Newsletter Archive

Open-source AI labs face unsustainable economics as frontier model training costs spiral; the author argues a multi-company consortium is the only viable long-term path to well-funded open models, pointing to recent departures at Qwen and Ai2 as signs of burnout under the current single-lab model.

Interconnects · 6 min read Community

A small cohort of AI infrastructure startups are building the plumbing that lets the Pentagon safely run LLMs on classified data—a $2B market largely invisible to public scrutiny. The work addresses a core deployment challenge: using powerful models without leaking secrets through training or inference, with companies like Ask Sage already serving thousands of DoD teams.

Fortune · 9 min read Industry

SiFive, an Nvidia-backed chip design startup, reached a $3.65B valuation in its latest funding round as it develops open-source AI accelerator designs to compete with proprietary alternatives.

TechCrunch · 3 min read Market

Developer built KubeezCut, an MIT-licensed in-browser video editor powered by WebGPU and WebCodecs, with optional AI-powered media generation—no server uploads or subscriptions required.

Dev.to · 4 min read Tools

Agent harnesses are replacing earlier scaffolding patterns (RAG chains, LangGraph) as the dominant way to build AI agents—and the critical lesson is that memory control determines lock-in. Open harnesses let you own your agent's memory; closed, proprietary ones hand that power to third parties.

LangChain · 7 min read Community

Step-by-step guide to deploying OpenClaw, an open-source agentic runtime, with loopback-only networking, API key authentication, and custom skill orchestration—showing how to build deterministic, local-first AI agents without cloud dependencies.

MarkTechPost · 6 min read Tools

MIT, NVIDIA, and Zhejiang University researchers introduced TriAttention, a KV cache compression method that achieves 2.5× higher throughput and 10.7× memory reduction while matching full attention accuracy on long-chain reasoning tasks—critical for deploying reasoning models like DeepSeek-R1 on consumer hardware.

MarkTechPost · 9 min read Research

Aria Networks raised $125M for AI-optimized networking infrastructure, targeting the data center bottleneck that limits multi-GPU training and inference scaling.

Pulse2 · 2 min read Market