Tech • AI • Robotics • Game
Huge Google DeepMind RSI Leaks! GPT-6 Astra Nerfed, Kimi K2.8 Code, & More! AI News
9.2/10TL;DR
Google DeepMind, Moonshot AI, OpenAI, and Anthropic all signaled important shifts this week, from reported work on recursive self-improvement to faster reasoning models, rapid bug fixes, and renewed calls to slow frontier AI development.
KEY POINTS
Google DeepMind and reported RSI work
New leaks tied to Google Vertex referenced an RSI model, fueling claims that Google DeepMind may be pursuing or nearing recursive self-improvement, where AI systems help improve future models. The idea aligns with reports that Sergey Brin has pushed resources toward RSI research while Demis Hassabis has focused on AGI. If accurate, the effort would mean Google is not only training flagship models but also building systems that help evaluate, refine, and accelerate the next generation.
Why the RSI claims matter
The speculation gained traction because Google has recently shipped model updates at an unusually fast pace, including closely spaced Gemini 3.7 and Gemini 3.8 releases. Google has also discussed agentic loops that recursively evaluate and refine models, reinforcing the theory that internal AI tools may already be part of the development pipeline. A mature RSI workflow could make Gemini 4 a larger leap than a standard version upgrade.
Moonshot AI’s Kimi K 2.8 appears in rollout
Moonshot AI quietly surfaced a Kimi K 2.8 preview in product code and early user access. Documentation indicates performance comparable to Kimi K3, but with more efficient reasoning, a change that could address one of K3’s biggest weaknesses: taking too long on simple tasks. The model reportedly supports a 1 million token context window, plus image and video input and multiple reasoning settings including low, high, and max.
Efficiency could be the real upgrade
Kimi K3 has built a reputation as a strong open model, but users have criticized its tendency to overthink routine prompts. If K 2.8 keeps similar capability while reducing latency and unnecessary reasoning, it could become more practical for coding and everyday use. The unusual version number, moving from K3 to K 2.8, also suggests a release aimed at optimization rather than a conventional headline-grabbing jump.
OpenAI responds to Astra quality complaints
Users comparing launch-day outputs from GPT-6 Astra with newer generations reported weaker image quality, including flatter, less photorealistic results and weaker complex 3D scenes. That prompted accusations that the model had been quietly downgraded after launch. OpenAI said the core model had not been intentionally nerfed and traced the regression to several technical issues instead.
What caused the Astra regression
The company identified legacy skill files that triggered too often, sometimes preventing Astra from properly checking its own work. It also found an opt-in context management experiment that caused early stops or replies to older messages, affecting an estimated 4,000 to 5,000 users. Most significantly, misconfigured inference engines degraded quality for part of Astra’s traffic, after which the issues were fixed or disabled.
Fast remediation and customer reset
The problems were reportedly addressed within roughly 24 to 36 hours, alongside smaller consistency improvements. Although only a limited share of users appeared affected, OpenAI reset usage limits across paid plans for the week, effectively compensating users broadly rather than only those directly hit. The response highlighted how model quality now depends not just on training, but also on orchestration layers and serving infrastructure.
GPT Live One reaches the API
GPT Live One is now available through the API, bringing more natural real-time voice interaction to developers. A key feature is turn-taking that allows the system to listen while speaking, making conversations feel less rigid than traditional voice bots. That opens the way for more capable assistants, customer support systems, and other real-time voice applications.
Anthropic urges restraint while predicting transformation
Anthropic chief executive Dario Amodei published a new argument for slowing frontier AI development even as he maintained an aggressive view of AI’s benefits. His proposal included giving third-party evaluators employee-level access to monitor safety and alignment. At the same time, he argued advanced AI could help cure major diseases, accelerate growth, and create broad abundance within 5 to 10 years, underscoring the industry’s tension between urgency and caution.
CONCLUSION
The latest developments show an AI sector accelerating on multiple fronts at once: faster model iteration, more efficient reasoning, real-time voice, and growing concern over safety and control. The biggest unanswered question is whether Google is merely catching up or building the tools that could speed up the entire race.
Explain this