HUGE Google DeepMind RSI LEAKS! GPT-6 Astra NERFED, Kimi K2.8 Code, & 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 Google DeepMind RSI LEAKS! GPT-6 Astra NERFED, Kimi K2.8 Code, & More! AI NEWS · Channel: WorldofAI · 11:10 · Watch on YouTube

Topics: AI agents, AI models, Multimodal models

Summary

The video covers reports that Google DeepMind may be developing an internal recursive self-improvement system to help train and refine future models, though the claims are unconfirmed. It also discusses Moonshot AI’s Kimi K2.8 preview, reportedly offering performance comparable to K3 with more efficient reasoning. The host explains reports of GPT-6 Astra quality problems as stemming from legacy skills, an optional context experiment, and misconfigured inference engines, which OpenAI reportedly addressed quickly. Finally, the video notes that GPT Live is available through the API, enabling voice agents to listen while speaking.

Key points

Leakers reportedly spotted an internal configuration called “RSI model” within Google Vertex, prompting speculation about recursive self-improvement at DeepMind.
The proposed Google system may use AI to evaluate, train, and refine subsequent generations of AI models; the video stresses that this is not confirmed.
The host says Sergey Brin is reportedly directing resources toward RSI, while Demis Hassabis is focusing on AGI.
Kimi K2.8 appeared in Kimi Code and is described as comparable to K3 while using more efficient reasoning.
Kimi K2.8 reportedly supports a 1-million-token context window, image and video input, and low, high, and max reasoning modes.
Reports that GPT-6 Astra had worsened were linked to legacy skills, an optional context-management experiment, and misconfigured inference engines.
OpenAI reportedly fixed or disabled the identified Astra issues within roughly 24 to 36 hours and reset limits for paid-plan users.
GPT Live became available through the API, allowing voice agents to listen while speaking and supporting developer-selected models and agent harnesses.
Anthropic CEO Dario is described as calling for slower frontier AI development while also expressing optimism about AI’s potential benefits.

Chapters

Start	Chapter	Summary
00:00	AI news roundup and reported Google RSI	The host introduces reports of Google DeepMind’s possible recursive self-improvement work, Kimi K2.8, and the alleged GPT-6 Astra regression.
01:37	Anthropic’s call to slow frontier development	The host discusses Dario’s proposed safety oversight and his simultaneous optimism about AI’s potential to transform society.
03:10	Why Google may be pursuing recursive self-improvement	The video examines leaks and reported internal research suggesting AI systems could help Google evaluate and improve future models.
05:49	Kimi K2.8 preview	The host outlines K2.8’s reported capabilities and its potential advantage over K3 in reasoning efficiency.
07:29	GPT-6 Astra issues and GPT Live API	The video explains the reported causes and fixes behind Astra’s quality issues, then describes GPT Live’s real-time voice features for developers.

Quotes

Google might slowly be making a huge comeback. — 00:00

Essentially, AI’s helping build a smarter AI. — 04:46

So, if K 2.8 can deliver similar while reasoning significantly faster, this could actually be a really useful model. — 06:22

they ended up actually giving a reset to every single OpenAI paid plan user — 09:36

voice agents can now listen while they’re speaking — 10:10

Concepts and entities

recursive self-improvement (RSI), Google DeepMind, Google Vertex, Sergey Brin, Demis Hassabis, Kimi K2.8, GPT-6 Astra, GPT Live 1, Anthropic

Full transcript (original language)

00:00 It’s getting wild in the AI space, and Google might slowly be making a huge comeback. Today, we’ve got some pretty interesting insights suggesting that Google DeepMind has officially achieved RSI, which also stands for recursive self-improvement. Basically, when an AI agent can improve on its own capabilities. Reliable leakers have pointed towards something called RSI model, and this is a checkpoint that was spotted just yesterday. And this is something that was spotted within Google Vertex. This might be more than just

00:32 rumors as Google says Demis Hassabis is focusing his attention on AGI, while reports suggest that Sergey Brin has been directing resources towards RSI. So, while everyone waits for Gemini 4, Google might actually be training the trainer behind the scenes. Moving over to Moonshot AI, the Gemini K 2.8 code has been spotted inside Gemini code, and it appears to be slowly rolling out. The early generations look comparable to Gemini K3, but with much

01:03 more efficient thinking, which is obviously going to be a huge upgrade considering how long the K3 can actually think. Over at OpenAI, people thought that GPT-6 Astra has been nerfed after generations started looking flatter and more synthetic. But apparently, Astra itself wasn’t downgraded. The regression came from legacy skill files and misconfigured inference engines. OpenAI listened though and reportedly fixed the issue within roughly 24 to 36 hours, which is pretty solid, and they actually ended up giving everyone a reset, even

01:37 though that this is something that didn’t affect everyone. Now, there’s a lot more to cover, so let’s just simply dive into it all. There’s a lot of stuff in the AI space that I don’t really put on to the YouTube channel, and you can actually access it through my free newsletter with the link in the description below, where you can subscribe completely for free. Starting things off with Dario, who is the CEO of Anthropic, who randomly dropped a new essay calling for the AI industry to slow down frontier development. I don’t know what this guy does or why he’s even mentioning this stuff. This is where you

02:09 have a full-on IPO incoming and you’re basically just nerfing yourself by instigating this sort of debate where AI should be slowed down. He proposed a three-part plan with Anthropic already committing to giving third-party evaluators per minute employee-level access to a system to monitor safety and alignment. What makes this interesting though is that Dario is simultaneously extremely bullish on AI, which is why I was kind of confused cuz he’s kind of

02:40 contradicting what he is trying to imply. He also is believing that it could cure most major diseases, massively accelerate economic growth, and create a world of abundance with within the next 5 to 10 years. So, essentially, Dario does believe AI could completely transform humanity, but also thinks we need to pace it. Moving forward, Google DeepMind may be getting surprisingly close to RSI or may have even achieved it, which is something that stands for recursive self-improvement for the people who do

03:10 not know. And if there’s actually something to this, it could be one of the biggest transitions that we have ever seen in AI. There have been reports that Sergey Brin has been heavily pushing resources towards RSI research. With people describing him as extremely invested in both RSI and AGI. Demis Hassabis is also reportedly putting his full attention towards AGI. But, this gets even more interesting when you take a look at Google and what they have been

03:41 doing with its models. Google has talked about using agentic loops that recursively evaluate and refine models. And DeepMind has reportedly described RSI as an important part of the overall investment thesis. Now, there is someone who is called Lyra, and essentially, this is where all the leaks actually came from. And he had pointed towards the rapid release of recent flash models that Google had released, and basically asked, “How else is Google shipping and improving these models this quickly?” Cuz if you think

04:15 about it, Gemini 3.7 and Gemini 3.8 had essentially released in basically a 1-month span. He is essentially proposing that they have an internal RSI model that is able to help them recursively think and improve on these models to release better ones faster. And he had essentially found a configuration called RSI model within the internal backlogs. And it had surfaced, which lines up surprisingly well with the previous comments

04:46 surrounding RSI as well as live RL. The theory is that Google could have an internal system where AI’s helping evaluate, train, and improve the next AI generation of models at Google. Essentially, AI’s helping build a smarter AI. Now, obviously, nothing is confirmed yet, but these leaks that had just arise is basically implying that Google has officially achieved RSI. Google might not just be training another frontier model. They could

05:17 potentially be training the trainer itself while everyone waits for the Gemini 4. And if they did manage to push that feedback loop significantly further internally, this Gemini 4 model could represent a much bigger jump than people are expecting. So, the big question is, is DeepMind quietly building the technology that makes the next generation of Gemini models fundamentally different, and could this finally be Google’s massive AI comeback? I truly think they are inching towards a

05:49 comeback, but who knows. Next up, we have Kimi K 2.8 preview, which randomly dropped inside Kimi code. And this was definitely a surprising release cuz we truly didn’t expect the Moonshot AI team to drop this model. According to the docs, Kimi K 2.8 has the same performance level as the K3, but apparently comes with more efficient thinking. It also supports the massive 1 million token context window with image and video input alongside with low, high, and max reasoning modes. And honestly, the efficient thinking part

06:22 might be the biggest upgrade here. Kimi K3, as we all know, is definitely a great model. It is one of the best open-source models that’s out there. But one of the biggest issues with this model is that it thinks way too long whenever you send in a request. It tends to overthink for basic tasks, which is some of the reasons why you would not want to use this model. So, if K 2.8 can deliver similar while reasoning significantly faster, this could actually be a really useful model. People subscribed to the Kimi coding plan are already starting to see this

06:56 appear within their code plan. The funny part is the naming though, cuz we went from K3 back down to K 2.8, which might be the first time I’ve seen a company actually make this version number smaller with a newer release. 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

07:29 content, plus a lot more. Next up is OpenAI because over the past few days, there have been a ton of reports claiming that GPT-6 Astra has gotten worse since at launch. And honestly, looking at a few of these demos, you can clearly see why people started questioning it. People tested launch day Astra versus what Astra is today using the exact same prompts and reasoning settings. And the launch version appeared noticeably more photorealistic, sharper, and stronger at generating these complex 3D scenes. Meanwhile, some

08:01 of the newer generations look flatter, more synthetic, and overall less impressive. This immediately had the community saying that GPT-6 Astra had been nerfed and questioned why Frontier Model Labs always seemed to incredibly nerf their model after the initial hype phase. However, that apparently did not happen. OpenAI investigated these reports that users had on X and found several actual issues. This is actually related to legacy skills that were built for older models

08:35 that triggered too often, sometimes preventing Astra from properly checking its own work. There was also an opt-in context management experiment, which was causing early stops or responses to older messages, which OpenAI estimated affected around 4,000 to 5,000 users. And probably most importantly, OpenAI discovered badly configured inference engines that were causing a measurable quality degradation for a portion of Astra’s traffic. Those issues have been

09:06 now fixed or disabled alongside some smaller improvements that designed Astra to be more consistent and better at checking its own work. So, these reports that had claimed that Astra had been degraded were partially true, but not actually the full-on case for most people. And they actually had only 5,000 people affected by this bug. And even though this was only a small portion or a small percentage of people that were affected, OpenAI listened to the

09:36 community, identified the problem, and pushed fixes within roughly 24 to 36 hours. But, that’s not all cuz they ended up actually giving a reset to every single OpenAI paid plan user, which is incredible cuz this is where they have now reset all of the limits for the full on week so that you can use Astra even further. Moving forward, GPT live one is officially available through the API, which brings ChatGPT these more natural voice conversations directly to developers. The big feature here is that

10:10 voice agents can now listen while they’re speaking, like how you would with Gemini live. Making conversations feel much more natural instead of constantly waiting for one side to finish. Developers can also pair GPT live one with models and agent harnesses they choose, opening the door to much more capable real-time voice agents. This is honestly a pretty big upgrade for anyone building AI assistants, customer support agents, or other voice-based applications. But, that’s about it, guys, for today’s video on

10:40 whatever is happening in the world of AI. Make sure you go ahead and take a look at the world of AI Insider Club. This is our school community that we had just recently launched, and you get access to free AI tools, you get access to the world of AI VIP Corner platform, as well as getting access to a couple of different courses where I teach you how to use a lot of these agents in the best manner. You also have the ability to get access to early models, and that is a way for you to test beta test these models. So, this is something that I’m

11:10 working with different labs to get you guys access, and you can easily get started with this with the links in the description below. Make sure you also take a look at the world of AI VIP Corner platform and benchmark. Take a look at the second channel, join the newsletter, join the Discord, follow me on Twitter, and lastly, make sure you guys subscribe, turn on the notification bell, like this video, and please take a look at our previous videos so that 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 fairly shortly. Peace out, fellas.