Restraint was the theme on the model side: Z.ai held back weights, OpenAI paused its largest frontier RL run. The two outage reports are dumber than that: retries that made things worse, and a CI signal nobody read.

AI & machine learning

Z.ai ships GLM-5.3 and delays the weights – GLM-5.3 landed 2026-08-14 on the same base model as GLM-5.2, so all gains come from post-training. CyberGym went 77.2 to 84.5, ExploitBench 24.4 to 54.4, and Z.ai reports 2,436 vulnerabilities found across 269 open source projects. Weights are held two weeks for safety work. Source

OpenAI puts an on-call process around training runs – On 2026-08-18 OpenAI described changes after the July sandbox escape that reached Hugging Face: stronger isolation for model-generated code and a two-week pause on RL training for deployment-track models. It now targets an alert within 30 minutes of concerning activity, and teams must pause the run if they can’t clear a false positive in 30 minutes. Source

Site reliability engineering

GitHub’s 7h47m outage, amplified by a VS Code retry bug – On 2026-08-17, 13:28 to 21:15 UTC, web and API errors peaked near 20%, raw-content downloads near 50%. An Istio sidecar hit its concurrency limit and didn’t scale: the autoscaling policy watched the host service, not the sidecar. Four HAProxy nodes then exhausted flow limits, and delayed replies to one internal endpoint tripped a latent VS Code retry bug that multiplied traffic 10x. Source

Tuist: a duplicate migration version CI had already flagged – A 21-minute ClickHouse outage on 2026-08-19 broke every read and write of the test_runs table. The migration adding two columns shared a version number with an already-applied one, so it was skipped silently everywhere. CI had failed with migration version 20260817120000 is duplicated two hours earlier. Source

Observability

Prometheus 3.14.0 – Released 2026-08-17. Duration expressions are on by default, promql-duration-expr is a no-op, first_over_time is stable, and native histogram scrape parsing allocates about 49% less. Colliding OTLP attribute names now warn and increment prometheus_api_otlp_translation_warnings_total. Source

Collector v1.65.0/v0.159.0 moves the batch size metrics – otelcol_exporter_queue_batch_send_size was recorded at enqueue time, measuring incoming requests rather than batches. Since 2026-08-17 it records after batching, the old values live in otelcol_exporter_enqueue_size, and neither appears unless sending_queue::batch is set. Source

  • Alertmanager 0.34.0 (2026-08-16) – alertmanager_notifications_failed_total splits authError and rateLimited out of clientError, so old matchers break.
  • Grafana 13.2.0 (2026-08-18) – adds trace export to file and commit authoring in Git Sync.
  • The Wording Effect – IBM Research finds meaning-preserving rephrasing flips answers both ways, and stronger models lose more than they gain.
  • OpenTelemetry entity events (2026-08-14) – what to build once entity events arrive.

My take

GitHub shipped a PR to cut gateway retry logic mid-incident. If you can’t state your retry budget and where it’s enforced, that’s the gap. Tuist’s lesson is cheaper: CI named the problem and the deploy went out anyway.

That’s the week. Tell me if your queue dashboards survived the upgrade.