article — General

Google DeepMind RSI Leaks, GPT-6 Astra Issues, Kimi K2.8 Code & More AI News

About this Resource

Look at the latest AI news, including Google DeepMind's reported RSI work, Kimi K2.8 Code, GPT-6 Astra, and new real-time AI voice capabilities.<h2> </h2><h2>Google DeepMind RSI Leaks, GPT-6 Astra Issues, Kimi K2.8 Code & More AI News</h2>The AI world is moving quickly, with new developments from Google DeepMind, Moonshot AI, OpenAI, and Anthropic. Some of the biggest stories involve reports of Google's work on recursive self-improvement (RSI), a new Kimi coding model, problems that made GPT-6 Astra appear weaker, and major improvements to AI voice agents. <h2>Google DeepMind May Be Working on Recursive Self-Improvement</h2>One of the biggest stories is a reported internal Google DeepMind system connected to recursive self-improvement, or RSI. RSI is the idea of an AI system helping improve AI systems themselves. Instead of humans doing all the work of evaluating, training, and refining a model, an AI could potentially help with parts of that process. Reports mentioned in the video point to an internal configuration called “RSI model” that was reportedly spotted within Google's systems. This has led to speculation that Google may be experimenting with AI systems capable of helping evaluate and improve future models. The idea is particularly interesting because Google has been releasing new Gemini models at a rapid pace. The possibility raised by the leaks is that an internal AI system could be helping Google improve models faster. However, the RSI achievement has not been officially confirmed, so it should be treated as a report rather than an established fact. If Google has genuinely developed a useful RSI system, it could become an important part of how future AI models are trained and improved. <h2>Kimi K2.8 Appears Inside Kimi Code</h2>Moonshot AI has also reportedly started rolling out Kimi K2.8, which appeared inside Kimi Code. The model is described as being around the performance level of Kimi K3 while focusing on more efficient reasoning. This is important because Kimi K3 is known for spending a long time thinking through tasks. While that can help with difficult problems, it can also make the model unnecessarily slow for simpler requests. Kimi K2.8 reportedly supports:<ul><li>A 1 million-token context window</li><li>Image and video input </li><li>Multiple reasoning modes </li><li>More efficient thinking </li><li>Coding-focused use through Kimi Code</li></ul> The goal appears to be getting similar levels of performance while reducing the amount of time the model spends reasoning. <h2>GPT-6 Astra Wasn't Actually Nerfed</h2>Another major discussion involved GPT-6 Astra appearing to become worse after its launch. Users compared earlier generations with newer ones using similar prompts and settings. Some tests appeared to show that the launch version produced sharper and more realistic results, particularly for complex 3D scenes. This led to claims that Astra had been deliberately “nerfed.” The explanation described in the video was different. OpenAI reportedly investigated the issue and discovered several technical problems affecting some users. These included:<ul><li>Legacy skills designed for older models triggering too often </li><li>A context-management experiment causing some conversations to stop early or reference older messages </li><li>Misconfigured inference engines reducing quality for part of Astra's traffic</li></ul> The context-management issue was estimated to affect roughly 4,000–5,000 users. OpenAI reportedly fixed or disabled the problematic systems within around 24–36 hours and made additional improvements to help Astra check its own work more consistently. So while some users genuinely experienced worse results, the explanation presented was that technical configuration problems, not a deliberate downgrade of Astra itself were responsible. <h2>GPT Live 1 Comes to the API</h2>Another important development is the availability of GPT Live 1 through the API. The major improvement is more natural real-time voice interaction. Voice agents can listen while they are speaking, making conversations feel less like a strict turn-by-turn exchange. Developers can also combine the voice system with different models and agent frameworks. This could make it more useful for applications such as:<ul><li>AI assistants </li><li>Customer support </li><li>Real-time voice agents </li><li>Voice-based applications</li></ul> The broader trend is clear, AI voice systems are moving toward conversations that feel more natural and less robotic. <h2>Anthropic Calls for Slower Frontier AI Development</h2>Anthropic CEO Dario Amodei also reportedly published an essay calling for the AI industry to slow down the development of frontier AI systems. The proposal includes stronger safety oversight, including Anthropic's commitment to giving third-party evaluators access to its systems to help monitor safety and alignment. At the same time, Amodei remains highly optimistic about what advanced AI could accomplish. The potential benefits discussed include accelerating scientific progress, improving medicine, increasing economic growth, and contributing to a world of greater abundance. The central argument is therefore not to stop AI development entirely, but to make sure increasingly powerful systems are developed with stronger safety measures. <h2>Final Thoughts</h2>The biggest takeaway is that AI development is increasingly becoming about how AI systems can improve themselves, reason more efficiently, and interact with people more naturally. Google's reported RSI work could point toward AI-assisted model development. Kimi K2.8 is focusing on getting strong performance without excessive reasoning time. OpenAI's Astra problems show how much model performance can depend on the systems surrounding the model itself, while GPT Live 1 demonstrates the growing importance of real-time voice interaction. Some of the most interesting developments are still unconfirmed, particularly the claims surrounding Google's RSI research. But if these systems continue advancing, the way AI models are trained, improved, and used could change significantly.