The data industry is still immature
The data industry is far from mature. Companies like Anthropic and OpenAI reportedly spend over $1M per individual RL environment, with annual totals in the hundreds of millions of dollars to push the RL frontier — which raises an interesting question: are Chinese labs buying the same environments from US data vendors, or is there a mirrored domestic ecosystem?
The answer isn't "there's no data industry" — it's that their experience is that the data industry's quality is relatively poor, and it's usually better to build environments and data in-house. Researchers spend a lot of time hand-crafting RL training environments; a few large companies (ByteDance, Alibaba) have internal annotation teams to support this. It's consistent with the broader "build, don't buy" mindset.
Dreaming: agents that learn between runs
Dreaming is Anthropic's name for a background process that reviews an agent's past sessions and memory, finds patterns, and rewrites memory so the agent improves between runs. OpenClaw shipped a similar feature in April, but Anthropic's framing is more focused on what teams of agents collectively learn than what a single agent remembers. The system learns from repeated corrections, recurring mistakes, and workflows that work well — over time, building an institutional knowledge base.
The feature currently lives inside Claude Managed Agents as a research preview. The obvious open question: when does this come to Claude Code?
GPT-Realtime-Translate
- Live streaming speech translation from 70+ input languages to 13 output languages
- OpenAI cofounder Greg Brockman noted that real-time voice-to-voice translation has been an anticipated OpenAI application since the company's early days, and is now available for anyone to build with
The asymmetry between 70+ inputs and 13 outputs is the interesting design choice: wide on input, focused on output — a sensible tradeoff between translation latency and quality.