17 summarised stories about Agent Training, each linking back to the original source. Browse all topics →
OpenAI Blog 8 years ago
DeepMind trained AI agents through self-play competition to discover athletic skills including tackling, ducking, and catching without explicit instruction. The approach automatically maintains optimal difficulty as agents improve, similar to mechanisms observed in their Dota 2 experiments. This suggests self-play will become a standard component in developing capable AI systems.
The daily briefing
TLDRocket reads 60+ sources, removes duplicate coverage, and summarises the day in two minutes. Free, no spam, unsubscribe anytime.
TLDRocket
23 July 2026
The industry is confronting a fundamental shift in how AI gets deployed: away from single expensive models toward intelligent routing systems that match tasks to optimal solutions. Cursor, Ramp, Meta, and Runway all launched model routers this week, each reducing costs by 30–50% by directing simple queries to cheaper models and reserving frontier systems like OpenAI's Opus 4.8 for genuinely complex work. The pattern reveals maturation—companies no longer treat AI as a monolithic service but as modular infrastructure where economics matter as much as capability. Simultaneously, Chinese labs are closing the efficiency gap with startling speed. Moonshot's Kimi K3 topped benchmarks three months after its predecessor, while open-weight models now trail frontier systems by merely 4–7 months instead of half a year. Industry estimates suggest Chinese labs extract 4–7 times more performance per compute dollar, potentially reshaping the business case for expensive model development.
Read the full briefing →