re talking about Qwen 4 27B, the smaller model that’s designed to be much more accessible for local AI enthusiasts as well as developers, alongside Qwen 4 Max, which is expected to be their flagship model. And here’s where things get really interesting cuz according to these latest claims, both

05:44 models will have their weights released publicly, potentially. Now, if that’s true, this would be a huge deal because imagine getting access to Alibaba’s latest flagship AI capabilities without being completely dependent on proprietary API. And for those of you running local AI models on your GPUs, the 27 billion parameter version could be particularly exciting, especially once optimized on sizes become available. Of course, running that much larger Max model would likely require considerably more hardware depending on

06:16 its architecture and final parameter count. But, the bigger picture here is that Chinese AI labs are continuing to push the boundaries of what developers can actually run and optimize themselves. Who knows? Qwen 4.0 could have us run Opus-level capabilities locally, which would be absolutely insane. But, that’s not even the craziest part because Alibaba also revealed that it’s experimenting with recursive self-improvement, also known as RSI, which is something that we also have been talking about lately a lot cuz

06:47 Gemini may have achieved it, and the AI labs in America seem to have all started working towards achieving or getting really close to having a prototype of it. Now, at its recent conference, Alibaba did reveal that Qwen 3.8 Max had reportedly been running through a automated self-improvement cycles for over a month, which is completing 33 iterations. Essentially, the system identifies its own weaknesses, generates new training data, validates improvements, and then feeds those

07:17 results back into training with significantly less human intervention. And the results were pretty insane cuz Alibaba claims that on artificial analysis intelligence index, it jumped from 40 to 45, which is a 5.8 jump in just 30 days. Now, with Qwen 4.0 reportedly on the horizon, there is speculation that these techniques could play a major role across the next generation of models, including the Max Plus, Flash, and even the 27 billion parameter local model. And speaking of

07:47 RSI, Z AI might be taking things in entirely new level with its upcoming GLM 5.5 model. Cuz according to new reports citing Z AI’s founder, Tang, the company is planning to scale its next generation GLM models from around 740 billion parameters to 1 trillion. But the parameter count isn’t even the most exciting part, cuz Z AI is reportedly working towards a fully closed-loop recursive self-improving system where AI actively participates in improving

08:17 itself throughout the entire training pipeline, from pre-training all the way to post-training. Essentially, the goal is to have AI increasingly automate its own development rather than relying on humans to manually manage every step. And if they actually pull this off, it would be a huge breakthrough in AI development. Of course, this is still early reports in their roadmap. We’ll find out a bit more when they end up releasing this model. So, this is absolutely insane that these Chinese labs are slowly trying to find breakthroughs in RSI. If you like this

08:48 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. Next up, we have some very exciting developments around Anthropic’s upcoming Claude Fable 5.5 model, cuz it looks like Claude Fable 5.1 might be once again routing certain requests to the newer Fable 5.5

09:19 checkpoint. Now, we actually saw reports of this happening previously, which is why I made a couple of videos on it. But then it appeared to stop for a while, and now recently, once again, it has started noticing the same behavior from before again. And the way people find out if this model is actually the Fable 5.5 is by trying to identify the model through the table test, which is essentially a fingerprinting prompt designed to distinguish model versions based on the responses. Now, I want to make this clear so we don’t confuse everyone, but this may not be actually a

09:50 newer checkpoint of Fable 5.5, but it is 100% the Fable 5.5. And regardless, the generations that we’re seeing are absolutely insane. Just take a look at a couple of these one-shot SVG generations of the PlayStation 5 and PlayStation 4 controllers. The level of detail here is genuinely mind-boggling. And what makes this really impressive is that SVGs aren’t traditional images. The model here is able to generate actual code that is defining every shape, curve,

10:20 gradient, shadow, and individual component. Recreating something as geometrically complex as a PS5 controller is incredibly difficult, especially when you’re trying to get the proportions, the symmetry, buttons, analog sticks, as well as subtle lighting effects emitting from the controller itself. And the fact that these were reportedly generated in a single shot makes it even more impressive. Like for example, if you were to just compare this with the original PS5 controller, it is very hard

10:50 to distinguish which one’s actually real. Obviously, this one is 100% the real one, and you can clearly tell, but it’s just incredible that this is something that was coded out. And if you even zoom in, the curvature, even the textures, the brittle of each of the analog sticks have been coded out, as well as the lighting cues. So, this is just nuts that we have gotten into a point where we’re getting these sorts of generations now. Now, this may be the craziest generation that I have seen from Fable 5.5, cuz someone essentially

11:21 challenged it to build a full first-person SWAT shooter game using 3GS, and it did it in a single shot. This is just insane. Cuz this is a fully playable 3D shooter game that kind of resembles the same characteristics as Call of Duty and it got the textures and everything nailed down perfectly featuring surprisingly tight gun play ragdoll physics intelligent enemy NPCs even got voice lines and so many other features. What’s interesting is that this isn’t just a static 3D environment.

11:52 The model was able to generate actual gameplay systems including combat mechanisms you have enemy behaviors physics and a lot of other things that makes this into a fully playable FPS shooter game. Now here is another incredible gameplay generation from the fable 5.5 and this is where it hit something that got it done within a few hours which is kind of insane with barely any usage. So this is the level of detail that you can expect but I want you to actually take a look at this

12:22 gameplay cuz it’s actually insane that an AI model is able to output this level of generation and it’s going to get to the point where we will actually see some crazy games being generated with these newer models especially when we get to the GPT 7 series as as well as the fable 6 series.

13:39 >> Guys, stable 5.5 made this in just a couple hours with barely any usage. I I’m so confused. I’m so shocked. >> [music] >> Bro, like look at this guys.

14:39 >> Moving forward, they both today had announced a new global AI usage reset for all open AI paid users. This is day six of their shipping initiative of 28 days of launching new features for open AI, and it looks like they didn’t have anything to ship today. And one of the promises they had is that if they don’t end up shipping anything, uh they end up actually giving everyone a free usage limit reset. But, it seems like this guy’s going to get this company bankrupt at this rate if they

15:10 don’t end up shipping new features. And speaking of usage resets, open AI has introduced an interesting new option for Codex users. And this is thanks to Testing Catalog for reminding me cuz you can apparently allow Codex to use a reset directly during a conversation without having to manually navigate through your account settings. But, there are a couple of conditions. You have to explicitly ask Codex to use a reset, and your remaining usage must be at 10% or lower. And importantly, these resets are never triggered

15:41 automatically, so you’re not always in control of when they’re actually used. You can find this new option under usage and billing section in your settings, and it is pretty convenient and a nice quality of life improvement for Codex users. Now, another important Codex tip that you should know that I mentioned a couple of days ago within my video is that you have this real-time steering feature. Essentially, instead of waiting for Codex to finish its current task before giving it a new instruction, you can now steer the agent while it’s actually working on your current

16:11 request. So, imagine Codex is building an app out, and halfway through you realize that you want to change the design or add another feature. Instead of waiting for it to finish, you can immediately send this new instruction, and Codex can adjust its approach in real time. It’s another great feature, and I highly recommend checking it out within the settings and enabling it. Moving forward, Microsoft CEO Satya has issued an important warning about AI security, arguing that companies should build AI systems assuming that the underlying model could already be

16:42 compromised. And having another AI monitor and agent isn’t necessarily enough, which is something that he’s emphasizing within this blog post because you’re essentially putting one black box in charge of another. And he puts it, “If it can’t be observed, it can’t be trusted.” He’s essentially calling for independent verification, transparent activity logs, strict access controls, and emergency shutdown mechanisms. And personally, I do think he is right. We’re entering this era where AI agents will be managing

17:12 infrastructure, executing transactions, and making real-world decisions autonomously. And the scary part is that an AI doesn’t even need to be malicious to cause serious damage. One mistake or prompt injection could be enough. So, while everyone is racing to build smarter AI agents, we’re seeing new AI models like Fable, as well as this new Bell model from OpenAI, it is definitely important that there are restrictions just in case if these models end up in the wrong hands or end up just

17:43 autonomously doing their own thing. Moving forward, Anthropic might be preparing a major upgrade to Claude’s voice capabilities, cuz there’s a hidden voice tab labeled preview. And it has reportedly been spotted alongside a new option that allows users to share voice data for training. Now, it’s unclear whether Anthropic is developing its own voice model, but it would be huge for vibe coders. Imagine just talking to Claude in real time while it builds your application, making changes, debugging code with you. And this is something that doesn’t even require you to type.

18:14 Now, I do remember they had a couple of checkpoints of different voice models that they were initially developing. So, it does look like they have a voice model coming soon, but this would be definitely huge. And in some pretty wild Anthropic news, reports suggest that the company has restricted internet access during its internal AI evaluations. This comes after a report of Claude exploiting real websites, bypassing safeguards, and even submitting a fake murder tip to police during testing. Now, these are reported incidents from

18:47 controlled evaluations, but they highlight just how unpredictable autonomous AI agents can become when they are given real access to real-world systems. And personally, this adds on to what I said about what the Microsoft CEO had stated about better safeguards, cuz it is kind of insane that these stricter restrictions weren’t entered sooner. So, it’s great that now companies are starting to realize they got to emphasize on the safeguards. Obviously, not putting too many safeguards that hinder the quality of the models, but at

19:19 least restricting some malicious stuff like having a model go rogue or how people actually end up using the model in a malicious way. Now, this next story is nuts because Nvidia could be preparing to compete directly against Frontier AI Labs like OpenAI, Anthropic, and Google. Cuz according to this new report attributed to the Financial Times, Nvidia is in talks to potentially acquire this new company called Reflection AI, which is an American AI startup reportedly valued at around 25

19:50 billion. Now, what’s fascinating is that Nvidia is already a major investor in that company reportedly committing around $800 million. Reflection AI is working on Frontier level open weight models positioning itself as a potential American competitor to Chinese labs like Deep Seek, Mind, and Zeyi. And here’s where this could actually become huge cuz Nvidia already dominates the AI hardware ecosystem supplying the GPUs that many Frontier labs depend on to train their AI models. But now with companies like

20:22 Google, Amazon, and other AI labs inventing investing heavily in custom AI chips, Nvidia has an incentive to expand beyond simply selling hardware. Imagine Nvidia combining its massive computer infrastructure, engineering talent, and software ecosystem with with a Frontier AI research lab. It could potentially become a bold major supplier in the AI industry while also being a direct competitor to its biggest customers. Now, this is still a reported deal and nothing has been finalized, but

20:53 personally I think this could be a huge new interesting strategic move in the AI space and it is something that would actually position Nvidia to be one of the leaders in the space. And lastly, here is something pretty wild happening in the world of AI assisted development and this is something related to the same package that we had talked about yesterday, which is where a developer known as Mihawk has been working on an experiment project called PS5 Proton or PP, which

21:23 aims to bring Windows PC game compatibility to PlayStation 5 hardware. And apparently, the project is evolving into something even bigger that is called Mihawk OS, which is a Windows 11 inspired environment with ambitions to support both 32-bit and 64-bit Windows applications to Proton alongside planned Linux compatibility. Now, what’s particularly interesting is that the developer has also hinted at Claude being involved in the project, which is kind of nuts. It’s insane how much

21:54 development was used in this creation of this software and personally, I think it’s fascinating and another example of how developers are experimenting with AI assisted programming to tackle increasingly complex software projects. But, that’s about it guys for today’s video. I hope you guys enjoyed today’s video and got some sort of value. Make sure you go ahead and take a look at the world of AI benchmark as well as our world of AI school community as well as our YouTube channel, the second channel where you can stay up-to-date with the latest AI news. Make sure you also take a look at the newsletter, join us on

22:25 Discord, follow me on Twitter and lastly, also take a look at these plushies. Vote on which one you like better cuz I do have a new store coming soon. And lastly, make sure you guys subscribe, turn on the notification bell, like this video and please take a look at our previous videos. You can stay up-to-date with the latest AI news. But with that thought guys, have an amazing day, spread positivity and I’ll see you guys very shortly. Peace out folks.