FIELD NOTES from the systems frontier
TypeSafe AI's Jev ditches free-text generation for typed, calibrated, parallel-sampled outputs trained with a new objective called RLCD. How RLCD differs from RLHF/RLVR, the workflow-evals numbers, the open-source adapter that makes the comparison reproducible, and where the "0% hallucination" claim needs an asterisk.
Astra, Claude 5.1, Gemini and Meta updates, Atlas and Solaris previews, the access debate, and ten new research papers.
Read edition 11 → LATEST · EDITION 12 · SEPTEMBER 7–13Astra reaches GA, DeepSeek undercuts on price, and ten papers on agents that build, audit, and correct their own procedures.
Read the latest briefing →FEATURED PATH · 14 ARTICLES
Follow Grok Build from its Rust composition root through the runtime loop, tools, workspace, context, safety, persistence, automation, ACP, and a practical harness blueprint.
Start Inside Grok Build →DISCOVER · WORKING REFERENCES
Pick the format that matches the moment: learn from a book, scan the commands, rehearse the interview, or choose the right tool.
Brendan Gregg's definitive guide made approachable—from CPUs, memory, disks, and networks to perf, ftrace, BPF, and a complete investigation.
Commands, flags, patterns, and gotchas for the middle of a task or the night before an interview.
Detailed Q&A from easy to hard, with the context behind what infrastructure and AI teams actually ask.
Short, honest reviews by role—DevOps, Security, AI Engineering, and SRE—with what to reach for and what to skip.