AI Engineering Signal #74
Amazon shutting Mechanical Turk to new customers signals the end of human-labeled data as a scalable commodity
Signals
Amazon shutting Mechanical Turk to new customers signals the end of human-labeled data as a scalable commodity
teams still sourcing annotation through MTurk need to audit pipelines and contract alternatives before access gates close.
TechCrunch
Zuckerberg admits AI agent progress slower than expected
timeline assumptions for autonomous agent deployment need recalibration in roadmaps.
Reuters
Llama-server discarding valid KV caches by default
fix available; audit your llama.cpp serving config to stop paying the recompute cost.
Qwen 3.6 27B vLLM benchmarks across BF16, FP8, NVFP4
quantization tier selection now has concrete throughput numbers to inform serving cost decisions.
AI tutor hits 0.71–1.30 SD effect size in Dartmouth trial
strongest controlled result yet; sets a real benchmark for evaluating educational AI deployments.
Web
Lilian Weng posts careful analysis of scaling laws
required reading before committing to next training run; challenges naive extrapolation assumptions.
Web
The Take
The commodity layer of AI infrastructure — cheap human labels, predictable agent timelines, naive scaling assumptions — is eroding simultaneously. Teams that built cost models on any of these three pillars need to rerun the numbers now.
Subscribe
Related Signals