AI digest: Agents get serious infrastructure
NVIDIA builds proper RL infrastructure for agents while everyone else scrambles to make AI systems that actually work in practice.
This week feels like a turning point. The focus has shifted from “look what our chatbot can do” to “how do we build AI systems that actually get things done.”
NVIDIA tackles the agent training bottleneck
NVIDIA released ProRL Agent, a “rollout-as-a-service” system that separates the I/O-heavy environment interactions from GPU-intensive training. This matters because current RL setups for multi-turn agents are woefully inefficient. NVIDIA’s essentially admitting that training capable agents requires proper infrastructure, not just throwing more compute at the problem.
JiuwenClaw promises self-evolving agents
The OpenJiuwen community launched JiuwenClaw, an agent that supposedly evolves its capabilities over time rather than just following pre-programmed workflows. The pitch is that most agents fail when real-world requirements change mid-task. Whether this actually works remains to be seen, but the problem they’re tackling is real.
Anthropic’s Claude Mythos leak hints at major capability jump
Leaked documents reveal Anthropic’s working on “Claude Mythos” with “dramatically higher scores” than existing models. The leak mentions a new model class above Opus and a deliberate slow release strategy focused on cybersecurity concerns. If true, this suggests we’re approaching another significant capability threshold.
Federal judge blocks Trump’s Anthropic ban
A San Francisco judge called the government’s security risk labelling of Anthropic models “Orwellian” and blocked the ban as “illegal First Amendment retaliation.” This sets an interesting precedent for how AI model restrictions might be challenged in court.