Development
- Netflix Tech Blog — Netflix lets Apache Spark workloads on managed compute swap the cloud provider’s identity for Netflix’s own internal workload identity, combining signed instance metadata with pre-signed-URL verification so managed-compute jobs get the same attested identity as self-run infrastructure.
AI
- arXiv cs.CL — KaliBench: 8,504 query-command pairs across 1,642 Kali Linux security tools; no open-weight model exceeds 42% exact-command accuracy unrestricted, but SFT+RL lifts an 8B model to parity with a 685B one.
- arXiv cs.CL — AutoCompact: training coding agents to decide for themselves when to compact context raises pass rate +9.2% on SWE-bench Verified and +5.0% on SWE-PolyBench Verified, holding across 16K-256K context windows.
- arXiv cs.CL — Hierarchical Continuous Diffusion LMs (HC-DLM): unifying discrete token generation with continuous latent denoising beats same-size discrete and continuous diffusion baselines on Sudoku, Countdown, and LM1B perplexity.
- Hugging Face Papers — ActiveSaddler: adapting the training curriculum jointly with agent-harness optimization adds +4.4 and +7.5 pp Pass@1 over fixed-curriculum baselines, framing curriculum design as a new axis of harness optimization.
- Hugging Face Papers — Adaptive Reward Routing: routing multi-reward RL updates across layers by cross-modal interaction, and rebalancing rewards so dominant ones don’t suppress weaker ones, improves audio-video diffusion quality, semantic consistency, and A/V sync over RL baselines.
General
- Stack Overflow Blog — Podcast episode with Julien Verlaguet (CEO, Skip Labs): deliberately chosen constraints, not maximal flexibility, make developers faster — argued across the human-tolerance/tooling-constraint balance, typed-language design, and cost-effective tooling for AI agents.