Engineering & Technical
The Engineer
Kubernetes 1.35.0 runs liveness probes before your startup probe finishes.
Reproduced on k3s, minikube and kind, which puts it in upstream kubelet rather than any one distro's patches. Anything with a slow warm-up burns failureThreshold 3 during init and cycles into CrashLoopBackOff toward the five-minute ceiling. The first symptom reads as a flaky pod, not a version bug, so the debugging hours go into your application code before anyone thinks to check the kubelet.
In Play
Kubernetes 1.35.0 And Go 1.27 Ship Behavior Changes You Never Opted Into
Kubernetes v1.35.0 evaluates liveness probes before the startup probe has succeeded, and Chris Short reproduced it on k3s, minikube and kind. Pods with a long warm-up restart forever instead of failing fast: the default failureThreshold is 3, and CrashLoopBackOff widens toward a five-minute ceiling. Go 1.27, released 19 August, rebuilds the original encoding/json on top of encoding/json/v2, so serialization behavior can drift in code that never imported v2.
Ask ClarityScaffolding Now Moves Scores More Than Weights Do
Nvidia's AVO wrapper took Claude Opus 5 from 30% to a perfect 183/183 on the public ARC-AGI-3 set with identical weights, per Techpresso, clearing it in 6,624 actions — about 12% fewer than the rival VISTA wrapper. Latent.Space reports a parallel result: one model scored anywhere from 52.4 to 76.2 across 106 identical tasks depending only on harness configuration. Roughly half your agent's measured capability lives in code you wrote and probably never versioned or benchmarked.
Ask ClarityFour Agents Debating Land At Or Below Chance
Anthropic ran the hidden-profile test — evidence split so that the facts every agent shares point at the wrong answer — on four-agent groups. Most model families decided correctly in only 17-36% of runs, per Exponential View, while a single agent handed the entire evidence base was right nearly every time. Blending several models' answers preserved only about a quarter of the good ideas one model produced, which makes a consensus stage a filter rather than redundancy.
Ask Clarity2027 Rack Prices And Memory Costs Reprice Your Efficiency Backlog
Server OEMs are telling customers that GB300 and VR200 systems with 2027 delivery dates will cost about 17% more, The Information reports — near $600-700K per rack and at least $5B per gigawatt of capacity. Amazon separately moved the Echo Dot from $50 to $80 and named memory and storage component costs. Every throughput point you win through caching, batching or quantization is now worth about 17% more in capacity avoided.
Ask ClarityRegulators Are Pricing The Missing Human-Review Step
The Dutch DPA fined Uber €825M ($966M) on 17 August for suspending driver accounts automatically with no warning and no human review, per Techpresso — second only to Meta's €1.2B GDPR penalty. Uber's own defense, that just 126 European drivers were deactivated on ratings in 2021, shows low volume bought nothing: the regulator priced the absence of the review path. Morning Brew reports TikTok separately agreed to pay $400M to settle DOJ child-privacy claims.
Ask Clarity
Deep Dives
- ●
Kubernetes 1.35.0 And Go 1.27: Same Interface, Different Behavior
Both bumps compile clean and pass every API check while changing how your pods restart and how your bytes serialize, so the first symptom is flakiness rather than an error.
Why it reads as flakiness, not as a version bug The timing arithmetic is what turns this from an afternoon into a sprint. The default failureThreshold is 3, the termination grace period is 30 seconds, and CrashLoopBackOff starts at 10…
3 action items
- ●
Fan Out For Capacity, Never For Judgment
Two agent results point opposite ways on scaffolding value, and the line between them decides which parts of your stack to fund and which to write for deletion.
The two different jobs people call "the harness" Scaffolding does two structurally different things, and these results only look contradictory until they are separated. Adding steps whose output a verifier can check compounds capability: search and backtracking policy, retained reasoning,…
3 action items
- ●
Eleven Months Of Depreciation Erases The 2027 Rack Increase
The hike is a planning input; the useful-life assumption sitting underneath your cost-per-token model is the term that actually decides the answer.
The dominant term is not the sticker Annualize a derived ~$3.9M rack and the ranking inverts. Five years of useful life carries about $780K a year; post-increase at the same life, about $912K. Six years lands near $760K, below the…
3 action items
The edition continues
Take the signal into the room.
Sign up or log in to read all 3 deep dives in full, plus the final take.
Read the full editionContinue with LinkedIn