The Download: AI’s self-improvement problem, and what’s driving the heat
| Source: MIT Technology Review AI
Tags: OpenAI, Astra, recursive self-improvement, AI safety, Unitree, Nvidia H200, AI agents
A new study finds AI agents still can't conduct open-ended research, casting doubt on near-term recursive self-improvement — and separately, OpenAI paused model work after its Astra model hit a "critical" safety risk threshold.
Details
The MIT Technology Review's daily newsletter surfaces two notable AI stories. A study found that current AI agents cannot conduct open-ended AI research — the kind requiring judgment, creativity, and free-form exploration — raising serious questions about whether recursive self-improvement is achievable soon. The key open question: whether AI systems can reach that capability by grinding through narrower tasks, or whether open-ended research is a prerequisite that must be solved directly. Separately, the newsletter reports that OpenAI paused some model work because its Astra model reached what the company called a critical safety risk threshold — a notable contrast with Anthropic, which has continued its release cadence. OpenAI also made security updates following the Hugging Face hack and launched ChatGPT for teenagers. Other items: Chinese humanoid robotics company Unitree surged 629% in its IPO debut, underscoring China's lead in the sector; China received roughly 10,000 Nvidia H200 GPUs each for ByteDance and Tencent; and Flock Safety launched an AI tool that identifies drivers by their movements, going beyond license plate tracking.