Developments that matter in production
What changed, how it works and what it means for the people running reliable systems.
18 stories · 13–18
How we reportSWE-Serve checks coding-agent patches through a running SGLang server. In NVIDIA’s experiment, adding live-serving checks reduced the pass rate for the same patches.
Background reading: AI in productionAn eight-B200 test retained 96.1% to 98.2% of baseline output-token throughput with confidential computing enabled. Its software and workload define what that comparison covers.
Background reading: Platform engineeringTopograph publishes discovered topology as Kubernetes labels or Slurm configuration. Placement policies can use that map to trade a shorter communication path against a longer queue.
Background reading: Platform engineeringThe alliance supports storage, generation and flexible AI workloads as ways to change grid demand. For a scheduled evaluation, a safe pause can still miss the deadline the job serves.
Background reading: Platform engineeringKubernetes attributes connect agent calls to pods, workloads and paging routes. A processor migration can break those connections while telemetry still flows.
Background reading: ObservabilityHost-memory throttling can keep an inference process alive while its requests slow down. Effective kubelet settings and request deadlines explain the tradeoff.
Background reading: Platform engineering