Age checks hit the web, CPU‑only exascale, and the new AI engineering playbook
NEWSLETTER | Amplifi Labs
KIDS Act Fast-Tracks Age Checks, Pressures Encryption and Moderation
Around the web • June 28, 2026
Congress is set to vote on the KIDS Act—a package combining a revised KOSA with the SAFE BOTS and SCREEN Acts—that would effectively push platforms to implement age verification under a negligence-style “knows or should have known” standard. The bill broadens moderation obligations (including around otherwise lawful discussions) and sets rules for direct, encrypted, and ephemeral messaging, pressuring services to collect IDs or use AI-based age estimation (with accuracy and bias risks) and potentially limit end-to-end encryption features. Developers should anticipate new compliance work across signup flows, parental controls, content policies, and chatbot capabilities, with outsized impact on startups.
AI engineering & tools
Proxy-KD distills black-box LLMs, beating classic white-box KD
Around the web •June 28, 2026
Researchers introduce Proxy-KD, a knowledge distillation approach that uses a proxy model to transfer capabilities from proprietary, black-box LLMs (e.g., GPT-4) to smaller student models. Experiments show Proxy-KD improves black-box distillation and surpasses traditional white-box KD, offering a practical path to train compact models without teacher internals—potentially reducing inference costs and enabling on-prem deployments.
Aleph Alpha codifies LLM training with Savanna, enabling one-click runs
Around the web •June 25, 2026
Aleph Alpha details Savanna, an internal “Model Training as Code” platform that encodes pretraining and post-training as typed, version-controlled functions, yielding hermetic, reproducible, one-click pipelines with provenance and caching. Integrated with GitHub CI (5‑minute PR-scale end-to-end tests, nightly larger validations) and orchestrated via Flyte on Kubernetes, it immutably versions artifacts in object storage with Weights & Biases lineage, automates evals, and deduplicates sweeps. For developers, the pattern reduces integration debt, accelerates iteration and cross-team ownership, and sets the stage for agent-driven auto-research.
Compute & performance shifts
China’s CPU-Only LineShine Tops TOP500 with 2.2 Exaflops, Leads HPCG
Around the web •June 28, 2026
China’s LineShine debuts at #1 on the TOP500 with a CPU-only Armv9 LX2 system delivering 2.198 FP64 exaflops (Rmax) across ~22k nodes/13M cores, and takes the HPCG crown at 22.004 PF/s—an exceptional result without GPUs. The machine draws 42.22 MW for 52.07 GF/W, underscoring rising CPU-side efficiency, while Italy’s MI300A-based HPC7 enters at #6 and Fugaku remains #3 on HPCG. Green500’s top 10 is unchanged, and TOP500 governance shifts to ACM SIGHPC with DOI-backed lists, signaling procurement and architecture shifts in exascale HPC.
Edera brings end-to-end NUMA awareness to Xen virtualization
Around the web •June 24, 2026
This piece explains how NUMA topology can impose 1.5x to 3x remote-memory latency (and worse under contention), why CPU and memory affinity must align, and how interleaving trades peak performance for predictability. Edera says it has shipped an end-to-end NUMA-aware stack for Xen—exposing real topology across guest, paravirt I/O, dom0, and hypervisor layers—to remove long-standing blind spots where dom0 and VMs lacked actionable vNUMA insight. For operators and performance-minded developers on modern EPYC/Xeon systems, this promises steadier tail latency and higher throughput, with next parts detailing automatic memory and vCPU placement.
Practical ops & tooling
Run and track AI dev agents in your terminal with Herdr
Around the web •June 29, 2026
Herdr is a Rust-based, tmux-like, agent-aware terminal multiplexer that organizes work into workspaces, tabs, and panes with mouse support and persistent detach/reattach—locally or over SSH. It auto-detects agent state (blocked/working/done/idle) for popular CLIs like Claude Code, GitHub Copilot CLI, Devin, Cursor, and more, and offers official integrations plus a socket API for session restore and programmatic control. Open-source under AGPL-3.0 with optional commercial licensing, it installs via curl/Homebrew/mise; Linux/macOS are stable and Windows builds are in preview.
Interactive Stanford tracker charts DRAM/NAND/HBM prices and AI accelerator costs
Around the web •June 28, 2026
Stanford’s DAM project publishes an interactive “Memory Prices” dataset tracking historical and current $/GB for DRAM, NAND, and HBM, with downloadable CSV and drill-downs by DRAM and HBM generations (HBM4 projected). It also models AI accelerator cost breakdowns (HBM, logic die, packaging/CoWoS, auxiliary) across Nvidia, AMD, Google, and Amazon; DRAM/NAND refresh monthly and HBM quarterly (last updated 2026-06-26). Useful for infra and AI teams budgeting and capacity planning, with caveats that figures reflect cheapest nominal retail listings and modeled estimates, not contract prices.
