Social Media Briefing

Boris Cherny announces that in the next Claude Code version subagents run in the background by default so users can keep talking to Claude while they work, with foreground available on request.

Rowan Cheung profiles Panthalassa, a startup building self-propelling offshore data centers that use ocean cooling and wave power to bypass electricity and water bottlenecks.

Ethan Mollick graphs Artificial Analysis AA-Briefcase scores (multi-week complex consulting tasks), highlighting rapid gains and a clear open-weights performance gap behind closed models.

The Rundown reports Meta pushed a non-invasive brain-to-text decoder (Brain2Qwerty v2) to 61 percent word accuracy without implants, up from a roughly 8 percent prior baseline, with training code released.

The vLLM project shares an engineering deep-dive on serving four TTS models, detailing model-specific optimizations like chunk decoupling, torch.compile, GPU-resident decode state, and custom Triton kernels with throughput and latency gains.

Santiago reports Cline experiments showing GLM 5.2 jumps from 57.3 to 68.5 percent on coding tasks when reasoning is turned up with the same harness, arguing harnesses, not open-weight models, are the bottleneck.