Meta, Google, and OpenAI Race to Release Powerful New AI Models as Safety Concerns Emerge Over Unreadable Reasoning Outputs
Summary
Meta, Google, and OpenAI are locked in a fierce AI arms race with powerful new model launches, but OpenAI's upcoming 'Astra' model is raising urgent safety alarms as its looping reasoning technique produces unreadable math-only outputs, threatening the industry's ability to monitor AI behavior.
Key Points
- Meta launches Muse Spark 1.3, ranking 62 on the Intelligence Index and trailing only top Claude models, with CEO Mark Zuckerberg teasing a larger upcoming model codenamed 'Watermelon' and open-weight release plans.
- Google ships Gemini 3.8 Flash with improvements in coding, reasoning, and agentic tasks, scoring a 59 on the Intelligence Index, though DeepMind acknowledges the model still sits below the frontier as pressure mounts ahead of Gemini 4.
- OpenAI's upcoming 'Astra' model uses a 'recurrent depth' looping technique that boosts performance but generates math-only reasoning outputs, raising industry-wide concerns about AI safety monitoring becoming increasingly fragile and unreadable.