HUGE Google DeepMind RSI LEAKS! GPT-6 Astra NERFED, Kimi K2.8 Code, & More! AI NEWS
The content argues that Google may be nearing recursive self-improvement in DeepMind, while also covering Anthropic’s call to slow frontier AI, OpenAI’s temporary GPT-6 Astra regression and fix, and a rumored Gemini K 2.8 rollout with more efficient reasoning.
MAIN POINTS FROM TRANSCRIPT
- Google DeepMind is rumored to have reached RSI, with a checkpoint spotted in Vertex and leadership reportedly prioritizing AGI and self-improvement research.
- Anthropic CEO Dario Amodei urged slowing frontier AI development, while proposing stronger third-party safety and alignment oversight.
- OpenAI’s GPT-6 Astra seemed degraded, but the issue was traced to legacy skill files and inference misconfiguration, then fixed quickly.
- Gemini K 2.8 code was reportedly found in Gemini, suggesting an early rollout with reasoning comparable to K3 but more efficient.
TAKEAWAYS
- Recursive self-improvement is emerging as a major strategic theme among top AI labs.
- Safety oversight and acceleration remain in tension, even within companies that strongly believe in AI’s benefits.
- Model quality regressions can come from infrastructure issues, not just model capability changes.
- Efficiency gains in reasoning models may matter as much as raw performance improvements.