AI digest: Infrastructure fights and agent breakthroughs
GitHub's pricing revolt, OpenAI's PC control ambitions, and the tools that might actually make AI agents work.
The infrastructure bills are coming due and developers aren’t happy about it.
GitHub Copilot’s token pricing sparks developer revolt
GitHub switched Copilot to token-based billing and developers are calling it “a joke”. The move ends the flat-rate subscription model that made Copilot predictable for teams. This feels like the end of AI tooling honeymoon period, where companies gave away compute to build market share.
OpenAI’s Codex takes control of Windows PCs
OpenAI’s Codex app now runs autonomously on Windows 11, controlling programmes and hunting bugs without human input. You can even start tasks remotely from your phone when away from the PC. This is proper computer-use capability, not the demo-ware we usually see.
Trajectory cracks the multi-LoRA training problem
Trajectory built a concurrent multi-LoRA training stack that delivers 2.81× faster experiment throughput for continual learning. Each RL experiment gets its own LoRA adapter on an always-hot engine, with the code open-sourced. This tackles a real bottleneck in agent development where training multiple skills simultaneously was painfully slow.
Genesis World 1.0 promises proper robot simulation
Genesis AI released Genesis World 1.0, a physics platform that cuts robot policy evaluation from 200+ hours to under 30 minutes. The key claim is 0.8996 Pearson correlation between simulation and real-world robot performance. If that correlation holds up, this could finally solve the sim-to-real gap that’s been plaguing robotics.
One company blew £400M on Claude in a month
An unnamed company reportedly spent $500 million on Claude after failing to implement usage caps. This highlights the infrastructure cost reality check hitting enterprises as they scale AI beyond proof-of-concept projects.