The Scientist

Nvidia's AVO wrapper carried Claude Opus 5 from 30% to 100% across all 183 levels of ARC-AGI-3's 25 public games, per Techpresso, with no weight change. Latent.Space reports the same effect isolated: Harness-Bench held one model fixed across 106 tasks and scores still ran 52.4 to 76.2. Any model ranking you built on end-task success measured model x harness and credited the model. The 100% is public-set only, so read it as evidence that scaffolding can saturate an eval.

Anthropic ran the hidden-profile test - a group task rigged so the evidence everyone shares points to the wrong answer - on four-agent groups. Most model families reached the right decision in 17-36% of runs, while one agent holding the same full evidence base got it right nearly every time, per Exponential View. Blending several models' answers preserved only about 25% of a single model's good ideas. The control arm your orchestration eval lacks is also the cheaper configuration.

One prompt sent to twelve frontier models, including GPT-5.6, Claude Opus 5, Gemini 3 Pro, DeepSeek V4 and Qwen 3.8 Max, returned substantively identical answers by the fifth reply, per Box of Amazing. At the implied error correlation, ten answering models collapse to roughly 1.1 independent judges under k/(1+(k-1)rho). Every consensus threshold in your eval, RAG verification and safety pipelines has been scoring training-distribution overlap. The sweep is n=1 with no seeds, so trust the direction only.

Kubernetes v1.35.0 fires liveness probes before the startup probe succeeds, so any pod that needs minutes to become ready restart-loops. Chris Short reproduced it on k3s, minikube and kind. Model servers pulling multi-gigabyte weights never reach ready, and CrashLoopBackOff's five-minute ceiling makes it look like a flaky deploy while accelerators bill in full. GitHub's 7h47m outage on August 17 is the paired lesson: Copilot's client-side retry loop kept pushing load into a recovering system.

Server OEMs have told customers that Nvidia's rack-scale GB300 and VR200 systems rise about 17% for 2027 delivery, adding at least $5B per gigawatt, The Information reports from two anonymous sources. Chip systems are 60-70% of fully loaded compute cost, so that arrives as roughly 10-12% on amortized $/GPU-hour. Useful life matters more: at four years instead of six, annual cost per unit of compute runs about 76% above today's baseline. Nobody published perf-per-dollar for the new generation.