My links of the wee 27.-31.07.

1. OpenChoreo (openchoreo.dev) is a CNCF sandbox project offering a complete, open-source developer platform for Kubernetes, ready to use from day one. It provides a modular multi-plane architecture (control/data/build/observability planes), a Backstage-powered developer portal, built-in SRE/FinOps AI agents, GitOps, observability, and MCP-based AI-assisted engineering tools.

2a. Kimi K3 Blog (kimi.com/blog/kimi-k3) Almost old, but Moonshot AI’s announcement of Kimi K3, a 2.8T-parameter model built on Kimi Delta Attention and Attention Residuals, with native vision and a 1-million-token context window. Also described as the world’s first open 3T-class model.

2b. Kimi-K3 Tech Report (github.com/MoonshotAI/K…); it’s the repo’s technical report accompanying the model, referenced from the HuggingFace card as the “Full Report.”

2c. Kimi-K3 Weights (huggingface.co/moonshot…) see the model card for moonshotai/Kimi-K3, 2.8T total params, 104B activated, 93 layers, 1,048,576-token context, MXFP4/MXFP8 quantization-aware training, released under the Kimi K3 License.

4. arXiv 2607.23809 “ACM: Agentic Context Management for Long Horizon Tasks” (arxiv.org/pdf/2607.23809) Proposes Agentic Context Management (ACM), giving agents purpose-built tools to autonomously compress context losslessly, offloading discarded content to external memory and querying it on demand. A teacher-guided post-training pipeline improves Qwen3.5-9B’s search/coding performance substantially over ReAct and prior summarization baselines (27% gain on BrowseComp-Plus).

5. github.com/saiyam1814/k… Kubernetes in Apple Container (KIAC) is a tool for local Kubernetes clusters on Apple’s container framework, where every node is its own lightweight VM (native on Apple silicon, no Docker Desktop/Lima/QEMU). Ships with metrics, storage, and LoadBalancer support out of the box, multi-node topologies, Cilium/eBPF option, node chaos testing, and a local web UI.

6. arXiv 2606.32026 (arxiv.org/abs/2606.32026) “AdaJEPA: An Adaptive Latent World Model” proposes a latent world model that performs test-time adaptation within the closed loop of model predictive control, using the observed next-state transition as a self-supervised signal to continuously recalibrate itself without additional expert demonstrations, substantially improving planning success on goal-reaching tasks.

7. AI development open letter (July 28, 2026) !—> More than 1,000 employees from frontier AI companies signed an open letter urging the US government to slow AI development pace, including OpenAI’s chief scientist, an OpenAI cofounder, Anthropic cofounders, and VPs at Meta and Google. edition.cnn.com/2026/07…

8. Graph-Engineering-Anthropic-Karpathy-Loop.pdf (drive.google.com/file/d…) — maps Andrej Karpathy’s work “autoresearch” (a single-agent loop that ran ~700 ML experiments in 2 days, retaining ~20 optimizations) and “AgentHub” (a sketched agent-first collaboration platform using a Git commit DAG instead of a main branch), onto Anthropic’s parallel infrastructure: the 2024 “Building Effective Agents” patterns (chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer), 2026 “Dynamic Workflows” (Claude-generated orchestration scripts spawning up to 1,000 sub-agents), and the Knowledge Graph Construction Cookbook (converting documents into typed entities/relations via structured-output prompts). Its core argument: each architecture externalizes a different bottleneck (loops externalize iteration, DAGs externalize lineage, knowledge graphs externalize shared/persistent memory), and it offers a staged build path with a single reflective loop (Day 1) through tool use, planning, multi-agent roles, a persistent knowledge graph, and finally swarm-scale workflows (Month 2), along with evaluation metrics, a decision framework, and a production checklist for when graph infrastructure is (and isn’t) worth the cost.

9. arXiv 2603.12277 (arxiv.org/abs/2603.12277) “Prompt Injection as Role Confusion” traces prompt injection to role confusion: models perceive who’s “speaking” from how text sounds rather than its labeled role, so injected text occupying a trusted role’s representational space can hijack an agent. Introduces “CoT Forgery,” a zero-shot attack injecting fake reasoning that achieves 60% attack success against frontier models with near-zero baselines